close

Вход

Забыли?

вход по аккаунту

?

Investigation into the use of microwave sensors to monitor particulate manufacturing processes

код для вставкиСкачать
*UDGXDWH6FKRRO(7')RUP
5HYLVHG 0114
PURDUE UNIVERSITY
GRADUATE SCHOOL
Thesis/Dissertation Acceptance
7KLVLVWRFHUWLI\WKDWWKHWKHVLVGLVVHUWDWLRQSUHSDUHG
%\ John Samuel Austin III
(QWLWOHG
Investigation into the Use of Microwave Sensors to Monitor Particulate Processes
)RUWKHGHJUHHRI
Doctor of Philosophy
,VDSSURYHGE\WKHILQDOH[DPLQLQJFRPPLWWHH
Harris
Michael
Gintaras Reklaitis
James Litster
Carl Wassgren
7RWKHEHVWRIP\NQRZOHGJHDQGDVXQGHUVWRRGE\WKHVWXGHQWLQWKHThesis/Dissertation Agreement.
Publication Delay, and Certification/Disclaimer (Graduate School Form 32)WKLVWKHVLVGLVVHUWDWLRQ
adheres to the provisions of 3XUGXH8QLYHUVLW\¶V³3ROLF\RQ,QWHJULW\LQ5HVHDUFK´DQGWKHXVHRI
FRS\ULJKWHGPDWHULDO
Michael Harris
$SSURYHGE\0DMRU3URIHVVRUVBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
BBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBBB
$SSURYHGE\ Michael Harris
+HDGRIWKHDepartment *UDGXDWH3URJUDP
4/14/2014
'DWH
i
INVESTIGATION INTO THE USE OF MICROWAVE SENSORS TO MONITOR
PARTICULATE MANUFACTURING PROCESSES
A Dissertation
Submitted to the Faculty
of
Purdue University
by
John Samuel Austin III
In Partial Fulfillment of the
Requirements for the Degree
of
Doctor of Philosophy
May 2014
Purdue University
West Lafayette, Indiana
UMI Number: 3635780
All rights reserved
INFORMATION TO ALL USERS
The quality of this reproduction is dependent upon the quality of the copy submitted.
In the unlikely event that the author did not send a complete manuscript
and there are missing pages, these will be noted. Also, if material had to be removed,
a note will indicate the deletion.
UMI 3635780
Published by ProQuest LLC (2014). Copyright in the Dissertation held by the Author.
Microform Edition © ProQuest LLC.
All rights reserved. This work is protected against
unauthorized copying under Title 17, United States Code
ProQuest LLC.
789 East Eisenhower Parkway
P.O. Box 1346
Ann Arbor, MI 48106 - 1346
ii
I would like to dedicate this work to Maggie. Your unyielding support and ardent love
are the anchors of my life. When I am feeling my worst, you are there with a lovely
smile, an encouraging word, and a helping hand to keep us moving towards our goals.
I would also like to thank my parents who have sacrificed so much for my happiness.
You have been there when I needed support and allowed me to find my way when I did
not. Thank you for all of you encouragement. Furthermore, I would like to thank my
sister. Liana, you have been an ideal big sister, always looking out for me; I deeply
appreciate it.
iii
ACKNOWLEDGEMENTS
I would like to convey my sincere gratitude to my committee chair and advisor, Professor
Michael T. Harris, who has given me the freedom to pursue what interests me and the
guidance to make sure I do not stray.
Furthermore, I would like to thank Professor Gintaras Reklaitis, who has on numerous
occasions gone out of his way to help me pursue my research and advertise my results to
outside parties. Lastly, I would like to thank my other committee members, Professors
Litster and Wassgren, who have taken time out of their busy lives to help guide me.
iv
TABLE OF CONTENTS
Page
LIST OF TABLES ........................................................................................................... viii
LIST OF FIGURES ........................................................................................................... ix
ABSTRACT……. ............................................................................................................ xvi
CHAPTER 1.
INTRODUCTION ................................................................................. 1
1.1
Determination of Chemical Composition Online...................................... 1
1.2
Determination of Moisture Content and Bulk Density Online ................. 3
1.3
Use of Microwave Sensors in the Pharmaceutical Industry ...................... 4
1.4
Benefits of Using Process Analytical Devices .......................................... 5
1.4.1
Continuous Manufacturing..................................................................5
1.4.2
Improved Process Understanding .......................................................6
1.5
Microwave Theory .................................................................................... 6
1.5.1
Maxwell’s Equations...........................................................................6
1.5.2
Interaction with Materials ...................................................................7
1.6
Microwave Moisture Monitoring .............................................................. 8
1.7
Microwave Density Monitoring ................................................................ 9
1.8
Advantages and Limitations of Microwave Sensors ............................... 11
1.9
Objectives ................................................................................................ 12
1.10
Significance of the Work......................................................................... 12
CHAPTER 2.
DEVELOPMENT OF A DENSITY INDEPENDENT MICROWAVE
MOISTURE CONTENT MODEL ................................................................................... 14
2.1
Objective ................................................................................................. 14
2.2
Materials and Methods ............................................................................ 14
2.2.1
Frequency Range of Sensor ..............................................................16
2.2.2
Determination of Reference Moisture Content by Loss on Drying ..16
2.2.3
Microwave Sensor Measurements ....................................................17
2.2.4
Determination of Moisture Content by Microwave Sensor ..............18
2.3
Results and Discussion ............................................................................ 23
v
Page
2.3.1
Microcrystalline Cellulose ................................................................23
2.3.2
α-Lactose Monohydrate ....................................................................24
2.3.3
Density Independence of Microwave Measurements .......................26
2.4
Conclusions of Study .............................................................................. 28
CHAPTER 3.
MICROWAVE MONITORING OF RAPIDLY FLOWING
POWDERS……… ........................................................................................................... 30
3.1
Objective ................................................................................................. 30
3.2
Materials and Methods ............................................................................ 30
3.2.1
3.3
Experimental Procedure ....................................................................33
Results and Discussion ............................................................................ 35
3.3.1
Bulk Density Determination .............................................................35
3.3.2
Moisture Content Determination ......................................................38
3.4
Conclusions of Study .............................................................................. 41
CHAPTER 4.
MONITORING ROLLER COMPACTED RIBBONS ONLINE
USING MICROWAVES AND NIR ................................................................................ 42
4.1
Objective ................................................................................................. 42
4.2
Roller Compaction Background .............................................................. 42
4.3
Materials and Methods ............................................................................ 43
4.3.1
Preconditioning of Materials .............................................................43
4.3.2
Roller Compaction ............................................................................44
4.3.3
Microwave Measurements ................................................................45
4.3.3.1
Development of Microwave Moisture Model ..................................... 46
4.3.3.2
Development of Microwave Density Model ....................................... 48
4.3.4
4.3.4.1
4.3.5
4.4
NIR Measurements............................................................................49
Development of Partial Least Squares Models.................................... 49
Ribbon Density and Moisture Content Reference ............................53
Results and Discussion ............................................................................ 54
4.4.1
4.4.1.1
Density Monitoring ...........................................................................54
NIR Spectroscopy ................................................................................ 54
vi
Page
4.4.1.2
4.4.2
Microwave Resonance ......................................................................... 56
Moisture Content Monitoring ...........................................................57
4.4.2.1
NIR Spectroscopy ................................................................................ 57
4.4.2.2
Microwave Resonance ......................................................................... 60
4.5
Conclusions of Study .............................................................................. 62
CHAPTER 5.
DEVELOPMENT OF A NOVEL MICROWAVE SENSOR TO
MONITOR CHEMICAL COMPOSITION ONLINE ...................................................... 63
5.1
Objective ................................................................................................. 63
5.2
Materials and Methods ............................................................................ 63
5.2.1
Development of Microwave Sensor ..................................................63
5.2.2
Comparison of Simulation to Realization .........................................68
5.2.3
Preconditioning of Materials .............................................................73
5.2.4
Test Setup ..........................................................................................74
5.2.5
Determination of Reference Moisture Content .................................76
5.2.6
NIR Measurements............................................................................76
5.2.7
Microwave Measurements ................................................................77
5.2.8
Model Development ..........................................................................77
5.3
Results and Discussion ............................................................................ 82
5.3.1
5.3.1.1
NIR Spectroscopy ................................................................................ 83
5.3.1.2
Microwave Resonance ......................................................................... 88
5.3.2
5.4
Determination of Acetaminophen Concentration .............................83
Determination of Moisture Content ..................................................93
5.3.2.1
NIR Spectroscopy ................................................................................ 93
5.3.2.2
Microwave Resonance ......................................................................... 98
Conclusions of Study ............................................................................ 102
vii
Page
CHAPTER 6.
DEVELOPMENT OF A REPLACEMENT ROLLER COMPACTION
SENSOR……….. ........................................................................................................... 103
6.1
Objective ............................................................................................... 103
6.2
Design of the Sensor ............................................................................. 103
6.3
Fabrication of the Sensor....................................................................... 105
6.4
Materials and Methods .......................................................................... 106
6.5
Results and Discussion .......................................................................... 107
6.6
Conclusions of Study ............................................................................ 110
CHAPTER 7.
CONCLUSIONS, INSIGHTS, AND FUTURE WORK................... 111
7.1
Conclusions ........................................................................................... 111
7.2
Insights into Microwave Sensor Design ............................................... 112
7.3
Future Work .......................................................................................... 113
7.3.1
Further Development of Novel Microwave Sensors .......................113
7.3.2
Interactions of Microwave Fields with Particulates ........................114
REFERENCES…. .......................................................................................................... 115
VITA…………… ........................................................................................................... 139
PUBLICATIONS. ........................................................................................................... 140
viii
LIST OF TABLES
Table ..............................................................................................................................Page
Table 1. Linear regressions through sets of different moisture contents in Figure 2........ 21
Table 2. Comparison of the accuracy of different MCC calibrations. .............................. 24
Table 3. Comparison of the accuracy of different α-lactose monohydrate calibrations ... 25
Table 4. Density independence of Meyer and Schilz (Eq. 3) and TRV (Eq. 7) methods . 27
Table 5. Summary of density models ............................................................................... 36
Table 6. Summary of moisture content models ................................................................ 40
Table 7. Operating range and normal operating conditions of various process parameters
........................................................................................................................................... 44
Table 8. Comparison of root mean square error of calibration (RMSEC), root mean
square error of prediction (RMSEP), and coefficient of determination (R2) for the various
models employed. ............................................................................................................. 83
ix
LIST OF FIGURES
Figure .............................................................................................................................Page
Figure 1. Schematic of the microwave setup for static powder measurements. The
powder is enclosed in a sealed acrylic container and placed on top of the microwavesensing surface [51]. ......................................................................................................... 17
Figure 2. Density independent lines of constant moisture content for microcrystalline
cellulose. ........................................................................................................................... 20
Figure 3. Linear relationship between slope and intercept of lines of constant moisture
content from Figure 2........................................................................................................ 21
Figure 4. Comparison of MC determination residuals using the Translated Raw Variable,
Meyer and Schilz, and the Trabelsi Universal Calibration methods on the same set of
MCC data. ......................................................................................................................... 24
Figure 5. Comparison of MC determination residuals using the Translated Raw Variable,
Meyer and Schilz, and the Trabelsi Universal Calibration methods on the same set of αlactose data. ....................................................................................................................... 26
Figure 6. Schematic of microwave sensor and stirring apparatus .................................... 32
Figure 7. Image of stirring apparatus without material to be tested ................................. 32
Figure 8. Quality of density calibration for rapidly stirred Avicel PH105 ....................... 36
x
Figure .............................................................................................................................Page
Figure 9. Relationship between (a) ε” and ε’-1 and (b) gr/fr and (fa/fr)2 for stirred MCC at
five different moisture contents and varied densities. As moisture content increases, the
slope of the resulting line of constant moisture content increases as well. Many more
moisture contents were evaluated than are shown here, for clarity. All other moisture
content curves showed the same trend. ............................................................................. 38
Figure 10. Quality of moisture calibration for rapidly stirred Avicel PH105. Using the
translated raw variable method, the magnitude and spread of the residuals is minimized.
........................................................................................................................................... 39
Figure 11. On-line Sensing Configuration to Monitor Roller Compacted Ribbons. (Blue
Arrow: Roller Compactor, Red Arrow: Microwave Sensor, Green Arrow: NIR Probe) . 46
Figure 12. Comparison of NIR Spectra of Microcrystalline Cellulose Taken at 4.65%
Moisture Content and at Varying Densities (Blue Dash: 1.09 g/cc, Red Solid: 0.977 g/cc,
Green Dot: 0.776 g/cc) ...................................................................................................... 51
Figure 13. Comparison of NIR Spectra of Microcrystalline Cellulose Taken At
Approximately 0.948 g/cc Envelope Density (Green Dash: 5.37 % MC, 0.949 g/cc;
Purple Long Dash: 4.81 % MC, 0.951 g/cc; Blue Solid: 3.39 % MC, 0.947 g/cc; Red Dot:
2.15 % MC, 0.944 g/cc) .................................................................................................... 52
Figure 14. Expanded View of 1440 - 1630 nm Region in Figure 3 to Highlight Most
Significant Region for Moisture Content Determination (Green Short Dash: 5.37 % MC,
0.949 g/cc; Purple Long Dash: 4.81 % MC, 0.951 g/cc; Blue Solid: 3.39 % MC, 0.947
g/cc; Red Dot: 2.15 % MC, 0.944 g/cc) ............................................................................ 53
xi
Figure .............................................................................................................................Page
Figure 15. Density residuals of NIR 2-factor PLS model based on raw spectral data and
density residuals of microwave model .............................................................................. 55
Figure 16. Moisture content residuals of two-factor model based on SNV pretreated data
........................................................................................................................................... 59
Figure 17. Moisture content residuals of microwave model............................................. 61
Figure 18. Top view of planar sensor as fabricated .......................................................... 65
Figure 19. COMSOL simulation results showing how the norm of the electric field varies
along the sensor surface when the bottom ring resonates at approximately 4.3 GHz. The
color-coding shown is on a log base 10 scale. Thus, the field strength around the top ring
is less than 1% of the strength in the bottom ring, indicating that the rings can resonate
independently of one another. ........................................................................................... 67
Figure 20. Comparison of COMSOL simulation results to physical realization .............. 69
Figure 21. COMSOL simulation result demonstrating the norm of the electric field at the
first resonance near 2.1 GHz. ............................................................................................ 71
Figure 22. COMSOL simulation result demonstrating how the norm of the electric field
varies across the sensor surface when neither the top nor the bottom ring resonates at 3.3
GHz. .................................................................................................................................. 72
Figure 23. COMSOL simulation result demonstrating how the norm of the electric field
varies across the sensor surface when the top and middle rings combine their path lengths
to form a figure-eight resonance at 4.8 GHz..................................................................... 73
xii
Figure .............................................................................................................................Page
Figure 24. Testing setup used to develop microwave and NIR sensor calibrations (blue
arrow: vibrating stage, green arrow: microwave sensor on second shelf, orange arrow:
NIR probe). Both shelves were designed to allow their slopes to be adjustable. In this
image, the top shelf is down to demonstrate this. ............................................................. 75
Figure 25. A comparison of NIR spectra with 15-point Savitzky-Golay smoothing and
subsequent first derivative pretreatment applied to them. Spectra were all recorded for
samples with approximately 2.4% moisture content. ....................................................... 78
Figure 26. A comparison of NIR spectra with 15-point Savitzky-Golay smoothing and
subsequent first derivative pretreatment applied to them. Spectra were all recorded for
samples with 25% acetaminophen but at varying moisture contents. .............................. 80
Figure 27. 10-fold cross validation residuals found when employing either the one-factor
PLS NIR model or the four-factor PLS microwave model for acetaminophen
concentration. No discernable bias is seen in the NIR model; however, an apparent bias
can be seen in the microwave model. ............................................................................... 84
Figure 28. Residuals from Figure 27 plotted against their observation number to ease in
readability. ........................................................................................................................ 85
Figure 29. Demonstration of how the RMSEC and RMSEP for both microwave and NIR
artificial neural network models for acetaminophen determination vary with the number
of neurons in their hidden layer. The dashed lines follow the combined RMSE of both
calibration and test sets. .................................................................................................... 86
Figure 30. Residuals found when employing both the NIR and microwave artificial
neural network models for acetaminophen determination. ............................................... 87
xiii
Figure .............................................................................................................................Page
Figure 31. Residuals from Figure 30 plotted against their observation number to ease in
readability. ........................................................................................................................ 88
Figure 32. Effect of acetaminophen concentration on the bandwidth of the microwave
sensor’s first resonance. .................................................................................................... 90
Figure 33. Effect of acetaminophen concentration on the quality factor of the microwave
sensor’s first resonance. .................................................................................................... 91
Figure 34. 10-fold cross validation residuals found when employing either the six-factor
PLS NIR model or the three-factor PLS microwave model for acetaminophen
concentration. .................................................................................................................... 94
Figure 35. Residuals from Figure 34 plotted against their observation number. .............. 95
Figure 36. Demonstration of how the RMSEC and RMSEP for both microwave and NIR
artificial neural network models for moisture content determination vary with the number
of neurons in their hidden layer. The dashed lines follow the combined RMSE of both
calibration and test sets. .................................................................................................... 96
Figure 37. Residuals found when employing both the NIR and microwave artificial
neural network models for moisture content determination. ............................................ 97
Figure 38 Residuals found when employing both the NIR and microwave artificial neural
network models for moisture content determination as a function of observation number.
........................................................................................................................................... 98
Figure 39. Effect of moisture content on the bandwidth of the microwave sensor’s first
resonance......................................................................................................................... 100
xiv
Figure .............................................................................................................................Page
Figure 40. Effect of moisture content on the quality factor of the microwave sensor’s first
resonance......................................................................................................................... 101
Figure 41. COMSOL Multiphysics simulation results indicating reflection coefficient
response as a function of frequency. ............................................................................... 104
Figure 42. Example microwave reflection coefficient spectra taken with a roller
compacted ribbon (15% reference APAP compacted at 30 bar) .................................... 108
Figure 43. Example of microwave phase spectra taken with a roller compacted ribbon (15%
reference APAP compacted at 30 bar) ............................................................................ 109
Figure 44. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a first order
resonance at approximately 2.4 GHz. ............................................................................. 123
Figure 45. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s bottom ring experiences a second order
resonance at approximately 4.3 GHz. ............................................................................. 124
Figure 46. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a second order
resonance at approximately 4.7 GHz. ............................................................................. 125
Figure 47. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s bottom ring experiences a third order
resonance at approximately 6.3 GHz. ............................................................................. 126
xv
Figure .............................................................................................................................Page
Figure 48. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a third order
resonance at approximately 7.0 GHz. ............................................................................. 127
xvi
ABSTRACT
Austin, John S. Ph.D., Purdue University, May 2014. Investigation Into the Use of
Microwave Sensor to Monitor Particulate Manufacturing Processes. Major Professor:
Michael Harris.
Knowledge of a material’s properties in-line during manufacture is of critical importance
to many industries, including the pharmaceutical industry, and can be used for either
process or quality control. Different microwave sensor configurations were tested to
determine both the moisture content and the bulk density in pharmaceutical powders
during processing on-line.
Although these parameters can significantly affect a
material’s flowability, compressibility, and cohesivity, in the presence of blends, the
picture is incomplete. Due to the ease with which particulate blends tend to segregate,
blend uniformity and chemical composition are two critical parameters in nearly all solids
manufacturing industries. The prevailing wisdom has been that microwave sensors are
not capable of or sensitive enough to measure the relative concentrations of components
in a blend. Consequently, it is common to turn to near infrared sensing to determine
material composition on-line. In this study, a novel microwave sensor was designed and
utilized to determine, separately, the concentrations of different components in a blend of
pharmaceutical powders. This custom microwave sensor was shown to have comparable
accuracy to the state-of-the-art for both chemical composition and moisture content
determination.
1
CHAPTER 1. INTRODUCTION
1.1
Determination of Chemical Composition Online
In many solids handling industries, it is very important to know material properties, such
as chemical composition, moisture content, and bulk density, in real time. Chemical
composition, most importantly, affects the uniformity as well as the purity of the product.
This research has a pharmaceutical focus, and in that industry, small changes in
composition can mean the difference between a dose being ineffective, efficacious, and
toxic. Several techniques have been investigated as a means of monitoring the chemical
composition of a particulate blend. Through the use of multivariate statistics, nuclear
magnetic resonance (NMR) has been used to characterize complex mixtures [1].
Hyperspectral imaging has been used to monitor the composition of manufactured food
products for quality and safety control [2].
Solid phase molecular fluorescence
spectroscopy has also been used to determine the chemical composition of complex
mixtures with minimal sample preparation and low cost [3, 4]. However, neither of these
techniques has shared the success of near infrared (NIR) spectroscopy. NIR has become
the method of choice to determine particulate blend composition on-line [5, 6].
