Conference PaperPDF Available

Fast measurement of metal laser powder bed fusion layer surfaces using light scattering and principal component analysis

Authors:

Abstract and Figures

To address the future challenges in quality monitoring of metal laser powder bed fusion, a novel method is proposed to detect topographic anomalies on layer surfaces, which may appear during the manufacturing process. The method combines light scattering and principal component analysis. Scattering patterns, experimentally generated from real surfaces representative of in-control processes and encoded as digital images, are collected and used to build a reference set, which is then further populated by simulation. Principal component analysis is then applied to the set. A certain number of principal components is extracted and used to define a transform to map any scattering pattern to principal component space. Using the created transform, any new scattering pattern can be transformed to principal component space and then back into the original space (reconstruction), with some reconstruction error. The error is expected to be low if a pattern from the reference set is processed. However, if a different pattern is processed, e.g. generated by an out-of-control layer topography, then the reconstruction error is larger. In this work, a layer monitoring system is proposed, capable of detecting out-of-control topographies through observation of the reconstruction error. The system was implemented and experimentally validated through application to a selected test case. Measurement, laser powder bed fusion, light scattering, principal component analysis
Content may be subject to copyright.
Joint Special Interest Group meeting between euspen and ASPE
Advancing Precision in Additive Manufacturing
Inspire AG, St. Gallen, Switzerland 2021
www.euspen.eu
Fast measurement of metal laser powder bed fusion layer surfaces using light
scattering and principal component analysis
Mingyu Liu1, Nicola Senin1,2, Richard Leach1
1Manufacturing Metrology Team, Faculty of Engineering, University of Nottingham, Nottingham, UK
2Department of Engineering, University of Perugia, Italy
mingyu.liu1@nottingham.ac.uk
Abstract
To address the future challenges in quality monitoring of metal laser powder bed fusion, a novel method is proposed to detect
topographic anomalies on layer surfaces, which may appear during the manufacturing process. The method combines light scattering
and principal component analysis. Scattering patterns, experimentally generated from real surfaces representative of in-control
processes and encoded as digital images, are collected and used to build a reference set, which is then further populated by
simulation. Principal component analysis is then applied to the set. A certain number of principal components is extracted and used
to define a transform to map any scattering pattern to principal component space. Using the created transform, any new scattering
pattern can be transformed to principal component space and then back into the original space (reconstruction), with some
reconstruction error. The error is expected to be low if a pattern from the reference set is processed. However, if a different pattern
is processed, e.g. generated by an out-of-control layer topography, then the reconstruction error is larger. In this work, a layer
monitoring system is proposed, capable of detecting out-of-control topographies through observation of the reconstruction error.
The system was implemented and experimentally validated through application to a selected test case.
Measurement, laser powder bed fusion, light scattering, principal component analysis
1. Introduction
With the rapid development of metal additive manufacturing
(AM) techniques [1], topography measurement of layer surfaces
has become increasingly important for in-process monitoring of
the quality of fabricated parts [2]. Any problem discovered in the
topographies of the layer surfaces may be indicative of problems
in the manufacturing process and affect the final product. In
metal AM processes, such as laser powder bed fusion (LPBF),
non-intrusive methods to monitor layer topography are
required, where the speed of measurement is essential to avoid
slowing down the manufacturing process and possibly altering
the physics of the process itself [3].
In previous work [4, 5], we have developed a method to
measure grating surfaces combining light scattering and
machine learning, which is suitable for fast and in-process
surface measurement. We then further developed the method
to monitor the quality of LPBF surfaces using an autoencoder [6].
In this paper, we present a fast method to measure
topographical changes of LPBF layer surfaces, which combines
light scattering and principal component analysis (PCA). In the
proposed method, laser light is projected onto the layer surface
and scattered light is captured by a camera. The scattering
pattern is then processed by a PCA-based monitoring system
which detects anomalous changes in the scattering pattern as an
indication of possibly detrimental changes in layer topography.
Experiments performed using a prototype implementation
based on the off-line measurement of test LPBF surfaces, show
that the proposed monitoring solution can be used to
discriminate between in-control and out-of-control LPBF
topographies. Data processing in the prototype implementation
is fast enough to warrant future in-process application without
the need for slowing down or temporarily halting the fabrication
process. Therefore, the proposed method has the potential to
be integrated into a commercial LPBF machine for real-time, in-
process quality monitoring.
