Upload
others
View
4
Download
1
Embed Size (px)
Citation preview
Sensor-based Online Process Monitoring in Advanced Manufacturing
David Mathew Roberson III
Thesis submitted to the faculty of the Virginia Polytechnic Institute and State University in partial fulfillment of the requirements for the degree of
Master of Science
in Industrial and Systems Engineering
Zhenyu (James) Kong, Committee Chair Ran Jin
Prahalad K. Rao
August 8, 2016 Blacksburg, VA
Keywords: Additive Manufacturing, Fused Deposition Modeling, Machining, Turning,
Sensor-based Monitoring
Sensor-based Online Process Monitoring in Advanced Manufacturing
David Mathew Roberson III
ABSTRACT
Effective quality improvement in the manufacturing industry is continually pursued.
There is an increasing demand for real-time fault detection, and avoidance of destructive
post-process testing. Therefore, it is desirable to employ sensors for in-process monitoring,
allowing for real-time quality assurance. Chapter 3 describes the application of sensor-
based monitoring to additive manufacturing, in which sensors are attached to a desktop
model fused deposition modeling machine, to collect data during the manufacturing
process. A design of experiments plan is conducted to provide insight into the process,
particularly the occurrence of process failure. Subsequently, machine learning
classification techniques are applied to detect such failure, and successfully demonstrate
the future potential of this platform and methodology. Chapter 4 relates the application of
online, image-based quantification of the surface quality of workpieces produced by
cylindrical turning. Representative samples of cylindrical shafts, machined by turning
under various conditions, are utilized, and an apparatus is constructed for acquiring images
while the part remains mounted on a lathe. The surface quality of these specimens is
analyzed, employing an algebraic graph theoretic approach, and preliminary regression
modeling displays an average surface roughness (Ra) prediction error of less than 8%.
Prediction occurs in less than 2 seconds, showing the capability for future application in a
real-time, quality control setting. Both of these cases, in additive manufacturing and in
turning, are validated using real experimental data and analysis, showing application of
sensor-based online process monitoring in multiple manufacturing areas.
Sensor-based Online Process Monitoring in Advanced Manufacturing
David Mathew Roberson III
GENERAL AUDIENCE ABSTRACT
Effective quality improvement in the manufacturing industry is continually pursued,
and there is an increasing demand for real-time quality monitoring. Therefore, it is
desirable to employ sensors for in-process monitoring, allowing for real-time quality
assurance. This is explored in two manufacturing areas. The first section of this work is in
the area of additive manufacturing (“3D printing”), in which sensors are attached to a
desktop model machine, to collect data during the printing process. Experiments are
conducted to provide insight into how the process behaves, particularly the occurrence of
printing failure. Machine learning classification techniques are then applied to detect such
failure, and successfully demonstrate the future potential of this platform and methodology,
for real-time monitoring of the process. The second section of this work relates to the
conventional machining process of turning, and describes the application of image-based
measurement of surface roughness. An apparatus is constructed for acquiring images, while
the cylindrically turned shaft remains mounted on the lathe. The surface roughness is
measured, and preliminary modeling displays an average surface roughness prediction
error of less than 8%. This prediction occurs in less than 2 seconds, showing the capability
for future application in a real-time, quality control setting. Both of these cases, in additive
manufacturing and in turning, show the application of sensor-based monitoring in various
manufacturing areas. This work provides a basis for future research and application,
demonstrating how this sensor-based monitoring approach may be used for real-time
quality monitoring in manufacturing.
iv
ACKNOWLEDGEMENTS
I would like to acknowledge and thank my advisor, Dr. Kong, for his guidance. He has kindly
believed in and encouraged me, and without his support I would not have been able to begin my
graduate career. I thank Dr. Rao for his valuable influence, motivating me to pursue independent
and critical thought. And I also thank Dr. Jin, for taking his time in being a part of my committee,
and for his thoughtful advice and input.
I would also like to thank my fellow lab members, Jia Liu, Kaveh Bastani, Chen-ang Liu, and
Babak Barazandeh, for their selfless friendship and help.
Most of all, I’d like to thank my family for their constant love and support throughout my entire
life, even when I can’t physically be with them. Without them I could not have gotten this far, and
I love them very much.
v
Table of Contents Chapter 1. Introduction ............................................................................................................................1
1.1 Motivation and Objective of the Proposed Research Related to Additive Manufacturing ..........2
1.2 Motivation and Objective of the Proposed Research Related to the Turning Operation .............3
Chapter 2. Literature Review ..................................................................................................................4
2.1 Review of the Literature Relating to Process Monitoring in Additive Manufacturing ................4
2.1.1 Typical defects in the FDM process ........................................................................................4
2.1.2 Prior experimental studies in FDM ........................................................................................10
2.1.3 Sensor-based monitoring in FDM ..........................................................................................11
2.2 Review of the Literature Relating to Surface Finish Monitoring in Turning .............................13
2.2.1 Optical methods for surface roughness estimation ................................................................13
2.2.2 Shortcomings in the existing literature ..................................................................................14
Chapter 3. Sensor-based Process Monitoring in Additive Manufacturing ........................................16
3.1 Process description of FDM .......................................................................................................17
3.2 Experimental Studies in FDM ....................................................................................................18
3.2.1 Design of standard test artifact ...............................................................................................18
3.2.2 Selection of process input and output variables .....................................................................19
3.2.3 Identification of process conditions leading to build failures ................................................21
3.2.4 Analysis of build failure in FDM ...........................................................................................24
3.3 Sensor-based Monitoring of FDM .............................................................................................26
3.3.1 Sensor platform for real-time monitoring of FDM process conditions .................................26
3.3.2 Discussion of acquired sensor data ........................................................................................30
3.4 Machine Learning Classification of FDM Process States ..........................................................32
3.4.1 Machine learning classification algorithm selection ..............................................................33
3.4.2 Sensor data selection ..............................................................................................................33
vi
3.4.3 Results of classification modeling for identifying FDM process states .................................34
3.4.4 Discussion of machine learning classification in FDM .........................................................40
Chapter 4. Image-based Online Monitoring of Surface Finish in Turning .......................................43
4.1 Limitations of the State-of-the-art ..............................................................................................43
4.2 Discussion of Surface Quality Characterization Metric .............................................................44
4.2.1 Methodology and background of surface quality metric .......................................................45
4.2.2 Case studies and simulations verifying surface quality metric ..............................................47
4.2.3 Fiedler number applicability to turning process ....................................................................47
4.3 Methodology of Image-based Approach for Surface Finish Monitoring ...................................48
4.3.1 Workpiece used for study ......................................................................................................48
4.3.2 Image acquisition setup ..........................................................................................................50
4.3.3 Processing and analysis of images .........................................................................................51
4.3.4 Results of image-based prediction of surface roughness .......................................................54
Chapter 5. Summary and Future Work ................................................................................................55
5.1 Research Contributions to Process Monitoring in Additive Manufacturing ..............................55
5.1.1 Future work in AM process monitoring .................................................................................56
5.1.2 Acknowledgements ................................................................................................................57
5.2 Research Contributions to Online Monitoring of the Turning Process ......................................57
5.2.1 Future work in image-based monitoring of turning ...............................................................58
5.2.2 Acknowledgements ................................................................................................................58
References 59
Appendix 66
vii
List of Figures Figure 3-1. Summary of the overall approach regarding process monitoring in additive manufacturing ...17
Figure 3-2. (a) Schematic of the FDM process; (b) Schematic of the FDM setup instrumented with multiple
sensors used for measuring process conditions in real-time [83] .......................................................18
Figure 3-3. Test part resembling the NAS 979 standard part used for experimental trials (note: dimensions
are in inches) [83] ................................................................................................................................19
Figure 3-4. Diagram showing the mean surface roughness (Ra, µm) for each TC, averaged over three
replications [83] ..................................................................................................................................23
Figure 3-5. Schematic representation of FDM failure due to nozzle clogging [83] ....................................24
Figure 3-6. Heterogeneous sensor array for real-time monitoring of FDM process [83] ............................27
Figure 3-7. System diagram denoting the data acquisition methodology ....................................................28
Figure 3-8. (Top) The three observed process states in FDM, demarcated as normal, abnormal, and failure;
(Bottom) The transition between normal, abnormal, and failure states, is evident from the time series
pattern of the non-contact IR temperature sensor [83] ........................................................................29
Figure 3-9. Summary of the machine learning classification approach .......................................................32
Figure 3-10. Model misclassification rates ..................................................................................................39
Figure 4-1. Summary of the overall approach regarding image-based surface finish prediction in turning
.............................................................................................................................................................48
Figure 4-2. Representative images of the 3 inch diameter workpiece, with the corresponding roughness
measurements (Ra) ...............................................................................................................................49
Figure 4-3. Experimental setup at Virginia Tech’s lab ................................................................................50
Figure 4-4. Experimental setup at CCAM’s lab ..........................................................................................50
Figure 4-5. Glare and toolmarks evident on the surface of the workpiece ..................................................51
Figure 4-6. Analysis procedure for surface roughness estimation ...............................................................52
viii
Figure 4-7. Fitted regression models using the 1 inch and 3 inch diameter workpieces; Error bars denote the
95% confidence interval bounds of the mean predicted Ra value, based upon the 30 acquired images
of each workpiece segment .................................................................................................................54
Figure B-1. Representative examples of raw data, from a normal condition and a failed condition ..........67
Figure C-1. Out-of-bag classification error vs. number of trees chosen, in Random Forest model type ....68
ix
List of Tables Table 2-1. Surface defects, with their respective causes and strategies for improvement .............................8
Table 2-2. Internal defects, with their respective causes and strategies for improvement ............................9
Table 3-1. Full factor design of experiments plan for FDM tests (blocked on replicates by filament color)
.............................................................................................................................................................22
Table 3-2. Sensor information and location .................................................................................................26
Table 3-3 List of possible features to be selected, along with the corresponding sensor ............................34
Table 3-4. Classification model results, using TC7 sensor data ..................................................................35
Table 3-5. Classification model results, using TC8 sensor data ..................................................................36
Table 3-6. Classification model results, using both TC7 and TC8 sensor data ...........................................38
Table 4-1. Workpiece conditions .................................................................................................................49
Table 4-2. Regression model test error ........................................................................................................54
Table A-1. Design of experiments plan, and corresponding experimental results ......................................66
Table D-1. Specifications of Point Grey Grasshopper 3 camera .................................................................69
Table E-1. Testing error of the regression model for the 1 inch diameter workpiece .................................70
Table E-2. Testing error of the regression model for the 3 inch diameter workpiece .................................70
1
Chapter 1. Introduction
In the modern manufacturing industry, as processes and technology evolve to become more
complex and more automated, the task of process monitoring continues to be applied for quality
and reliability improvement. There is an increasing demand for real-time fault detection and
quality monitoring, remaining an ongoing area of research [1-4]. Offline quality control testing
may be destructive and expensive, highlighting the need for prediction of product quality in a
timely manner [5]. A real-time approach to fault detection allows for maintenance to be performed,
and to detect abnormalities and avoid process failure, which may damage the product or system
components [2-4, 6, 7]. Such monitoring and control poses a challenge, requiring knowledge of
the process, and development of systems to aid in this critical pursuit of quality improvement [8].
Therefore, this research investigates sensor-based solutions to these questions of online, real-
time process monitoring. This work examines application of this sensor-based monitoring
approach in two areas of manufacturing; one is the relatively novel technology of additive
manufacturing, and the other is the well-established, conventional machining process of turning.
The rest of this thesis is organized as follows: the remainder of Chapter 1 presents the
motivation and objective underlying the investigation of these two manufacturing technologies.
Following this is a review of the related literature, in Chapter 2. The research methodology and
experimental studies for the additive manufacturing process are detailed in Chapter 3, and for the
turning process in Chapter 4. Finally, Chapter 5 summarizes the results and contribution of this
work, and describes the potential future research for application of real-time, online process
monitoring.
2
1.1 Motivation and Objective of the Proposed Research Related to Additive Manufacturing
At this time, additive manufacturing processes, such as fused deposition modeling (FDM), are
relatively novel with respect to challenges concerning process reliability and consistency [9-12].
Quality control in FDM is largely limited to offline techniques, leading to high scrap rates [13].
Therefore, there exists a need to develop effective process control and monitoring techniques in
FDM [14], leading to more than merely failure diagnosis, but eventually to real-time prediction of
process failure [15].
Chapter 3 of this research describes the successful implementation of several sensors on a
desktop model FDM machine, for the purpose of online process monitoring. This is shown in the
following tasks:
1) Equipping an FDM machine (MakerBot Replicator 2X) with multiple types of online
sensors, including vibration and temperature sensors.
2) Using the sensor data for construction of machine learning classification models, to
demonstrate the capability for detection of process drifts in FDM.
This setup allows real-time monitoring and detection of FDM process drifts from ideal
conditions, in order to pursue corrective action and reduction of the occurrence of product defects.
This research demonstrates that the constructed platform and data collection is suitable for
potential in-process monitoring applications in the future.
3
1.2 Motivation and Objective of the Proposed Research Related to the Turning Operation
Surface finish is critical for the functional integrity of engineering components, as it influences
aspects such as fatigue, thermal behavior, service life, etc. [16-20]. Quantification of the surface
finish is essential not only from a quality assurance perspective, but also for tracking the
performance of the manufacturing process [19, 21].