Despite its popularity, NIR sensing of particulates suffers from several significant
drawbacks. Most importantly, NIR radiation can only penetrate into a material a very
short distance, on the order of millimeters [7]. This is not a significant drawback in
liquid systems, where solutions can often be assumed well mixed and thus homogenous.
2
Particulate blends can segregate as a result of many different phenomena, most notably
due to differences in particle size and cohesivity [8]. With a penetration depth of only a
few millimeters, segregation can render NIR results to be inaccurate. Other notable
drawbacks of NIR spectroscopy include sensitivity to physical properties, such as particle
size and roughness, and the necessity to develop models using complex chemometric
software [9-11].
Microwave sensors do not suffer from these drawbacks. First, microwave sensors are
able to monitor bulk material properties due to their increased penetration depths [12].
Second, microwave sensors are able to monitor moisture contents with great accuracies,
due to large differences in dielectric properties between solids and polar liquids.
Commonly, empirical and semi-empirical linear regression models are employed to
calibrate microwave sensors. Therefore, complex chemometric software is usually not
needed. Microwave interactions with particulate mixtures can be complex and are often a
function of at least frequency, temperature, moisture content, crystallinity, and
composition. Due to the highly nonlinear behavior of some systems, artificial neural
networks have also been employed to monitor particulate properties with high levels of
accuracy [13].
Microwave rotational spectrometers have been used off-line in laboratories for the
determination of chemical composition of gases [14]. Often, these off-line sensors make
use of highly sensitive Fabry-Perot resonators. The design and use of these resonators
requires that they be very precisely controlled and are thus not suitable for on-line
analysis during particulate manufacture. To the author’s knowledge, microwave sensors
3
have not been used on-line to determine the chemical composition of a particulate
product.
1.2
Determination of Moisture Content and Bulk Density Online
Moisture content and bulk density are important material properties that are often
overlooked. Among other things, they affect a material’s flowability, cohesivity, and
compressibility. Traditional methods cannot measure these properties in real-time and do
not meet the FDA's new Process Analytical Technology guidelines, where the use of the
most up-to-date manufacturing technologies for real time, continuous quality assurance is
encouraged [15, 16]. Deviations from process parameters can adversely affect product
quality and increase processing time.
Most traditional methods to determine moisture content are invasive, time consuming,
and labor intensive.
The most common method to determine moisture content
analytically is through the Loss on Drying (LOD) method, wherein the moisture content
is determined from the weight change of the sample upon evaporation of the absorbed
water in an oven. Other methods include Karl Fischer titration and differential scanning
calorimetry [17, 18]. These methods lack the ability to determine the moisture content of
a sample in situ. Additionally, each of these methods destroys the sample being tested.
Traditionally, density is determined gravimetrically through knowledge of the sample’s
weight and volume. This is very difficult to implement on-line without disturbing the
process.
Other techniques such as NIR spectroscopy and NMR have been used to monitor the
moisture content of a sample in real-time.
However, NMR is often prohibitively
expensive, while NIR techniques suffer from the same issues when trying to determine
4
chemical composition. Microwave sensors can be employed that easily measure these
bulk quantities in real time with minimal error and at a reasonable cost [19-21]. Several
different sensor designs and configurations have been proposed to measure moisture
content and bulk density since their first use in the 1940s. Many studies have been
published investigating microwave sensing of moisture content in static seeds and grains
[13, 22-27]. Others have investigated the use of both portable and satellite microwave
sensing to monitor soil properties [28-30]. In the food industry, they are used to monitor
product quality [31]. Microwaves have also been employed to monitor the curing of
resins [32]. The introduction of microwave sensing to the pharmaceutical industry has
been more recent, driven by new PAT guidelines.
1.3
Use of Microwave Sensors in the Pharmaceutical Industry
The first reported use of microwave sensors to monitor a pharmaceutical fluid bed
process was in a patent published in 2002 by Tondar et al [33]. This patent claimed the
use of microwave resonance sensors to monitor the moisture content in fluidized bed
granulation processes. In their design, a resonant microwave sensor was placed in the
wall of a fluid bed granulator. The sensor can then monitor bulk properties during
processing. This patent is currently held by Glatt.
Subsequently, Buschmüller et al. examined the use of microwave resonance sensors to
monitor a fluidized bed process and published their work in 2008 [34]. They were able to
accurately correlate microwave measurements to reference moisture content values, with
a coefficient of determination of 0.976. In large part, this was due to the use of a densityindependent moisture content function; the independence of which was verified by
varying the air inlet rate while attempting to maintain the moisture content. Using the
5
microwave sensor, they reduced their drying time by 75%. This study demonstrated that
microwave resonance sensors could be used to eliminate the need for periodic
interruptions to determine the moisture content of the bed from inhomogeneous samples.
Building on the previous study into fluidized bed monitoring, Lourenco et al. used
multivariate analysis combined with microwave sensing to better understand fluid bed
wet granulation [35]. They operated at an industrial scale, with over 500 kg of material
per batch. From their analysis, they were able to conclude that granule size can affect the
resonant frequency, and thus property measurements. They were also able to use their
analysis to help predict the trajectory of a batch from initial in situ measurements.
Sung et al. published a study investigating whether or not microwaves could be used to
monitor crystallographically bound water in pharmaceutical excipients, specifically
alpha-lactose monohydrate [36]. The study demonstrated that microwaves could not be
used to monitor water of crystallization. The rotational mechanism was inhibited by the
high degree of hydrogen bonding. Monitoring the transformation of samples of betalactose without any bound water to alpha-lactose over time with a microwave sensor
yielded results similar to powders drying out, even though the moisture content remained
the same.
1.4
Benefits of Using Process Analytical Devices
Once a process analytical technology has been selected, it can provide value information
to both improve process control and understanding.
1.4.1
Continuous Manufacturing
The pharmaceutical industry has long been dominated by batch manufacturing, while
traditional chemical and pharmaceutical manufacturing is usually performed using
6
continuous manufacturing techniques. Batch manufacturing allows easy separation of
lots, which can be beneficial to track and segregate problems.
Furthermore, batch
manufacturing allows easy sharing of process equipment, which can significantly reduce
capital expenditures. However, there are many benefits to continuous manufacturing that
may outweigh the benefits of operating in under batch.
For instance, one of the largest problems encountered during batch manufacturing occurs
during scale up of the process to production scale. Often increasing the volume of
product requires nonlinear changes to process variables. This can lead to longer lead
times to market.
One can avoid many of these issues by switching to continuous
manufacturing [37]. To operate in continuous manufacturing, the process needs to be
instrumented and controlled. Furthermore, through proper instrumentation, faults and
out-of-spec conditions can be detected before they contaminate an entire batch.
1.4.2
Improved Process Understanding
Both continuous and batch processes can benefit from process analytical devices. When
a problem occurs, if the process has been fully instrumented, one can often use the data to
diagnose what occurred and use the data to predict or illuminate future faults. If however
the process has not been instrumented, lengthy tests need to be done off-line. Since the
exact conditions cannot be reproduced offline, diagnosis is not always guaranteed.
1.5
1.5.1
Microwave Theory
Maxwell’s Equations
Maxwell’s equations are shown in Eqs. (1)-(4), where E is the electric field intensity, D is
the electric flux density, H is the magnetic field intensity, B is the magnetic flux density, t
is the unit of time, r is the position vector, J is the current density, and ρv is the charge
7
density. These combined equations form the foundation for electromagnetic field
analysis. The solution of these equations can describe the complete electromagnetic field
in a space at any given instant.
(1)
(2)
(3)
(4)
1.5.2
Interaction with Materials
The relative permittivity of a material can concisely describe how a material interacts
with an electric field at a particular frequency. It is a complex quantity given by Eq. (5):
(5)
where j2 = -1. The ability of the material to store electrical energy is described by the
dielectric constant, ε'. Microwaves are sensitive to the presence of small quantities of
water in solids due to a large difference in their dielectric constants. At 20 °C and 700
MHz, the dielectric constant of water is approximately 80 [38]. At the same temperature
and frequency, the dielectric constant of powdered MCC is approximately 1.6 [39]. Thus,
small quantities of water can be detected when added to MCC due in large part to a
significant change in the effective dielectric constant of the bulk material. The loss factor,
ε", describes the material's ability to dissipate electrical energy in the form of heat. For a
vacuum, the relative dielectric constant and loss factor are one and zero, respectively. The
complex permittivity is a function of several parameters, most notably: chemical
composition, moisture content, density, temperature, and frequency [40].
8
Microwave radiation is typically considered to range in frequency from 300 MHz to 300
GHz. In this frequency range, microwaves lack the quantum energy to excite chemical
bonds or ionize molecules. Microwaves primarily interact with a material through two
mechanisms. The force of the applied electric field induces a dipole in the molecules of
the material. The extent of dipole formation is heavily dependent on the chemistry of the
molecules. Additionally, as the electromagnetic waves propagate and the electric field
reverses direction, the field attempts to rotate the molecule by acting on a permanent or
induced dipole.
Although it is impossible to directly measure interactions of microwaves with materials,
when sensing one monitors how either a reflected or transmitted signal is altered by the
presence of a material and uses this information to infer properties of the material. This
is also how NIR and many other high frequency electromagnetic sensing techniques are
employed.
1.6
Microwave Moisture Monitoring
Moisture content is the most common material property determined using online
microwave sensors, due to the large contract between the dielectric properties of most
liquids and solids. A moisture content model suitable for online use in a particulate
manufacturing environment must be independent of density. Often a density independent
parameter is sought that is linearly proportional to moisture content.
Two of the most common calibrations of this type were proposed by Meyer and Schilz
[41] and by Trabelsi et al. [40, 42].
Meyer and Schilz observed that a material’s
dielectric constant (ε’) and loss factor (ε”) are commonly functionally dependent on
density in the same way. Thus, their ratio (ε”/ε’) is independent of density and is solely a
9
function of moisture content. Commonly, the ratio ε”/(ε’-1) is used; this form allows
calibration curves to intersect at the origin, representing a point of zero density. Trabelsi
et al. proposed a universal calibration function of the form of Eq. (6). Furthermore, they
showed that the density independent parameter Ψ is a linear function of moisture content.
√
(6)
Both of these parameters require the determination of a material’s dielectric properties
prior to the determination of moisture content.
It was determined that this step
introduced unnecessary errors into a calibration. Thus, it was necessary to develop a
more accurate moisture model that would rely solely on raw sensor measurements.
Moisture content can also be determined from a polynomial surface with fitted constants.
In this dissertation, a polynomial 2,1 model was used for comparison. This nomenclature
indicates that the polynomial surface is second order in (ε’-1) and first order in ε”, as
shown in Eq. (7). The fitted constants were determined by MATLAB’s surface fitting
toolbox.
(7)
1.7
Microwave Density Monitoring
Several methods to determine density from microwave measurements have been reported
[43-45]. One is the density compensated method demonstrated by King [46]. In this
method, the mass of the dry powder and the mass of the water are both thought to
contribute linearly to the values of (fa/fr)2 and (gr/fr) through Eqs. (8) and (9). fa is the
resonant frequency of the microwave sensor in with no material overlaid; fr is the
10
resonant frequency with material present; and gr is the conductance of the sensor at
resonance with material overlaid.
( )
(8)
(9)
By solving the linear regression problem, the above constants can be obtained; from the
rearranged equations, the powder and water densities, and thus the bulk density, can also
be obtained.
(( )
)
(
)
(10)
(( )
)
(
)
(11)
(12)
Additionally, the moisture content can be determined simultaneously using this model.
Moisture content on a wet basis can be found from Eq. (13).
(13)
Another method to determine density was published by Trabelsi et al. By normalizing
the relative complex permittivity with respect to density, a linear regression can be found
whose slope and intercept are nearly independent of MC and density [47]. This linear
regression takes the form of Eq. (14).
11
(14)
If only one frequency is used, the slope and intercept of this line are sufficient to
determine the bulk density. However, with a resonant sensor, the frequency at which the
measurements are taken varies with moisture content and bulk density. As suggested by
Trabelsi et al., a linear relationship was used to compensate for the slope’s dependence on
frequency.
By substituting the proportional raw data variables for the complex
permittivity and by assuming linear frequency dependence, the bulk density can be found
from a least squares fit to Eq. (15).
( )
(15)
Under rapid stirring, it was assumed that the powder uniformly distributed itself inside of
the cylinder. By definition, density is found by dividing the mass of powder by the
volume it occupies. The total volume was found by subtracting the volume of the stirrer
from the volume of the cylinder. As will be shown, results indicate that a uniform
distribution of particles while stirring is a reasonable approximation.
1.8
Advantages and Limitations of Microwave Sensors
Compared to many other process analytical technologies, microwave sensors offer
several key advantages. Firstly, microwave sensors are capable of taking very rapid
measurements, as is demonstrated in CHAPTER 3. One of the main reasons why
microwave sensors are used to monitor moisture content is that there is a large difference
in the dielectric properties of liquids and solids in the microwave region. Microwave can
interact with molecules through polarization and rotation, if they have enough freedom to
rotate in their current state. Liquids are able to rotate in microwave fields, whereas solid
12
molecules cannot. Furthermore, microwave sensors are relatively inexpensive process
monitoring devices.
There are a few drawbacks, however. For instance, microwave sensors often have to be
tailored to each application to achieve the highest level of accuracy.
The biggest
drawback though has been that microwaves had only really been employed to monitor
moisture content and bulk density, whereas other techniques such as NIR sensing have
been used to measure a variety of other material parameters.
1.9
Objectives
This study has three main objectives: (1) to incorporate a microwave sensor into a
process line to determine process parameters in real-time, (2) to design novel microwave
sensors capable of measuring chemical composition, and (3) to develop more useful
calibration and detection schemes from first principles that allow for more accurate
determination of process parameters.
1.10 Significance of the Work
There are numerous process analytical technologies available to process engineers today.
In particulate manufacturing, NIR spectroscopy is commonly employed, due to the
versatility of the probes and the numerous properties that can be measured by it.
However, this boon can also be a significant hindrance. The ability to measure many
properties simultaneously conversely means that measurements can be deteriorated by
variations in properties not of interest.
Microwave analytical devices are very sensitive to only a few properties of interest. Prior
to this study, microwave sensors had been primarily used to monitor either the moisture
content or the bulk density of a material. To measure the critical attribute of chemical
13
composition necessitated the use of an NIR probe. By expanding the use of microwave
sensors to other particulate processes and to monitoring chemical composition, one no
longer needs to use NIR probes and does not have to be concerned with complex
chemometric software, poor sampling, nor interference with material properties not of
interest.
There exists little diversity in the types of research studies completed using microwave
sensors. In particular, most studies completed thus far have focused on determining
either the moisture content or the bulk density of static or slow moving materials [48].
This has fostered the creation of many academic papers, but most with little practical use.
To be of practical use, research on microwave sensors needs to move from controlled
laboratory environments to pilot- or full-scale processing lines. Some work had been
done in this field; however, there was still much important research to be done.
14
CHAPTER 2. DEVELOPMENT OF A DENSITY INDEPENDENT MICROWAVE
MOISTURE CONTENT MODEL
2.1
Objective
Microwave sensor calibrations to determine moisture content often make use of universal
parameters that can easily transfer between formulations or sensors. However, these
methods sacrifice accuracy for versatility.
In sensing, the most accurate method is
usually sought. In this study, an attempt was made to develop a novel calibration scheme
to provide the highest level of accuracy.
2.2
Materials and Methods
An MDA-1000/MDA-1100L open-reflection microwave resonance sensor was fabricated
for this study by the KDC Technology Corporation in Livermore, California and is now
commercially available. Resonance sensors have been shown to be more accurate than
non-resonance sensors in determining bulk properties of materials [49]. The sensor was
calibrated by KDC using dielectric standards. The sensor determines its resonant
frequency under load as well as the amplitude in decibels of the reflection coefficient (Γ).
Each sensor must be designed for the material type and moisture content range it will be
used for. If designed correctly, maximum power transfer to the sensor can be achieved;
this allows critical coupling to occur, which in turn allows very precise prediction of the
resonant frequency. Critical coupling occurs when the relative conductance, given by g in
Eqs. (16)-(18), is equal to one.
15
(16)
(17)
| |
( )
(18)
Near critical coupling, the material's dielectric constant and loss factor can be determined
from measurements of the sensor's resonant frequency and conductance at resonance
under load from Eqs. (19) and (20):
((
(
)
)
(19)
)
(20)
where Co/k and Kε" are calibration coefficients specific to the sensor, fa is the resonant
frequency in air without the test material (approximately 737 MHz), and fr is the resonant
frequency of the sensor with the test material overlaid on the sensor aperture, gr is the
normalized conductance with the test material overlaid on the sensor aperture, and ga is
the normalized radiation conductance in air without material present [50].
When a material to be tested is overlaid on the sensing surface, the resonant frequency of
the sensor decreases and the bandwidth of the response curve decreases. Additionally, if
the sensor is designed correctly for the application, the relative capacitance of the sensor
should approach one.
16
2.2.1
Frequency Range of Sensor
The sensor used operates in the 600-750 MHz region. Within this frequency range, the
dielectric constant and loss factor of water remain relatively constant, simplifying the
analysis. Moreover, the frequency range used in this study is above that used for
detection of ionic compounds, minimizing errors from the presence of ionic impurities
[48]. Lastly, at this frequency, the wavelength of the microwaves is significantly larger
than the size of the material being tested, which helps to minimize the effects of
scattering. For these reasons, this frequency range is well suited for this application.
2.2.2
Determination of Reference Moisture Content by Loss on Drying
To determine a reference moisture content of the powder tested, three small (2-4 gram)
quantities of powder were taken from three different locations in the powder bed after
testing. After weighing the samples before and after baking at 205 ± 10 °F with MCC
(235 ± 10 °F with α-lactose) for 36 hours, the moisture content on a wet basis was
calculated from Eq. (21).
(21)
Where mw is the mass of water and mp is the mass of dry powder. The subsequent
average of the three samples was taken to be the reference moisture content of the
powder. The accuracy of using an oven for LOD was verified by thermogravimetric
analysis using a TA Instruments Q500.
Additionally, conventional oven LOD
measurements closely matched those obtained using a Mettler Toledo HG63.
17
2.2.3 Microwave Sensor Measurements
Powder, either MCC (Avicel PH105 obtained from the FMC Biopolymer corporation in
Philadelphia, PA) or α-lactose monohydrate (Sigma-Aldrich in St Louis, MO), was
prepared two days prior to testing according to the method presented in Sung et al [36].
Prior to testing the material, an acrylic sample holder obtained from McMaster-Carr in
Santa Fe Spring, California was placed around the sensing area as shown in Figure 1.
Figure 1. Schematic of the microwave setup for static powder measurements. The
powder is enclosed in a sealed acrylic container and placed on top of the microwavesensing surface [51].
A frequency sweep was completed in air to establish reference values. The powder to be
tested was then sifted (mesh size approximately 1 mm) into the sample holder until it was
full. Sifting the wetted powders avoids granule formation and significant particle size
variations that could cause local density fluctuations. The powder must come into direct
contact with the sensor surface.
The sifting allowed the powder to fill the container
uniformly at a low density. Excess powder was then carefully removed and an acrylic lid
18
was placed on top of the powder to ensure more accurate volume measurements as well
as a sealed environment.
The first set of resonant frequency and conductance measurements was taken with the
powder completely filling the acrylic cylinder. The container was then gently tapped
from each of the sides to help settle and densify the powder. The decrease in height of
the powder was measured from at least four locations around the rim using a caliper
(ProMax S225 from the Fowler Company in Newton, MA). This process was repeated
until the powder would not densify further. All experiments were conducted in a climatecontrolled area. The temperature was maintained between 21-24 °C. Considerable care
was taken during measurements to minimize contact with air. Most notably, the material
under test was contained in a closed acrylic cylinder during measurements, minimizing
the effect of ambient humidity changes. Each set of measurements consisted of five data
points for each moisture content/density combination and were recorded to an external
compact flash card (SanDisk Ultra CompactFlash 4 GB).
2.2.4
Determination of Moisture Content by Microwave Sensor
Although many methods to determine moisture content from microwave measurements
have been proposed, moisture content is often determined through a density-independent
method. As previously mentioned, the Meyer and Schilz method is commonly used.
This method has been shown to be accurate under ideal conditions; however, the model
relies on the assumption that the real and imaginary components of the relative complex
permittivity have identical dependencies upon density. Thus, when this assumption is
valid, the ratio of the two parameters should be nearly independent of density. One
significant benefit to using this or other methods based on the material's dielectric
19
properties is that calibrations much more easily transfer between sensors. This method
has two significant drawbacks, however. First, the assumption of identical dependencies
on density is not always a good assumption. It has been shown that at different regimes
of moisture content, the microwave response of a material may vary depending on which
physical phenomena dominates the interactions with the electromagnetic field [27]. Nonlinearity in the behavior of ε"/(ε-1) may not be observed within the range of moisture
contents to be monitored; however, if significant nonlinearities exist between the range to
be monitored and 0% MC, as is sometimes the case, the Meyer and Schilz method cannot
be used accurately. Additionally, precise knowledge of the fitting parameters for ε” and ε’
is required to ensure accurate results, and extra time and care are needed to build a
second calibration set using dielectric standards.
