Machine Learning Algorithms For Defect Detection in Metal Laser-Based Additive Manufacturing: A Review

Version of Record: https://www.sciencedirect.
com/science/article/pii/S1526612522000093
Manuscript_6b0e53f4d601f56ae3a69ea3aab37270
1 Machine learning algorithms for defect detection in metal

2 laser-based additive manufacturing: A review
3 Yanzhou Fua,∗ , Austin Downeya,b , Lang Yuana , Tianyu Zhanga , Avery Prattc and
4 Yunusa Balogunc
5 a Department of Mechanical Engineering, University of South Carolina, Columbia, SC 29208, United States
6 b Department of Civil and Environmental Engineering, University of South Carolina, Columbia, SC 29208, United States
7 c Naval Surface Warfare Center (Crane Division) Crane, IN 47522, United States
109 ARTICLE INFO ABSTRACT

11
12 Keywords: Laser-based additive manufacturing (LBAM), a series of additive manufacturing technologies,
13 Laser-based additive manufacturing has unrivaled advantages due to its design freedom to manufacture complex parts with a
14 Machine learning wide range of applications. Although advancements in LBAM processes and materials have
15 Defect detection led to increased manufacturing capabilities, the printing process’s repeatability, durability, and
16 Product quality reliability still face significant challenges. Therefore, a defect detection system for the LBAM
17 Artificial intelligence processes is essential, as it promises to guarantee product quality and increase the efficiency of
18 the printing process. As a practical and widely applied technology, machine learning methods
19 have been providing novel insights into the manufacturing process, which has proven advantages
20 for defect detection in LBAM. This paper summarizes the machine learning algorithms for defect
21 detection in the metal LBAM processes. To have a comprehensive and systematic summary,
22 machine learning algorithm, material type, defect type, dataset type, and algorithm accuracy for
23 various LBAM technologies are described.
24
25
26 Nomenclature
27 AI artificial intelligence
28 AM additive manufacturing
29 ANN artificial neural network
30 BC Bayesian classifier
31 CNN convolutional neural network
32 DBN deep belief network
33 DCNN deep convolutional neural network
34 DED directed energy deposition
35 DL deep learning
36 DMLS direct metal laser sintering
37 DNN deep neural network
38 DT decision Tree
39 GMM Gaussian mixture model
40 GP genetic programming
∗ Corresponding author
yanzhouf@email.sc.edu (Y. Fu)
ORCID (s):
Y. Fu et al.: Preprint submitted to Elsevier Page 1 of 31

© 2022 published by Elsevier. This manuscript is made available under the Elsevier user license
https://www.elsevier.com/open-access/userlicense/1.0/
Machine learning algorithms for defect detection: A review
41 KMC K-means clustering

42 KNN K-Nearest neighbor
43 LBAM laser-based additive manufacturing
44 LDA linear discriminant analysis

45 LENS laser engineered net shaping
46 LMD laser metal deposition
47 LPBF laser powder bed fusion
48 LR linear regression
49 LS-SVM least square support vector machine
50 LSTM long short-term memory
51 ML machine learning
52 PBF powder bed fusion
53 PCA principal component analysis
54 QDA quadratic discriminant analysis
55 RF random forests
56 RL reinforcement learning
57 RMSE root-mean-square error
58 RNN recurrent neural network

59 SLM selective laser melting
60 SOMs self-organized maps
61 SVM support vector machine
62 1. Introduction
63 Additive manufacturing (AM) is a rapidly growing technology that has proven advantages for unlocking design
64 freedom for lightweight components with complex geometries and as a disruptive technology, it offers exciting
65 new manufacturing capabilities [1, 2]. AM has been actively utilized in various industries, and it shows enormous
66 economic potential [3, 4]. According to the Wohlers 2021 Report on AM, in 2020, the AM industry grew 7.5%,
67 or nearly $ 12.8 billion, despite the global COVID-19 pandemic [5]. Among various AM techniques, laser-based
68 additive manufacturing (LBAM) displays considerable potential for industrial adoption and has already changed the
69 manufacturing process by enabling complex design and innovative application development [6]. The principle of
70 LBAM technologies lie in using a laser beam to yield thermal energy for sintering/melting and consolidating additive
71 materials or emitting light quanta of a particular wavelength to induce a chemical curing response in vat polymerization.
72 The materials utilized in the LBAM processes can be in the form of powders (metals, ceramics, and polymers), solids
73 (paper, plastics, and metals), or liquids (resin). In LBAM, according to the difference in material feeding approach,
74 there are three major systems: powder blown system, powder bed system, and wire-feed system. In the powder blown
75 system, the material is carried out through single or multiple nozzles and melted by a laser beam. The powder bed
76 system generally has two chambers, a build chamber for printing and a powder chamber with a coating roller to spread
77 the powder material across the build chamber. In the wire-feed system, the wire is fed and melted to manufacture the
78 metal parts.