Chapter 4 of this research implements a non-contact, image-based surface quality estimation
method for machined surfaces. This is done for workpieces machined by turning on a lathe, as
turning is a matured manufacturing method in which improved productivity and quality control
are desirable. Specifically, this is done by:
1) Incorporating an optical imaging setup in a lathe, to acquire images of cylindrically
turned workpieces.
2) Conducting analysis of the processed images using an algebraic graph theoretic
approach, and demonstrating the ability of regression modeling to predict the arithmetic
mean surface roughness (Ra) of the workpiece, in near real-time.
This research shows the application of an existing graph theoretic quantifier [22] to non-contact,
image-based surface quality estimation of conventionally machined workpieces, which
traditionally are limited to contact-based methods. Such a non-contact approach shows potential
for future online process monitoring applications.
4
Chapter 2. Literature Review
This chapter describes the research background and reviewed literature relating to process
monitoring in the two manufacturing technologies of fused deposition modeling (a type of additive
manufacturing), and the conventional machining process of turning.
2.1 Review of the Literature Relating to Process Monitoring in Additive Manufacturing
Here are reviewed typical defects which occur in the fused deposition modeling (FDM) process,
as well as prior experimental studies in FDM, and sensor-based monitoring in FDM, in order to
identify shortcomings in the current state-of-the-art.
2.1.1 Typical defects in the FDM process
The purpose of this section of this literature review is to present a general overview of the most
common predictable defects which occur in fused deposition modeling (FDM), as well as methods
to minimize or avoid such defects. There are two main categories of defects which may occur in
parts produced using FDM, namely surface and internal defects.
Surface defects are visible on the surface of the fabricated part, and include the staircase effect,
chordal errors, and overfill and underfill errors. The so-called staircase effect occurs due to the
layer by layer build method of the FDM process [23-25]. Any sloped region of a part, once
converted into slices, results in the predictable staircase effect. This is seen from a profile view of
the part as distinct edges where each layer is visible in a pattern resembling stairs, instead of the
ideal geometry of a smooth, sloped surface. Although a smaller layer or slice thickness will reduce
this effect, the build time will necessarily increase due to the increased number of layers. Choice
of careful part orientation [26-28], as well as adaptive slicing which varies the layer thickness [29,
30], help to decrease the staircase effect without negatively affecting build time of the part as
5
severely. Chordal errors may occur from approximation of the CAD model which the final part is
based on [23-25]. The original CAD file is converted to an STL file, and approximated by a surface
of triangles, which result in the curved surfaces becoming visible as a series of line segments in
the fabricated part. A proposed solution is to use a positive offset for the dimensions of the part,
and then to use post-processing to smooth the constructed part [23, 24]. This may be deemed
unnecessary if the software used to slice the model maintains an acceptably high resolution [25].
Overfill and underfill errors result when there is lack of coordination between motion of the nozzle
and the extrusion motors, and are visible in the part as excess material or a void, respectively [23-
25, 31, 32]. These may also occur if the material is not extruded at a constant rate due to slippage
or inconsistent filament diameter. If filament continues to be extruded while having reached the
end of the toolpath segment, excess material is deposited (overfill). Conversely, if extrusion stops
before reaching the end of the toolpath, a void appears (underfill). This problem is made more
difficult due to residual pressure of the melted filament within the extrusion chamber, causing
extrusion even after the motor stops. Avoiding these errors requires careful determination of the
point at which extrusion should be stopped, by optimization of the process parameters [33]. When
fabricating parts that include overhangs, we should consider the removal of the support structures
which can leave burrs on the surface of the part [23, 24]. Most surface defects in FDM may be
minimized through the use of post-processing techniques. Examples include CNC machining
operations [34], hot cutter machining [35], an abrasive jet machining based method [36], and
chemical treatment [37, 38]. Although such post-processing has proven to be beneficial in
improving surface finish, we must decide whether the time, cost, and effort are worthwhile.
Internal defects are the second category of defects in FDM, which include the occurrence of
voids and weak bonding of the deposited material within the part. These internal defects may be
6
considered more detrimental than surface defects, as they affect the structural integrity of the part
and cannot be easily eliminated by post-processing [23]. Sub-perimeter voids are seen as
incomplete filling of the area adjacent to the perimeter of the part [23, 24, 32, 39]. Such voids often
occur due to the raster pattern, when the toolpath includes sharp turns as it reaches the perimeter
of the part. To prevent these voids, a negative offset can be given in the toolpath so that the internal
fill pattern overlaps slightly with the perimeter path, or an increased flow rate can be used for
material deposited at points of intersection between the internal fill path and perimeter path [23,
24]. Inter-road voids occur when there is no physical contact between adjacent roads, most likely
due to insufficient flow rate, and may be evident when there is inconsistent material flow, due to
slipping in the extrusion mechanism or variation in filament diameter [23, 24]. These voids may
also be due to shrinkage of the cooling filament during long lengths of the toolpath, before the
adjacent road is deposited. Both inter-road and sub-perimeter voids may be reduced by a small
overlap of adjacent roads [23, 39]. Qiu and Langrana [40] also show that by careful toolpath
generation to minimize the number of discontinuous fill patterns, voids can be minimized. Core
voids can be thought of as a particular case of the inter-road void, resulting specifically from a
contour fill pattern, when an incorrect road width choice leads to a void in the center of the layer
[23, 24]. Weak bonding between roads shares similar causes with inter-road voids. Although the
deposited roads may be physically touching, the bond is not strong, which negatively affects the
physical strength of the produced part [23, 24, 41-43]. The temperature within the build chamber
has a significant effect on bond strength between deposited material filaments, and so should be
taken into account when part strength is desirable [44]. Chen et al. [45] conducted a study using
nano-focus computed tomography to examine the integrity of parts produced using FDM,
identifying the presence of air bubbles occurring within deposited filament (internal bubbles) and
7
within the bonding region between adjacent roads (necking bubbles). These necking bubbles in
particular can compromise the strength of the finished part, and the authors caution that the
presence of moisture on the filament can cause these air bubbles. Warping of a part may occur due
to temperature changes as the filament is extruded and solidifies, and the internal stresses lead to
part distortions. The longer the individual deposited line of material within the toolpath, the greater
the internal stress [25, 28, 46]. Zhang and Chou [47] show through FEA modeling and
experimental validation that deposition speed (of the extruded filament), followed by layer
thickness, are the most significant factors affecting residual stresses in the part which lead to
warping. Wang et al. [48] find that the presence of support structures may restrict warping of the
part. In addition, fabrication of a part in an enclosed chamber with a relatively high temperature
may alleviate the thermal stresses existing in the part, similar to an annealing process. Guerrero-
de-Mier et al. [49] present an approach to combat warping by splitting the part into square or
hexagonal bricks, effectively reducing the accumulated internal stresses within the part, and
thereby minimizing warping.
Peng and Xiao [25] demonstrate that while careful choice of process parameters can improve
part accuracy and eliminate defects, they may have contradictory effects. For example, a higher
build chamber temperature may decrease part warping, but also may affect the surface finish. And
while voids may be decreased by using an overlap in the toolpath, overfill errors may increase.
Therefore, when considering how to avoid the described defects in FDM produced parts, we must
develop a deeper understanding of these defects and their respective causes. A summary of the
surface and internal defects are shown in Table 2-1 and Table 2-2, respectively.
8
Table 2-1. Surface defects, with their respective causes and strategies for improvement
Surface defect Caused by Solutions
Staircase effect [23-25]
Sloping part geometry Part orientation
Large layer thickness Adaptive slicing
Post-processing
Chordal errors [23-25] Software approximation of part model Positive dimensional offset with post-processing
Overfill [23-25, 31, 32]
Lack of coordination between extruder movement and filament extrusion Parameter optimization
Filament slippage Post-processing
Inconsistent filament diameter
Underfill [23-25, 31, 32]
Lack of coordination between extruder movement and filament extrusion Parameter optimization
Filament slippage
Inconsistent filament diameter
Support structure burrs [23, 24] Removal of support structures Post-processing
9
Table 2-2. Internal defects, with their respective causes and strategies for improvement
Internal defect Caused by Solutions
Sub-perimeter voids [23, 24, 32, 39]
Sharp turns in fill pattern toolpath Toolpath overlap
Increased material flow rate
Fill pattern optimization
Inter-road voids [23, 24]
Filament slippage Toolpath overlap
Inconsistent filament diameter Increased material flow rate
Filament cooling and shrinkage Fill pattern optimization
Core voids [23, 24] Incorrect road width in contour fill pattern Improved toolpath planning
Weak bonding [23, 24, 41-43]
Filament slippage Toolpath overlap
Inconsistent filament diameter Increased material flow rate
Filament cooling and shrinkage Increased build chamber temperature
Air bubbles [45] Presence of moisture Keep filament dry
Warping [25, 28, 46]
Slow deposition speed Increased build chamber temperature
Large layer thickness Support structures
Low build chamber temperature "Bricking"
Long toolpaths
10
2.1.2 Prior experimental studies in FDM
Mahesh et al. [50] compared different additive manufacturing (AM) techniques, namely,
stereolithography (SLA), laminated object manufacturing (LOM), selective laser sintering (SLS),
and fused deposition modeling (FDM) based on their ability to adhere to geometric dimensions
and tolerance for a standardized test part. They reported that FDM is typically less suitable for
crafting parts with fine features compared to the other additive manufacturing processes studied;
defects were prominently visible on the FDM-crafted workpiece.
Peng et al. [51] investigated the effect of some of the process variables in FDM described in
this work, namely, extrusion temperature, temperature of the enclosure (ambient chamber
temperature), scanning speed of the extruder mechanism (feed rate), and extrusion speed (flow
rate) (see Section 3.2.2 for a brief explanation of the variables). They discussed the effect of
extruder (nozzle) temperature, and concluded that a nozzle temperature below 220 °C is inadequate
for melting the ABS plastic filament and leads to clogging, poor layer adhesion, and delamination.
The authors also made observations regarding feed rate and flow rate settings, noting that part
defects are minimized when their ratio approaches 1. Significant deviation from the 1:1 feed rate
to flow rate ratio can lead to clogs and scraping of the part by the nozzle. This work is valuable for
providing some of the physical basis for the experimental observations in this current research,
especially the reasons for explaining part build failures resulting from suboptimal feed rate to flow
rate ratios (see Section 3.2.4).
Wang et al. [48] proposed and experimentally validated a mathematical model for warp
deformation in FDM parts. The authors acknowledged the significance of numerous process
variables, such as ambient chamber temperature, layer thickness, extrusion temperature, extruder
11
head speed, geometric structure of the part, and tool path. However, quantitative correlations
between these factors and part deformation were unexplored.
Anitha et al. [52] used a Taguchi (L18) approach to specifically correlate certain process
parameters with the surface roughness (Ra) of an unspecified test part. The chosen parameters were
road width, layer thickness, and extrusion speed, each at three levels, while other variables such
as temperature were maintained at a constant (unspecified) level. They found all three chosen
factors to be statistically significant contributors in determining surface roughness. However, one
of the major shortcomings in this work is the absence of temperature variables, which have been
known to dominate build quality in FDM [53].
Armillotta [54] and Agarwala et al. [23] found that FDM errors may result from frequent start-
stop motion of the extruder head, warping of the part [55], or simply a failure to extrude the build
material (nozzle clog). Other studies have explored methods of improving part quality by
examining error compensation of the extrusion path [40, 56], determining the optimal part
orientation in relation to the build direction [57, 58], and consideration for the balancing of build
speed with accuracy by selectively decreasing the size of outer build layers to increase the part’s
precision [59, 60].
From the above review, it can be seen that although some work has been done in studying the
quality control issues in the additive manufacturing process, there is a lack of experimental work
and analysis for capturing process defects and failures using online real-time sensors.
2.1.3 Sensor-based monitoring in FDM
From a sensor-based monitoring perspective, Bukkapatnam et al. [61] investigated vibration
that occurs during the printing of a test part, by comparing mechanistic models with experimentally
obtained sensor data. The sensors used were two accelerometers affixed to the extruder head, and
12
two more on the frame of the machine. They validated their model using a 3-level factorial design,
varying extrusion speed (flow rate) and extruder mechanism movement speed (feed rate), resulting
in a total of 15 trial conditions. Bukkapatnam et al. demonstrate the ability to distinguish process
abnormalities from sensor data.
Online defect detection in extrusion-based layered manufacturing (resembling FDM) of
advanced multiple ceramics was investigated by Fang et al. [62]. They used machine vision
techniques to detect defects in layered manufacturing based on optical imaging of each layer during
the build. The optical images were used to visually identify various defects. Additionally, the
authors were able to evaluate the geometrical integrity of the build by comparing the optical image
of a layer to its original CAD design. In a related research, Fang et al. [31] further augmented their
monitoring approach using online signature analysis. They first mathematically represented (in the
form of a function) the ideal build morphology for each layer from the CAD file, which is termed
the process signature. Thereafter, the process signature was compared with the physical layer-by-
layer build pattern after image analysis to detect process anomalies.
Cheng and Jafari [63] further examined the surface pattern using image intensity information.
Build defects were classified into two types, namely, randomly occurring defects, and anomalies
due to assignable causes, e.g., improper extruder tool path. The randomly occurring defects were
detected by 2D texture analysis, while the assignable defects were identified using the process
signature technique developed previously by Fang et al. [31].