It can be argued that the intrinsic dielectric properties better represent the interaction of
the material and the electromagnetic radiation and thus are more desirable parameters to
use for a calibration. However, as the dielectric properties can only be determined
indirectly through their own calibration equation, this will inevitably introduce some
errors, even if only slight under ideal conditions. Additionally, only effective dielectric
properties can be measured. Therefore, it is desirable to build a calibration that does not
rely on the determination of the real and imaginary components of the material's
permittivity as an intermediate step prior to the determination of moisture content.
To avoid these issues, a different method was proposed. As is demonstrated later in
subsection 2.3.3, both the ratios (gr/fr) and (fair/fr)2 have similar dependencies on density.
Thus, a plot can be created using these parameters that shows lines of constant moisture
content independent of density (Figure 2) that follow the form of Eq. (22). As with
20
traditional ε"/(ε'-1) calibration curves, increased MCs and densities resulted in increased
x- and y-coordinate values. Data taken with constant moisture content were found to be
linear (Table 1).
( )
(22)
0.002
0.0018
0.0016
gr/fr
0.0014
0.0012
0.001
1.05
1.055
1.06
1.065
1.07
1.075
1.08
1.085
1.09
(fa/fr)2
6.163% MC
4.949% MC
3.978% MC
2.528% MC
0.584% MC
Figure 2. Density independent lines of constant moisture content for microcrystalline
cellulose.
21
Table 1. Linear regressions through sets of different moisture contents in Figure 2.
Moisture Content Equations of Regression Lines Coefficient of Determination
6.16 %
4.95 %
3.98 %
2.53 %
0.58 %
y = 0.0148x - 0.0142
y = 0.0136x - 0.0130
y = 0.0124x - 0.0118
y = 0.0102x - 0.0096
y = 0.0078x - 0.0072
0.9994
0.9995
0.9998
0.9992
0.9998
Traveling along a curve of constant moisture content in the negative x-direction,
consecutive points represent powders of the same moisture content but with decreased
densities. This is true for all curves of constant moisture content. Thus, out of necessity,
all of the curves of constant moisture content must intersect at one point, representing
zero density. It is thus necessary to find this point.
-0.006
-0.008
Intercept
y = -1.0074x + 0.0007
-0.01
-0.012
-0.014
-0.016
0.006
0.01
0.014
0.018
Slope
Figure 3. Linear relationship between slope and intercept of lines of constant moisture
content from Figure 2.
22
By plotting the intercepts of these fitted lines against their slopes (Figure 3), another
linear relationship is found with a very high coefficient of determination (r2 greater than
0.99999) that follows the form of Eq. (23).
(23)
The negative slope of this regression line represents the (fair/fr)2 value at which all of the
constant moisture content lines intersect, while the y-intercept of this line represents the
(gr/fr) value at which the constant moisture content lines intersect. Thus, a translation can
be made where all of the lines of constant moisture content shown in Figure 2 intersect at
the origin and follow the form of Eq. (24).
(
( ) )
(24)
As can be seen in Figure 2, the only differences between lines of constant moisture
content are their slopes. Furthermore, the moisture content is a linear function of the
slope, A. The moisture content can thus be determined from Eq. (25) by fitting two
additional constants, c3 and c4. For clarity, the use of the method described in this section
will be known as the Translated Raw Variable (TRV) method in later sections of this
paper.
(
( )
(25)
)
23
2.3
2.3.1
Results and Discussion
Microcrystalline Cellulose
From tests carried out on MCC (MC range 0.568-6.163% and density range 0.280-0.499
g/cm3), a least squares best fit was found to a set of 96 data points. A comparison of the
average relative errors and standard error of calibrations (SECs) of four different
moisture prediction models can be seen in Table 2. For further information regarding
Trabelsi's Material Independent method, see Trabelsi et al. [21]. Relative error and SEC
were calculated from the equations below and were found by applying the calibrated
models to all of the points collected. The relative error of each data point collected is
calculated from Eq. (26). In the following sections, the average relative error reported is
the average of the absolute values of the relative errors for the entire set of data points.
Also shown are the 95% confidence bounds on the relative error.
(26)
√
(27)
Over this moisture content and density range, the TRV method was found to predict an
arbitrary MCC sample’s moisture content more accurately compared to other methods.
The relative errors found with the TRV method and the Meyer and Schilz method at each
data point was compared. Of the 96 points, 66 points exhibited a lower relative error
with the TRV method, corresponding to 69 % of the total. A comparison of the residuals
of each method is shown in Figure 4. As can be seen, both the magnitude and the
variance of the data are minimized when using the TRV method.
24
Microwave Moisture Content Residuals
Table 2. Comparison of the accuracy of different MCC calibrations.
Average Relative
Standard Error of
Model
Error
Calibration
Translated Raw Variable Method
3.71 ± 0.74 %
0.133 %
Meyer and Schilz Method
6.69 ± 1.02 %
0.197 %
ε"/((ε')1/2-1)
8.95 ± 1.47 %
0.253 %
Trabelsi's Material Independent Method
9.23 ± 2.32 %
0.311 %
0.6
0.4
0.2
0
0.00%
-0.2
1.00%
2.00%
3.00%
4.00%
5.00%
6.00%
7.00%
-0.4
-0.6
-0.8
-1
Moisture Content Determined by LOD
TRVM
Meyer and Schilz
Trabelsi Universal Method
Figure 4. Comparison of MC determination residuals using the Translated Raw Variable,
Meyer and Schilz, and the Trabelsi Universal Calibration methods on the same set of
MCC data.
2.3.2 α-Lactose Monohydrate
α-lactose was tested in the MC range 5.59-8.69% and the density range 0.346-0.691
g/cm3. Seventy-one data points were collected and used for this analysis. Moisture
contents below 5 % were not chosen because below this point water in α-lactose exists
only as crystallographically bound water that is hard to detect with a microwave sensor.
25
The ability of both models to predict moisture content is demonstrated in Table 3,
alongside that of Trabelsi's universal calibration.
With α-lactose, the TRV method
possesses a lower average relative error and a reduced SEC within this MC and density
range. Of the 71 points evaluated, 49 points exhibited a lower relative error with this
method, corresponding to 69 % of the total. α-lactose results were chosen to graphically
compare the traditional Meyer and Schilz method presented in this paper. The magnitude
and spread of the residuals in Figure 5 were minimized when using the TRV method.
Table 3. Comparison of the accuracy of different α-lactose monohydrate calibrations
Average Relative
Standard Error of
Model
Error
Calibration
Translated Raw Variable Method
2.13 ± 0.36 %
0.203 %
Meyer and Schilz Method
2.98 ± 0.50 %
0.270 %
Trabelsi's Material Independent Method
2.80 ± 0.60 %
0.301 %
26
Microwave Moisture Content Residuals
0.8
0.6
0.4
0.2
0
5.0%
5.5%
6.0%
6.5%
7.0%
7.5%
8.0%
8.5%
9.0%
-0.2
-0.4
-0.6
-0.8
Moisture Content Determined by LOD
TRVM
Meyer and Schilz
Trabelsi Universal Calibration
Figure 5. Comparison of MC determination residuals using the Translated Raw Variable,
Meyer and Schilz, and the Trabelsi Universal Calibration methods on the same set of αlactose data.
2.3.3
Density Independence of Microwave Measurements
The ratio of ε" to ε' has been found to be nearly independent of density and linear over
small moisture content ranges [41]. This is due to the observation that both ε' and ε" both
depend on density in a similar way for many materials. Similarly, the ratios (gr/fr) and
(fair/fr)2 can be shown to have similar dependencies on density. This has been observed
for both MCC and α-lactose by Austin et al.; the Trabelsi method can also be shown to be
nearly independent of density [51].
27
The improved accuracy of the Translated Raw Variable method can be explained by
increased density independence. As can be seen in Table 4, lines of constant moisture
content fit using Eq. (25) consistently exhibit larger coefficients of determination than
those fit using the Meyer and Schilz method. The increase is not always large, but is
consistently higher. The only difference between consecutive points on lines of constant
moisture content is the densities they represent.
Table 4. Density independence of Meyer and Schilz (Eq. 3) and TRV (Eq. 7) methods
Coefficient of
Coefficient of
Moisture
Confidence Level of
Determination
Determination
Content
Heteroscedasticity
Using Meyer and Schilz
Using TRV
6.16 %
0.99934
0.99941
98.5 %
5.92 %
0.99904
0.99923
99.6 %
4.95 %
0.99928
0.99946
97.7 %
4.27 %
0.99893
0.99920
99.2 %
3.98 %
0.99963
0.99981
97.9 %
3.45 %
0.99988
0.99991
97.4 %
2.53 %
0.99868
0.99917
76.7 %
1.99 %
0.99928
0.99950
86.5 %
0.58 %
0.99926
0.99977
95.4 %
To test whether the difference in coefficients of determination was statistically significant,
a series of F-tests were performed that compared the variances from each set of moisture
content data using both the TRV method and the Meyer and Schilz method. Nearly every
test showed that the variance of the TRV method was statistically different from the
28
variance of the Meyer and Schilz method at a 95% confidence level. Only two tests
failed to show the residuals were heteroscedastic at that confidence level.
Furthermore, because both models assume a linear form, independent of density, both
calibrations should yield lines of constant moisture content that intersect at a single point.
The accuracy of this assumption can be measured by the coefficient of determination of
an equation relating each of the resulting line's slope to its intercept, as in Eq. (23).
Using the moisture contents listed in Table 4, this coefficient of determination from the
Meyer and Schilz method was found to be 0.89580; from Eq. (25), it was found to be
0.99997. The linear assumption of the Meyer and Schilz method may not hold outside of
the density range tested when using the ratio of ε" to ε'-1. By comparing the coefficients
of determination from the two models, it was reasoned that lines of constant moisture
content found from the TRV method, could be assumed independent of density over a
greater range than those found with the Meyer and Schilz method for the materials tested.
For these reasons, it was concluded that the ratio of (gr/fr) to (fair/fr)2 was more
independent of density than was the ratio of ε" to ε'-1 when applied to wetted MCC and
α-lactose particles.
2.4
Conclusions of Study
The method presented in this study was able to determine accurately the moisture
contents of many different samples of MCC and α-lactose. When applied to wetted MCC
and α-lactose, the translated raw variable method consistently offered both the lowest
relative error and SEC. Additionally, the method was shown to be more independent of
density than the method presented by Meyer and Schilz, already shown to be nearly
independent of density, when applied to MCC and α-lactose. As a result, the TRV
29
method has been shown to be a viable alternative to other common microwave calibration
schemes. This method can be used to reduce measurement error when microwave-based
resonator
sensors
are
used
to
monitor
the
moisture
content
of
powders.
30
CHAPTER 3. MICROWAVE MONITORING OF RAPIDLY FLOWING POWDERS
3.1
Objective
In pharmaceutical manufacturing, it is often necessary to monitor particulate properties
while the material is in motion, such as during fluid bed granulation, high shear mixing,
and roller compaction. The complexity of particulate flow makes this type of monitoring
difficult for most sensing methods. A study was initiated to investigate if microwaves
could be used to effectively monitor the moisture content and bulk density of rapidly
flowing powders.
3.2
Materials and Methods
In order to analyze powders while they are in constant rapid motion, an experimental
apparatus (Figure 6) was constructed in which powders are uniformly mixed above the
sensing surface. The cylinder used to contain the powder was made from acrylic and had
an inner diameter of 123.1 mm, a wall thickness of 4.8 mm, and a height of 70.0 mm.
Stainless steel, which is most commonly used to fabricate impeller and mixing blades,
was not used; the high conductivity of metals in close proximity severely influences
electromagnetic fields. In order to study the ability of a microwave sensor to monitor
bulk material properties of rapidly flowing powders, care was taken to ensure that the
powder in the local vicinity of the sensor surface was in rapid motion. The mixer was
placed approximately one millimeter above the sensing surface. This close proximity
prohibited the use of stainless steel. Instead, a stirring shaft and blades were made from
31
nylon and PVC, respectively. Unlike metals, these materials exhibit very small dielectric
losses and have dielectric constants near to those of many solid excipients. To use a
microwave sensor in a real piece of process equipment made from metal, such as a high
shear granulator, stainless steel impeller blades would need to be located outside of the
sensing area. For instance, the sensor should be placed on the wall of the vessel, instead
of the bottom where the impeller is spinning. In a high shear granulator, this is also the
most desirable position for sensing, as it allows effective monitoring while the powder is
flowing in a spiral pattern, which effectively transmits shear stresses through the whole
wet-mass [52]. The sensitivity of planar resonance sensors, such as the one used in this
study, drops off rapidly with distance. The minimum distance required between any metal
impellers and a microwave sensor will change between geometries and operating
frequencies, but will usually be on the order of centimeters when material is present. The
length of each of the stirrers seen in Figure 7 was 119.0 mm. Thus, the ratio of the
stirrer’s length to the inner diameter was 0.967. The bottom stirrer blade was made from
26.67 mm outer diameter schedule 40 PVC pipe. The top stirrer blade was made from
17.14 mm outer diameter schedule 40 PVC pipe. The pipes were cut to create a blade, as
shown in Figure 7. The axis of rotation of the impeller blades was positioned exactly over
the center of the sensor.
32
Figure 6. Schematic of microwave sensor and stirring apparatus
Figure 7. Image of stirring apparatus without material to be tested
To further ensure that the powder at the very bottom of the sample was mixed, a low
dielectric loss brush was attached to the bottom of the stirrer. This 6.35 mm long brush
33
swept the powder that was just on top of the microwave sensor without contacting the
sensor surface, ensuring there were no dead-zones at the sensor surface. As a driving
force to rotate the stirrer, a commercial drill press (Ryobi model number DP121L) was
used because it had a constant speed motor that had high initial torque and was easily
adjustable. Several experiments were conducted, and it was concluded that the PVCNylon stirrer did not appreciably affect the sensor readings when it was rotated with the
drill press directly above the sensor. Thus, air reference values could be taken while the
stirrer was in motion above the sensor surface in the absence of powders.
3.2.1
Experimental Procedure
Three hundred gram quantities of MCC were either dried or humidified to achieve
varying moisture contents by the method presented in Sung et al [36]. The moisture
content of prepared powders ranged from 0.27 - 6.37 % on a wet basis. Above a certain
moisture content, when MCC granules absorb moisture, they swell. This allows water to
infiltrate previously unavailable regions of the particles. In these regions, the water
becomes tightly bound to anhydroglucose units [53, 54]. Moisture contents up to 10.2 %
were tested; however, above 6.4 % MC no model tested could accurately predict the
moisture content within the frequency range used. This was concluded to be a result of
the loss of mobility of the excess water over 6.4 % MC, as also reported previously [39].
Though this limits the use of microwaves to monitor only the unbound moisture content
of microcrystalline cellulose systems, it should not be a detriment when monitoring wet
granulation processes. In wet granulation, the water needed to be monitored as a process
variable is free, surface water and can easily be detected by microwaves. Microwave
sensing in general is not limited by the amount of water present; microwave sensors have
34
been employed to measure the dielectric properties of pure water as well as other liquids
[55]. Microwave sensing, however, is usually limited to sensing only unbound water.
Once the samples had been prepared, they were placed in two layers of sealed Ziploc
bags for two days to ensure uniform distribution of water and thermal equilibrium. After
these two days, tests on the powders were conducted. First, air reference values were
taken with the empty cylinder and stirrer in place; this testing arrangement is shown in
Figure 7. Then, 180.0 grams of microcrystalline cellulose (Avicel PH105 obtained from
the FMC Biopolymer Corporation in Philadelphia, PA, USA) was added to the
cylindrical acrylic sample holder (obtained from McMaster-Carr in Santa Fe Spring, Ca,
USA). The drill press was turned on, stirring the powder at 800 RPM. The resulting
angular velocity corresponded to a maximum tangential velocity of 5.1 m/s. The most
sensitive region of the sensor coincided with this high-speed region. After a few seconds,
the resonant frequency and conductance measurements would stabilize. Data was then
recorded to an external compact flash card (SanDisk Ultra CompactFlash 4 GB) at a rate
of approximately eight measurements per second. After 10 seconds of recording data, the
drill was turned off. For analysis, the data collected over each 10 second interval was
averaged together to yield one data point. Twenty grams of MCC were then added and
more measurements were taken. For each prepared sample at a particular moisture
content, this process was repeated four times more until six different densities of powder
had been tested.
Experiments were carried out in a climate-controlled environment. The temperature was
maintained between 21-24 °C. Temperature fluctuations of this magnitude do not
significantly affect the dielectric properties of a microcrystalline cellulose-water system.
35
Complications due to humidity fluctuations were minimized by both storing and
conducting tests in sealed environments.
3.3
3.3.1
Results and Discussion
Bulk Density Determination
As can be seen in Table 5, the microwave sensor was able to detect small changes in
density even when the powder was flowing rapidly. One hundred stirred powder samples
of varying moisture contents in the density range of 0.225 – 0.375 g cm-3 were
investigated in this study. Many experiments were carried out to ensure that the results
were statistically significant and reproducible. Relative error and standard error of
calibration were found by applying the calibrated model to the data points collected.
Both the density compensated model as well as the normalized relative complex
permittivity with frequency compensation method provided good fits to the experimental
data. Plots of the residuals found with both methods can be compared in Figure 8. There
is a discernible pattern present in the residuals found from the normalized method not
present in the other method’s residuals. This pattern arises from the use of identical 20gram additions between all tests (i.e. for each data set at a particular moisture content,
there were always 180-, 200-, and 220-gram measurements taken). The normalized
method employs a single linear regression to find the bulk density, while the density
compensated method combines multiple equations to arrive at its predicted density value.
The greater complexity of the density compensated method increased the spread of the
residuals compared to the simpler normalized method.
36
Table 5. Summary of density models
Model
Average Relative Error Standard Error of Calibration
Normalized Model
4.1 %
0.0139 g cm-3
Density Compensated
4.6 %
0.0156 g cm-3
0.04
Density Residual (g cm-3)
0.03
0.02
0.01
0
-0.01
-0.02
-0.03
-0.04
-0.05
0.15
0.2
0.25
0.3
0.35
0.4
0.45
Predicted Density (g cm-3)
Compensated Density
Normalized
Figure 8. Quality of density calibration for rapidly stirred Avicel PH105
These results indicate that the bulk density of rapidly flowing powders can be monitored
with the same degree of accuracy as can be the bulk density of static powders, using
microwaves. For instance, using the normalized model, the SEC for monitoring the bulk
density of rapidly flowing powders was found to be 0.0139 g cm-3, compared to 0.0167 g
cm-3 for static tests on the same material. The slightly reduced accuracy in the static case
37
is a result of small errors made determining the exact free surface of the reference static
beds.
An analysis of a moving material using an NIR probe usually exhibits significantly
different variability than one conducted on a static sample [56]. Undesired motion during
processing, such as from equipment vibration, can adversely affect NIR results.
Additionally, temperature variations during processing can lead to abnormal spectra.
From the similarity of the SECs for both in-situ and off-line results, it is apparent that this
limitation is not present when using microwaves. This is primarily a result of three
important features of microwave sensing. First, the wavelength of microwaves is
relatively long, especially compared with infrared waves; this significantly reduces the
effects of scatter. Second, modern vector network analyzers can complete frequency
sweeps around the resonant frequency much faster than NIR probes can complete a single
scan. A microwave instrument can complete several hundred complete sweeps in one
second. Over the fraction of a second that it takes to complete one measurement, the
material barely changes position. For instance, the maximum tangential velocity of 5.1
m/s seen in this study would correspond to the powder moving a distance only on the
order of millimeters during a single sweep. Compared to the size of the sensing area, this
is miniscule. Third, microwaves can penetrate into a material a distance on the order of
centimeters. This penetration helps to significantly reduce wall effects that hinder higher
frequency detection schemes. For these reasons, microwave sensors represent an
attractive alternative to other established methods, such as NIR analysis.
38
3.3.2
Moisture Content Determination
Linear relationships were found between ε” and ε'-1, as well as between gr/fr and (fa/fr)2
(Figure 9). Both models fit the data very well, with average coefficients of determination
of 0.9973 and 0.9983 for the Meyer and Schilz and TRV methods, respectively. As
expected, both x- and y-coordinate values increased with either increasing moisture
content or density in both figures.
Even though the lines of constant moisture content in Figure 9 (a) have been shown to be
nearly independent of density, they do not intersect at the origin. When all of the powder
is removed, the values of ε” and ε’ should be zero and one, respectively. This discrepancy
is likely due to the more complex flow system used in this study compared to standard
static measurements. Though some error may be introduced from local temperature
deviations, the majority of the error was likely introduced by the introduction of the
stirrer into the measurements.