2. Methodology
The schema of the proposed method is shown in Figure 1.
Scattering patterns from reference surfaces (manufactured
under in-control states using optimal parameters) are collected
experimentally as digital images, and used to populate a
reference dataset, which is then augmented by simulation
(algorithmic shifts and rotations applied to the measured
images). PCA is then applied to the reference dataset. The
principal components represent the inherent multidimensional
features of the scattering patterns from the reference surfaces.
A PCA-based encoding and decoding system is then created
using a certain number of principal components (the first 50%
components). Using the encoding/decoding system, scattering
images can be encoded into principal component space and
then back into images, with a small reconstruction error. As the
PCA-based encoding and decoding system has been tuned
specifically for in-control surfaces (i.e. the reference dataset),
any out-of-control scattering pattern processed through the
Figure 1. Schema of the proposed method
same system will result in a larger reconstruction error. Thus,
reconstruction error itself can be used to detect out-of-control
patterns.
Figure 2 shows the experiment setup to evaluate the proposed
method. Collimated laser light with a wavelength of 633 nm and
an approximate beam diameter of 0.8 mm is projected to a
mirror and reflected onto the measured LPBF sample. The
sample is mounted on a rotation stage, to simulate different
surface orientations. Scattering light is reflected to a 150 mm ×
150 mm screen. The scattering pattern can then be captured by
a camera and further processed by a PC.
Figure 2. Experiment setup
3. Results and discussions
Two LPBF samples were used for the experiment; one was a
reference surface produced by an in-control LPBF process, whilst
another was representative of an out-of-control process. The
topographies of the two surfaces are shown in Figure 3 and their
manufacturing parameters are shown in Table 1. The reference
surface was manufactured using optimal parameters, resulting
in evenly distributed textures, as shown in Figure 3(a). The
defective surface was manufactured using significantly lower
energy density, which resulted in large humps on the surface
(due to insufficient melting energy), as shown in Figure 3(b).
Both samples were used to perform the scattering experiment
in the setup shown in Figure 2. For each sample, thirty-six
scattering patterns were measured by rotating the surface every
10°. The scattering patterns were then further populated in
simulation by algorithmically shifting the digital images by six
steps in both the x and y directions and by rotating ten steps with
per step. As a result, there were 36 × 6 × 6 × 10 = 12960
datasets for each sample. A circular mask was applied for each
dataset to make the effective area rotationally symmetric,
eliminating the corner effect due to rotating the square-shaped
dataset. The original pixel densities of the measured images
were 6000 × 4000. The images were cropped according to the
size of the screen and were eventually resized to 20 × 20 pixels.
As a result, the data size was significantly reduced. The intensity
values in the pixels of the reference set were then processed by
mapping to the standard normal distribution (zero mean, unit
variance) [7] and used to perform the PCA. In total, there were
20 × 20 = 400 principal components. In this study, we used the
first half of the principal components, i.e., 200 principal
components, to reduce the dimension of the datasets and
establish the PCA-transform.
Figure 3. Topographies of LPBF samples, (a) reference surface, and (b)
defective surface, measured by Zygo NexView NX2 with 20× objective
lens
Table 1 Manufacturing parameters for the LPBF samples
Surface Laser
power/W
Scan
speed/m s-1
Energy
density/J mm-2
Reference 170 1.1 2.1
Defective 120 1.1 1.5
Figure 4 shows the results for one dataset from the reference
surface. Figure 4(a), Figure 4(b) and Figure 4(c) are the input
scattering pattern, reconstructed scattering pattern and the
reconstruction error, respectively. The results show that the
reconstructed scattering pattern is visually similar to the original
one. The reconstruction error is determined by the deviations
from the reconstructed scattering pattern to the input scattering
pattern. The root mean square (RMS) value of the
reconstruction error is 0.089, which is relatively small, indicating
that the PCA-transform can efficiently transform and
reconstruct the scattering pattern measured from the reference
surface.