A key shortcoming evident from this literature review is the lack of real-time process
monitoring in order to detect the status of the build. Prior studies have rarely used multiple sensors
to evaluate the FDM process or sought to conduct process analysis in real-time. This research seeks
13
to address this gap by means of experimental studies on an FDM machine, which has been
integrated with multiple sensors for real-time process monitoring.
2.2 Review of the Literature Relating to Surface Finish Monitoring in Turning
Here are reviewed studies of optical methods for characterization of the surface finish of
machined workpieces, with respect to a desire for online application, and a focus on surfaces
machined by turning. Existing studies of surface roughness prediction note the benefits of optical
methods versus common stylus-based measurement. The most notable of which is the non-contact
nature of optical measurement, as well as less variation among measurements, due to the
evaluation of an area of the surface instead of the line segment obtained by a stylus. Optical and
image-based roughness measurement is especially suitable for online application, opening the door
to successful in-process quality control.
2.2.1 Optical methods for surface roughness estimation
This review of existing literature is therefore focused on optical methods for surface roughness
estimation of machined workpieces, specifically those that utilized computer-vision as the core
method of acquisition. A common approach in optical techniques is to obtain information by light
scattering - reflecting a laser or focused light off a surface and analyzing the distribution pattern
[64-67]. These methods deal with very smooth surfaces (<1 µm) machined by polishing or lapping,
and thus are not suitable for the applications pursued by this work, as the common roughness range
for turning is generally between 1 and 10 µm (Ra).
Computer vision techniques focus on obtaining image features that can be directly or indirectly
correlated to the quality of the surface being examined, whose chosen indicator is usually the
arithmetic surface roughness (Ra). The information extracted from the image is most often
classified as a first order or second order feature, dealing with direct single-pixel features or
14
pairwise pixel comparisons, respectively. The most common method of surface quality estimation
is to obtain gray level information from the image, and then to use this value (or a derived feature)
as a direct correlation with surface roughness [68-76]. Several works utilize neural or polynomial
networks for the construction of a prediction algorithm [74, 77-80]. A few noted works use features
derived through the Fourier or Wavelet transform of the image, as well [80, 81].
2.2.2 Shortcomings in the existing literature
The most commonly examined surface roughness range is between 0.10 and 4.0 µm (Ra), with
another common range being from 2-16 µm (Ra). By far the most prominent and limiting drawback,
often in studies with successful prediction, is in examining only a very narrow, smooth roughness
range (<2 µm) [70, 73, 74, 76, 82]. Several studies dealt with very wide-ranging roughness (45-
200, 25-60 µm), but often had large gaps of ~3-10 µm between each tested roughness value. These
works with a wider roughness range show a possible lack of precision for distinguishing 1-2
micron changes [68, 80, 81]. This lack of precision is undesirable.
In addition, some of the most successful prediction techniques require prior knowledge of the
cutting parameters for the workpiece, which limits application in a process control setting [72, 77-
79]. Methods which rely exclusively on image-derived data are able to be used without prior
process information, making these solely image-based approaches more attractive.
Two of the most notable successful approaches are those developed by Tsai et al. [80] and
Chang et al. [81]. Tsai et al. used the Fourier transform of the image in conjunction with a neural
network, for shaped and milled steel specimens with roughness of 6.3-100 and 1.6-50 µm,
respectively. Chang et al. utilized features extracted from the Wavelet transform of the image for
use in regression modeling of the surface roughness, examining shaped, ground, and polished steel
surfaces. Not all roughness values were explicitly listed, however the shaped steel specimens
15
varied from 2.032-15.469 µm. The speed of Tsai et al.’s method is stated to be ~2 sec, while the
speed of Chang et al.’s method is unspecified. However, the authors of these approaches did not
examine workpieces machined by turning.
In conclusion of this review of the existing literature, it is noted that very few prior studies
examined surfaces machined using turning. The most common material examined in prior work is
ground or milled steel, with only few examples of turned surfaces, which is the application of this
work’s method. Also, an almost universal shortcoming in the existing literature’s methodology is
a lack of online application of these acquisition techniques. The most notable differences between
this work and existing prediction methods are a more precise prediction range, and application in
turning. This research also demonstrates a clear potential for real-time, in-process implementation.
16
Chapter 3. Sensor-based Process Monitoring in Additive Manufacturing
The objective of this chapter is to demonstrate online monitoring of process conditions in
additive manufacturing using multiple sensors. The overall approach and structure of this research
has the following main parts (detailed also in Figure 3-1):
1) Experimental studies: Section 3.2 describes the investigation of the effect of three process
variables: extruder temperature, feed rate to flow rate ratio, and layer thickness, on the
build quality, evaluated as the arithmetic average surface roughness (Ra) of the workpiece,
using a design of experiments plan. It is notable to identify those process settings that lead
to build failures, and the physical root cause for these failures.
2) Sensor-based monitoring: The following Section 3.3 explains the design of a custom
sensor apparatus to record various aspects of the FDM process, using thermocouples,
accelerometers, a non-contact infrared (IR) temperature sensor, and a real-time video
borescope. Sensor data is acquired under various process conditions to learn how sensor
signal patterns are associated with different process states and the occurrence of build
failures.
3) Machine learning classification: Then, in Section 3.4 is shown the application of several
machine learning techniques for classification of the process state, based on the acquired
sensor data. This concludes the research by examining the potential of using such a
constructed sensor system in an online process monitoring scenario.
The experimental work of this research has also been utilized in an already published work [83],
in which a novel method using a nonparametric Bayesian DP mixture model and evidence theoretic
17
approach was used for classification of the process state. This research instead investigates the use
of existing machine learning classification algorithms for process state monitoring.
Figure 3-1. Summary of the overall approach regarding process monitoring in additive manufacturing
3.1 Process description of FDM
FDM is an additive manufacturing process in which an object is manufactured by depositing
progressive layers of extruded molten material [10, 13, 53, 84] (see Figure 3-2(a)). In FDM, the
material used for printing is most often ABS or PLA plastic filament. A spool of this plastic
material is fed through a heated mechanism (the extruder), and deposited in a layer by layer fashion
onto a heated platform (the “bed” or “table”). Figure 3-2(a) shows a diagram of this process; and
Figure 3-2(b) shows the attached sensors used for monitoring, described in detail later (see Section
3.3). Thermoplastic polymers, such as ABS and PLA, given their relatively low glass transition
temperatures (approximately 105°C and 65°C, respectively), are the materials of choice in FDM.
Although thermoplastic materials are prevalent in FDM, the process has evolved to accommodate
novel materials and application areas, e.g. mortar, clay, and lunar regolith [85, 86].
18
Figure 3-2. (a) Schematic of the FDM process; (b) Schematic of the FDM setup instrumented with multiple sensors used for measuring process conditions in real-time [83]
3.2 Experimental Studies in FDM
This section presents the conducted experimental studies. The objective of these experiments
is to understand the functioning of the process, and to identify part build failures (Section 3.2.3)
and understand their root causes (Section 3.2.4). The sensor data acquired during such build
failures is then used for machine learning classification (Section 3.4), to explore the potential for
online process monitoring.
3.2.1 Design of standard test artifact
Prior to conducting experiments using the FDM machine, it is important to select a standard
artifact. Different types of standard test artifacts have been suggested for evaluating AM processes,
e.g. by Mahesh et al. [50] and Kruth et al. [87]. Recently, the test part developed by Moylan has
been proposed for universal standardization for benchmarking different AM processes and
machines [88]. However, it was found that these parts could be cumbersome to produce on a
desktop FDM machine, such as the MakerBot Replicator 2X used in this work. The MakerBot has
a maximum layer resolution of 0.1 mm. Consequently, intricate features are difficult to produce
on this experimental setup. Therefore, it is necessary to design a custom test artifact that functions
19
well with the available machine. The test workpiece designed in this research is shown in Figure
3-3. It resembles the NAS 979 standard test artifact, and also is popularly known as the circle-
diamond-square test, which is widely used for assessing accuracy of CNC machine tools [89]. At
0.3 mm layer height this part takes between 25 to 35 minutes, while at 0.2 mm the build time is
close to 50 to 55 min; the exact times are contingent on the chosen feed rate, which for this case is
in the range of 40 to 50 mm/sec.
Figure 3-3. Test part resembling the NAS 979 standard part used for experimental trials (note: dimensions are in inches) [83]
3.2.2 Selection of process input and output variables
The experimental tests were conducted to understand the functioning of the process, and to
observe the range of feasible operating conditions of the parameter settings. The process output
variable was chosen to be the overall arithmetic mean surface roughness (Ra) of the workpiece,
which is the average of surface roughness (Ra) measurements taken at 10 different predetermined
locations. These measurements were recorded using a profilometer (Precision Devices Analog
Surfometer). Based on prior literature, the process input variables were selected to be the ratio of
20
feed to flow rates, layer height, and extruder temperature [48, 51, 52, 90]. Here is explained the
physical reason for choosing the particular settings for each of the process input variables.
1) Feed rate (fr) to flow rate (fw) ratio (𝑣): Feed rate (fr) is the movement speed of
fabrication; it is the velocity at which the extruder travels. Flow rate (fw) refers to the rate
at which material is being pulled through the extruder nozzle (Figure 3-2(a)). As described
in the literature review (Section Chapter 2), Peng et al. [51] have found that veering farther
away from a 1:1 ratio for the feed rate to flow rate leads to extruder clogs and failed builds.
This ratio is denoted by the symbol 𝑣(=fr/ fw). Preliminary studies were conducted to find
the appropriate operating range for 𝑣. The exploratory experimental trials indicate that
choosing 𝑣< 0.75 and 𝑣 > 1.25 invariably leads to build failures irrespective of the setting
of other variables. At the lower end, i.e. 𝑣< 0.75, the deposited layer is thick enough to
make contact with the extruder head on subsequent passes. While at higher values (𝑣>
1.25) the deposited layers are of uneven thickness and the extruded material is inconsistent;
the build frequently features voids, and the material often peels off and sticks to the
extruder causing the nozzle to clog. Therefore, the ratio of feed rate to flow rate (𝑣) is
selected in the range of 0.8 ≤ 𝑣≤ 1.2 for further investigation.
2) Layer height (h0): Layer height (ho) refers to the nominal distance setting between each
deposited material layer; it is effectively the distance by which the build table retracts
(moves down) between layers of the build. The lower the layer height the better the
resolution of the build, and thus the finer the feature detail. However, there is a lower limit
to the layer height, which is primarily governed by machine parameters, such as the
resolution of the motor coupled to the table (bed), the quality of the lead screws driving the
table, the stiffness of the machine structure, and the skew of the table with respect to the
21
nozzle. It was found that ho less than 0.2 mm on this machine often led to the nozzle making
contact with the workpiece. If the layer height is set too high (ho > 0.4 mm in this case),
the thickness of the extruded material will be less than the movement of the table, and as a
result the material is more liable to be deposited unevenly on the previous layer. Therefore,
ho was maintained in the range of 0.20 mm ≤ ho ≤ 0.30 mm.
3) Extruder temperature (te): The extruder temperature (te) governs the viscosity of the
extruded material [51]. At higher te, occurring in the 250 °C – 260 °C range on the
MakerBot machine, the extruded material is far too liquefied, and therefore may not
solidify quickly and will not adhere well to the previous layer. Consequently, there is a
tendency for the material to adhere to the nozzle on subsequent passes over an
insufficiently solidified part of the surface. This may cause the nozzle to have extraneous
material at the tip which then scrapes fresh layers, in the process damaging the part quality
and sometimes leading to nozzle clogs. On the other hand, at te < 225 °C, the extruded
material is quite viscous, and as a result, the material can fail to extrude from the nozzle,
leading to a failed build. Consequently, the extruder (nozzle) temperature was maintained
in the range of 230 °C ≤ te ≤ 245 °C for the experimental tests.
3.2.3 Identification of process conditions leading to build failures
The feed to flow rate ratio (𝑣) was chosen to vary at 3 levels, the layer height (h0) at 2 levels,
and the temperature of the extruder (te) at 2 levels (see Table 3-1). The effect of these three factors
(process input variables) on build integrity was tested using a full factorial design with 12 (= 3 ×
2 × 2) treatment conditions (TCs) plus one center point (repeated three times) for a total of 13 TCs
and 15 runs (= 3 × 2 × 2 + 3 × center point), as shown in Table 3-1. In all trials, the flow rate was
set at 50 mm/s, while the feed rate varied between 40 mm/s, 50 mm/s, and 60 mm/s, resulting in 𝑣
22
of 0.8, 1.0, and 1.2, respectively. The layer heights (h0) tested were 0.2 mm, 0.25 mm (center
point), and 0.3 mm. The temperature of the extruder (te) varied with settings at 230 °C, 237.5 °C
(center point), and 245 °C. For these experiments the ambient temperature of the build chamber
was maintained at 32.5 ± 2.5 °C.
It should be noted that the experimental design does not have equal levels for all three factors.
This is because preliminary exploratory experiments had two levels for all three factors, giving a
total of 8 (= 2 × 2 × 2) treatment conditions. These are conditions TC1 to TC8 in Table 3-1.