Figure 9. Relationship between (a) ε” and ε’-1 and (b) gr/fr and (fa/fr)2 for stirred MCC at
five different moisture contents and varied densities. As moisture content increases, the
slope of the resulting line of constant moisture content increases as well. Many more
moisture contents were evaluated than are shown here, for clarity. All other moisture
content curves showed the same trend.
39
Moisture Content Residual (%)
3
2
1
0
-1
-2
-3
0
1
2
3
4
5
6
7
8
Predicted Moisture Content (%)
Translated Raw Variable
Polynomial 2,1
Meyer and Schilz
Figure 10. Quality of moisture calibration for rapidly stirred Avicel PH105. Using the
translated raw variable method, the magnitude and spread of the residuals is minimized.
This discrepancy led the authors to conclude that in between the densities tested and zero
density, there is a non-linear region as the contribution to fr and gr from the stirrer begins
to dominate those contributions from the MCC. This non-linearity makes calibrating the
sensor using the ratio ε”/(ε’-1) relatively inaccurate, as can be seen by Figure 10, where
absolute moisture content residuals are plotted for both the Translated raw variable
method and the Meyer and Schilz method. Several of the residuals from the Meyer and
Schilz and Polynomial 2,1 methods are larger in magnitude than 1.00%, which is greater
than any of the residuals found using the Translated raw variable method. Furthermore,
40
as seen in Figure 10, the Translated raw variable residuals are more tightly clustered
around 0% than are the residuals found using the other two methods.
As can be seen in Table 6, the Translated raw variable method and the approximation
using a polynomial of the form Eq. (7) both resulted in the most accurate predictions.
Comparing their average relative errors and standard errors of calibration demonstrates
that the polynomial surface was better able to predict the moisture content of lowmoisture containing powders than was the translated raw variable method; this can also
be observed in Figure 10. However, the latter method was better able to approximate the
water content of higher moisture content powders. Both models provide viable means of
monitoring the moisture content of rapidly flowing powders.
Table 6. Summary of moisture content models
Average Relative
Model
Error
Translated Raw Variable
14.5 %
Method
Polynomial 2,1
7.9 %
ε”/(ε’-1)
28.4 %
Standard Error of
Calibration
0.319 %
0.459 %
0.567 %
Compared to measurements conducted on static MCC powder, the SEC for rapidly
flowing MCC powder is larger. For example, using the Translated raw variable method,
the SEC for monitoring rapidly flowing powders was found to be 0.319 %, compared to
0.133 % for static tests on the same material. This was to be expected, as the
experimental scheme was much more complicated. There are two likely sources of error.
41
First, errors could be a result of local temperature fluctuations caused by friction (though
the overall temperature of the powder did not raise more than 3 °C during any one set of
tests). Secondly and most likely, they could have resulted from the introduction of the
stirrer within a millimeter of the sensor. To reduce this effect in practical applications, a
microwave sensor should be placed at least one to two centimeters away from a stirring
apparatus, as the sensitivity of the sensor decays with distance. To test this, a metal plate
was moved near the microwave sensor while material filled the cylinder; the proximity of
the metal plate had a negligible effect on the microwave sensor. In this study, intimate
contact between the sensor and the stirrer was made to investigate whether a microwave
sensor could accurately measure the bulk density of rapidly flowing materials.
3.4
Conclusions of Study
Using an open-reflection, resonance sensor, both the moisture content and the effective
bulk density of rapidly flowing MCC were monitored with a high degree of accuracy.
These methods of monitoring were shown to be accurate for powders flowing at least 5.1
m/s. Although the sensor’s sensitivity to moisture content diminished compared to the
static case somewhat, the microwave sensor was able to monitor the bulk density of the
flowing material with equal accuracy to the static case. This investigation enables the use
of microwave resonance technology to accurately monitor the bulk density and the
moisture content of pharmaceutical powders in high flow and shear rate unit operations.
42
CHAPTER 4. MONITORING ROLLER COMPACTED RIBBONS ONLINE USING
MICROWAVES AND NIR
4.1
Objective
After demonstrating that it is possible to monitor rapidly flowing particulates using
microwaves with similar accuracy as static powders, it was necessary to demonstrate that
microwave sensors can be viable process analytical tools on-line. Additionally, it was
necessary to compare the accuracy and robustness of microwave sensors to the more
established method, NIR sensing.
4.2
Roller Compaction Background
Roller compaction is a dry granulation unit operation commonly used in the
pharmaceutical and nutraceutical industries to increase the size of a particulate material in
a controlled manner, making it easier to process. Compared to wet granulation, dry
granulation offers several benefits.
Most notably, these include the absence of an
additional costly drying unit operation and allowing the processing of heat and moisture
sensitive materials.
Furthermore, roller compactors require less space, energy, and
person-hours to operate compared to wet granulation processes. In order to increase the
particle size of a particulate material while also maintaining a narrow size distribution,
the process parameters need to be precisely controlled.
Both the moisture content and the envelope density of roller compacted ribbons
significantly affect the roller compaction process and thus the final quality of produced
tablets. Over densification of the ribbon yields undesired tablets with both low hardness
43
and high friability. In contrast, under densification usually yields granules with a wide
size distribution, which results in tablets with a large variability in weight [57]. The
Johanson model has been used extensively to predict the density of roller compacted
ribbons, but it can only serve as a simplified basis and is thus not ideal for accurate
process control or exceptional events management if one is interested in density or
normal stress magnitude predictions [58, 59].
The moisture content of granules
significantly affects their flowability, cohesivity and compressibility [60]. The presence
of moisture can lead to the formation of both liquid and solid bridges between particles.
In most cases, this leads to increased cohesion and friction, which reduces the flowability
of the material [61, 62]. If the powder cannot flow predictably into the roller compaction
region, then it will likely not compact uniformly. It is therefore desirable to monitor both
the envelope density as well as the moisture content in real-time as accurately as possible.
Historically, near infrared spectroscopy has been the method of choice to monitor these
parameters in roller compacted ribbons [63, 64]. Instrumented rolls containing force
transducers have also been used to monitor density distributions in roller compacted
ribbons [65].
To the author’s knowledge, no studies thus far have investigated
microwave resonance technology as a viable means of monitoring a roller compaction
process.
4.3
4.3.1
Materials and Methods
Preconditioning of Materials
Microcrystalline cellulose (MCC) (Avicel PH200 obtained from FMC BioPolymer
Corporation in Philadelphia, PA), with an average particle size of 180 µm and loose bulk
density ranging from 0.29 to 0.36 g/cc, was used for this study. Two days prior to testing,
44
MCC was either humidified with steam or dried in an oven. This preconditioned the
samples’ moisture content. Subsequently, the powders were tumbled in a drum to help
ensure moisture content uniformity. Between conditioning and testing, the powders were
stored in two layers of Ziploc bags for two days to ensure moisture content and
temperature uniformity.
4.3.2
Roller Compaction
A model WP120x40 roller compactor on loan to Purdue University from Alexanderwerks
in Remscheid, Germany was used for this study. Flat rolls with a diameter of 12 cm and
a width of 4 cm were used. The roll gap was not set by the operator but was controlled
by the roller compactor to maintain a constant compaction pressure. The operating range
and normal operating conditions (NOC) for each process parameter is shown in Table 7.
Table 7. Operating range and normal operating conditions of various process parameters
Process Parameter
Operating Range
NOC
Roller Compactor (RC) Roll Gap (mm)
1-3
Variable
RC Hydraulic Pressure (bar)
0-230
Variable
RC Roll Speed (rpm)
3-13
6
RC Feed Screw Speed (rpm)
19-102
26
Each preconditioned batch of microcrystalline cellulose was tested separately.
The
hydraulic pressure was changed during each batch, and ranged from 20 to 80 bars. At
least two ribbons were created for each permutation of moisture content and compaction
pressure. In total, 119 ribbons were tested and analyzed. The ambient temperature was
45
maintained at approximately 22 °C while the ambient humidity was maintained using a
dehumidifier.
4.3.3
Microwave Measurements
A 167PA-0501-01000 forked microwave sensor, controlled by a PMD320PA, on loan
from Sartorius Stedim in Bohemia, NY was placed in-line with the produced roller
compacted ribbon, as shown in Figure 1. Although a prototype at the time of testing, this
model is now commercially available. This model sensor was chosen because it allowed
complete transmission of the microwave signal through the sample during testing.
Additionally, the sensor cavity operated as a resonator, and microwave resonant sensors
have been shown to be more accurate compared to other types of microwave sensors [49].
At resonance, a standing wave pattern forms in the cavity as the microwaves are reflected
off the top and bottom walls of the cavity. When a material is placed inside of the
sensor’s resonant cavity, the resonant frequency decreases. This phenomenon results as
the speed of the electromagnetic waves temporarily reduces as they pass through the
material. Additionally, the half bandwidth of the resonant curve increases. This is
largely due to an increase in the signal’s attenuation [66]. By measuring the frequency
and bandwidth shifts during resonance, it is possible to accurately measure both moisture
content and density [12]. Approximately 530 measurements were recorded over each
one-second interval and averaged together to yield one set of microwave measurements.
This time interval was chosen to correspond with the length of sample tested using NIR
analysis, discussed below.
46
Figure 11. On-line Sensing Configuration to Monitor Roller Compacted Ribbons. (Blue
Arrow: Roller Compactor, Red Arrow: Microwave Sensor, Green Arrow: NIR Probe)
4.3.3.1 Development of Microwave Moisture Model
Many empirical and semi-empirical models to predict moisture content using microwave
sensors have been proposed [36, 41, 47, 51]. Most have been developed for planar
sensors, such as ring resonators, or transmission measurements using horn antennas;
these models could not predict the moisture content of ribbon samples as they passed
through the open cavity resonator with the desired accuracy.
For this study, a model of the form of Eq. (28) was used to predict moisture content.
Although the constants were determined by fitting the semi-empirical Eq. (28) to
experimental data, the choice of the functional form of the equation was based on theory.
The derivation of Eq. (28) is shown below. The quality factor (Q) of a resonant cavity is
a dimensionless parameter that indicates the relative magnitude of the energy stored to
47
the average power lost at resonance. Mathematically, this can be expressed as the ratio of
the resonant frequency (fr) to the bandwidth (B) of the resonant curve. The measured
quality factor (Qmeas) can be broken into three components, as shown in Eq. (29), [30]
where the contribution from the material’s dielectric losses (Qdiel) is inversely
proportional to its loss tangent (Eq. (30)).
(28)
(29)
(30)
As the resonant frequency of the cavity did not change drastically when the material was
inserted, the contributions to the measured quality factor from radiation (Qext) and losses
due to the finite conductivity of the walls (Qwalls) were assumed constant. With this
assumption, and after insertion of Eq. (30) into Eq. (29), Eq. (31) was established.
Rearrangement yielded Eq. (32).
(31)
(32)
The loss tangent of a material quantifies its ability to dissipate electromagnetic energy.
In most particulate-water systems, the majority of these losses are due to heating in the
presence of water. It has been shown that in some materials, the loss tangent is nearly a
linear function of moisture content (Eq. (33)) [22]. As the microwave device used in this
work outputs the measured values of the changes in resonant frequency and bandwidth
48
instead of absolute values, Eqs. (34) and (35) were combined with Eqs. (32) and (33) to
yield the final microwave moisture model upon rearrangement (Eq. (28)). The constants
were fit empirically using JMP 10 from microwave measurements and moisture reference
values.
(33)
(34)
(35)
4.3.3.2 Development of Microwave Density Model
As has been stated previously, the resonant frequency of a resonator decreases as material
is loaded into the electric field. The shift in the resonant frequency is directly related to
the amount of mass present in the electric field. If the sample is contained in a known
volume, the density of the sample can be monitored. Alternatively, if the sample volume
is significantly larger than the electric field, the material can be assumed to be infinite
and once again, the density can be determined. In this study, roller compacted ribbons
with similar volumes were monitored; this allowed microwave measurements to be
correlated with density.
Although the magnitude of the resonant frequency shift is the primary indicator of the
amount of mass in the field, if multiple components are present with significantly
different dielectric properties, an extra factor must be included, as shown in Eq. (36).
The second term shown below corrects for the presence of water, which has a large
dielectric constant and loss factor compared to most particulates.
49
(36)
The form of Eq. (36), especially the choice of the second term, was chosen because it
minimized the error and bias when it was used to predict the envelope density of roller
compacted ribbons. Eq. (36) could likely be used to accurately monitor the bulk density
of loose particulate materials as well.
4.3.4
NIR Measurements
A commercially available model Turbido OFS-12S-120H NIR sensor obtained from
Solvias AG in Switzerland was placed directly downstream of the microwave sensor, as
also shown in Figure 11. An NIR256L-1.7T1 spectrometer from Control Development
Inc. in South Bend, IN was used to collect the spectra and was then recorded using
Spec32 (version 1.5.4.8). The integration time was found to be 0.067 seconds; a 16sample average was used. The length of the ribbons scanned during each test was thus 4
cm. The tip of the NIR probe was measured to be 0.5 cm in width. Analyses were
carried out over the wavelength range of 904-1687 nm.
4.3.4.1 Development of Partial Least Squares Models
The spectrum from each ribbon was exported to Unscrambler X, distributed by CAMO
Software in Woodbridge, NJ, for pretreatment. The spectra were analyzed in either JMP
10, distributed by SAS in Cary, NC, or Unscrambler X using principal component
analysis (PCA) and partial least squares (PLS) regression with cross validation. PCA is
an algorithm that transforms a set of correlated variables into a set of linearly
uncorrelated ones. This type of analysis is best used as a qualitative means of better
understanding patterns in complex data. PLS regression is similar to PCA analysis;
50
however, this quantitative type of analysis seeks to find the multidimensional
transformation of the data set that minimizes the variance in the response variable.
Example spectra taken at similar moisture contents and varying envelope densities can be
seen in Figure 12. When moisture content remains fixed and the density is allowed to
vary, a baseline shift is observed for the NIR spectra. As the density decreases, the
curves shift downwards; a vertical shift was also reported by Acevedo et al. when
monitoring roller compacted ribbons of microcrystalline cellulose [63]. Thus, a good
model to predict density should rely on this vertical shift while minimizing effects due to
other process conditions.
There is a peak present at the 1685 nm region strongly
correlated to the presence of water. This region was therefore truncated when modeling
the effects of density variations.
51
Figure 12. Comparison of NIR Spectra of Microcrystalline Cellulose Taken at 4.65%
Moisture Content and at Varying Densities (Blue Dash: 1.09 g/cc, Red Solid: 0.977 g/cc,
Green Dot: 0.776 g/cc)
As has been reported previously, standard normal variate (SNV) pretreatment helps to
remove physical information from spectra, such as particle size and the multiplicative
interferences of scatter [67, 68].
This usually allows for clearer observation of
differences in chemical composition. Example spectra pretreated with SNV taken at
similar envelope densities and varying moisture contents can be seen in Figure 13. The
most significant region for moisture content determination was found to be between 1440
nm and 1630 nm; this region has been highlighted in Figure 14. This region is important
for moisture content determination due to the presence of a strong absorbance band for
water present at 1450 nm [69]. Although the peak found near 1685 nm was strongly
52
correlated to the presence of water, it was not used to aid in the determination of moisture
content. For some of the spectra, the peak maximum was shifted beyond the upper
detection range of 1687 nm.
Figure 13. Comparison of NIR Spectra of Microcrystalline Cellulose Taken At
Approximately 0.948 g/cc Envelope Density (Green Dash: 5.37 % MC, 0.949 g/cc;
Purple Long Dash: 4.81 % MC, 0.951 g/cc; Blue Solid: 3.39 % MC, 0.947 g/cc; Red Dot:
2.15 % MC, 0.944 g/cc)
53
Figure 14. Expanded View of 1440 - 1630 nm Region in Figure 3 to Highlight Most
Significant Region for Moisture Content Determination (Green Short Dash: 5.37 % MC,
0.949 g/cc; Purple Long Dash: 4.81 % MC, 0.951 g/cc; Blue Solid: 3.39 % MC, 0.947
g/cc; Red Dot: 2.15 % MC, 0.944 g/cc)
4.3.5
Ribbon Density and Moisture Content Reference
Reference envelope density measurements were carried out using a GeoPyc 1360
Envelope Density Analyzer obtained from Micromeritics in Norcross, GA. Samples
were weighed prior to testing using an XS104 model microbalance, obtained from
Mettler Toledo in Columbus, OH. Because the NIR probe could only monitor a 0.5 cm
wide area at the center of the ribbon, the centerline was cut out of the ribbon prior to
testing. The center and edge pieces were tested separately. The centerpiece was used to
calibrate the NIR probe while the entire density used to calibrate the microwave sensor
54
was found using a weighted average of the center and edge densities. An array of NIR
probes could be used to sample the density of the entire ribbon. The moisture content of
each ribbon was determined as using the Loss on Drying method from an HG63 Halogen
Moisture Analyzer obtained from Mettler Toledo in Columbus, OH.
4.4
Results and Discussion
Roller compacted ribbons of MCC in the density range of 0.675 g/cc to 1.216 g/cc and
the moisture content range of 2.1 % to 5.5 % were tested. Although this moisture content
range is relatively narrow, especially compared to wet granulation processes, it covers the
expected range of values during roller compaction for MCC when exposed to different
relative humidity levels. Over this moisture content range, significant changes to MCC’s
flowability and yield strength can still be observed [60], emphasizing the need to monitor
its moisture content during the roller compaction process.
4.4.1
Density Monitoring
4.4.1.1 NIR Spectroscopy
Pretreatments, including first and second derivatives, SNV, and baseline shift were
evaluated, but as the largest observable change in the spectra of different density ribbons
was a shift in their baselines (Figure 12), no spectral pretreatment was used. A twofactor model using mean centering and uniform weighting was developed using JMP 10
from 119 raw spectra. The model was developed over the truncated frequency range of
904 – 1675 nm. The first principal component accounted for 98.3 % of the variation in
the spectral data; it quantified changes most closely correlated to density, such as
compaction pressure and roll thickness. The second principle component was most
55
strongly correlated to moisture content variations and accounted for approximately 1.6 %
of the variance in the spectral data.
Using “leave one out” cross validation, a root mean squared error (RMSE) of calibration
of 0.073 g/cc was found. This result is similar to that found by both Gupta et al. [60] and
Acevedo et al. [63] when monitoring the density of MCC compacts and helps to validate
the model . The magnitude and scattering of the density residuals can be seen in Figure
15 as the red triangles. As is shown, they are randomly scattered and do not indicate a
discernible pattern or bias in the data not accounted for by this two-factor model.
Figure 15. Density residuals of NIR 2-factor PLS model based on raw spectral data and
density residuals of microwave model
56
By combining NIR spectroscopic data with the pressure set point of the roller compactor,
a more accurate PLS model was developed. As predicted by the Johanson equation,
under normal operation the pressure of the roller compactor is directly proportional to the
envelope density [59]. To weight the pressure information equally with the NIR spectral
data, scaling in JMP 10 was used. This resulted in a two-factor PLS model with a RMSE
of calibration of 0.068 g/cc.
4.4.1.2 Microwave Resonance
JMP 10 was used to find the calibration coefficients in Eq. (36) for the microwave sensor
with the same 119 ribbons used for the NIR analysis.
As the microwave sensor
monitored the density of the entire ribbon, compared to the NIR probe that monitored
only the center density, the whole ribbon densities reference values were used. A RMSE
of 0.044 g/cc was found. This error is approximately half of the error encountered using
NIR analysis. A plot of the residuals from Eq. (36), shown in Figure 15 as blue squares,
indicates that they are randomly distributed. Furthermore, the magnitude of the residuals
was significantly reduced when employing the microwave model as compared to the NIR
PLS model.
This increased accuracy is likely due to the direct relationship between the magnitudes of
microwave sensor measurements and the amount of mass present in the field. This
differs from NIR analysis, where physical properties such as density need to be filtered
from chemical effects using a technique like principal component analysis.
As mentioned previously, a known or semi-infinite volume is needed to monitor very
precisely the density of a material in the microwave field. Although the ribbon volume
57
did not change with pressure significantly, by including the pressure set point (P), a more
accurate microwave calibration can be obtained (Eq. (37)).
(37)
The pressure set point takes the place of the compensation term for moisture content in
Eq. (36). This substitution occurs because the majority of the density variation with
pressure is due to changes in moisture content.
As noted previously, moisture
significantly affects the compressibility of most particulate materials. The inclusion of
compaction pressure information in Eq. (37) resulted in a RMSE of 0.034 g/cc. The
RMSE after inclusion of roller compaction pressure using a microwave sensor was once
again about half of that found when using NIR sensing.
It is important to note that the inclusion of resonance bandwidth into Eq. (37) did not
significantly improve the results. This suggests that a less expensive scalar network
analyzer could be used to monitor density of roller compacted ribbons with similar
accuracy to vector network analyzers.