Figure 4. Results for the reference surface, (a) original scattering pattern,
(b) reconstructed scattering pattern, and (c) reconstruction error. All
subfigures are 20 × 20 pixels
The results for one dataset from the defective surface are
shown in Figure 5. Comparing to the results for the reference
surface, the RMS value of the reconstruction error is significantly
larger, which is 0.214. The large reconstruction error is due to
the low efficiency of the encoding/decoding process for the
defective surface, whose datasets were not used to establish the
PCA-transform. The results of the reconstruction errors for the
reference surface and the defective surface are also summarised
in Table 2.
Figure 5. Results for the defective surface, (a) original scattering pattern,
(b) reconstructed scattering pattern, and (c) reconstruction error. All
subfigures are 20 × 20 pixels
Table 2 Reconstruction errors for the reference surface and defective
surface
Surface RMS of reconstruction error/A.U.
Reference 0.089
Defective 0.214
The RMS values for the reconstruction error for all datasets
from both reference and defective surfaces are summarised in
Figure 6. The mean value for those from the reference surface is
0.055 whilst it is 0.155 for the defective surface. These two types
of surfaces can be easily discriminated by thresholding the
reconstruction error. In this study, we set the threshold to be
0.1, i.e., if the RMS error is less than 0.1, the measured surface
is classified as a non-defective surface, otherwise a defective
surface. As a result, 12861 and 99 datasets from the reference
surface were classified as non-defective and defective,
respectively. On the other hand, 12938 and 22 datasets from the
defective surface were classified as defective and non-defective
surface, respectively. The confusion matrix can then be
summarised as shown in Table 3. The overall accuracy of the
classifier is 0.995, indicating that the proposed method has good
performance.
Figure 6. Histogram of results for all the datasets from the reference
surface and defective surface
Table 3 Confusion matrix
Predicted:
Non-defective
Predicted:
Defective
Actual:
Non-defective
12861 99
Actual:
Defective
22 12938
4. Conclusions
The paper presents a fast method to measure the LPBF layer
surfaces combining light scattering and PCA. A PCA-based
encoding/decoding system is developed and tuned on a
reference dataset made of scattering patterns experimentally
acquired from reference surfaces and further augmented by
simulation. The PCA-based encoding/decoding system can then
be used to convert scattering pattern images into the principal
component space, and then back into images. The
reconstruction error can be used to classify whether the
measured surface has significantly different topography from
the reference surfaces, possibly produced from an out-of-
control process. The accuracy of the classifier was
experimentally determined to be as high as 0.995, which
indicates the good performance of the proposed method,
although more and more diverse datasets are needed to obtain
a more comprehensive assessment of performance. The
computational burden of the PCA-based encoding/decoding
process is relatively low, which makes the system able to achieve
fast response times, an essential prerequisite for in-process
utilisation. The relatively simple and low-cost design makes the
proposed method potentially suitable for implementation in
modern LPBF machines.
Acknowledgements
We acknowledge the support of the Engineering and Physical
Sciences Research Council [EP/R028826/1], European Union’s
Horizon 2020 Research and Innovation Staff Exchange
Programme [MNR4SCell, 734174].
References
[1] Leach R K and Carmignato S 2020 Precision Metal Additive
Manufacturing (CRC Press)
[2] Leach R K, Bourell D, Carmignato S, Donmez A, Senin N and Dewulf
W 2019 Geometrical metrology for metal additive manufacturing
CIRP Annals 68 677-700
[3] Grasso M, Remani A, Dickins A, Colosimo B M and Leach R K 2021
In-situ measurement and monitoring methods for metal powder
bed fusion – an updated review Meas. Sci. Technol. in press
[4] Liu M, Cheung C F, Senin N, Wang S, Su R and Leach R K 2020 On-
machine surface defect detection using light scattering and deep
learning Journal of the Optical Society of America A 37
[5] Liu M, Senin N, Su R and Leach R K 2020 Cascaded machine
learning model for reconstruction of surface topography from light
scattering Proc. SPIE 113520Q
[6] Liu M, Senin N and Leach R K 2021 Intelligent quality monitoring
for additive manufactured surfaces by machine learning and light
scattering Proc. SPIE 1178206
[7] Abdi H and Williams L J 2010 Principal component analysis Wiley
Interdisciplinary Reviews: Computational Statistics 2 433-459
... We have developed a new method for surface characterisation combining light scattering and machine learning (ML). The method has been successfully applied to the binary classification (acceptable vs. defective) of micro-structured [3,4] and additive manufacturing (AM) surfaces [5][6][7][8], particularly laser-based powder bed fusion (PBF-LB) surfaces. The general idea is to impinge laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. ...