However, the failure region of TC7 and TC8 was encountered, which led to an extension of the
feed to flow rate ratio (𝑣) to a third level in order to facilitate further exploration of the process
(seen as conditions TC9 to TC12). Three replications of these 13 TCs were conducted. Each
replication used a different filament color (𝒸), namely, red, white, and blue. Essentially, this is akin
to blocking the design on replicates based on the filament color [91].
Table 3-1. Full factor design of experiments plan for FDM tests (blocked on replicates by filament color)
Treatment Condition
(TC)
Feed/Flow Rate Ratio (mm/sec)
(𝑣)
Layer Height (mm) (h0)
Extruder Temperature
(°C) (te)
Remarks
TC1 50/50 0.3 245 TC2 50/50 0.3 230 TC3 50/50 0.2 245 TC4 50/50 0.2 230 TC5 40/50 0.3 245 TC6 40/50 0.3 230 TC7 40/50 0.2 245 Failed build in five
out of eight trials TC8 40/50 0.2 230 TC9 60/50 0.3 245
TC10 60/50 0.3 230 TC11 60/50 0.2 245 Gives lowest surface
roughness (Ra) TC12 60/50 0.2 230
TC13 50/50 0.25 237.5 Center Point with three repetitions
23
Results from the experimentation (see Figure 3-4) show a clear optimal region (TCs 11 and 12),
as well as a consistent failure region (TCs 7 and 8). Based on Figure 3-4 the least surface roughness
(Ra) is obtained at = 1.0 to 1.2, h0 0.20 mm, and te = 237.5 °C to 245 °C. The complete results
are tabulated in Table A-1. It is straightforward to build a regression model to predict Ra based on
this data using a technique such as response surface methodology [91]. Many examples of such
statistical modeling and optimization studies are available in the AM literature [48, 51, 52, 92].
However, the goal of this present research is not to build a statistical model in order to optimize
the FDM process; instead it is desired to predict the onset of process anomalies, such as TC7 and
TC8, using sensor data (see Section 3.3). In the failed build conditions TC7 and TC8, the nozzle
clogged during the process for five out of eight trials. The physical causes for these failures are
discussed in detail in the following Section 3.2.4.
Figure 3-4. Diagram showing the mean surface roughness (Ra, μm) for each TC, averaged over three replications [83]
24
3.2.4 Analysis of build failure in FDM
The failure region at TC7 and TC8 provides important information regarding the FDM process
characteristics. Understanding the physical effects of the process input variables will allow one to
seek the optimal region and avoid build failure. Here are discussed the physical reasons for the
observed build failures, which are primarily determined by feed to flow rate ratio ( ) and layer
height (h0) settings.
A lower value for the feed to flow rate ratio ( indicates extrusion of material (flow rate) at a
faster rate than movement of the extruder head (feed rate). This results in thicker layers and a
gradual buildup of extruded material around the extruder head, typically leading to clogs (Figure
3-5). This condition, if allowed to continue, will lead to filament breakage. Because of breakage,
a portion of the filament will remain lodged inside the extruder. Clearing out such clogs is an
involved process, requiring disassembly of the extruder mechanism. From a productivity
perspective, it is valuable for such failure to be prevented.
Figure 3-5. Schematic representation of FDM failure due to nozzle clogging [83]
25
In contrast, the observed optimal region (see Figure 3-4) has an increased 𝑣(= 1.2), and the
faster movement of the extruder (feed rate) allows more room for the material to be deposited on
the layer, resulting in relatively smoother layers and no clogging. As mentioned previously in
Section 3.2.3, the varying effect of 𝑣 on extrusion success is explained by Peng et al. [51].
Layer height (h0) is the second significant process parameter relating to build failure. A higher
layer height setting (= 0.3 mm) means that at each successive layer the table moves down by 0.3
mm, causing the filament to be extruded farther from the previous layer, therefore increasing the
average height of each road (leading to a rougher appearance). Conversely, the lower layer height
of 0.2 mm results in a lower Ra, as the material is deposited closer to the previous layer and the
between-layer road width is decreased. However, h0 = 0.2 mm typically leads to failed builds at a
feed to flow rate ratio (𝑣) of 0.8, as indicated in Figure 3-4. This is due to the close proximity of
the extruder nozzle to the layer, increasing the possibility of clogging.
26
3.3 Sensor-based Monitoring of FDM
3.3.1 Sensor platform for real-time monitoring of FDM process conditions
The MakerBot Replicator 2X FDM machine was instrumented with multiple sensors, which
give quantitative information about the process in real-time (see Figure 3-2(b)). The technical
details of the various sensors used in this research are provided in Table 3-2.
Table 3-2. Sensor information and location
Temperature and vibration information was acquired at several different locations on the machine:
i. Extruder Temperature: A thermocouple was placed directly on the extruder using a high
temperature adhesive (Figure 3-6 (a)).
ii. Meltpool Temperature: A non-contact infrared (IR) temperature sensor (Figure 3-6 (a))
was oriented to measure the temperature at the contact point where material is deposited
onto the workpiece surface, i.e., the so-called material meltpool.
iii. Borescope Video: Also visible in Figure 3-6 (a) is a borescope camera (15 fps), which is
used to record the entire printing process for each run. This camera gives up-close, live
video, which allows monitoring of the extrusion process visually from start to finish.
Sensor Placement Measurement Model Vendor
Thermocouple (x4) Table Table temperature 5TC-GG-K-20-36 Omega
Thermocouple Extruder head Extruder temperature 5TC-GG-K-20-36 Omega
Thermocouple Inside machine Ambient temperature 5TC-GG-K-20-36 Omega
Accelerometer (tri-axis) Table Table vibration ADXL335 Analog Devices
Accelerometer (tri-axis) Extruder arm Extruder arm vibration ADXL335 Analog Devices
IR temperature sensor Extruder head Temperature of deposited
material UIRT/C.4-440F Exergen
Borescope Extruder head Real-time video N006 Supereyes
27
iv. Extruder and Table Vibration: Two 3-axis miniature MEMS accelerometers were used
to measure vibration on the table and on the extruder head. The accelerometers on the
extruder head and table are shown in Figure 3-6 (b) and (c), respectively.
v. Table Temperature: Four K-type thermocouples were attached to the table with two on
each side of the table as shown in Figure 3-6 (c). A thermocouple was also mounted inside
the machine enclosure to obtain ambient temperature measurements.
Figure 3-6. Heterogeneous sensor array for real-time monitoring of FDM process [83]
(a) Non-contact IR temperature sensor for measuring meltpool temperature, borescope camera
for real-time video capture capability, and thermocouple to measure extruder temperature.
(b) MEMS accelerometer to measure extruder vibration.
(c) MEMS accelerometer to measure table vibration, and thermocouples (x4) to measure table
temperature.
To acquire data from the sensors (excluding the IR temperature sensor and borescope) an
Arduino Mega 2560 microcontroller is used. Data from the six thermocouples and two vibration
sensors (3 channels each) is collated in real-time and streamed using MATLAB for visualization.
These 12 different sensor channels are synchronized and recorded at a sampling rate of ~ 2.5 Hz.
The IR temperature sensor is pre-amplified separately, and has a sampling rate of 1 Hz. In all,
including the borescope video data, 14 channels of sensor data are acquired. Figure 3-7 provides
28
the system diagram; including the FDM machine, the sensors, data acquisition layout, and
connection to the computer for data collection and subsequent analysis. Representative examples
of collected sensor data are shown in Figure B-1.
Figure 3-7. System diagram denoting the data acquisition methodology
29
Figure 3-8. (Top) The three observed process states in FDM, demarcated as normal, abnormal, and failure; (Bottom) The transition between normal, abnormal, and failure states, is evident from the time series pattern
of the non-contact IR temperature sensor [83]
By observing video recordings of the process from the borescope camera (Figure 3-8), three
typical working process states are identified from the experimental studies, these are:
i. normal operation of the process (Figure 3-8 (a));
ii. abnormal operation due to insufficient extrusion (Figure 3-8 (b)); and
iii. failure to extrude due to nozzle clog (Figure 3-8(c)).
The documented failures, TC7 and TC8, both exhibited clogging of the extruder nozzle (see
Figure 3-4). However, the nozzle clog progresses gradually. The process remains in the normal
30
state for the initial third of the layers (~ 15 minutes) with stable extrusion. Thereafter the extrusion
tends to be intermittent; at this stage the extruded material is relatively less consistent (almost
stringy), and this is termed the abnormal process state. Finally, the extrusion process completely
stops or fails on account of a nozzle clog. It would be desirable to detect the process drift as it
progresses from normal to abnormal, so that corrective action can be taken and failure can be
avoided. It should be noted that further investigation of such corrective action, although valuable
for future study, is outside the scope of this research. Each of these process states manifests in a
pattern recognizable on the IR temperature sensor data (see Figure 3-8).
3.3.2 Discussion of acquired sensor data
Having understood the process states from the results of the previously described experiments,
focus is turned to those factors (i.e. sensor data channels) which are or are not suitable for use in
constructing a classification model.
First, the IR temperature is clearly observed as changing with respect to the change in process
state (seen in Figure 3-8). This coincides with understanding of the process, being that if the
process fails, material is no longer deposited onto the surface of the part being constructed. The IR
temperature is a measurement of the area directly underneath the extruder mechanism, therefore it
follows that IR temperature would decrease if no material is currently being extruded. No other
sensors have such a clearly visible change. For this reason, IR temperature was chosen as the initial
feature to be included in all classification models.
When considering the mechanics of the process, as well as the conducted experiments, certain
factors are not beneficial to use in attempts to build a classification model. These factors and
parameters are described here:
31
1) The feed/flow rate ratio and layer height parameters are equal for all failed experiments
(TC7 and TC8), therefore these parameters are not considered as factors for
classification.
2) Replication number essentially describes the different color of material used in the
process. Attempting to use this as a factor would not yield useful information. This is
because replication 1 contains two failed and two normal conditions, replication 2
contains one normal and one failed condition, and replication 3 contains two failed
conditions. Due to the low number of experiments within each replication, it is
reasonable to suspect that the existing data is not an accurate representation of the real
population of experiments from each replication. There are an uneven number of
experiments from each condition and each replication, due to time constraints when the
experiments were originally conducted.
3) Extruder temperature, as a parameter, was set at different levels for the TC7 and TC8
experiments. Therefore, we expect the sensor recording this data to show this consistent
difference. Also, similarly to considering the replication number, there are an unequal
number of TC7 and TC8 experiments (five of TC7, and three of TC8).
4) When examining the process in general, the following conclusions can be drawn:
o A failed experiment will begin at normal, then progress to abnormal, then to failure.
o Ambient temperature will tend to increase over time.
o Bed temperature will tend to increase over time.
The machine has an enclosed chamber, which naturally heats up as the process is running. This
means that failed conditions will appear to indicate that failure is more likely if ambient or bed
32
temperature increases, which is not necessarily a valid conclusion, as temperature increases over
time in successful experiments as well.
The candidate features that remain, in addition to the IR temperature sensor data, are the six
channels of sensor data from the accelerometers. The objective of the subsequent section (Section
3.4) is to investigate classification of the three process states defined above, using the data acquired
from different sensors, and then to examine application of this classifier for real-time process
monitoring.
3.4 Machine Learning Classification of FDM Process States
This section describes the application of machine learning classification methods to the process
and data introduced previously, in order to show the potential of this platform for use in future
development of online, real-time process monitoring in AM. The following Figure 3-9 shows a
diagram summarizing the overall classification approach.
Figure 3-9. Summary of the machine learning classification approach
33
3.4.1 Machine learning classification algorithm selection
The chosen techniques for model construction and classification are:
i. SVM (linear)
ii. SVM (using radial basis kernel function)
iii. LDA
iv. QDA
v. Logistic Regression
vi. Naïve Bayes
vii. Random Forest
These are chosen as representative, mature classification methods that currently exist in the
field of machine learning. Although different or novel approaches may be applicable, as seen in
the already published work utilizing these experimental results [83], as well as a separate
publication utilizing the concept of sparse estimation of sensor data [93], the intended purpose of
this research is to construct a platform which demonstrates potential to be used for future process
monitoring of defects and failure. Therefore, analysis and modeling was limited to the above listed
techniques.
3.4.2 Sensor data selection
Selection of sensor data to include as model features was conducted by sequential forward
selection, using 10-fold stratified cross validation, and observing the overall misclassification rate
of each model. Features were added while observing the decrease in overall misclassification rate,
stopping when improvement was less than 0.1%. In addition, the tuning parameter for the Random
Forest method (the number of trees) is seen to reach stability at 100 trees (see Figure C-1). The
candidate features (described in Section 3.3.2) are listed in the following table.
34
Table 3-3 List of possible features to be selected, along with the corresponding sensor
Sensor Feature
IR Thermocouple IR Temperature (initial included feature)
Extruder Accelerometer Extruder Vibration X axis, Extruder Vibration Y axis, Extruder Vibration Z axis
Bed Accelerometer Bed Vibration X axis, Bed Vibration Y axis, Bed Vibration Z axis
3.4.3 Results of classification modeling for identifying FDM process states
For each model type, classification was conducted using 10-fold cross validation. This was
done in three main divisions, (i) data from the TC7 experiments, (ii) data from the TC8
experiments, and (iii) all data including both TC7 and TC8 experiments. These results are shown
in Table 3-4, Table 3-5, and Table 3-6, respectively. Relative model success is seen through
examination of the misclassification rates and F-scores (also known as F1 scores).