4.4.2
Moisture Content Monitoring
4.4.2.1 NIR Spectroscopy
As was shown previously, significant changes in the spectra can be observed in the 14401630 nm region that correspond to moisture content differences once physical effects
have been removed by pretreatment. Thus, this region was chosen to construct the PLS
model for moisture content. Although spectral differences were most readily observed
using SNV pretreatment, as shown in Figure 14, principal components found using first
derivative pretreatment were more strongly correlated with moisture content differences.
58
Therefore, a three-factor model was constructed using Unscrambler X after first
derivative pretreatment utilizing a 15-point Savitzky-Golay algorithm. This PLS model
employed mean centering and uniform weighting; this resulted in a RMSE of calibration
of 0.115 % moisture content using “leave one out” cross validation. There was no
significant improvement in the RMSE of calibration using a four-factor model (0.105 %
moisture content). Furthermore, the residuals, as shown in Figure 16 are randomly
distributed and do not indicate a pattern that could be explained by including further
factors. Lastly, there was a strong linear correlation (greater than 0.8) between the T and
U scores of the first three factors. This correlation dropped significantly for factors
greater than three.
Based on these observations, it was concluded that only three
significant principle components were present in the first derivative pretreated spectra.
59
Figure 16. Moisture content residuals of two-factor model based on SNV pretreated data
The first principal component, which accounted for 88 % of the variance in the spectra
and 68 % of the variance in the model’s response, quantified the moisture content
variations in the ribbons. The second principle component, which accounted for 10 % of
the variance in the data, quantified variations in density and accounted for 26 % of the
variance in the model’s response. The third principle component accounted for 1 % of
the variance in the spectral data and 5 % of the variance in the model’s response. Not
enough parameters were measured to properly identify what property of the ribbons the
third principle component quantified.
However, the strong linear correlation (0.90)
between the T and U scores of factor three, as well as the significant decrease in the
RMSE when employing a three factor model compared to a two factor model (0.115 % vs.
60
0.249 %), both indicated that the third principle component was significant and not due to
noise and should therefore have been included. As moisture content is not a function of
compaction pressure, the addition of the pressure set point to the NIR model did not
significantly improve the accuracy of the model.
4.4.2.2 Microwave Resonance
JMP 10 was used to find the constants in Eq. (28) by fitting microwave sensor
measurements to the same 119 ribbon moisture content values used for NIR analysis.
With four fitted constants, a RMSE of 0.065 % moisture content was found. The
microwave model’s residuals are randomly distributed, as shown in Figure 17.
61
Figure 17. Moisture content residuals of microwave model
The RMSE found using microwave sensing was again approximately half of the error
seen using NIR analysis. Microwave sensors are very sensitive to even tiny amounts of
moisture, capable of detecting water concentrations in the ppm range [50]. This is due to
the large contrast in dielectric properties of water and most solid materials. For instance,
the dielectric constant, which characterizes a material’s ability to store electrical energy,
of water is approximately 80 in the frequency range of interest while it is approximately
two for microcrystalline cellulose. In contrast, NIR detection schemes must use complex
62
chemometric software and spectral pretreatments to remove physical effects from spectra
before they can accurately monitor chemical compositions.
4.5
Conclusions of Study
Both microwave resonance and NIR sensing have been shown to be viable process
analytical tools for monitoring the moisture content and envelope density of roller
compacted ribbon. In this study, microwave resonance technology was shown to be
markedly more accurate than NIR, with approximately half of the RMSE for both density
and moisture determination. Furthermore, as was also noted by Corredor et al. [12],
microwave resonance benefits from working without sophisticated chemometric software
and
having
the
ability
to
transfer
between
formulations
more
easily.
63
CHAPTER 5. DEVELOPMENT OF A NOVEL MICROWAVE SENSOR TO
MONITOR CHEMICAL COMPOSITION ONLINE
5.1
Objective
Rapid and accurate determination of process variables is valuable in nearly every
manufacturing industry. In those that deal with the manufacture of particulates, where
segregation is always a concern, chemical composition is one of the most important
process and quality control variables.
This study seeks to demonstrate that microwave
sensors can be viable measurement devices for determining the chemical composition of
a particulate blend on-line.
5.2
Materials and Methods
5.2.1 Development of Microwave Sensor
To detect material composition on-line, a novel microwave sensor was designed and
developed. Most resonant microwave sensors operate around a single resonance, each
providing two independent variables to the user.
For instance, many sensors are
calibrated to provide both the real and imaginary components of the effective dielectric
constant at the frequency of resonance [36]. To discern the concentrations in a system
with several components, one needs more independent variables, and thus the microwave
sensor designed needs to have multiple resonances. Ideally, microwave sensors used to
determine chemical composition should have resonances spread out over the available
range of frequencies, since the dielectric properties of a material do not normally differ
significantly with small changes in frequency.
In addition, it is necessary for the
64
components in the material under test to have noticeably different valued dielectric
properties in the microwave region. For instance, it would be very difficult to tell the
difference between two similar polymers or between polymers based on the same unit,
but of different chain lengths.
A sensor with six resonances over the 2-8 GHz region was developed using COMSOL
Multiphysics, available from COMSOL Inc. (Burlington, MA). This software used the
finite element procedure to approximate the solution to the frequency domain vector
wave equation, shown in Eq. (38), where E is the electric field intensity, D is the electric
flux density, H is the magnetic field intensity, B is the magnetic flux density, ⃡ is the
permeability of the material, ⃡ is the permittivity of the material, J is the electric current
density, M is the magnetic current density, ko is the wave number, and Zo is impedance.
This equation can be derived from Maxwell’s Equations (Eqs. (1)-(4)) and after
separation of variables (Eq. (39)) to make the final equation time-invariant.
[⃡
̃
]
⃡
̃
̃
̃
(⃡
̃
)
(38)
(39)
As shown in Figure 18, the final sensor design consisted of three parallel ring resonators
made from 62 mil Rogers 4350 with 1 oz. copper by Advanced Circuits (Aurora, CO).
To match the impedance of the 50-ohm coaxial feed cable, the microstrip lines were
manufactured to be 3.67 mm wide. A solder mask (PSR-4000BN from Taiyo America,
Inc; Carson City, NV) was applied to the sensor surface to make it more rugged and to
reduce the risk of material sticking to the surface. As reported by the vendor, the solder
mask had an approximate thickness of 25 microns. Due to this negligible thickness, the
65
solder mask was not simulated in COMSOL Multiphysics to significantly reduce
computation time.
Figure 18. Top view of planar sensor as fabricated
The sensor was designed with three parallel ring resonators for several reasons. Firstly,
as was discovered from simulations, this design allows the quality factors of the
resonances to be very high. The accuracy and sensitivity of a resonant sensor both
increase with its quality factor [70]; it is a measure of the energy stored in a resonator to
the energy dissipated per cycle. A standard design using a single ring resonator would
normally yield a quality factor in the range of 100-300. In this design, the middle ring
couples to the outer rings magnetically and allows them to reach quality factors in the
thousands. Thus, this design allows for significantly more accurate sensing.
66
Secondly, it allows material properties to be determined uniquely from either side of the
sensor, allowing the user to determine if segregation is occurring. By slightly varying the
radii of the three rings, they can be made to resonate independently of one another. The
outer diameters of the three rings were, from left to right (Figure 18): 2.78 cm, 2.9 cm,
and 3.02 cm. These diameters were chosen based on trial and error from simulations and
produced the largest quality factors while maintaining unique resonances. A larger
difference in the diameters would result in less defined resonances while a smaller
difference would allow the rings to resonate together. How the norm of the electric field
varies at resonance can be seen in Figure 19. Keeping in mind that the color-coding is on
a log base 10 scale, one can observe that the bottom ring is able to resonate independently
of the top ring, with more than 100x the electric field strength around the bottom ring
than around the top ring.
67
Figure 19. COMSOL simulation results showing how the norm of the electric field varies
along the sensor surface when the bottom ring resonates at approximately 4.3 GHz. The
color-coding shown is on a log base 10 scale. Thus, the field strength around the top ring
is less than 1% of the strength in the bottom ring, indicating that the rings can resonate
independently of one another.
Furthermore, ring resonators were chosen because they have both higher quality factors
and are more compact than their linear, microstrip resonator counterparts are. The
68
quality factor increases because fringing fields at the ends of the linear microstrip are
removed when it is made into a circle. Other designs, such as high-k dielectric sensors
[71] and cavity resonators [24, 72] were also considered. Although high-k dielectric
resonators have very large quality factors, a microstrip design was chosen because it is
simpler to design and cheaper to produce. Furthermore, dielectric resonators are very
compact, which reduces the sensing area. In contrast, cavity resonators are large, with
dimensions on the order of half a wavelength. Moreover, cavity resonators can be
difficult to use and clean with cohesive powders due to buildup in the cavity.
It is important to note that sensors of the type used in this study are designed to operate at
a very low power level (approximately 1 mW) and thus do not pose a health risk to an
operator. Furthermore, they do not heat or cause deterioration of the material under test.
5.2.2
Comparison of Simulation to Realization
After fabrication, the sensor was connected to an Agilent E5071B vector network
analyzer obtained from TRS-RenTelco (DFW Airport, TX). A scattering parameter trace
was recorded with no test material present, which was then converted to the reflection
coefficient (Γ) using Eq. (40) where S11’ and S11” are the real and imaginary components
of the reflection scattering parameter, respectively.
√
(40)
A comparison of the trace obtained from the physical sensor to one generated from the
COMSOL simulation can be seen in Figure 20.
When the sensor resonates, the
magnitude of the reflection coefficient drops sharply.
As is shown, the simulation
accurately modeled both the location and magnitude of the resonances. The downshift of
69
the resonances in the physical response is due to the presence of the soldermask, which
was not thick enough to easily simulate across the entire domain. From these results, one
can conclude that the simulation can accurately model the electromagnetic field the
sensor produces.
1
0.9
Reflection Coefficient
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
1
3
5
7
Frequency (GHz)
COMSOL Simulation
Physical Sensor
Figure 20. Comparison of COMSOL simulation results to physical realization
70
Thus, the COMSOL simulations can be used to better understand what the
electromagnetic field looks like during operation at key places in the spectra. At the first
resonance, near 2.1 GHz, one can see that the bottom ring is able to resonate
independently of the other rings in Figure 21. Once again, the norm of the electric field
is shown on a log base 10 scale. As is shown, the sensor resonates when one wavelength
can constructively interfere around the circumference of the bottom ring. Similarly, the
top ring’s first resonance occurs when one wavelength can constructively interfere
around the circumference of the top ring. At higher order resonances, this occurs when
‘n’ resonances can constructively interfere around the circumference of a ring. Examples
of top ring and higher order resonances can be seen in Appendix A, Figure 45 through
Figure 49.
71
Figure 21. COMSOL simulation result demonstrating the norm of the electric field at the
first resonance near 2.1 GHz.
When no ring resonates, and the magnitude of the reflection coefficient is near one at 3.3
GHz, the norm of the electric field is as shown in Figure 22. The small intensity in the
center ring is much smaller in magnitude than when the outer rings resonate and cannot
be used for sensing.
72
Figure 22. COMSOL simulation result demonstrating how the norm of the electric field
varies across the sensor surface when neither the top nor the bottom ring resonates at 3.3
GHz.
An interesting phenomena occurs 0.1 GHz above the top ring’s second order resonance
(Figure 47).
As shown in Figure 20, at 4.8 GHz, there occurs a small resonance
phenomenon that does not correspond to either the top or the bottom rings resonating
alone. This resonance phenomenon is demonstrated in Figure 23 as a result of a figure –
eight path taken by the resonating field between the top and middle rings. Events like
this are hard to predict and underscore the importance of using simulations to better
understand the physics involved.
73
Figure 23. COMSOL simulation result demonstrating how the norm of the electric field
varies across the sensor surface when the top and middle rings combine their path lengths
to form a figure-eight resonance at 4.8 GHz.
5.2.3
Preconditioning of Materials
Two days prior to testing, mixtures of different concentrations of microcrystalline
cellulose (Avicel PH200 obtained from FMC Biopolymer in Philadelphia, PA; average
particle size of 180 μm and approximate microwave dielectric constant and loss factor of
1.7 and 0.1, respectively [36]) and acetaminophen (APAP) obtained from Mallinckrodt
Inc. (Raleigh, NC; microwave dielectric constant and loss factor previously unreported)
were blended in a BP4 blender from Dott. Bonapace & C. (Limbiate, Italy) for 7 minutes.
Once mixed, some of the blends were either humidified using steam or dried in an oven
74
at 90 degrees Celsius. After the samples were preconditioned to different moisture
contents, they were tumbled in a drum to help ensure moisture content uniformity.
Subsequently, samples were stored in two layers of Ziploc bags for two days to further
ensure moisture content and temperature uniformity.
5.2.4
Test Setup
Testing was carried out using the system shown in Figure 24. The vibrating stage from a
Rotary Sample Divider PT-100 obtained from Retsch (Haan, Germany) was used to
discharge the blended powders from a funnel into the custom designed acrylic chute with
the NIR and microwave sensors attached. This configuration was chosen because it
allowed powders to flow evenly over the sensors’ surfaces. Furthermore, it allowed the
flow of the powders to be altered by changing the slopes of the shelves.
As
measurements were taken while the material was in motion, a very large amount of
unique data could be recorded.
75
Figure 24. Testing setup used to develop microwave and NIR sensor calibrations (blue
arrow: vibrating stage, green arrow: microwave sensor on second shelf, orange arrow:
NIR probe). Both shelves were designed to allow their slopes to be adjustable. In this
image, the top shelf is down to demonstrate this.
76
During testing, powder flowed from the vibrating stage onto the first slanted shelf. The
powder then flowed from the top shelf onto the second one to which the sensors were
attached. Using two shelves helped to disperse the powders evenly across the width of
the chute before making the measurement. After the material had flown down the chute,
it was collected for subsequent moisture content analysis.
5.2.5
Determination of Reference Moisture Content
Reference moisture contents were determined on a wet basis, as described in section 2.2.2.
5.2.6
NIR Measurements
A commercially available Turbido OFS-12S-120H NIR sensor obtained from Solvias AG
in Basel, Switzerland was used during this study.
commercially
available
NIR256L-1.7T1
To collect the NIR spectra, a
spectrometer
obtained
from
Control
Development Inc. in South Bend, IN was used. Spectra were recorded using Spec 32
(version 1.7.3.6) over the wavelength range of 904-1687 nm.
A 0.00453-second
integration time was used with an 8-sample average; 1155 NIR spectra were accumulated.
The NIR probe was measured to be 5 mm in width. Data points were collected at each
integer wavelength.
Partial least squares (PLS) regression seeks the projection directions that maximize the
covariance between the predictor and response variables. It is widely used to build
models from large data sets of highly correlated data, such as NIR spectra, and was used
to develop models for both moisture content and acetaminophen concentration. To
investigate non-linear effects and for comparison to microwave measurements, artificial
neural network models were developed after dimensionality reduction using principal
component analysis (PCA).
77
5.2.7
Microwave Measurements
A custom MATLAB script was written to control the E5071B vector network analyzer
over USB. The code initialized the instrument, started frequency sweeps, and gathered
the results for analysis. From laboratory measurements, 2397 independent data sets were
accumulated. This large amount of data allowed for accurate model development.
Most microwave sensors can be calibrated using a simple linear regression based on
either a resonance’s center frequency, bandwidth, or quality factor, or a combination of
those parameters [73]. However, most of those resonant sensor designs only make use of
one resonant peak [34, 39]. The sensor designed for this research makes use of six
resonances. Due to the nonlinear relationship between the six resonant peak parameters
and the pharmaceutical components contents, artificial neural networks with a single
hidden layer were used to develop calibration models for both acetaminophen
concentration and moisture content. For comparison to NIR PLS models, microwave
PLS models were also constructed from microwave peak properties.
5.2.8
Model Development
Unscrambler X, distributed by CAMO Software, was used to pretreat raw NIR spectra.
Pretreatment of NIR spectra has been shown to remove undesirable physical information
from spectra, such as particle size and multiplicative interference of scatter [68, 74]. The
pretreated spectra were used to develop partial least squares regression models for
material composition and moisture content from NIR measurements in Unscrambler X.
Although the spectrometer was capable of measuring over the wavelength range of 9041687 nm, NIR models were developed using only the wavelength range of 930-1650 nm
to avoid artifacts at the detector’s limits.
78
Example spectra with first derivative pretreatment after 15-point, second-order, SavitzkyGolay smoothing applied for samples with significantly different acetaminophen
concentrations and similar moisture contents can be seen in Figure 25. An increase in
acetaminophen resulted in an increase in the intensity of the pretreated spectra at many
locations, for instance near 1130 nm, 1350 nm, 1430 nm, 1550 nm, and 1630 nm. A
good model to predict acetaminophen concentration should rely on this trend while
minimizing the effect of moisture content.
0.002
0.0015
Intensity
0.001
0.0005
0
-0.0005
-0.001
930
1030
1130
1230
1330
Wavelength (nm)
1430
0% APAP 2.6% MC
20% APAP 2.2% MC
30% APAP 2.6% MC
40% APAP 2.3% MC
1530
1630
Figure 25. A comparison of NIR spectra with 15-point Savitzky-Golay smoothing and
subsequent first derivative pretreatment applied to them. Spectra were all recorded for
samples with approximately 2.4% moisture content.
79
The same pretreatment used for APAP determination has also been shown to be a viable
pretreatment method for determining the moisture content of microcrystalline cellulose
samples [75]. Shown in Figure 26 are three example spectra from samples with the same
acetaminophen concentration but different moisture contents. These spectra have also
had Savitzky-Golay smoothing and subsequent first derivative pretreatment applied to
them. A large difference in the spectra can be observed in the 1330-1530 nm region,
where a combination band of OH stretching modes can be seen; this region is critical for
moisture content determination due to the presence of a strong water absorbance band
[76].
80
0.0014
0.0012
0.001
0.0008
Intensity
0.0006
0.0004
0.0002
0
-0.0002
-0.0004
-0.0006
930
1030
1.9% MC 25% APAP
1130
1230
1330
Wavelength (nm)
1430
3.1% MC 25% APAP
1530
1630
5.6% MC 25% APAP
Figure 26. A comparison of NIR spectra with 15-point Savitzky-Golay smoothing and
subsequent first derivative pretreatment applied to them. Spectra were all recorded for
samples with 25% acetaminophen but at varying moisture contents.
Development of PLS models for microwave measurements was also carried out using
Unscrambler X. As the sensor was not developed to export raw spectral information,
peak properties, such as center frequency, bandwidth, and quality factor, were used as
model inputs. Therefore, no spectral pretreatments were used.
MATLAB’s Neural Network Toolbox, distributed by MathWorks (Natick, MA) was used
to develop artificial neural networks for both NIR and microwave data. Artificial neural
networks are capable of modeling highly nonlinear data. To use NIR spectra, which
81
consist of large sets of highly correlated data, as model inputs, dimensionality reduction
needs to be done. Principal component analysis was used to reduce the dimensionality of
the data, while still maintaining the largest sources of variability in the data.
For
microwave and NIR models, the principal components that each accounted for more than
0.01 % of the variance in the inputs were used. For NIR models, this corresponded to
three principal components, while for microwaves this resulted in 18 principal
components due to the highly uncorrelated nature of the microwave peaks. Hyperbolic
tangent transfer functions were used for each hidden neuron.
To train the weights in the networks, the Levenberg-Marquardt method was used. Over
fitting is the most significant problem one encounters when developing a neural network
model [77]. To help reduce this possibility, Bayesian regularization was used, which
modifies the error function to include a term to help minimize the weights. This method
was developed based on the observation that over-fitted functions show a high degree of
curvature and thus require large weights. Furthermore, to reduce over fitting, the data set
was randomly broken up into a training set, comprising 83% of the total data, and an
independent test set, comprising the remainder of the data. The test set was not used by
the training algorithm, so it was used as an independent metric to determine if the
network over fitted the data.
In the case of microwave model development, this
corresponded to 1990 randomly chosen data points used for training and 407 randomly
chosen data points for independent testing. NIR models made use of 959 randomly
chosen training points and 196 randomly chosen testing ones. If the independent test set
has a root mean square error of prediction (RMSEP) similar to the root mean square error
82
of calibration (RMSEC) of the training set, it can be assumed that the model is not overfitting the data.
5.3
Results and Discussion
Blends tested varied in acetaminophen concentration from zero to forty percent; moisture
content varied from 1.59 – 6.47 %. This moisture content range is relatively narrow;
however, it covers the expected range of moisture contents one would expect to observe
for powders exiting a mixing unit operation at different relative humidity levels. Critical
particulate properties, such as flowability and yield strength, still vary significantly over
this narrow moisture content range [78]. For ease of comparison, all of the RMSEP and
RMSEC for the following models can be seen in Table 8. The RMSEP values for the PLS
models were determined using 10-fold cross validation, whereas the RMSEP values for
the artificial neural network models were determined from an independent test set.
83
Table 8. Comparison of root mean square error of calibration (RMSEC), root mean
square error of prediction (RMSEP), and coefficient of determination (R2) for the various
models employed.