... The general idea is to impinge laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. We have developed the following ML models to characterise PBF-LB surfaces: convolutional autoencoder (CNN) [5], autoencoder based on a multilayer, fully connected network (ANN) [6], classifier based on principal component analysis (PCA) [7], and classifier based on a one-class support vector machine (SVM) [8]. All these methods share a common approach where the ML system is internally used to learn an optimal strategy to encode/decode only the scattering patterns of acceptable surfaces. ...
... We use the PCA-based method [7] as an example to study the influence of scattering image resolution. The schema of the method is shown in Figure 1 and is related to implementing a binary classifier (i.e., acceptable vs. defective surface). ...
... We have recently developed a new method for surface characterisation combining light scattering and machine learning (ML). The method has been successfully applied to the binary classification (acceptable vs. defective) of micro-structured [1-4] and additive manufacturing (AM) surfaces [5][6][7][8]. The general idea is to shine laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. ...
... The general idea is to shine laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. We have developed the following machine learning models: convolutional autoencoder (CNN) [5], autoencoder based on a multilayer, fully connected network (ANN) [6], classifier based on principal component analysis (PCA) [7], and classifier based on a oneclass support vector machine (SVM) [8]. All these methods share a common approach where the ML system is internally used to learn an optimal strategy to encode/decode only the scattering patterns of acceptable surfaces. ...
Conference Paper
Full-text available
We have recently developed a new method for surface characterisation combining light scattering and machine learning (ML). The method has been successfully applied to the binary classification (acceptable vs. defective) of micro-structured [1-4] and additive manufacturing (AM) surfaces [5-8]. The general idea is to shine laser light onto a surface and process the scattered reflection pattern by a ML binary classifier. We have developed the following machine learning models: convolutional autoencoder (CNN) [5], autoencoder based on a multilayer, fully connected network (ANN) [6], classifier based on principal component analysis (PCA) [7], and classifier based on a one-class support vector machine (SVM) [8]. All these methods share a common approach where the ML system is internally used to learn an optimal strategy to encode/decode only the scattering patterns of acceptable surfaces. Defected surfaces are then detected during operation, when poor encoding/decoding performance is observed on a new pattern. The single-class approach is advantageous for AM because the ML model can be trained without the need of examples from defective surfaces. In this paper, the performance of the ANN, CNN, PCA and SVM-based classifiers is compared by processing multiple scattering patterns experimentally acquired from two laser powder bed fusion surfaces (one acceptable, one defected). The performance indicators are the accuracy of the binary classifier; the time to classify each newly acquired pattern; and the internal separation of acceptable vs defective observations, measured by the difference in the RMS reconstruction errors of the encoding/decoding process associated with the acceptable (reference) and unacceptable surfaces. For the SVM, internal separation is measured as the difference of likelihood for an observation to belong to either one of the two classes.
... During the PBF-LB fabrication process, structures on the surfaces are formed, commonly known as process signatures. These signatures can often be used to evaluate the fabrication quality [2]. ...
... We have shown that using machine learning can solve the complex inverse scattering problem efficiently (compared to the traditional library search method [26]) and our solution based on combining light scattering and machine learning is relatively fast (compared to fringe projection methods [15,16]), thus suitable for application to in-process monitoring, and introduces minimal concerns in terms of process disturbance or accessibility issues. In addition, we have shown that the training for machine learning can be accomplished using numerical simulation to generate artificial scattering patterns, which is more efficient than collecting a large number of real ones [27,28]. However, the previously proposed solution could only rely on 2D scattering simulation for training (i.e. ...