The results shown in Figure 3-10 are in terms of each model type’s misclassification rate with
respect to the number of features included in the model. These results are divided into columns
based on the data used (TC7, TC8, or both TC7 and TC8); and rows based on specific
misclassification for separate labels (overall misclassification, misclassification rate for data
labeled as normal, then abnormal, then failure).
35
Table 3-4. Classification model results, using TC7 sensor data
36
Table 3-5. Classification model results, using TC8 sensor data
37
38
Table 3-6. Classification model results, using both TC7 and TC8 sensor data
39
Figure 3-10. Model misclassification rates
40
3.4.4 Discussion of machine learning classification in FDM
This section provides a discussion of the classification results, and notable conclusions that can
be drawn, as well as considerations for future applications of machine learning classification in
this particular process setting.
Modeling results
i. Using the TC7 sensor data, the SVM model (using radial basis kernel) provides the
highest average F-score of 0.6087, using three features (IR temperature, Z axis bed
vibration, and X axis bed vibration).
ii. Using the TC8 sensor data, the SVM model (using radial basis kernel) provides the
highest average F-score of 0.5746, using one feature (IR temperature).
iii. Using both the TC7 and TC8 sensor data, the Random Forest model provides the
highest average F-score of 0.5328, using three features (IR temperature, Z axis bed
vibration, and Y axis bed vibration).
It should be noted that these results (i.e. the F-scores) differ from those found for the same
model types in the published work [83]. This is due to the fact that the published work constructed
its models using individual experimental runs, whereas in this current research models were
constructed using multiple runs (e.g. multiple runs from TC7, as opposed to a single TC7 run).
Also, labeling of the sensor data (i.e. as normal, abnormal, and failure) was conducted
independently in this work, introducing some variation in which data was and was not labeled
accordingly, particularly for the difficult to identify abnormal process state.
Sensor data selection
In addition to the originally chosen IR temperature, subsequent additional features did not
generally appear to significantly improve the average F-scores. Although significant improvement
is not evident in every case, the marginally improved model classification F-scores, due to
41
inclusion of additional features, follows from a physical understanding of the process. As
previously discussed, during process failure, the deposited material begins to block the extruder
nozzle. As well as stopping the extrusion of new material, this can lead to the extruder nozzle
scraping the part itself, which can perhaps be seen in vibration of the bed, and of the extruder itself.
However, this effect appears to be slight, and in addition the accelerometers used have a relatively
low resolution and sensitivity. In the future, higher quality sensors may provide more useful data.
Data selection
A major question to ask is whether to combine or separate the TC7 and TC8 data, as they only
differ in a single parameter setting (extruder temperature). Ideally a single model could be
constructed which applies to all cases of parameter settings in the FDM process. However, based
upon the current results, it seems that models tend to perform better when trained and tested using
separated data sets, particularly in the failure case. In the case of this research, it is not wise to
combine the TC7 and TC8 data sets because there are not an equal number of experiments from
each. When combining all of the data, it could be the case that a model performs extremely well
on TC7 but quite poorly on TC8. If the quantity of data is weighted in favor of TC7, this would
skew the apparent accuracy of the model. However, for future work, testing of classification
methods will need to be conducted for data collected during any parameter conditions.
Data preprocessing
In this research, the raw sensor data was utilized, in order to avoid the loss of potentially useful
information. In future work, data preprocessing may be investigated, such as smoothing of the
data, or selecting particular features to be used, such as the mean and variance. However, care must
be taken to avoid any loss of information.
42
Abnormal process state
A key conclusion of this research is recognition of the abnormal process state. Recognition of
the abnormal state is potentially very important from a quality monitoring perspective, since if we
can detect this in real-time, it provides the opportunity for corrective action or maintenance. It is
observed that constructed models perform very poorly when attempting to classify the abnormal
state. There are multiple reasons for this. First, it should be noted that of all collected experimental
data only 5.6% is designated to be from the abnormal state. This means there is relatively very
little data that can be used for construction of a model. Also, classifying the process state as
abnormal is difficult to do from a human perspective, as this is done by observing the recorded
video and noting roughly at what time the extrusion of material seems to deviate from normal. It
is quite difficult to determine the exact time the process transitions between the states. In this
regard, identification of the abnormal state is a key area that should be investigated in any future
work.
43
Chapter 4. Image-based Online Monitoring of Surface Finish in Turning
The objective of this chapter is to demonstrate near real-time monitoring of the surface finish
of turned workpieces, using an image-based approach. The overall structure of this research has
the following main parts:
1) Discussion of Surface Quality Quantifier: Section 4.2 details the background of the
graph theoretic quantifier used to characterize the surface morphology of turned
workpieces. It provides a summary of the reference work published by Rao et al. [22],
which demonstrates its capability for use in characterization of the morphology of ultra-
precision engineering components, specifically copper wafers processed by chemical
mechanical planarization.
2) Methodology: The following Section 4.3 describes the overall methodology of this
research, discussion of the workpieces chosen for analysis, the physical setup of the image
acquisition system, processing of the acquired images, and the regression modeling for
prediction of the surface finish.
4.1 Limitations of the State-of-the-art
Online surface finish assessment is challenging, particularly more for conventional machining
than for precision surfaces. This is because high-resolution optical techniques, e.g. laser light
scattering and laser speckle techniques, are limited by the wavelength of visible light [16, 64-67,
94-97]. These approaches are not physically capable of measuring surfaces whose root mean
square roughness (i.e. Rq, Sq) is greater than the wavelength of light (~ 400 to 700 nanometers)
[94, 96-99]. The typical roughness range for conventional machining processes, e.g. turning,
44
milling, shaping, planing, drilling, etc., is generally between 0.5 and 10 micrometers [20, 95, 96,
100]. Therefore, most high-resolution optical methods are more suitable for measurement of
precision, sub-micron (<1 micrometer) and ultraprecision (<50 nanometer) surface finish, such as
those obtained from polishing, lapping, and diamond turning [94-97, 101].
In addition, it is infeasible to incorporate these techniques into a conventional machining setup,
with its relatively harsh operating conditions (vibration, chip formation, coolant splashes, etc.).
Consequently, non-contact assessment of surface morphology for conventional machining
remains an open research problem. To address this, researchers have sought to use indirect sensor-
based techniques, e.g. vibration, force, and acoustic emissions, for online assessment of surface
quality in various machining operations [94-97, 101, 102]. This current research uses captured
images, and therefore has the potential to overcome the common difficulties in online monitoring
of conventional machining operations.
4.2 Discussion of Surface Quality Characterization Metric
This section details the background and basis of the surface characterization approach used in
this research, by providing a summary of the relevant portions of the reference work published by
Rao et al. [22]. The authors utilize optical micrograph images for quantification of surface
morphology variations of copper semiconductor wafers. They discuss key shortcomings of current
analytical approaches used for surface morphology quantification, and note the need for an
approach that is non-contact, preferably optical, and has the ability to capture the underlying
morphology of the surface. Rao et al. therefore propose an algebraic graph-theoretic approach for
interpreting the morphology of a surface based on its topographical connectivity. (The robustness
of this approach may be noted in that this graph theoretic quantifier has also been used recently,
45
in quantifying the dimensional integrity of AM constructed parts, through the use of 3D point
cloud coordinate measurements [103].)
The background and focus of their work is on the surface morphology of ultraprecision
engineering components, specifically semiconductor wafers produced by chemical mechanical
planarization (CMP). Rao et al. demonstrate that traditional statistical quantifiers can be
inadequate for capturing surface morphology variations.
4.2.1 Methodology and background of surface quality metric
There are two phases in the approach used by Rao et al. The first regards representation of the
raw input measurement (e.g. a 2D image) as a network graph, and the second uses the Fiedler
number to quantify the connectivity of this network graph.
The first phase is representation of the image as a network graph. Each row (or column) of the
image is considered to be an individual node of the graph. Row-wise pixel comparisons, as
opposed to individual pixel comparisons, are chosen in order to reduce the computational load.
First, pairwise comparison metrics wij (of the nodes) are computed using a kernel function. Then,
a threshold function converts wij to binary form, i.e. either 1 or 0. This is defined as the similarity
matrix S. The threshold function used to construct the similarity matrix is a manually tuned
heuristic. Rao et al. note that using a threshold function which is set equal to the average of wij
yields superior results, based upon both numerical and experimental studies. Two parallel
approaches are proposed for pairwise comparison, to obtain the similarity matrix S.
The first method is termed “ɛ-neighborhood graph representation”, and involves a preliminary
step of converting the original image, using Canny edge detection, into a binary image. Once the
Canny filter is applied, pairwise comparison of the binary image is conducted using the Euclidean
kernel function, yielding a similarity matrix in which the nodes are connected relative to their
46
topological closeness. This results in the ɛ-neighborhood graph being suitable for capturing surface
defects, such as scratches, pits, voids, etc. However, surface texture information is lost due to the
initial filtering of the image.
The second approach is termed “edge-weighted graph representation”, which uses the raw
grayscale image directly. In this case, the pairwise comparison is computed using a Gaussian radial
basis function, meaning that nodes are connected directly based upon their similarity of texture
(i.e. their relative pixel contrast). A similarity matrix constructed using this method compares the
actual surface texture between pixels of the image.
Once the similarity matrix of the image is obtained, the second phase uses the graph-theoretic
topological invariant Fiedler number to quantify the connectivity of the network graph. First, the
diagonal degree matrix of the graph is computed. The degree of an individual node is simply the
count of the number of edges that are connected to that node (e.g. an isolated node has degree 0).
The similarity and degree matrices are used to form the combinatorial Laplacian matrix. Then the
combinatorial Laplacian and degree matrices are used to form the normalized Laplacian matrix.
The smallest non-zero eigenvalue of the normalized Laplacian is called the Fiedler number. For
further detail regarding the formulation of the Fiedler number itself, the reader may refer to the
work by Rao et al. [22], as well as the original work by Fiedler [104].
Intuitively, the Fiedler number measures how easy it is to isolate a subset of nodes in a graph,
i.e. the “weakest links”. A more homogeneous image results in a smaller Fiedler number,
indicating that the nodes are difficult to isolate. And conversely, a relatively larger Fiedler number
indicates a more heterogeneous image, with nodes that are easier to isolate. The obtained Fiedler
number can be used to quantify the morphological characteristics of a surface, by using an optical
image.
47
4.2.2 Case studies and simulations verifying surface quality metric
The authors use several simulations of different surface morphologies to show that the Fiedler
number is capable of reflecting the changes, where conventional statistics are unsuitable. They
examine: 1) surfaces with different morphologies but identical feature density, 2) surfaces with
identical morphology but varying feature density, and 3) surfaces with different autocorrelation
lengths (Sal), but are statistically identical with respect to the first four statistical moments of their
peak heights (Sa, Sq, Ssk, and Sku). The Fiedler number’s superiority was also verified using real
CMP wafer surface data.
4.2.3 Fiedler number applicability to turning process
Having understood the capability of the Fiedler number in its application to ultraprecision
surfaces such as CMP processed wafers, this current research shows the Fiedler number’s
capability in application to the conventional machining process of turning.
There are two techniques used for the first phase of the authors’ method, termed the ɛ-
neighborhood graph representation, and the edge-weighted graph representation. As previously
mentioned, the edge-weighted approach is noted to be suitable for characterizing surface texture;
and the ɛ-neighborhood graph representation is useful for capturing prominent surface defects such
as scratches, voids, etc. This makes it appropriate for use in characterization of the surface of
turned workpieces, as the toolmarks are quite evident, resulting from the machining process. These
toolmarks are analogous to surface defects, and this essentially allows the turned workpiece to be
evaluated based upon the density of these toolmarks. From this perspective, we may use the Fiedler
number technique as the basis for the methodology described in the following sections of this
work.
48
Rao et al. demonstrates that, in high-precision applications such as CMP, the Fiedler number
outperforms conventional surface roughness parameters. In the case study examining three types
of surfaces with different autocorrelation lengths (Sal), these conventional parameters are identical,
while the Fiedler number captures the changes in surface morphology.
For the turning process, one can imagine that using a simplistic metric (e.g. counting the number
of toolmarks) may be an approach that perhaps yields results, but there is no guarantee such an
approach would be feasible in more complex scenarios. However, the Fiedler number has been
shown to be flexible and sensitive to subtle changes in surface morphology. In other words,
demonstrating the capability of the Fiedler number in this case provides a foundation for future
application in more challenging situations.
4.3 Methodology of Image-based Approach for Surface Finish Monitoring
This section, describing the methodology of this work, is organized as follows: discussion of
the workpieces chosen for analysis, the setup of the image acquisition system, and the processing
and analysis of the acquired images. Steps of the methodology are detailed in Figure 4-1.
Figure 4-1. Summary of the overall approach regarding image-based surface finish prediction in turning
4.3.1 Workpiece used for study
The chosen workpieces are turned steel shafts (AISI 4340) with 1 inch and 3 inch diameters.
These two workpieces were provided by the Commonwealth Center for Advanced Manufacturing
49
(CCAM), and various roughness levels were achieved by varying the feed rate, as shown in Table
4-1. Representative images of the workpiece segments are shown in Figure 4-2.