Model
RMSEC RMSEP R2
NIR APAP
(One-Factor PLS)
NIR APAP
(Four-Node ANN)
Microwave APAP
(Four-Factor PLS)
Microwave APAP
(Seven-Node ANN)
NIR Moisture
(Six-Factor PLS)
NIR Moisture
(Five-Node ANN)
Microwave Moisture
(Three-Factor PLS)
Microwave Moisture
(Seven-Node ANN)
5.3.1
4.7 %
4.7 %
0.88
1.9 %
1.8 %
0.98
11.7 %
11.7 %
0.31
2.6 %
2.6 %
0.96
0.47 %
0.47 %
0.90
0.21 %
0.21 %
0.98
0.94 %
0.94 %
0.57
0.18 %
0.19 %
0.99
Determination of Acetaminophen Concentration
5.3.1.1 NIR Spectroscopy
Based on the behavior of the RMSEP, the weights of the components, as well as the
percent variation explained in both the first derivative pretreated NIR spectra (X-block)
as well as the acetaminophen concentration (Y-block), a one-factor PLS model was
developed to model the concentration of APAP from 1155 first derivative pretreated
spectra. The RMSEC and RMSEP were both found to be 4.7 % APAP concentration
using 10-fold cross-validation with an R2 value of 0.88. Increasing the model complexity
to two factors had a negligible effect on both of the RMSE values and the R 2 value. The
84
residuals found from employing this model can be seen in Figure 27. The scattering of
the NIR residuals do not indicate a bias not accounted for by the one-factor model. By
plotting the residuals against their observation number, as in Figure 28, the spread of the
residuals is readily apparent.
APAP Concentration Residual (%)
30
20
10
0
-10
-20
-30
0
5
10
15
20
25
30
APAP Concentration Reference (%)
Microwave PLS Model
35
40
NIR PLS Model
Figure 27. 10-fold cross validation residuals found when employing either the one-factor
PLS NIR model or the four-factor PLS microwave model for acetaminophen
concentration. No discernable bias is seen in the NIR model; however, an apparent bias
can be seen in the microwave model.
85
30
APAP Concentration Residual (%)
20
10
0
-10
-20
-30
Observation
Microwave PLS Model
NIR PLS Model
Figure 28. Residuals from Figure 27 plotted against their observation number to ease in
readability.
A neural network with four hidden nodes was developed to investigate if the use of a
non-linear model could significantly increase the accuracy of the model. Four hidden
nodes were chosen after careful examination of the RMSEC and RMSEP values as a
function of the number of hidden nodes (Figure 29). Using this artificial neural network,
the RMSEC and RMSEP were found to be 1.9 % and 1.8 % APAP concentration,
respectively; an R2 value of 0.98 was calculated for this model. An F-test showed that
above a 95% confidence level, the NIR PLS and neural network models have statistically
86
different variances.
A plot of the residuals as a function of the reference APAP
concentration can be found in Figure 30, along with the residuals plotted against their
observation number in Figure 31.
8
APAP Concentration Error (%)
7
6
5
4
3
2
1
0
1
3
Microwave RMSEP
5
7
9
11
Number of Hidden Neurons
Microwave RMSEC
NIR RMSEP
13
15
NIR RMSEC
Figure 29. Demonstration of how the RMSEC and RMSEP for both microwave and NIR
artificial neural network models for acetaminophen determination vary with the number
of neurons in their hidden layer. The dashed lines follow the combined RMSE of both
calibration and test sets.
87
10
8
APAP Concentration Residual (%)
6
4
2
0
-2
-4
-6
-8
-10
0
5
NIR Testing Set
10
15
20
25
30
Reference APAP Concentration (%)
NIR Training Set
Microwave Testing Set
35
40
Microwave Training Set
Figure 30. Residuals found when employing both the NIR and microwave artificial
neural network models for acetaminophen determination.
88
8
APAP Concentration Residual (%)
6
4
2
0
-2
-4
-6
-8
Observation
Microwave Model
NIR Model
Figure 31. Residuals from Figure 30 plotted against their observation number to ease in
readability.
5.3.1.2 Microwave Resonance
Unscrambler X was used to develop a four-factor PLS model with mean centering
pretreatment to predict APAP concentrations. An increase in the number of factors had a
minimal impact on both the variance explained in the APAP concentration as well as the
RMSEs. Both the RMSEC and the RMSEP of the models were found to be 11.7 %
APAP concentration with an R2 value of 0.31. An F-test confirms that the NIR PLS
model better fits the data than the microwave PLS model, to greater than a 99%
confidence level. The residuals found when employing this model can be seen in Figure
89
27 alongside those found when using the NIR probe and a one-factor PLS model. This
model’s low accuracy is likely due to PLS regression being a linear regression. The
response of this sensor is highly nonlinear.
The response of the sensor’s first resonant bandwidth and quality factor as a function of
APAP concentration can be seen in Figure 32 and Figure 33, respectively. There is no
readily apparent pattern between bandwidth and APAP concentration.
Figure 33
demonstrates that an increase in APAP concentration may correspond to an increase in
the quality factor.
Although acetaminophen’s dielectric response in the microwave
region has not been published, this result likely indicates that APAP does not attenuate
microwaves to the same extent as MCC.
90
9
Bandwidth of First Resonance (MHz)
8
7
6
5
4
3
2
1
0
0
5
10
15
20
25
30
APAP Concentration (%)
35
40
45
Figure 32. Effect of acetaminophen concentration on the bandwidth of the microwave
sensor’s first resonance.
91
4000
Quality Factor of First Resonance
3500
3000
2500
2000
1500
1000
500
0
0
5
10
15
20
25
30
APAP Concentration (%)
35
40
45
Figure 33. Effect of acetaminophen concentration on the quality factor of the microwave
sensor’s first resonance.
An artificial neural network was developed to predict acetaminophen concentration based
on the principal components of the microwave sensor data. As was previously mentioned,
over fitting is a significant problem encountered when training neural networks. Shown
in Figure 29, both the RMSEC and RMSEP reduce when more neurons are added to the
hidden layer. Some of the random variation in RMSE values is due to the way the neural
network algorithm randomly divides the training and testing sets each time the network is
trained. As seen in this figure, the root mean squared errors of the testing and training
sets both begin to stabilize around nine hidden nodes. One could construct a model based
92
off nine hidden nodes to yield the lowest RMSEC. However, there might be issues with
over-fitting at this number of hidden neurons. Erring on the side of caution, a criterion
was developed that the RMSE of the complete data set needed to reduce by more than 10%
for each additional node to justify adding an additional node. Based on this criterion, a
model using seven hidden neurons was developed for acetaminophen determination. As
shown in Figure 29, there is a negligible difference between the RMSEC value for the
training set and the RMSEP value for the independent testing set with this number of
hidden neurons.
The model produced both RMSEC and RMSEP values of 2.6 % APAP concentration,
with an R2 value of 0.963. This is significantly reduced compared to the 11.7% RMSEP
found using the PLS microwave model, indicating the microwave sensor data needs to be
fit using a flexible, non-linear method. Although an F-test demonstrated that the mean
variances found when using the NIR and microwave neural network models were
statistically different above a 95% confidence level, comparing the RMSEP of the NIR
PLS model and the NIR artificial neural network model to the RMSEP of the microwave
artificial neural network model demonstrates that the microwave sensor is also capable of
producing accurate chemical composition measurements in-line.
This is further
confirmed by examining the residuals for the microwave model in both Figure 30 and
Figure 31 and comparing them to the residuals for the NIR artificial neural network
model.
By operating over a large frequency range, the sensor can detect multiple
components by monitoring how the dielectric response changes differently at each
frequency. Likely, by expanding the frequency range further, one could construct a
93
microwave sensor that is even more accurate for the determination of a blend’s
composition.
Although the NIR probe was shown to be more accurate for the determination of APAP
concentration than the novel microwave sensor, these results should not be discouraging
for further microwave sensor research in this area. Research to improve the accuracy of
NIR probes for chemical composition determination has been ongoing for decades. As a
first attempt, the microwave sensor’s accuracy is encouraging.
Furthermore, the
microwave sensor demonstrated does not suffer from some of the most significant
drawbacks of NIR sensing. Most importantly, the demonstrated microwave sensor can
sample bulk material properties, whereas the NIR probe used can only sample the first
few millimeters of a material. In this experiment, material was only allowed to build up
to approximately one millimeter.
However, in more common bulk flows that are
centimeters or more thick, the microwave sensor demonstrated here would be capable of
taking much more representative measurements.
5.3.2
Determination of Moisture Content
5.3.2.1 NIR Spectroscopy
A six-factor PLS model was developed using Unscrambler X to predict the moisture
content of the blends based on the behavior of the RMSEP, the weights of the
components, as well as the percent variation explained in both the first derivative
pretreated NIR spectra as well as the moisture content. Figure 34 demonstrates that there
is no bias or discernible pattern in the residuals not accounted for by this model. Both the
RMSEC and the RMSEP for the NIR PLS model from 10-fold cross-validation were
94
found to be 0.47 % moisture content with an R2 value of 0.90. Although in many
applications this RMSE would be acceptable, over this relatively narrow moisture content
range of 4.88 % it is not very accurate.
3
2.5
Moisture Content Residual (%)
2
1.5
1
0.5
0
-0.5
-1
-1.5
-2
1
2
3
4
5
Moisture Content Reference (%)
Microwave PLS Model
6
7
NIR PLS Model
Figure 34. 10-fold cross validation residuals found when employing either the six-factor
PLS NIR model or the three-factor PLS microwave model for acetaminophen
concentration.
95
3
Moisture Content Residual (%)
2
1
0
-1
-2
-3
Observation
Microwave PLS Model
NIR PLS Model
Figure 35. Residuals from Figure 34 plotted against their observation number.
To improve upon the accuracy, an artificial neural network with five hidden nodes was
developed. The RMSEC and the RMSEP were both 0.21 % moisture content with an R2
value of 0.980. Five hidden nodes were chosen because both the RMSEC and the
RMSEP changed negligibly when six hidden nodes were used, as shown in Figure 36.
The training and testing set residuals found when employing this artificial neural network
model can be seen in Figure 37 and do not indicate any bias or obvious pattern not
accounted for by the model. By comparing the NIR PLS and neural network results, it is
96
apparent that the latter is better able to model the data; this is further confirmed using an
F-test, which demonstrates that the two mean variances are not equal above a 95%
confidence level. As the neural network uses the principal components of the spectra,
both make use of dimensionality reduction. However, in the PLS case, the regression is
linear, while the neural network can model non-linear behavior, which is likely why it is
better able to fit the data.
1.2
Moisture Content Error (%)
1
0.8
0.6
0.4
0.2
0
1
3
Microwave RMSEP
5
7
9
11
Number of Hidden Neurons
Microwave RMSEC
NIR RMSEP
13
15
NIR RMSEC
Figure 36. Demonstration of how the RMSEC and RMSEP for both microwave and NIR
artificial neural network models for moisture content determination vary with the number
of neurons in their hidden layer. The dashed lines follow the combined RMSE of both
calibration and test sets.
97
1
0.8
Moisture Content Residual (%)
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
1
NIR Training Set
2
3
4
5
Reference Moisture Content (%)
NIR Testing Set
Microwave Training Set
6
7
Microwave Testing Set
Figure 37. Residuals found when employing both the NIR and microwave artificial
neural network models for moisture content determination.
98
1
0.8
Moisture Content Residual (%)
0.6
0.4
0.2
0
-0.2
-0.4
-0.6
-0.8
-1
Observation
Microwave Model
NIR Model
Figure 38 Residuals found when employing both the NIR and microwave artificial neural
network models for moisture content determination as a function of observation number.
5.3.2.2 Microwave Resonance
Unscrambler X was used to develop a three-factor PLS model to predict moisture content
from the microwave spectra’s peak information; mean centering pretreatment was used.
Increasing the number of factors had a negligible impact on the explained variance in the
moisture content. Both the RMSEC and the RMSEP were found to be 0.94 % moisture
content with an R2 value of 0.57. An F-test confirms that the NIR PLS model better fits
99
the data at a greater than 95% confidence level. Once again, the low accuracy of this
microwave PLS model is likely a result of the sensor’s non-linear behavior. This is
demonstrated in Figure 39 and Figure 40, where the response of the sensor’s first
resonance’s bandwidth and quality factor can be seen as a function of moisture content,
respectively. As shown in these figures, there are some discernable patterns between the
sensor parameters and the moisture content. For instance, a trend of an increase in
bandwidth with moisture content can be seen in Figure 39; likewise, a decrease in quality
factor can be seen with an increase in the moisture content (Figure 40). These results are
expected [66]; however, the inclusion of different chemical compositions complicates the
figures and demonstrates that a complex, non-linear model is necessary for accurate
measurements.
100
9
Bandwidth of First Resonance (MHz)
8
7
6
5
4
3
2
1
0
1
2
3
4
Moisture Content (%)
5
6
7
Figure 39. Effect of moisture content on the bandwidth of the microwave sensor’s first
resonance.
101
4000
Quality Factor of First Resonance
3500
3000
2500
2000
1500
1000
500
0
1
2
3
4
5
Moisture Content (%)
6
7
Figure 40. Effect of moisture content on the quality factor of the microwave sensor’s first
resonance.
To capture the nonlinear behavior of the microwave sensor, an artificial neural network
was developed. How both the RMSEC and the RMSEP vary as a function of the number
of hidden nodes in the artificial neural network can be seen in Figure 36. Based on the
previous criterion that the total RMSE must be reduced by more than 10% to justify using
additional hidden neurons, the neural network model for moisture content was trained
using seven hidden neurons. This resulted in both an RMSEC of 0.18 % and an RMSEP
of 0.19 % moisture content with an R2 value of 0.985. This RMSEP is nearly identical to
102
the RMSEP of the NIR artificial neural network; however, an F-test showed that the
variances are not equivalent at a 95% confidence level.
5.4
Conclusions of Study
Both microwave resonance and NIR sensing have been shown to be viable process
analytical tools for monitoring the moisture content of particulate materials. However,
prior to this study, the prevailing wisdom was that microwaves could not be used for online determination of blend uniformity. This study has shown that microwave sensors
also be used to monitor chemical composition of particulate blends accurately.
In
addition to this high degree of accuracy, microwave sensors are capable of measuring
true bulk material properties, which NIR probes cannot.
103
CHAPTER 6. DEVELOPMENT OF A REPLACEMENT ROLLER COMPACTION
SENSOR
6.1
Objective
As the sensor used in CHAPTER 4 was on a temporary loan from Sartorius Stedim,
another sensor that could be used to continue monitoring roller compacted ribbons online
needed to be designed for later use. In the replacement sensor, it was desirable to be able
to measure the chemical composition of the ribbons online.
6.2
Design of the Sensor
COMSOL Multiphysics was once again chosen for sensor design. The sensor was
chosen to be of the same type of design as that used in CHAPTER 4, a microwave cavity
resonator. After some analysis of different materials, 6101-T61 aluminum alloy was
chosen for construction, due to a low resistivity of 2.92E-08 ohm-meters [79]. Low
resistivity allows for large quality factors of the cavity’s resonances, which is necessary
for accurate sensing. A design with many resonances over the sensing range of the same
Agilent E5071B used in CHAPTER 5 would be necessary to determine chemical
composition online. Thus, in developing a new cavity resonator, each of the x-, y-, and zdimensions of the cavity needed to be unique. After some trial and error with different
dimensions in COMSOL, the x-, y-, and z-dimensions were chosen to be 3.15 inches,
4.00 inches, and 2.75 inches, respectively. Half-inch thick 6101-T61 aluminum was
chosen for construction. A 3.00 cm x 1.00 cm rectangular opening was cut into the
center of the y-z plane faces of the box to allow the ribbon to pass through the box. In
104
this design, the opening was much smaller than the opening found in the sensor detailed
in CHAPTER 4. This was done to increase the quality factors of the resonator. A
COMSOL simulation produced the reflection coefficient curve shown in Figure 41. As is
shown, many small resonance phenomena are visible. Additionally, there are several
strong resonances above 6 GHz, indicated by large sudden decreases in the reflection
coefficient. In addition, as shown in Figure 41, the resonances are very sharply defined,
indicating very large quality factors.
1
Reflection Coefficient
0.9
0.8
0.7
0.6
0.5
0.4
3
4
5
6
Frequency (Ghz)
7
8
Figure 41. COMSOL Multiphysics simulation results indicating reflection coefficient
response as a function of frequency.
105
6.3
Fabrication of the Sensor
Fabrication was carried out using Purdue University’s Artisan and Fabrication Laboratory.
The 6101-T61 aluminum alloy cavity pieces were cut using a water jet obtained from
Flow International Corporation in Kent, WA. A 4.10 mm hole was drilled into the center
of one of the x-z aluminum cavity walls using a CNC lathe obtained from Haas
Automation Inc in Oxnard, CA for later insertion of a coaxial antenna. The interior sides
of each of the aluminum pieces was then polished with successively finer grit sandpaper
sheets until a nearly mirror finish was produced. A flatter surface disperses fewer
microwaves during sensing. This results in higher quality factors of resonances and thus
improved accuracy. The pieces were then welded together to form a cavity resonator.
Four 1/8” sheets of Teflon were cut and placed inside the cavity resonator with plastic
bonding glue to create a channel through which roller compacted ribbons could pass
during testing. The final fabricated sensor can be seen in Figure 42. A PE4128-SF SMA
adapter obtained from Pasternack in Irvine, CA was then inserted into the drilled hole.
During testing, a coaxial cable was attached to this SMA adapter to excite the cavity.
106
Figure 42. Replacement cavity resonator sensor as fabricated
6.4
Materials and Methods
Blends ranging from 0-40% acetaminophen in microcrystalline cellulose (Avicel PH200)
were prepared two days prior to testing through blending for five minutes. Samples were
then stored in two layers of Ziploc bags to ensure moisture content uniformity during
testing. Roller compaction parameters and testing methodology were as established in
section 4.3, except in this study, compaction force varied from 50-80 bars. Fifty-two
roller compacted ribbons were tested. The reference chemical composition of the ribbons
107
was assumed constant throughout each sample and equal to the concentration it was
prepared to. Reference moisture contents and densities were not measured for this study,
as this type of sensor has already been shown to be very accurate for the determination of
these parameters online (4.4).
6.5
Results and Discussion
Shown in Figure 43 is an example spectrum taken online using the microwave sensor.
By comparison with Figure 41, it is apparent that many of the simulated resonance modes
are not excited; this is likely due to inevitable slight fabrication flaws, such as deviations
from perfect right angles during welding. Furthermore, many of the modes exhibit
significant broadening of their resonances when material is inserted into the sensor. This
results as the material dissipates some of the microwave energy in the form of heat.
Additionally, the surfaces could not be made perfectly smooth and some oxidation
occurred, both effects decreasing the quality factors of the resonances. Moreover, as the
electromagnetic waves slowdown in the material, the resonant frequency (center
frequency of each resonance) shifts to a lower frequency.
108
1
0.9
Reflection Coefficient
0.8
0.7
0.6
0.5
0.4
0.3
0.2
3
4
5
6
Frequency (GHz)
7
8
Figure 43. Example microwave reflection coefficient spectra taken with a roller
compacted ribbon (15% reference APAP compacted at 30 bar)
Using partial least squares regression, a calibration model was developed for the
detection of chemical composition online using this sensor from the phase response
without pretreatment of the data (an example of which is shown in Figure 44). Phase was
chosen when constructing the PLS regression model because different materials will slow
down the microwaves inside of the resonator to different extents, thus modulating the
phase of the returned signal to different extents.
Furthermore, phase based PLS
regression models were found to better predict chemical composition than magnitude
109
(reflection coefficient) based models. With four factors, 97.4 % of the variation in the
acetaminophen concentration could be accounted for by the model. An RMSEC value of
2.02 % acetaminophen concentration was found, and an RMSEP value of 2.53 %
acetaminophen concentration was calculated using four factors.
90
60
Phase (Degrees)
30
0
-30
-60
-90
3
4
5
6
Frequency (GHz)
7
8
Figure 44. Example of microwave phase spectra taken with a roller compacted ribbon (15%
reference APAP compacted at 30 bar)
110
Although there is a slight discrepancy between the predicted and calibrated errors,
indicating slight over fitting, reduction to three factors significantly reduced both RMSE
values, so four factors was chosen. Factor one explained 17% of the variance in the
microwave spectra, yet 91 % of the variation in acetaminophen
concentration and
therefore appeared to most closely be related to acetaminophen concentration. All pure
MCC samples exhibited very low values of factor one, and all high APAP concentration
samples exhibited very large corresponding values. Since reference moisture content and
density tests were not conducted, it is impossible to describe what factors two through
four are most closely related to; however, the results indicate that they are necessary
factors and are not fitting noise.