Article
Full-text available
Quality monitoring for laser powder bed fusion (L-PBF), particularly in-process and real-time monitoring, is of importance for part quality assurance and manufacturing cost reduction. Measurement of layer surface topography is critical for quality monitoring, as any anomaly on layer surfaces can result in defects in the final part. In this paper, we propose a surface measurement method, based on the use of scattered light patterns and a convolutional autoencoder-based unsupervised machine learning method, designed and trained using a large set of scattering patterns simulated from reference surfaces using a scattering model. The advantage of using an autoencoder is that the monitoring model can be trained using solely data from acceptable surfaces, without the need to ensure the presence of representative observations for all the types of possible surface defects. The advantage of using simulated data for training is that we can obtain an effective monitoring solution without the need for a large collection of experimental observations. Here we report the results of a preliminary investigation on the performance of the proposed solution, where the trained autoencoder is tested on experimental data obtained off-process, using a dedicated experimental apparatus for generating and collecting light scattering patterns from manufactured L-PBF surfaces. Our results indicate that the proposed monitoring solution is capable of detecting both acceptable and anomalous surfaces. Although further validation is required to fully assess performance within an on-machine and in-process setup, our preliminary results are encouraging and provide a glimpse of the potential benefits of using our surface measurement solution for L-PBF in-process monitoring.
... melt-pool, scan track, powder bed) [1]. To improve measurement speed, to meet the requirements of real-time performance, the authors have developed methods based on light scattering combined with different machine learning methods, including conventional, perceptron-like fully connected artificial neural networks (ANNs) [2], principal component analysis (PCA) [3] and convolutional neural networks (CNNs) [4]. In this paper, we propose a method combining light scattering with a one-class support vector machine (SVM). ...
Conference Paper
Full-text available
Measurement of laser-based powder bed fusion (PBF-LB) surfaces provides a promising solution for closed-loop quality control of the final parts. This paper presents a light scattering method to measure PBF-LB surfaces combined with one-class support vector machines (SVMs) for anomaly (defect) detection during the manufacturing process. With the one-class SVM method, datasets from solely reference (acceptable) surfaces are used to fit a classification model. Experimental results show that the method is fast and has higher accuracy than our previous work, which is promising for integration into next-generation PBF-LB machines for process quality monitoring.
Article
Full-text available
The possibility of using a variety of sensor signals acquired during metal powder bed fusion processes, to support part and process qualification and for the early detection of anomalies and defects, has been continuously attracting an increasing interest. The number of research studies in this field has been characterised by significant growth in the last few years, with several advances and new solutions compared with first seminal works. Moreover, industrial powder bed fusion systems are increasingly equipped with sensors and toolkits for data collection, visualisation and, in some cases, embedded in-process analysis. Many new methods have been proposed and defect detection capabilities have been demonstrated. Nevertheless, several challenges and open issues still need to be tackled to bridge the gap between methods proposed in the literature and actual industrial implementation. This paper presents an updated review of the literature on in-situ sensing, measurement and monitoring for metal powder bed fusion processes, with a classification of methods and a comparison of enabled performances. The study summarises the types and sizes of defects that are practically detectable while the part is being produced and the research areas where additional technological advances are currently needed.
Article
Full-text available
This paper presents an on-machine surface defect detection system using light scattering and deep learning. A supervised deep learning model is used to mine the information related to defects from light scattering patterns. A convolutional neural network is trained on a large dataset of scattering patterns that are predicted by a rigorous forward scattering model. The model is valid for any surface topography with homogeneous materials and has been verified by comparing with experimental data. Once the neural network is trained, it allows for fast, accurate and robust defect detection. The system capability is validated on micro-structured surfaces produced by ultra-precision diamond machining.
Article
The needs, requirements, and ongoing and future research issues in geometrical metrology for metal additive manufacturing are addressed. The infrastructure under development for specification standards in AM is presented, and the research on geometrical dimensioning and tolerancing for AM is reviewed. Post-process metrology is covered, including the measurement of surface form, texture and internal features. In-process requirements and developments in AM are discussed along with the materials metrology that is pertinent to geometrical measurement. Issues of traceability, including benchmarking artefacts, are presented. The information in the review sections is summarized in a synthesis of current requirements and future research topics.
  • H Abdi
  • L J Williams
Abdi H and Williams L J 2010 Principal component analysis Wiley Interdisciplinary Reviews: Computational Statistics 2 433-459