Table 4-1. Workpiece conditions
Feed (in/rev) 1 inch diameter Roughness (Ra)
(µm)
3 inch diameter Roughness (Ra)
(µm)
0.0140650 10.2345 10.0000 0.0135800 9.2870 9.8920 0.0126100 8.2630 8.6565 0.0114460 6.5600 7.2665 0.0104760 5.2200 6.2115 0.0095060 4.1840 5.1340 0.0087300 3.5705 4.2875 0.0077600 2.6315 3.4600 0.0067900 2.1140 2.6935 0.0055775 1.3585 1.8335
Figure 4-2. Representative images of the 3 inch diameter workpiece, with the corresponding roughness measurements (Ra)
50
4.3.2 Image acquisition setup
The image acquisition system consists of a CCD camera (specifications shown in Appendix D)
and a light source (AI Diffuse Axial LED Illuminator) affixed to the slide on a small manual lathe,
for images captured in Virginia Tech’s lab (of the 1 inch diameter workpiece), shown in Figure
4-3. Images of the 3 inch diameter workpiece have been acquired at the CCAM research lab by
attaching the camera to the turret of an industrial CNC lathe, using the built-in lighting of the
machine (see Figure 4-4).
Figure 4-3. Experimental setup at Virginia Tech’s lab
Figure 4-4. Experimental setup at CCAM’s lab
51
Positioning of the camera with respect to the workpiece was carefully determined, with the
desire to achieve the highest possible image resolution. Due to the nature of the workpiece, distinct
glare was reflected off of the surface, rendering certain portions of the image unusable for analysis.
In addition, portions of the image demonstrated the apparent distortion of the toolmarks along the
surface, noted in Figure 4-5. We ideally desire a flat surface, meaning that cropping and rotation
of the image are needed to produce a final image with parallel toolmarks and even illumination.
Figure 4-5. Glare and toolmarks evident on the surface of the workpiece
4.3.3 Processing and analysis of images
Acquired images are processed and analyzed using MATLAB, with the overall procedure
shown in Figure 4-6. For each workpiece (i.e. any new camera position or lighting condition),
calibration is required to choose the crop area for an evenly illuminated portion of the image, the
rotation angle needed to achieve near-vertical toolmarks, and the proper parameters for the Canny
edge detection filter. Once calibration is finished, the final Fiedler number calculation can be
performed. For image acquisition, the workpiece is mounted on a lathe and rotated to simulate
capturing images in an actual turning process. For image acquisition on the lathe in Virginia Tech’s
52
lab, the camera was within 1.0-1.5 inches of the surface of the workpiece, while the position of the
camera on the CCAM lathe was 1.5-2.0 inches from the surface. The camera’s field of view is
aligned for each segment (i.e. each Ra condition), and 30 images are captured, manually moving
the camera to each subsequent segment. The Fiedler numbers calculated from these 30 images are
correlated with the surface roughness (Ra) and used for construction of a regression model to
evaluate the prediction capability (see Section 4.3.4).
Figure 4-6. Analysis procedure for surface roughness estimation
Determination of the crop area is done heuristically, avoiding areas of the workpiece surface
affected by glare from the light source. Crop size is limited partially by computation time, notable
because the intended goal of this research is to develop a method which can in the future be applied
as an online, real-time measurement system. Although a larger image size generally improves
53
roughness prediction, it also directly increases computation time. Maximum size of the cropped
image is also limited by the need to avoid areas obscured by glare.
Image rotation is conducted based upon the need to align workpiece toolmarks vertically, as the
nature of the Fiedler number calculation is a column-by-column pixel comparison. The rotation
angle was chosen by assigning the sum of column variance as a “cost” for each particular angle; a
minimum variance score indicates near-vertical toolmarks.
Canny edge detection was used to extract texture primitives from the acquired image, in order
to prepare the image for the final step in analysis. Parameter choice for the Canny filter appears to
have the largest effect on the final successful prediction of surface roughness. These parameters
are the sigma value, designating the size of the Gaussian blur, and the threshold values for edge
detection. Canny parameter values are chosen which suitably limit noise in the images, while
preserving the characteristic toolmarks on the workpiece surface. In this research, these parameters
have been chosen heuristically, adjusted until the output roughness prediction is considered
sufficient. This initial heuristic calibration, although time consuming, only needs to be conducted
once for a given camera placement or lighting condition.
The key element in this approach for surface roughness prediction is the use of the
characterization parameter termed the Fiedler number (see Section 4.2). This quantifier is
calculated from the Canny filtered image, giving an associated output for each particular image.
This output value is directly used for comparison, and correlation with the obtained surface
roughness (Ra) of the workpiece. The computation time for this metric is highly dependent on
image size. For a typical 500x300-pixel image, total computation time is less than 2 seconds,
indicating suitability for real-time application. The results of analysis are discussed in the
following section.
54
4.3.4 Results of image-based prediction of surface roughness
Figure 4-7 plots the output Fiedler number (mean of 30 images) versus the respective measured
surface roughness (Ra) of each corresponding workpiece segment. Regression analysis is
conducted, with these final models constructed using the average of the 10-fold cross validation
model parameters. Table 4-2 shows the prediction results of the regression models, demonstrating
an overall average prediction error of ~7.85%. Further detailed results can be found in Appendix
E. Equations (1) and (2) show the constructed regression models, which are graphed in Figure 4-7.
Figure 4-7. Fitted regression models using the 1 inch and 3 inch diameter workpieces; Error bars denote the 95% confidence interval bounds of the mean predicted Ra value, based upon the 30 acquired images of each workpiece segment
Table 4-2. Regression model test error
Error 1 inch 3 inch Relative error (mean) (%) 8.1807 7.5269 Absolute error (mean) (µm) 0.2853 0.3956
RMSE (µm) 0.3321 0.4976
𝑅+ = 1.3103 +9.1735
(1 + 1044.5678(6.9558:;<=>?=@))6.A7BC (1)
𝑅+ = 2.0384 +11.0191
(1 + 10B8.G7BG(6.566G:;<=>?=@))6.48A6 (2)
55
Chapter 5. Summary and Future Work
This chapter summarizes the contributions of this research in sensor-based, online process
monitoring of additive manufacturing and turning, as well as the direction of future work in this
area. Multiple types of sensors are integrated, and these two methodologies are validated through
real experimental data and analysis.
5.1 Research Contributions to Process Monitoring in Additive Manufacturing
In order to successfully monitor the FDM process, this research develops and implements an
array of 14 in situ sensors on the machine, which provide data quantifying some of the most
important factors in FDM. Specific contributions are as follows:
1) Implementation of sensor platform for online process monitoring in FDM: A
heterogeneous sensor platform, consisting of accelerometers, thermocouples, an IR
temperature sensor, and close proximity video, is integrated with the MakerBot Replicator
2X FDM machine. This acquires the following process information: (i) bed (or table)
temperature; (ii) extruder temperature; (iii) ambient temperature; (iv) meltpool
temperature; (v) extruder vibration; (vi) table vibration; and (vii) video. This
comprehensive monitoring of the FDM process has not been attempted before.
2) Correlation of process input variables to workpiece quality: Experiments are conducted
to study the effect of three process input variables, namely, feed/flow rate ratio (𝑣), layer
height (h0), and extruder temperature (te), with the part build quality in terms of the surface
roughness (Ra). Also described are the physical effects of these variables on build failure.
With the current experimental data set, it is found that 𝑣 = 1.2, h0 = 0.2, te = 230 °C gives
the best surface roughness (Ra ~ 7 µm).
56
3) Classification of process state using machine learning techniques: Representative,
mature machine learning classification methods are applied, using the FDM process data
collected using the constructed platform. This provides useful conclusions which help to
understand the considerations that need to be made regarding future research and
application in FDM process monitoring. In particular, it shows the difficulty, and
importance, of recognizing the abnormal process state; it shows the value of the sensor
data, in particular the importance of the IR temperature data; and it shows that existing
methods, such as machine learning classification techniques, can be used in conjunction
with this platform for online process monitoring of AM.
5.1.1 Future work in AM process monitoring
A key conclusion resulting from this work is recognition of the so-called abnormal process
state, which describes the critical transition period between the normal and failure states. Detection
of process failure is essential, and it would be beneficial to avoid such failure if possible. This
would require detection of the abnormal state as soon as possible. For this purpose, further data
collection and improved sensor resolution are desirable, and are suspected to provide superior
results in future work. With respect to classification and analysis, future improvement may be
sought by considering the time series nature of the sensor data, as well as by exploring additional
classification methods. The next step of future study may be in replication of this sensor platform
using a different FDM machine. Even with the same sensors, and same experimental studies, this
would give more and better understanding of the process, and of nozzle clogging in particular, as
machines differ mechanically, even though the process is the same.
57
5.1.2 Acknowledgements
This research uses work conducted as part of a project funded by the following grant from the
National Science Foundation (NSF): CMMI 1436592. We would like to thank the machinists and
technicians at the Harris Manufacturing Lab at Virginia Tech, and we also thank the ISE senior
design team, Wade Andersen, Nick Bambino, Jake Snyder, and Diego Valdez, who helped with
the experiments.
5.2 Research Contributions to Online Monitoring of the Turning Process
This research presents a method for near real-time, online analysis of the surface finish of
workpieces finished through turning operations, using an image-based approach in conjunction
with a graph-theoretic quantifier. Benefits and key features of this proposed method are as follows:
1) Online incorporation of an image-based, non-contact monitoring method: A high
resolution grayscale camera is attached to both small and industrial-scale lathes, and utilized
to capture images of turned steel shafts. These images are recorded while the machines are
running, simulating actual operating conditions of the pre-machined workpieces. This
online, image-based monitoring of the turning process has not been attempted before.
2) Capability for real-time implementation, featuring rapid computation (~1-2 sec): The
captured images are quantified using the Fiedler number metric. The entire process occurs
in less than 2 seconds, demonstrating the capability for use in real production and quality
monitoring scenarios.
3) Ability to predict a wide and precise Ra range (1-10 µm): Regression modeling is utilized
to demonstrate the ability of the Fiedler number to be used to predict surface roughness of
a workpiece, with an average relative error of less than 8%. This is accomplished solely
using captured images, requiring no prior process knowledge.
58
5.2.1 Future work in image-based monitoring of turning
We note that calibration of the image analysis procedure, for the acquisition setups at the
Virginia Tech lab and the CCAM lab, suggests that this approach is stable to use in different
environments. Initial calibration is time-consuming, and must be conducted for each different
camera position and lighting condition. In particular, choosing the Canny filter parameters is
heuristic, and the prediction results are very sensitive to changes of these Canny filter parameters.
In future study, it may be worth investigating application of the Fiedler number method which
does not use the Canny filter.
Future work in applying this proposed prediction method can easily be pursued concerning
implementation in a live turning process, needing to account for process-specific features such as
coolant and chips, which may obscure the field-of-view of the camera. We may also pursue
application of this method to other materials finished by turning, other types of workpieces, and
surfaces machined by other manufacturing processes. This has already been demonstrated, in part,
in the measurement of copper wafers produced by chemical-mechanical planarization [105].
5.2.2 Acknowledgements
This research comes from a study sponsored by the Commonwealth Center for Advanced
Manufacturing (CCAM), as part of the project, “In-Process Surface Finish Measuring”. This work
was carried out under the supervision of Dr. James Kong, associate professor, Grado Department
of Industrial & Systems Engineering (ISE), and director of the Sensing and Analytics for
Innovative Manufacturing (SAIM) Laboratory at Virginia Tech. Finally, we would also like to
thank the machinists and technicians at the Harris Manufacturing Lab: M/s. P. Ratcliff, R.
Waldron, W.R. Monk, K. Snidow, and J. Linkous, at Virginia Tech for allowing use of their
machines and equipment.
59
References
[1] Ge, Z., Song, Z., and Gao, F., 2013, "Review of recent research on data-based process monitoring," Industrial & Engineering Chemistry Research, 52(10), pp. 3543-3562.
[2] Kano, M., and Nakagawa, Y., 2008, "Data-based process monitoring, process control, and quality improvement: Recent developments and applications in steel industry," Computers & Chemical Engineering, 32(1–2), pp. 12-24.
[3] Qin, S. J., 2012, "Survey on data-driven industrial process monitoring and diagnosis," Annual Reviews in Control, 36(2), pp. 220-234.
[4] Yin, S., Ding, S. X., Xie, X., and Luo, H., 2014, "A review on basic data-driven approaches for industrial process monitoring," IEEE Transactions on Industrial Electronics, 61(11), pp. 6418-6428.
[5] Milo, M. W., Roan, M., and Harris, B., 2015, "A new statistical approach to automated quality control in manufacturing processes," Journal of Manufacturing Systems, 36, pp. 159-167.
[6] Gao, Z., Ding, S. X., and Cecati, C., 2015, "Real-time fault diagnosis and fault-tolerant control," IEEE Transactions on Industrial Electronics, 62(6), pp. 3752-3756.
[7] Xiao, H., Raj, S., Bonissone, P., Hai, Q., and Iyer, N., 2008, "Multivariate anomaly detection in real-world industrial systems," Proc. IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence), Hong Kong, pp. 2766-2771.
[8] Yin, S., Gao, H., and Kaynak, O., 2014, "Data-driven control and process monitoring for industrial applications - part i," IEEE Transactions on Industrial Electronics, 61(11), pp. 6356-6359.
[9] Cooper, K. P., 2002, "Layered manufacturing: Challenges and opportunities," Proc. Materials Research Society Symposia, Cambridge University Press, Boston, MA, pp. 23-34.