6.6
Conclusions of Study
A microwave sensor replacement was developed using finite elements. The particular
design chosen allowed the sensor to exhibit several well-defined resonances over the 28.5 GHz region. Through PLS regression, a linear model was fit to calibrate the sensor to
changes in moisture content over the relatively large roller compaction range of 30 to 50
bars. The results indicate that the sensor developed can very accurately monitor the
chemical composition of roller compacts online. The use of a resonant cavity structure
allows for very large quality factors and thus very accurate measurements.
111
CHAPTER 7. CONCLUSIONS, INSIGHTS, AND FUTURE WORK
7.1
Conclusions
The goal of this study was to expand the capabilities and improve the accuracy of
microwave sensors through model development and sensor design. A moisture content
model was developed that improved upon the accuracy of previous methods, such as the
Meyer and Schilz and Trabelsi Universal calibrations. This model was then employed to
accurately monitor the moisture content in rapidly flowing materials. To achieve the
highest level of accuracy, temperature compensation should be used to correct for
changes in the bulk dielectric constant with frictional heating. Furthermore, the density
was monitored during rapid mixing without loss of accuracy compared to static methods,
opening the door to using microwave sensors as PAT devices in high shear mixers and
granulators and pneumatic conveying lines.
Additionally, a prototype microwave sensor was used in-line to monitor both the bulk
density and the moisture content of roller compacted ribbons with improved accuracy
compared to an NIR probe. This was in part due to the increased penetration depth and
thus improved sample size of the microwave sensor. Moreover, microwave sensors are,
in general, more sensitive to the presence of water than NIR probes.
To demonstrate that microwave sensors can be used to monitor chemical composition
online, a novel microwave sensor was designed using finite elements. This procedure
allowed many variations to be tested without having to pay for prototypes to be fabricated.
112
Once a final design was chosen and fabricated, the sensor was used with flowing material
alongside an NIR probe. Calibration using a neural network model proved to capture the
nonlinear dynamics of the system and significantly increased the accuracy compared to a
linear PLS model. Results indicate that the microwave sensor is capable of measuring
both chemical composition and moisture content accurately on-line. Furthermore, the
replacement roller compact sensor was capable of measuring chemical composition in
blends with a very high degree of accuracy.
When picking a sensor for a given application, it is important to know the properties of
the materials under test, as well as the properties that are the most critical quality
attributes and to utilize the most appropriate sensing methods available. Microwave
sensors excel at the detection of bulk density, moisture content, and as demonstrated,
chemical composition, in powder blends.
7.2
Insights into Microwave Sensor Design
When designing a microwave sensor, the first key decision is whether to design a
transmission or reflectance sensor. Transmission offers the benefit of sampling larger
quantities of material, yet design and calibration increases in complexity. Furthermore,
transmission requires the use of two ports per sensor on a connected vector network
analyzer. Reflection sensors offer the opposite and only require the use of a single port
on a vector network analyzer. In general, unless one needs to sample decimeters or more
of material, reflection sensors are the preferred choice.
Next, one must make a choice between a resonant and non-resonant design. As was
shown previously, resonant sensors are usually more accurate than non-resonant designs
113
[49]. In many transmission configurations, though, resonant designs may not fit within
the process as desired. All of the sensor designs used in this work were resonant sensors.
If one wishes to monitor chemical composition or another parameter that requires very
precise measurements, it is necessary to develop a sensor that can operate over a wide
wavelength range. If one is only interested in moisture content or bulk density, operating
over a much narrower frequency range will suffice.
There is a wide variety of possible sensor configurations, covering microstrips,
waveguides, dielectric resonators, cavity resonators, transmission using pairs of horn
antennas, etc… Proper design of a microwave sensor necessitates the use of creativity
and process knowledge.
Some general guidelines should be observed though.
In
developing a resonant sensor, care should be taken to maximize the quality factors of the
resonances as much as possible. Additionally, at the center frequency of the resonance,
the reflection coefficient should be made as small as possible. Care should be taken to
avoid overlapping resonances, as this complicates measurements and makes discerning
individual resonance values very difficult.
When working with metal materials to
construct a resonator, choose materials with high conductivities to further increase the
quality factor of the resonances. To monitor how the dielectric properties of materials
under test change over the frequency spectrum available for sensing, employ a design that
maximizes the number of possible non-overlapping resonances.
7.3
7.3.1
Future Work
Further Development of Novel Microwave Sensors
Microwave sensors differ from higher frequency electromagnetic techniques in that the
wavelength of the field is on the order of the size of the sensor. Therefore, the geometry
114
of the sensor is critical to its performance. Further research into novel designs could
produce new sensors with further increased accuracy and capabilities.
7.3.2
Interactions of Microwave Fields with Particulates
From experiments conducted during the course of this study, it was apparent that
microwave sensors were less affected by extraneous material properties than other
methods, such as NIR spectroscopy. However, it is still unclear to what extent these
other properties affect measurements, and under what conditions, if any, other properties
may have a large influence on microwave measurements. Both physical characterization
studies and numerical simulations could help shed more light on this.
REFERENCES
115
REFERENCES
[1]
Charlton AJ, Farrington WHH, Brereton P. Application of 1H NMR and
Multivariate Statistics for Screening Complex Mixtures: Quality Control and
Authenticity of Instant Coffee. Journal of Agricultural and Food Chemistry. 2002
2002/05/01;50(11):3098-103.
[2]
Gowen AA, O'Donnell CP, Cullen PJ, Downey G, Frias JM. Hyperspectral
imaging – an emerging process analytical tool for food quality and safety control. Trends
in Food Science & Technology. 2007;18(12):590-8.
[3]
Alves JCL, Poppi RJ. Simultaneous determination of acetylsalicylic acid,
paracetamol and caffeine using solid-phase molecular fluorescence and parallel factor
analysis. Anal Chim Acta. 2009;642(1–2):212-6.
[4]
Alves JCL, Poppi RJ. Pharmaceutical analysis in solids using front face
fluorescence spectroscopy and multivariate calibration with matrix correction by
piecewise direct standardization. Spectrochimica Acta Part A: Molecular and
Biomolecular Spectroscopy. 2013;103(0):311-8.
[5]
Roggo Y, Chalus P, Maurer L, Lema-Martinez C, Edmond A, Jent N. A review of
near infrared spectroscopy and chemometrics in pharmaceutical technologies. J Pharm
Biomed Anal. 2007;44(3):683-700.
[6]
Vanarase AU, Alcalà M, Jerez Rozo JI, Muzzio FJ, Romañach RJ. Real-time
monitoring of drug concentration in a continuous powder mixing process using NIR
spectroscopy. Chemical Engineering Science. 2010;65(21):5728-33.
[7]
Saeed M, Probst L, Betz G. Assessment of diffuse transmission and reflection
modes in near-infrared quantification, part 2: DIFFuse reflection information depth. J
Pharm Sci. 2011;100(3):1130-41.
[8]
Harwood CF. Powder segregation due to vibration. Powder Technol.
1977;16(1):51-7.
[9]
Mark J, Karner M, Andre M, Rueland J, Huck CW. Online Process Control of a
Pharmaceutical Intermediate in a Fluidized-Bed Drier Environment Using Near-Infrared
Spectroscopy. Anal Chem. 2010 May;82(10):4209-15.
[10] Bugay DE. Characterization of the solid-state: spectroscopic techniques. Adv
Drug Deliv Rev. 2001 May;48(1):43-65.
[11] Green RL, Thurau G, Pixley NC, Mateos A, Reed RA, Higgins JP. In-line
monitoring of moisture content in fluid bed dryers using near-IR spectroscopy with
consideration of sampling effects on method accuracy. Anal Chem. 2005;77(14):4515-22.
[12] Corredor CC, Bu DS, Both D. Comparison of near infrared and microwave
resonance sensors for at-line moisture determination in powders and tablets. Anal Chim
Acta. 2011;696(1-2):84-93.
116
[13] Bartley PG, Jr., Nelson SO, McClendon RW, Trabelsi S. Determining moisture
content of wheat with an artificial neural network from microwave transmission
measurements. Instrumentation and Measurement, IEEE Transactions on.
1998;47(1):123-6.
[14] Balle TJ, Flygare WH. Fabry--Perot cavity pulsed Fourier transform microwave
spectrometer with a pulsed nozzle particle source. Rev Sci Instrum. 1981;52(1):33-45.
[15] FDA. Guidance for Industry: PAT - A Framework for Innovative Pharmaceutical
Development, Manufacturing, and Quality Assurance. 2004.
[16] Baker-Jarvis J, Geyer RG, Domich PD. A nonlinear least-squares solution with
causality constraints applied to transmission line permittivity and permeability
determination. Instrumentation and Measurement, IEEE Transactions on.
1992;41(5):646-52.
[17] Fisher K. Neues Verfahren zur massanalytischen Bestimmung des Wassergehaltes
von Flussigkeiten und festen Korpern. Agnewandte Chemie. 1935;48:394-6.
[18] Slade L, Levine H. Beyond water activity - Recent advances based on an
alternative approach to the assessment of food quality and safety. Crit Rev Food Sci Nutr.
1991;30(2-3):115-360.
[19] Trabelsi S, Nelson SO, Lewis M, Ieee. Practical Microwave Meter for Sensing
Moisture and Density of Granular Materials. 2008 Ieee Instrumentation and
Measurement Technology Conference, Vols 1-5. New York: Ieee 2008:1021-5.
[20] Kraszewski AW, Trabelsi S, Nelson SO. Comparison of density-independent
expressions for moisture content determination in wheat at microwave frequencies. J Agr
Eng Res. 1998 Nov;71(3):227-37.
[21] Trabelsi S, Kraszewski AW, Nelson SO. New calibration technique for
microwave moisture sensors. IEEE Trans Instrum Meas. 2001 Aug;50(4):877-81.
[22] King RJ, King KV, Woo K. MICROWAVE MOISTURE MEASUREMENT OF
GRAINS. IEEE Trans Instrum Meas. 1992;41(1):111-5.
[23] Kraszewski AW, You TS, Nelson SO. MICROWAVE RESONATOR
TECHNIQUE FOR MOISTURE-CONTENT DETERMINATION IN SINGLE
SOYBEAN SEEDS. IEEE Trans Instrum Meas. 1989 Feb;38(1):79-84.
[24] Kraszewski AW, Nelson SO, You TS. USE OF A MICROWAVE CAVITY FOR
SENSING DIELECTRIC-PROPERTIES OF ARBITRARILY SHAPED BIOLOGICAL
OBJECTS. IEEE Trans Microw Theory Tech. 1990 Jul;38(7):858-63.
[25] Abegaonkar MP, Karekar RN, Aiyer RC. A microwave microstrip ring resonator
as a moisture sensor for biomaterials: application to wheat grains. Meas Sci Technol.
1999 Mar;10(3):195-200.
[26] Joshi KK, Abegaonkar M, Karekar RN, Aiyer RC. Microstrip ring resonator as a
moisture sensor for wheat grains. In: Koepf GA, Jelks C, Huebner D, Donovan JJ, eds.
1997 Ieee Mtt-S International Microwave Symposium Digest, Vols I-Iii - High
Frequencies in High Places. New York: I E E E 1997:1679-82.
[27] Abegaonkar MP, Karekar RN, Aiyer RC. Miniaturized nondestructive microwave
sensor for chickpea moisture measurement. Rev Sci Instrum. 1999 Jul;70(7):3145-9.
117
[28] Mironov VL, Komarov SA, Kleshchenko VN, Ieee. Microwave dielectric
spectroscopy for bound water in saline soil. IGARSS 2005: IEEE International
Geoscience and Remote Sensing Symposium, Vols 1-8, Proceedings. New York: Ieee
2005:3196-9.
[29] Shu Y, Wong TY. Perturbation of dielectric resonator for material measurement.
Electronics Letters. 1995;31(9):704-5.
[30] Sarabandi K, Li ES. Microstrip ring resonator for soil moisture measurements.
IEEE Trans Geosci Remote Sensing. 1997 Sep;35(5):1223-31.
[31] Reh C, Bhat SN, Berrut S. Determination of water content in powdered milk.
Food Chem. 2004 Jul;86(3):457-64.
[32] King RJ, Werner MJ, Mayorga GD. MICROWAVE DYNAMIC DIELECTRIC
ANALYSIS OF CURING NEAT RESINS. Journal of Reinforced Plastics and
Composites. 1993;12(2):173-85.
[33] Tondar M, Luy B, Prasch A. Method for monitoring and/or controlling a
granulation, coating and drying process. Google Patents 2002.
[34] Buschmuller C, Wiedey W, Doscher C, Dressler J, Breitkreutz J. In-line
monitoring of granule moisture in fluidized-bed dryers using microwave resonance
technology. Eur J Pharm Biopharm. 2008 May;69(1):380-7.
[35] Lourenco V, Herdling T, Reich G, Menezes JC, Lochmann D. Combining
microwave resonance technology to multivariate data analysis as a novel PAT tool to
improve process understanding in fluid bed granulation. European journal of
pharmaceutics and biopharmaceutics : official journal of Arbeitsgemeinschaft fur
Pharmazeutische Verfahrenstechnik eV. 2011 Feb 17;78(3):513-21.
[36] Sung P-F, Hsieh Y-L, Angonese K, Dunn D, King RJ, Machbitz R, et al.
Complex Dielectric Properties of Microcrystalline Cellulose, Anhydrous Lactose, and
alpha-Lactose Monohydrate Powders Using a Microwave-based Open-reflection
Resonator Sensor. J Pharm Sci. 2011 Jul;100(7):2920-34.
[37] Betz G, Junker-Bürgin P, Leuenberger H. Batch And Continuous Processing In
The Production Of Pharmaceutical Granules#. Pharm Dev Technol. 2003;8(3):289-97.
[38] Nelson SO. Measurement of microwave dielectric-properties of particulate
materials. J Food Eng. 1994;21(3):365-84.
[39] Austin J, Rodriguez S, Sung P-F, Harris M. Utilizing microwaves for the
determination of moisture content independent of density. Powder Technol. 2013;236:1723.
[40] Trabelsi S, Nelson SO. Nondestructive sensing of physical properties of granular
materials by microwave permittivity measurement. IEEE Trans Instrum Meas. 2006
Jun;55(3):953-63.
[41] Meyer W, Schilz W. A microwave method for density independent determination
of the moisture-content of solids. J Phys D-Appl Phys. 1980;13(10):1823-30.
[42] Trabelsi S, Krazsewski AW, Nelson SO. New density-independent calibration
function for microwave sensing of moisture content in particulate materials. IEEE Trans
Instrum Meas. 1998 Jun;47(3):613-22.
[43] Trabelsi S, Kraszewski AW, Nelson SO. Microwave dielectric sensing of bulk
density of granular materials. Meas Sci Technol. 2001 Dec;12(12):2192-7.
118
[44] Heng PWS, Loh ZH, Liew CV, Lee CC. Dielectric Properties of Pharmaceutical
Materials Relevant to Microwave Processing: Effects of Field Frequency, Material
Density, and Moisture Content. J Pharm Sci. 2010 Feb;99(2):941-57.
[45] Nelson SO, You TS. RELATIONSHIPS BETWEEN MICROWAVE
PERMITTIVITIES OF SOLID AND PULVERIZED PLASTICS. J Phys D-Appl Phys.
1990 Mar;23(3):346-53.
[46] King RJ. CONTINUOUS MOISTURE AND DENSITY MEASUREMENT OF
FOODS USING MICROWAVES. Food Processing Automation III. 1994;94(1):231-40.
[47] Trabelsi S, Kraszewski AW, Nelson SO. A microwave method for on-line
determination of bulk density and moisture content of particulate materials. IEEE Trans
Instrum Meas. 1998 Feb;47(1):127-32.
[48] Kaatze U. Techniques for measuring the microwave dielectric properties of
materials. Metrologia. 2010 Apr;47(2):S91-S113.
[49] Chen L. Microwave electronics : measurement and materials characterisation.
Chichester: John Wiley 2004.
[50] King RJ, inventor KDC Technology Corporation, assignee. Microwave Reflection
Resonator Sensors. United States patent 5334941. 1994.
[51] Cheng WH, Fedotov AV, Gluckstern RL. Frequency dependence of the
penetration of electromagnetic fields through a small coupling hole in a thick wall. Phys
Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics. 1995 Sep;52(3):3127-42.
[52] Mort PR. Scale-up and control of binder agglomeration processes — Flow and
stress fields. Powder Technol. 2009;189(2):313-7.
[53] Zografi G, Kontny MJ, Yang AYS, Brenner GS. Surface-area and water-vapor
sorption of microcrystalline cellulose. Int J Pharm. 1984;18(1-2):99-116.
[54] Luukkonen P, Maloney T, Rantanen J, Paulapuro H, Yliruusi J. Microcrystalline
cellulose-water interaction - A novel approach using thermoporosimetry. Pharm Res.
2001 Nov;18(11):1562-9.
[55] Barthel J, Bachhuber K, Buchner R, Hetzenauer H. Dielectric spectra of some
common solvents in the microwave region. Water and lower alcohols. Chemical Physics
Letters. 1990;165(4):369-73.
[56] Feng T, Wang F, Pinal R, Wassgren C, Carvajal MT. Investigation of the
Variability of NIR In-line Monitoring of Roller Compaction Process by Using Fast
Fourier Transform (FFT) Analysis. AAPS PharmSciTech. 2008 2008/06/01;9(2):419-24.
[57] Bourdel E, Pasquet D, Denorme P, Roussel A. Measurement of the moisture
content with a cylindrical resonating cavity in TM<sub>010</sub> mode.
Instrumentation and Measurement, IEEE Transactions on. 2000;49(5):1023-8.
[58] Muliadi AR, Litster JD, Wassgren CR. Modeling the powder roll compaction
process: Comparison of 2-D finite element method and the rolling theory for granular
solids (Johanson's model). Powder Technol. 2012;221(0):90-100.
[59] Ching-Lei L, Kun-Mu C. Determination of electromagnetic properties of
materials using flanged open-ended coaxial probe-full-wave analysis. Instrumentation
and Measurement, IEEE Transactions on. 1995;44(1):19-27.
[60] Krupka J, Derzakowski K, Riddle B, Baker-Jarvis J. A dielectric resonator for
measurements of complex permittivity of low loss dielectric materials as a function of
temperature. Measurement Science and Technology. 1999;9(10):1751.
119
[61] Gentili G, Riminesi C, Tesi V. Low Cost Microwave Sensor for Moisture Content
Measurement in Paper Milling Industry. Sensing and Imaging: An International Journal.
2006 2006/12/01;7(4):155-73.
[62] Anlage S, Talanov V, Schwartz A. Principles of Near-Field Microwave
Microscopy. In: Kalinin S, Gruverman A, eds. Scanning Probe Microscopy: Springer
New York 2007:215-53.
[63] Komarov V, Wang S, Tang J. Permittivity and measurements. Encyclopedia of
RF and microwave engineering. 2005.
[64] Krupka J, Weil C. Recent advances in metrology for the electromagnetic
characterization of materials at microwave frequencies. Microwaves and Radar, 1998
MIKON '98, 12th International Conference on; 1998 20-22 May 1998; 1998. p. 243-53
vol.4.
[65] Nesarikar VV, Vatsaraj N, Patel C, Early W, Pandey P, Sprockel O, et al.
Instrumented roll technology for the design space development of roller compaction
process. Int J Pharm. 2012;426(1–2):116-31.
[66] Knochel R, Taute W, Doscher C. Stray field ring resonators and a novel trough
guide resonator for precise microwave moisture and density measurements. Meas Sci
Technol. 2007 Apr;18(4):1061-8.
[67] Zoughi R, Ganchev S. Microwave Nondestructive Evaluation: State-of-the-Art
Review: DTIC Document; 1995.
[68] Candolfi A, De Maesschalck R, Jouan-Rimbaud D, Hailey PA, Massart DL. The
influence of data pre-processing in the pattern recognition of excipients near-infrared
spectra. J Pharm Biomed Anal. 1999 Oct;21(1):115-32.
[69] Kajfez D, Chebolu S, Abdul-Gaffoor MR, Kishk AA. Uncertainty analysis of the
transmission-type measurement of Q-factor. Microwave Theory and Techniques, IEEE
Transactions on. 1999;47(3):367-71.
[70] Cerny P, Piksa P, Korinek T, Zvanovec S, Mazanek M. Broadband axial
excitation of Fabry-Perot resonator. Antennas and Propagation (EuCAP), 2010
Proceedings of the Fourth European Conference on; 2010 12-16 April 2010; 2010. p. 1-4.
[71] Bansal R. Fundamentals of Engineering Electromagnetics. CRC Press 2006:24952.
[72] Nyfors E, Vainikainen P. Industrial microwave sensors. Microwave Symposium
Digest, 1991, IEEE MTT-S International; 1991 10-14 Jul 1991; 1991. p. 1009-12 vol.3.
[73] Austin J, Gupta A, McDonnell R, Reklaitis GV, Harris MT. The use of nearinfrared and microwave resonance sensing to monitor a continuous roller compaction
process. J Pharm Sci. 2013;102(6):1895-904.