[10] Bourell, D., Beaman, J., Leu, M., and Rosen, D., 2009, "A brief history of additive manufacturing and the 2009 roadmap for additive manufacturing: Looking back and looking ahead," Proc. US-Turkey Workshop on Rapid Technologies, Istanbul, Turkey, pp. 5-11.
[11] Scott, J., Gupta, N., Weber, C. L., Newsome, S., Wohlers, T., and Caffrey, T., 2012, "Additive manufacturing: Status and opportunities," Science and Technology Policy Institute, pp. 1-29.
[12] 2013, "Roundtable forum on additive manufacturing: Opportunities and constraints," Royal Academy of Engineering, London, UK.
[13] Kruth, J.-P., Leu, M., and Nakagawa, T., 1998, "Progress in additive manufacturing and rapid prototyping," CIRP Annals-Manufacturing Technology, 47(2), pp. 525-540.
[14] 2013, "Measurement science roadmap for metal-based additive manufacturing," National Institute of Standards and Technology, Gaithersburg, MD.
[15] Cheng, C., Sa-Ngasoongsong, A., Beyca, O., Le, T., Yang, H., Kong, Z., and Bukkapatnam, S. T. S., 2015, "Time series forecasting for nonlinear and non-stationary processes: A review and comparative study," IIE Transactions, 47(10), pp. 1053-1071.
60
[16] De Chiffre, L., Kunzmann, H., Peggs, G., and Lucca, D., 2003, "Surfaces in precision engineering, microengineering and nanotechnology," CIRP Annals - Manufacturing Technology, 52(2), pp. 561-577.
[17] De Chiffre, L., Lonardo, P., Trumpold, H., Lucca, D., Goch, G., Brown, C., Raja, J., and Hansen, H. N., 2000, "Quantitative characterisation of surface texture," CIRP Annals -Manufacturing Technology, 49(2), pp. 635-652.
[18] Peters, J., Bryan, J. B., Estler, W. T., Evans, C., Kunzmann, H., Lucca, D. A., Sartori, S., Sato, H., Thwaite, E. G., Vanherck, P., Hocken, R. J., Peklenik, J., Pfeifer, T., Trumpold, H., and Vorburger, T. V., 2001, "Contribution of CIRP to the development of metrology and surface quality evaluation during the last fifty years," CIRP Annals - Manufacturing Technology, 50(2), pp. 471-488.
[19] Jiang, X., 2012, "Precision surface measurement," Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 370(1973), pp. 4089-4114.
[20] Sherrington, I., and Smith, E. H., 1986, "The significance of surface topography in engineering," Precision Engineering, 8(2), pp. 79-87.
[21] Jiang, X., Scott, P. J., Whitehouse, D. J., and Blunt, L., 2007, "Paradigm shifts in surface metrology. Part 2 : The current shift," Proceedings of the Royal Society A: Mathematical, Physical and Engineering Science, 463(2085), pp. 2071-2099.
[22] Rao, P. K., Beyca, O. F., Kong, Z., Bukkapatnam, S. T., Case, K. E., and Komanduri, R., 2015, "A graph-theoretic approach for quantification of surface morphology variation and its application to chemical mechanical planarization process," IIE Transactions, 47(10), pp. 1088-1111.
[23] Agarwala, M. K., Jamalabad, V. R., Langrana, N. A., Safari, A., Whalen, P. J., and Danforth, S. C., 1996, "Structural quality of parts processed by fused deposition," Rapid Prototyping Journal, 2(4), pp. 4-19.
[24] Van Weeren, R., Agarwala, M., Jamalabad, V., Bandyophadyay, A., Vaidyanathan, R., Langrana, N., Safari, A., Whalen, P., Danforth, S., and Ballard, C., 1995, "Quality of parts processed by fused deposition," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 314-321.
[25] AnHua, P., and XingMing, X., 2012, "Investigation on reasons inducing error and measures improving accuracy in fused deposition modeling," Advances in Information Sciences & Service Sciences, 4(5).
[26] Pandey, P., Reddy, N. V., and Dhande, S., 2007, "Part deposition orientation studies in layered manufacturing," Journal of Materials Processing Technology, 185(1), pp. 125-131.
[27] Thrimurthulu, K., Pandey, P. M., and Reddy, N. V., 2004, "Optimum part deposition orientation in fused deposition modeling," International Journal of Machine Tools and Manufacture, 44(6), pp. 585-594.
[28] Sahu, R. K., Mahapatra, S., and Sood, A. K., 2013, "A study on dimensional accuracy of fused deposition modeling (FDM) processed parts using fuzzy logic," Journal for Manufacturing Science & Production, 13(3), pp. 183-197.
[29] Mani, K., Kulkarni, P., and Dutta, D., 1999, "Region-based adaptive slicing," Computer-Aided Design, 31(5), pp. 317-333.
[30] Sabourin, E., Houser, S. A., and Helge Bøhn, J., 1997, "Accurate exterior, fast interior layered manufacturing," Rapid Prototyping Journal, 3(2), pp. 44-52.
61
[31] Fang, T., Jafari, M. A., Danforth, S. C., and Safari, A., 2003, "Signature analysis and defect detection in layered manufacturing of ceramic sensors and actuators," Machine Vision and Applications, 15(2), pp. 63-75.
[32] Jafari, M., Han, W., Mohammadi, F., Safari, A., Danforth, S., and Langrana, N., 2000, "A novel system for fused deposition of advanced multiple ceramics," Rapid Prototyping Journal, 6(3), pp. 161-175.
[33] Han, W., Jafari, M. A., Danforth, S. C., and Safari, A., 2002, "Tool path-based deposition planning in fused deposition processes," Journal of Manufacturing Science and Engineering, 124(2), pp. 462-472.
[34] Kulkarni, P., and Dutta, D., 2000, "On the integration of layered manufacturing and material removal processes," Journal of Manufacturing Science and Engineering, 122(1), pp. 100-108.
[35] Pandey, P. M., Reddy, N. V., and Dhande, S. G., 2003, "Improvement of surface finish by staircase machining in fused deposition modeling," Journal of Materials Processing Technology, 132(1), pp. 323-331.
[36] Leong, K. F., Chua, C. K., Chua, G. S., and Tan, C. H., 1998, "Abrasive jet deburring of jewellery models built by stereolithography apparatus (SLA)," Journal of Materials Processing Technology, 83(1), pp. 36-47.
[37] Galantucci, L., Lavecchia, F., and Percoco, G., 2009, "Experimental study aiming to enhance the surface finish of fused deposition modeled parts," CIRP Annals -Manufacturing Technology, 58(1), pp. 189-192.
[38] Galantucci, L., Lavecchia, F., and Percoco, G., 2010, "Quantitative analysis of a chemical treatment to reduce roughness of parts fabricated using fused deposition modeling," CIRP Annals - Manufacturing Technology, 59(1), pp. 247-250.
[39] Clancy, R., Jamalabad, V., Whalen, P., Bhargava, P., Dai, C., Rangarajan, S., Wu, S., Danforth, S., Langrana, N., and Safari, A., 1997, "Fused deposition of ceramics: Progress towards a robust and controlled process for commercialization," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 421-429.
[40] Qiu, D., and Langrana, N. A., 2002, "Void eliminating toolpath for extrusion-based multi-material layered manufacturing," Rapid Prototyping Journal, 8(1), pp. 38-45.
[41] Martínez, J., Diéguez, J., Ares, E., Pereira, A., Hernández, P., and Pérez, J., 2013, "Comparative between FEM models for FDM parts and their approach to a real mechanical behaviour," Procedia Engineering, 63, pp. 878-884.
[42] Sood, A. K., Ohdar, R. K., and Mahapatra, S. S., 2012, "Experimental investigation and empirical modelling of FDM process for compressive strength improvement," Journal of Advanced Research, 3(1), pp. 81-90.
[43] Yardimci, M. A., Guceri, S. I., Danforth, S., and Agarwala, M., 1996, "Part quality prediction tools for fused deposition processing," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 539-548.
[44] Sun, Q., Rizvi, G., Bellehumeur, C., and Gu, P., 2008, "Effect of processing conditions on the bonding quality of FDM polymer filaments," Rapid Prototyping Journal, 14(2), pp. 72-80.
[45] Chen, R. K., Lo, T. T., Chen, L., and Shih, A. J., 2015, "Nano-CT characterization of structural voids and air bubbles in fused deposition modeling for additive manufacturing," Proc. ASME 2015 International Manufacturing Science and Engineering Conference, Charlotte, NC, pp. V001T002A071-V001T002A071.
62
[46] Kantaros, A., and Karalekas, D., 2013, "Fiber bragg grating based investigation of residual strains in ABS parts fabricated by fused deposition modeling process," Materials & Design, 50, pp. 44-50.
[47] Zhang, Y., and Chou, K., 2008, "A parametric study of part distortions in fused deposition modelling using three-dimensional finite element analysis," Proceedings of the Institution of Mechanical Engineers, Part B: Journal of Engineering Manufacture, 222(8), pp. 959-968.
[48] Wang, T.-M., Xi, J.-T., and Jin, Y., 2007, "A model research for prototype warp deformation in the FDM process," International Journal of Advanced Manufacturing Technology, 33(11-12), pp. 1087-1096.
[49] Guerrero-de-Mier, A., Espinosa, M., and Domínguez, M., 2015, "Bricking: A new slicing method to reduce warping," Procedia Engineering, 132, pp. 126-131.
[50] Mahesh, M., Wong, Y., Fuh, J., and Loh, H., 2004, "Benchmarking for comparative evaluation of RP systems and processes," Rapid Prototyping Journal, 10(2), pp. 123-135.
[51] Peng, A. H., and Wang, Z. M., 2010, "Researches into influence of process parameters on FDM parts precision," Applied Mechanics and Materials, 34, pp. 338-343.
[52] Anitha, R., Arunachalam, S., and Radhakrishnan, P., 2001, "Critical parameters influencing the quality of prototypes in fused deposition modelling," Journal of Materials Processing Technology, 118(1–3), pp. 385-388.
[53] Gibson, I., Rosen, D. W., and Stucker, B., 2010, Additive manufacturing technologies: Rapid prototyping to direct digital manufacturing, Springer, New York, NY.
[54] Armillotta, A., 2006, "Assessment of surface quality on textured FDM prototypes," Rapid Prototyping Journal, 12(1), pp. 35-41.
[55] Pei, E., Campbell, R. I., and de Beer, D., 2011, "Entry-level RP machines: How well can they cope with geometric complexity?," Assembly Automation, 31(2), pp. 153-160.
[56] Tong, K., Joshi, S., and Lehtihet, E. A., 2008, "Error compensation for fused deposition modeling (FDM) machine by correcting slice files," Rapid Prototyping Journal, 14(1), pp. 4-14.
[57] Lin, F., Sun, W., and Yan, Y., 2001, "Optimization with minimum process error for layered manufacturing fabrication," Rapid Prototyping Journal, 7(2), pp. 73-82.
[58] Xu, F., Loh, H., and Wong, Y., 1999, "Considerations and selection of optimal orientation for different rapid prototyping systems," Rapid Prototyping Journal, 5(2), pp. 54-60.
[59] Sabourin, E., Houser, S. A., and Bøhn, J. H., 1997, "Accurate exterior, fast interior layered manufacturing," Rapid Prototyping Journal, 3(2), pp. 44-52.
[60] Ziemian, C., and Crawn III, P., 2001, "Computer aided decision support for fused deposition modeling," Rapid Prototyping Journal, 7(3), pp. 138-147.
[61] Bukkapatnam, S., and Clark, B., 2007, "Dynamic modeling and monitoring of contour crafting: An extrusion-based layered manufacturing process," Journal of manufacturing Science and Engineering, 129(1), pp. 135-142.
[62] Fang, T., Bakhadyrov, I., Jafari, M. A., and Alpan, G., 1998, "Online detection of defects in layered manufacturing," Proc. IEEE International Conference on Robotics and Automation, Leuven, Belgium, pp. 254-259.
[63] Cheng, Y., and Jafari, M. A., 2008, "Vision-based online process control in manufacturing applications," Automation Science and Engineering, IEEE Transactions on, 5(1), pp. 140-153.
63
[64] Brodmann, R., Gast, T., Thurn, G., Wirtz, A., and Rodenstock, O. W. G., 1984, "An optical instrument for measuring the surface roughness in production control," CIRP Annals - Manufacturing Technology, 33(1), pp. 403-406.
[65] Brodmann, R., Gerstorfer, O., and Thurn, G., 1985, "Optical roughness measuring instrument for fine-machined surfaces," OPTICE, 24(3), pp. 243408-243408.
[66] Mitsui, K., 1986, "In-process sensors for surface roughness and their applications," Precision Engineering, 8(4), pp. 212-220.
[67] Novák, A., and Colding, B., 1981, "Sensing of workpiece diameter, vibration and out-off-roundness by laser — way to automate quality control," CIRP Annals - Manufacturing Technology, 30(1), pp. 473-476.
[68] Kiran, M. B., Ramamoorthy, B., and Radhakrishnan, V., 1998, "Evaluation of surface roughness by vision system," International Journal of Machine Tools and Manufacture, 38(5–6), pp. 685-690.
[69] Al-Kindi, G. A., Baul, R. M., and Gill, K. F., 1992, "An application of machine vision in the automated inspection of engineering surfaces," International Journal of Production Research, 30(2), pp. 241-253.
[70] Gadelmawla, E. S., 2004, "A vision system for surface roughness characterization using the gray level co-occurrence matrix," NDT & E International, 37(7), pp. 577-588.
[71] Khalifa, O., Densibali, A., and Faris, W., 2006, "Image processing for chatter identification in machining processes," Int J Adv Manuf Technol, 31(5-6), pp. 443-449.
[72] Kumar, R., Kulashekar, P., Dhanasekar, B., and Ramamoorthy, B., 2005, "Application of digital image magnification for surface roughness evaluation using machine vision," International Journal of Machine Tools and Manufacture, 45(2), pp. 228-234.
[73] Luk, F., Huynh, V., and North, W., 1989, "Measurement of surface roughness by a machine vision system," Journal of Physics E: Scientific Instruments, 22(12), p. 977.
[74] Zhisheng, Z., Zixin, C., Jinfei, S., Ruhong, M., and Fang, J., 2009, "A neural network-based machine vision method for surface roughness measurement," Proc. International Conference on Mechatronics and Automation, Changchun, China, pp. 3293-3297.
[75] Dutta, S., Kanwat, A., Pal, S. K., and Sen, R., 2013, "Correlation study of tool flank wear with machined surface texture in end milling," Measurement, 46(10), pp. 4249-4260.
[76] Chen, Z., and Xu, F., 2014, "Application of gray level co-occurrence matrix method in characterization of cylindrical grinding surface roughness," Applied Mechanics and Materials, 433, p. 2113.
[77] Ho, S.-Y., Lee, K.-C., Chen, S.-S., and Ho, S.-J., 2002, "Accurate modeling and prediction of surface roughness by computer vision in turning operations using an adaptive neuro-fuzzy inference system," International Journal of Machine Tools and Manufacture, 42(13), pp. 1441-1446.
[78] Lee, B. Y., Juan, H., and Yu, S. F., 2002, "A study of computer vision for measuring surface roughness in the turning process," International Journal of Advanced Manufacturing Technology, 19(4), pp. 295-301.
[79] Lee, B. Y., Yu, S. F., and Juan, H., 2004, "The model of surface roughness inspection by vision system in turning," Mechatronics, 14(1), pp. 129-141.
[80] Tsai, D. M., Chen, J. J., and Chen, J. F., 1998, "A vision system for surface roughness assessment using neural networks," International Journal of Advanced Manufacturing Technology, 14(6), pp. 412-422.
64
[81] Chang, S. I., and Ravathur, J. S., 2005, "Computer vision based non-contact surface roughness assessment using wavelet transform and response surface methodology," Quality Engineering, 17(3), pp. 435-451.
[82] Sarma, P. M. M. S., Karunamoorthy, L., and Palanikumar, K., 2008, "Surface roughness parameters evaluation in machining GFRP composites by PCD tool using digital image processing," Journal of Reinforced Plastics and Composites, 28(13), pp. 1567-1585.
[83] Rao, P. K., Liu, J., Roberson, D., Kong, Z., and Williams, C., 2015, "Online real-time quality monitoring in additive manufacturing processes using heterogeneous sensors," Journal of Manufacturing Science and Engineering, 137(6), pp. 061007-061007.
[84] Levy, G. N., Schindel, R., and Kruth, J.-P., 2003, "Rapid manufacturing and rapid tooling with layer manufacturing (LM) technologies, state of the art and future perspectives," CIRP Annals-Manufacturing Technology, 52(2), pp. 589-609.
[85] Khoshnevis, B., Hwang, D., Yao, K.-T., and Yeh, Z., 2006, "Mega-scale fabrication by contour crafting," International Journal of Industrial and Systems Engineering, 1(3), pp. 301-320.
[86] Khoshnevis, B., Bodiford, M. P., Burks, K. H., Ethridge, E., Tucker, D., Kim, W., Toutanji, H., and Fiske, M. R., 2005, "Lunar contour crafting–a novel technique for ISRU-based habitat development," Proc. American Institute of Aeronautics and Astronautics Conference, Reno, NV.
[87] Kruth, J.-P., Vandenbroucke, B., Van Vaerenbergh, I. J., and Mercelis, P., 2005, "Benchmarking of different SLS/SLM processes as rapid manufacturing techniques," Proc. International Conference of Polymers and Moulds Innovations, Ghent, Belgium.
[88] Moylan, S., Slotwinski, J., Cooke, A., Jurrens, K., and Donmez, M. A., 2012, "Proposal for a standardized test artifact for additive manufacturing machines and processes," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 902-920.
[89] Cooke, A., and Soons, J., 2010, "Variability in the geometric accuracy of additively manufactured test parts," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 1-12.
[90] Vasudevarao, B., Natarajan, D. P., Henderson, M., and Razdan, A., 2000, "Sensitivity of RP surface finish to process parameter variation," Proc. Solid Freeform Fabrication Symposium, Austin, TX, pp. 251-258.
[91] Schmidt, S. R., and Launsby, R. G., 1989, Understanding industrial designed experiments, Air Academy Press, Coloraod Springs, CO.
[92] Weheba, G., and Sanchez-Marsa, A., 2006, "Using response surface methodology to optimize the stereolithography process," Rapid Prototyping Journal, 12(2), pp. 72-77.
[93] Bastani, K., Rao, P. K., and Kong, Z., 2016, "An online sparse estimation-based classification approach for real-time monitoring in advanced manufacturing processes from heterogeneous sensor data," IIE Transactions, 48(7), pp. 579-598.
[94] Bennett, J. M., 1992, "Recent developments in surface roughness characterization," Measurement Science and Technology, 3(12), p. 1119.
[95] Shiraishi, M., 1981, "In-process measurement of surface roughness in turning by laser beams," Journal of Engineering for Industry, 103(2), pp. 203-209.
[96] Shiraishi, M., 1989, "Scope of in-process measurement, monitoring and control techniques in machining processes—part 2: In-process techniques for workpieces," Precision Engineering, 11(1), pp. 27-37.
65
[97] Vorburger, T., and Teague, E., 1981, "Optical techniques for on-line measurement of surface topography," Precision Engineering, 3(2), pp. 61-83.
[98] Sherrington, I., and Smith, E. H., 1988, "Modern measurement techniques in surface metrology: Part ii; optical instruments," Wear, 125(3), pp. 289-308.
[99] Sherrington, I., and Smith, E. H., 1988, "Modern measurement techniques in surface metrology: Part i; stylus instruments, electron microscopy and non-optical comparators," Wear, 125(3), pp. 271-288.
[100] DeGarmo, E., Black, J. T., and Kosher, R., 2003, Material and processing in manufacturing, John Wiley and Sons, New York, NY.
[101] Rao, P., Bukkapatnam, S., Beyca, O., Kong, Z., and Komanduri, R., 2014, "Real-time identification of incipient surface morphology variations in ultra-precision machining process," Transactions of the ASME, Journal of Manufacturing Science and Engineering, 136(2), p. 021008.
[102] Damodarasamy, S., and Raman, S., 1991, "Texture analysis using computer vision," Computers in Industry, 16(1), pp. 25-34.
[103] Rao, P. K., Kong, Z., Duty, C. E., Smith, R. J., Kunc, V., and Love, L. J., 2015, "Assessment of dimensional integrity and spatial defect localization in additive manufacturing using spectral graph theory," Journal of Manufacturing Science and Engineering, 138(5), pp. 051007-051007.
[104] Fiedler, M., 1973, "Algebraic connectivity of graphs," Czechoslovak mathematical journal, 23(2), pp. 298-305.
[105] Rao, P. K., 2013, "Sensor-based monitoring and inspection of surface morphology in ultraprecision manufacturing processes," Ph.D. thesis, Oklahoma State University, Stillwater, OK.
66
Appendix
Appendix A. Experimental studies in AM
Shown below are the results from the FDM experimental tests. The arithmetic average surface
roughness (Ra) is measured at ten different predetermined locations on the workpiece surface. The
average Ra over these ten locations is reported along with the standard deviation (shown in
parenthesis) in the table below.
Table A-1. Design of experiments plan, and corresponding experimental results
Treatment Condition
(TC)
Process Input Variables Process Output Variable (Surface Roughness, Ra)
Feed/Flow Rate Ratio (mm/sec)
(𝑣)
Layer Height (mm) (h0)
Extruder Temperature
(°C) (te)
Replication 1 Red filament
(mean Ra, µm) (std. deviation, µm)
Replication 2 White filament (mean Ra, µm)
(std. deviation, µm)
Replication 3 Blue filament (mean Ra, µm)
(std. deviation, µm)
TC1 50/50 0.3 245 15.56 (5.30) 13.69 (4.00) 12.82 (3.68)
TC2 50/50 0.3 230 16.60 (5.33) 15.08 (3.12) 14.75 (4.04)
TC3 50/50 0.2 245 8.61 (3.07) 7.63 (2.85) 6.64 (1.52)
TC4 50/50 0.2 230 6.72 (2.27) 7.63 (2.04) 10.77 (4.37)
TC5 40/50 0.3 245 24.88 (5.23) 21.38 (8.66) 25.14 (6.54)
TC6 40/50 0.3 230 27.99 (5.19) 29.61 (6.23) 26.43 (4.36)
TC7 40/50 0.2 245 21.80 (5.83) 19.92 (7.72) Failed to build
TC8 40/50 0.2 230 20.91 (6.77) Failed to build Failed to build
TC9 60/50 0.3 245 15.77 (3.78) 15.10 (3.26) 12.01 (3.69)
TC10 60/50 0.3 230 13.70 (3.15) 14.23 (3.55) 12.59 (3.54)
TC11 60/50 0.2 245 6.72 (1.90) 7.06 (1.52) 6.50 (2.59)
TC12 60/50 0.2 230 6.99 (0.85) 7.02 (0.79) 5.49 (1.28)
TC13 (Center Point)
50/50 0.25 237.5 11.96 (4.35) 12.71 (5.13) 9.84 (4.04)
50/50 0.25 237.5 12.02 (4.62) 11.87 (3.21) 7.91 (2.72)
50/50 0.25 237.5 12.43 (4.08) 13.31 (5.91) 9.63 (4.82)
67
Appendix B. Sensor-based monitoring in AM
Figure B-1. Representative examples of raw data, from a normal condition and a failed condition
68
Appendix C. Machine learning classification in AM
Figure C-1. Out-of-bag classification error vs. number of trees chosen, in Random Forest model type
69
Appendix D. Image acquisition in turning
Table D-1. Specifications of Point Grey Grasshopper 3 camera
Specification
Model Number GS3-U3-28S4M-C Imaging Device Progressive Scan Camera Sensor Format 1/1.8" Type of Sensor Sony ICX687 CCD Pixels (H x V) 1928 x 1448 Pixel Size, H x V (µm) 3.69 Pixel Depth 8, 12, 16 and 24-bit digital data Frame Rate (fps) 26 Type of Shutter Global shutter Electronic Shutter 0.03ms - 30s Image Format Mono8, Mono12, Mono16 Transfer Speed (Gbit/s) 5 Image Buffer 128MB frame buffer Memory (MB) 2MB non-volatile Flash memory Synchronization External or Via Software Video Output USB 3.0 interface with screw locks Mount C-Mount Dimensions (mm) 44 x 29 x 58 Operating Temperature (°C) 0 to +50 Storage Temperature (°C) -30 to +60 Weight (g) 90
70
Appendix E. Surface quality estimation in turning
Table E-1. Testing error of the regression model for the 1 inch diameter workpiece
Workpiece condition: 1 2 3 4 5 6 7 8 9 10
Actual Ra (µm) 1.3585 2.1140 2.6315 3.5705 4.1840 5.2200 6.5600 8.2630 9.2870 10.2345
Predicted Ra (µm) 1.6697 1.8517 2.4371 4.0805 4.0586 4.8144 7.0844 7.8456 9.2827 10.3329
Predicted Ra 95% CI (±) (µm) 0.0721 0.2310 0.2987 0.6639 0.6069 0.6229 0.6006 0.6074 0.4889 0.0446
Absolute Error (µm) 0.3112 0.2623 0.1944 0.5100 0.1254 0.4056 0.5244 0.4174 0.0043 0.0984
Relative Error (%) 22.9076 12.4078 7.3874 14.2837 2.9971 7.7701 7.9939 5.0514 0.0463 0.9615
Table E-2. Testing error of the regression model for the 3 inch diameter workpiece
Workpiece condition: 1 2 3 4 5 6 7 8 9 10
Actual Ra (µm) 1.8335 2.6935 3.4600 4.2875 5.1340 6.2115 7.2665 8.6565 9.8920 10.0000
Predicted Ra (µm) 2.0928 2.4549 3.1436 4.5552 5.0650 7.0239 6.7509 9.0341 9.8434 11.0508
Predicted Ra 95% CI (±) (µm) 0.2840 0.2131 0.3032 0.9566 1.0092 0.8334 0.6525 0.7350 0.4231 0.3204
Absolute Error (µm) 0.2593 0.2386 0.3164 0.2677 0.0690 0.8124 0.5156 0.3776 0.0486 1.0508
Relative Error (%) 14.1424 8.8584 9.1445 6.2437 1.3440 13.0790 7.0956 4.3620 0.4913 10.5080