[74] Reich G. Near-infrared spectroscopy and imaging: Basic principles and
pharmaceutical applications. Adv Drug Deliv Rev. 2005 Jun;57(8):1109-43.
[75] Acevedo D, Muliadi A, Giridhar A, Litster JD, Romanach RJ. Evaluation of three
approaches for real-time monitoring of roller compaction with near-infrared spectroscopy.
AAPS PharmSciTech. 2012 Sep;13(3):1005-12.
[76] Carter GA, McCain DC. Relationship of leaf spectral reflectance to chloroplast
water content determined using NMR microscopy. Remote Sens Environ.
1993;46(3):305-10.
120
[77] Jain AK, Jianchang M, Mohiuddin KM. Artificial neural networks: a tutorial.
Computer. 1996;29(3):31-44.
[78] Gupta A, Peck GE, Miller RW, Morris KR. Influence of ambient moisture on the
compaction behavior of microcrystalline cellulose powder undergoing uni-axial
compression and roller-compaction: A comparative study using near-infrared
spectroscopy. J Pharm Sci. 2005 Oct;94(10):2301-13.
[79] Davis JR, Associates JRD, Committee ASMIH. Aluminum and Aluminum Alloys:
ASM International 1993.
APPENDICES
121
Appendix A COMSOL Simulations
COMSOL simulations were carried out using an I7 Intel 3770 processor with access to 32
GB of RAM. The RF package was used to study the electromagnetic fields produced by
the sensor. The following boundary conditions were employed:
1) Perfect Electric Conductor:
on metal surfaces far from high concentrations of the electric field, such as on
the exterior of the coaxial cable.
2) Impedance Boundary Condition:
√
⁄
on metal surfaces near where the
electric field was concentrated, which attempts to model how a small fraction of an
incident electromagnetic field dissipates when it reflects off of a metal surface.
3) Perfectly Matched Layer:
A perfectly matched layer is a numerical approximation to the real-life anechoic chamber
used in sound studios and EE labs. This layer is designed to absorb waves coming
towards it while minimizing the reflection from the surface and was used to truncate the
infinite domain.
122
4) Absorbing Boundary Condition:
(
)
(
⁄ )
(
)
( (
)
⁄ )
on the exterior of the perfectly matched layer, to further simulate an infinite domain.
5) Lumped Port:
and
feed cable to excite the sensor.
(
)
on the end of the coaxial
123
Figure 45. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a first order
resonance at approximately 2.4 GHz.
124
Figure 46. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s bottom ring experiences a second order
resonance at approximately 4.3 GHz.
125
Figure 47. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a second order
resonance at approximately 4.7 GHz.
126
Figure 48. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s bottom ring experiences a third order
resonance at approximately 6.3 GHz.
127
Figure 49. COMSOL simulation result demonstrating how the norm of the electric field
varies on a log base 10 scale when the sensor’s top ring experiences a third order
resonance at approximately 7.0 GHz.
128
Appendix B Standard Operating Procedures to Use Vector Network Analyzer
Experimental Procedure to Calibrate the E5071B
A calibration uses known standards to eliminate the effects of systematic errors. The
calibration should be performed under the same condition as testing will occur. Perform
a calibration if the temperature or humidity in the room has changed notably (i.e. when
seasons change). Likewise, a calibration will need to be performed if any of the cables or
sensors is moved appreciably. When calibrating, leave the cable in the position it will be
in during testing. Attach the standards, when prompted, to the end of the cable where the
sensor will be placed. Ensure you only use the correct female standards.
1. Ensure that the VNA is plugged in to an outlet and is connected to a PC via the square
USB port on the back of the device.
2. Ensure that the power switch on the back is set to on. DO NOT SWITCH THIS TO
OFF UNLESS IN AN EMERGENCY. USE THE FRONT PANEL POWER
BUTTON TO POWER ON AND OFF THE DEVICE.
3. Power on the E5071B using the power button located in the bottom left corner of the
front panel of the device; wait until the operating system and software have
completely loaded.
4. Allow around 45 minutes for the VNA’s internal temperature to reach steady state.
On screen, you will see a READY indicator light up with a blue background on the
bottom bar when it is all right to proceed.
129
5. Press Channel Next or Channel Prev to select the channel for which you want to
perform the calibration. Channel 1 will be for the planar sensor connected to port 1.
Channel 2 will be for the box sensor connected to port 2. If only one channel is
active, you can enable the second channel by pressing DISPLAY and then allocating
a second channel by showing two channels at once.
6. From the instrument panel, press the button labeled: CAL.
7. Press Calibrate
8. Press 1-Port Cal
9. Press Select Port
10. Select a test port (1 for calibration of sensor at port 1, 2 for calibration of sensor at
port 2).
11. Connect an OPEN calibration standard to the test port. BE INCREDIBLY
CAREFUL WHEN THREADING ON THE STANDARDS. DO NOT TWIST THE
STANDARDS. Tighten the standard using the supplied torque wrench by holding
the wrench at the end. DO NOT OVER TIGHTEN.
12. Press OPEN
13. Repeat steps 11 and 12 using a SHORT instead.
14. Repeat steps 11 and 12 using a LOAD instead.
15. Press DONE to terminate the calibration process. Upon pressing this key, calibration
coefficients will be calculated and saved. The error correction function will also be
automatically enabled.
130
16. Save the calibration by pressing the Save/Recall button. Then select Save State.
Next, select State01. You need to save it as State01, as this is the state that will be
called by Matlab. This will override the previous data/calibration permanently.
Experimental Procedure for Using the E5071B Connected to a PC Running Matlab to
Gather Spectroscopic and Resonance Data
1. Ensure that the VNA is plugged in to an outlet and is connected to a PC via the square
USB port on the back of the device (if using the DeltaV system to store data, the
vector network analyzer must be connected to the right USB port of the first DeltaV
computer).
2. Ensure that the power switch on the back is set to on. DO NOT SWITCH THIS TO
OFF UNLESS IN AN EMERGENCY. USE THE FRONT PANEL POWER
BUTTON TO POWER ON AND OFF THE DEVICE.
3. Power on the E5071B using the power button located in the bottom left corner of the
front panel of the device; wait until the operating system and software have
completely loaded.
4. Allow around 45 minutes for the VNA’s internal temperature to reach steady state.
On screen, you will see a READY indicator light up with a blue background on the
bottom bar when it is all right to proceed.
5. Open the Matlab file: “VNA_File.m”, contained in Appendix C, and ensure that the
function file “WriteToDeltaV.m” is in the same directory.
131
6. Run the Matlab script “VNA_File.m”. Reply as to whether you want to take a single
measurement or take measurements continuously. If a single measurement is chosen,
the instrument will prompt you to enter the number of sweeps to average. (Increasing
this increases the accuracy, but takes longer. The default is three.) This will setup
the instrument correctly and begin taking measurements from the sensor. Sweeping
and recording data takes approximately 1 second per iteration per port. Information
on the resonance peaks are downloaded to the matrix Peaks after each peak. A
complete spectra is downloaded each sweep and can be found in the Matlab matrix
SParameters for sensor 1; for sensor 2, the s-parameter data will be stored in
TwoSParameters.
At any time, the taking of measurements can be stopped by pressing CTRL+C. However,
one must then enter in the command window: “fclose(obj1)” to disconnect Matlab from
the instrument correctly.
132
Appendix C Matlab VNA Files
VNA_File.m
% This file is the main file for controlling the sensor.
clear all;
clc;
%% Define Parameters
global obj1 apapNN;
% Load APAP neural network from memory. The neural network must be located
% in the same folder as this file. It will load as a network object
% labeled apapNN and be made a global variable.
load NETforAPAP;
% Sets up empty frequency vector to hold frequency values
freq = zeros(1601,1);
start = 1.8E9;
stop = 7.5E9;
numOfPoints = 1601;
% Define table to hold peak search values
Peaks = zeros(18,4); % Bandwidth, Center Frequency, Q Value, Loss
% Define Sweep Table for Sensor 1
Sweep1Start = 1.8E9;
Sweep2Start = 2.6E9;
Sweep3Start = 3.7E9;
Sweep4Start = 4.8E9;
Sweep5Start = 5.8E9;
Sweep5End = 7.5E9;
Sweep2Points = 40;
Sweep3Points = 450;
Sweep4Points = 40;
Sweep5Points = 350;
Sweep1Points = 1601 - (Sweep2Points + Sweep3Points + Sweep4Points +
Sweep5Points);
GapOffset1 = (Sweep3Start - Sweep2Start)/(Sweep2Points + 1);
GapOffset2 = (Sweep5Start - Sweep4Start)/(Sweep4Points + 1);
numOfAverages = 3;
133
% Ask the user what type of measurement to take
choice = input('Would you like to take a single measurement (1) or enter continuous
mode for logging to DeltaV system (2)? Until a PLS model is developed for roller
compactor sensor, it is recommended to only do single measurements for sensor two.
DeltaV transfer from sensor 2 is not yet enabled. SParameter data must be transfered to
excel file after each grab. ');
if choice == 1
numOfAverages = input('Enter the number of sweeps to average together: ');
end
% Find a VISA-USB object.
obj1 = instrfind('Type', 'visa-usb', 'RsrcName',
'USB0::0x0957::0x0509::my42402527::0::INSTR', 'Tag', '');
% Create the VISA-USB object if it does not exist
% otherwise use the object that was found.
if isempty(obj1)
obj1 = visa('AGILENT', 'USB0::0x0957::0x0509::my42402527::0::INSTR');
else
fclose(obj1);
obj1 = obj1(1);
end
% Set a sufficiently large input buffer size to store the S-Parameter data.
% You need 40 bytes per data point.
set(obj1, 'InputBufferSize', numOfPoints*40);
% Increase Output Buffer Size
set(obj1, 'OutputBufferSize', 3000);
% Set large timeout in the event of long s-parameter measurement
set(obj1, 'Timeout', 30);
% Connect to instrument object, obj1.
fopen(obj1);
%% Set up instruments, obj1
fprintf(obj1, '%s', ':MMEM:LOAD "D:State01.sta"'); % Set instrument state to State01
saved on machine
fprintf(obj1, '%s', ':DISP:SPL D1_2');
% Setup Channel 1 (Channel 1 is for Sensor 1).
fprintf(obj1, '%s', ':CALC1:PAR1:SEL'); % Set Chanel 1, Trace 1 to be source of data
fprintf(obj1, '%s', ':SENS1:AVER ON'); % Turns on averaging
fprintf(obj1, '%s', [':SENS1:AVER:COUN ', num2str(numOfAverages)]); % Changes to
using numOfAverage sweeps per averaging
134
fprintf(obj1, '%s', ':TRIG:AVER ON'); % Turns on trigger averaging, which resets the
average each loop
fprintf(obj1, '%s', ':SENS1:CORR:STAT ON'); % Turns on error correction
fprintf(obj1, '%s', [':SENS1:SWE:POIN ', num2str(numOfPoints)]); % Set number of
points
fprintf(obj1, '%s', ':CALC1:FORM MLOG'); % Set data to be displayed in log-magnitude
format
fprintf(obj1, '%s', ':SENS1:SWE:TIME:AUTO ON'); % Set sweep time to auto (shortest
possible time)
fprintf(obj1, '%s', ':CALC1:SMO OFF'); % Set smoothing off - smoothing destroys
resolution of sharp peaks
fprintf(obj1, '%s', ':INIT1:CONT OFF'); % Turns off continuous measurements and sets
the trigger state to hold
fprintf(obj1, '%s', ':TRIG:SOUR INT'); % Sets the trigger source to internal
fprintf(obj1, '%s', ':FORM:DATA ASC'); % Sets data to be output in ASCII format
fprintf(obj1, '%s', ':CALC1:PAR1:DEF S11'); % Set the measured parameter to S11
fprintf(obj1, '%s', [':SENS1:SEGM:DATA 6,0,1,0,0,1,0,5,', num2str(Sweep1Start), ',',
num2str(Sweep2Start), ',', num2str(Sweep1Points), ',350E3,0,', num2str(Sweep2Start +
GapOffset1), ',', num2str(Sweep3Start - GapOffset1), ',', num2str(Sweep2Points),
',70E3,2,', num2str(Sweep3Start), ',', num2str(Sweep4Start), ',', num2str(Sweep3Points),
',70E3,2,', num2str(Sweep4Start + GapOffset2), ',', num2str(Sweep5Start - GapOffset2),
',', num2str(Sweep4Points), ',70E3,2,', num2str(Sweep5Start), ',', num2str(Sweep5End), ',',
num2str(Sweep5Points), ',70E3,2']); % Create a sweep segment table
fprintf(obj1, '%s', ':SENS1:SWE:TYPE SEGM'); % Sets the sweep type to segmented
sweep
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:PEXC 3');% Set Peak excursion value to 3
%Setup marker for bandwidth search
fprintf(obj1, '%s', ':CALC1:MARK1 ON'); % Turn marker 1 on
fprintf(obj1, '%s', ':CALC1:MARK1:BWID ON'); % Turn bandwidth search on
fprintf(obj1, '%s', ':CALC1:MARK1:BWID:THR 3'); % Set bandwidth search value to 3
dB
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM ON'); % Turn marker search range on
fprintf(obj1, '%s', ':CALC1:MARK1:DISC OFF'); % Turn on marker ability to sit
between points
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:PPOL NEG'); % Sets the peak search
polarity to negative (i.e. looking for valleys)
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:TYPE PEAK'); % Sets the marker search
type to peak search
% Gather time required to take a sweep in seconds (Placed here because it
% takes some time
SweepTime = str2double(query(obj1, ':SENS1:SEGM:SWE:TIME?', '%s')); % Retrieves
the sweep time of the segmented sweep to be used for wait()
135
% Pause and let machine calibration take effect
pause(4);
if choice == 1
FINALNUMBER = 1;
elseif choice == 2
FINALNUMBER = inf;
else
FINALNUMBER = 0;
end
%% Take measurements
for j = 1:FINALNUMBER
pause(0.25);
fprintf(obj1, '%s', ':INIT1');
wait(numOfAverages,SweepTime);
% Find Location/Depth/Bandwidth of different peaks
% Peak 1
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 1.8E9'); % Start search at 1.8
GHZ
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 2.2E9'); % Stop search at 2.2
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(1,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
% Peak 2
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 2.2E9'); % Start search at 2.2
GHZ
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 2.5E9'); % Stop search at 2.5
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(2,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
% Peak 3
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 3.8E9'); % Start search at 3.8
GHZ
136
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 4.3E9'); % Stop search at 4.3
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(3,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
% Peak 4
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 4.3E9'); % Start search at 4.3
GHZ
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 4.8E9'); % Stop search at 4.8
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(4,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
% Peak 5
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 5.8E9'); % Start search at 5.8
GHZ
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 6.5E9'); % Stop search at 6.5
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(5,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
% Peak 6
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STAR 6.5E9'); % Start search at 6.5
GHZ
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:DOM:STOP 7.2E9'); % Stop search at 7.2
GHz
fprintf(obj1, '%s', ':CALC1:MARK1:FUNC:EXEC'); % Conducts a peak search
Btmp1 = query(obj1, ':CALC1:MARK1:BWID:DATA?', '%s'); % Finds the bandwidth,
the center frequency, the q value, and the loss in that order
Btmp2 = regexp(Btmp1,'([^ ,:]*)','tokens');
Btmp3 = cat(2,Btmp2{:});
Peaks(6,:) = str2double(Btmp3); % Converts strings in scientific notation to doubles
and places in peaks array
137
% Query the instrument to get all of the S-parameter data.
RawSParameters = query(obj1, ':CALC1:DATA:SDATA?', '%s');
tmp1 = regexp(RawSParameters,'([^ ,:]*)','tokens');
tmp2 = cat(2,tmp1{:});
% Converts strings in scientific notation to doubles
tmp3 = str2double(tmp2);
% Changes from a 1X3202 to 2X1601 with real and imaginary s-parameters
% paired, then changes to 1601X2
SParameters = reshape(tmp3, [2 numOfPoints]);
SParameters = SParameters.';
%Create frequency values to go with S-Parameters
counter = 1;
for i=1:Sweep1Points % First Segment
freq(counter) = Sweep1Start + (i-1)*(Sweep2Start - Sweep1Start)/(Sweep1Points-1);
counter = counter + 1;
end
for i=1:Sweep2Points % Second Segment
freq(counter) = (Sweep2Start + GapOffset1) + (i-1)*((Sweep3Start - GapOffset1) (Sweep2Start + GapOffset1))/(Sweep2Points-1);
counter = counter + 1;
end
for i=1:Sweep3Points % Third Segment
freq(counter) = Sweep3Start + (i-1)*(Sweep4Start - Sweep3Start)/(Sweep3Points-1);
counter = counter + 1;
end
for i=1:Sweep4Points % Fourth Segment
freq(counter) = (Sweep4Start + GapOffset2) + (i-1)*((Sweep5Start - GapOffset2) (Sweep4Start + GapOffset2))/(Sweep4Points-1);
counter = counter + 1;
end
for i=1:Sweep5Points % Fifth Segment
freq(counter) = Sweep5Start + (i-1)*(Sweep5End - Sweep5Start)/(Sweep5Points-1);
counter = counter + 1;
end
SParameters = [freq, SParameters];
% Write bandwidth and center frequency information to DeltaV
WriteToDeltaV(Peaks);
end
%% Disconnect from instrument object, obj1.
fclose(obj1);
138
WriteToDeltaV.m
function [] = WriteToDeltaV(Peaks)
% This function writes over OPC to Delta V. It transfers the peak
% information for control purposes.
global apapNN
Inputs = [(Peaks(:,1));(Peaks(:,2));(Peaks(:,3))];
Comp = 1 - sim(apapNN,Inputs);
% Create and open OPC object called OpcObj
OpcObj = opcda('128.210.180.238', 'OPC.DeltaV.1');
connect(OpcObj);
grp = addgroup(OpcObj, 'OPCWRITE');
% Define items to send to server
BW1 = additem(grp,'MWSENSOR1/BANDWIDTH1.CV');
BW2 = additem(grp,'MWSENSOR1/BANDWIDTH2.CV');
BW3 = additem(grp,'MWSENSOR1/BANDWIDTH3.CV');
BW4 = additem(grp,'MWSENSOR1/BANDWIDTH4.CV');
BW5 = additem(grp,'MWSENSOR1/BANDWIDTH5.CV');
BW6 = additem(grp,'MWSENSOR1/BANDWIDTH6.CV');
CFREQ1 = additem(grp,'MWSENSOR1/CFREQUENCY1.CV');
CFREQ2 = additem(grp,'MWSENSOR1/CFREQUENCY2.CV');
CFREQ3 = additem(grp,'MWSENSOR1/CFREQUENCY3.CV');
CFREQ4 = additem(grp,'MWSENSOR1/CFREQUENCY4.CV');
CFREQ5 = additem(grp,'MWSENSOR1/CFREQUENCY5.CV');
CFREQ6 = additem(grp,'MWSENSOR1/CFREQUENCY6.CV');
FracMCC = additem(grp, 'MWSENSOR1/FRACTIONMCC.CV');
% Write data to DeltaV
start(grp);
DataToWrite = num2cell([cat(1,Peaks(:,1),Peaks(:,2));Comp]);
write(grp, DataToWrite);
% Disconnect and clear memory
disconnect(OpcObj);
delete(OpcObj);
clear OpcObj grp BW1 BW2 BW3 BW4 BW5 BW6 CFREQ1 CFREQ2 CFREQ3
CFREQ4 CFREQ5 CFREQ6 FracMCC
end
13
VITA
139
VITA
John Austin received his BS degree with a double major in Chemical Engineering and
Chemistry from Worcester Polytechnic Institute in Worcester, Massachusetts. While
there, he worked on the development of an improved protein biosensor utilizing aptamer
ligands and iron oxide nanoparticles. He enrolled in the Chemical Engineering PhD
program at Purdue University in 2010. His current research area is focused on the
development of novel process analytical devices used to monitor particulate processes.
He will receive his PhD degree at Purdue University in May 2014.
PUBLICATIONS
140
PUBLICATIONS
Austin J, Rodriguez S, Sung P-F, Harris M. Utilizing microwaves for the determination
of moisture content independent of density. Powder Technol. 2013;236:17-23.
Austin J, Gupta A, McDonnell R, Reklaitis GV, Harris MT. The use of near-infrared and
microwave resonance sensing to monitor a continuous roller compaction process. J
Pharm Sci. 2013;102(6):1895-904.
Austin J, Harris M. In-Situ Monitoring of the Bulk Density and the Moisture Content of
Rapidly Flowing Microcrystalline Cellulose Using a Microwave Resonance Sensor.
Sensors Journal, IEEE. 2013;PP(99).
Austin J, Gupta A, McDonnell R, Reklaitis GV, Harris MT. A novel microwave sensor to
determine particulate blend composition on-line. Anal Chim Acta. 2014.
Документ
Категория
Без категории
Просмотров
0
Размер файла
4 649 Кб
Теги
sdewsdweddes
1/--страниц
Пожаловаться на содержимое документа