79 Due to the convenience of fabrication and the practicability of mechanical properties enhancement, LBAM is
80 regarded as a highly favored manufacturing technology for Industry 4.0 over conventional technologies [7]. However,
81 its full potential is held back by variabilities inherent to the process. The LBAM processes are dominated by
82 complicated physics processes, including laser energy absorption and transmission, material evaporation, remelting and
83 solidification, melt pool fluid dynamics, and microstructure evolution via epitaxial growth and nucleation. Defects, such
84 as cracks, pores, distortion, and insufficient melt frequently occur in the printing process, impacting the manufactured
85 parts’ mechanical and functional properties. The severity and density of these defects can be seen as a semi-stochastic
86 function of material parameters and printing parameters related to laser energy density and its associated parameters,
87 such as power, speed, and spot size [8]. In addition to the above factors, the LBAM process consists of different print
88 technologies. Even though they share the same basic principle, individual characteristic are exhibited by different
89 LBAM processes, which makes defect detection harder. Also, part geometry and the specific powder utilized for
90 fabrication influence defect formation dramatically, making defect detection in LBAM complicated and challenging.
91 To accelerate the industrialization of the LBAM processes, a novel and effective defect detection system to detect and
92 eliminate defects and guarantee product quality is essential.
93 Machine learning (ML) has progressed from a laboratory curiosity to a fundamental process used in various
94 industries and fields, including smart manufacturing [9, 10], civil engineering [11], and biomedical science [12]. For the
95 LBAM processes, ML has proven itself as a useful way to monitor product quality or detect defects [13]. Applying ML
96 to a defect detection system offers new insights into LBAM processes due to its ability to discover implicit knowledge
97 and build the relationship between printing parameters and product quality [14].
98 This paper reviews and summarizes the latest ML algorithms that have been utilized for LBAM defect detection
99 systems, examines the achievement of these systems, and discusses the current and future research direction. It is
100 important to point out that this paper only focuses on metal LBAM processes. In the author’s search, no defect detection
101 ML algorithms for laser-based wire-feed systems were found; all research focused on powder blown systems and
102 powder bed fusion systems. Powder blown system is commonly known as directed energy deposition (DED), laser
103 metal deposition (LMD), or laser engineered net shaping (LENS), while powder bed fusion system, is commonly
104 known as laser powder bed fusion (LPBF), direct metal laser melting (DMLM), direct metal laser sintering (DMLS),
105 or selective laser melting (SLM). As these technologies essentially are the same, DED will be used for all powder
106 blown systems and LPBF for all powder bed fusion processes to unify the naming conventions. To have a systematic
107 and comprehensive summary, ML algorithm, material type, defect type, input data type, and algorithm accuracy for
108 various LBAM technologies are described. In this paper, the background is described in section 1, and section 2 presents
109 different defect types in the LBAM processes. In section 3, ML algorithms used in defect detection systems for the
110 LBAM processes are discussed; a conclusion to the review and future trends are discussed in section 5.
111 2. Defect types in LBAM

112 Although different LBAM processes have their own characteristics, they share similar manufacturing methodolo-
113 gies. The two most common technologies for LBAM are DED and LPBF and are shown in Fig. 1. As shown in Fig. 1(a),
114 the center of a typical DED system is the nozzle head, which consists of the energy source (e.g., laser beam, electron
115 beam) to melt additive material at the point of deposition and the material delivery nozzle for feeding powder or wire
116 [15]. Typically, the nozzle head is fixed on either an articulated arm or a multi-axis computer numerical control (CNC)
117 head. The laser beam melts the powder/wire and forms the melt pool on the substrate at the start spot along the build
118 track. This process continues until the whole part is completed. The LPBF system is shown in Fig. 1(b). In the LPBF
119 printing process, the laser beam scans at a controlled speed on the powder bed, and the selected locations of the powder
120 are fused to form a solid track. The powder bed is sunk by the predefined layer thickness, and a new powder layer is
121 spread and leveled after the previous layer is finished. The process repeats until the part is completely built. As the
122 LBAM processes are complicated; many printing parameters, such as laser scan speed, laser power, scanning pattern,
123 material type and size, and chamber environment, are involved in this process. Any improper settings in printing will
124 produce defects. These defects can be divided into four types: product geometric and dimensional defect, porosity,
125 incomplete fusion, and cracks. This section describes defect types and is organized as follows: section 2.1 shows the
126 geometric and dimensional defect, section 2.2 presents the porosity defect, incomplete fusion is described in section 2.3,
127 and section 2.4 discusses the crack defect.

Figure 1: Schematic illustration of LBAM processes with two dierent feeding systems: (a) the DED technology of the
powder blown system; and (b) the LPBF technology of the powder bed fusion system. This gure is an original work by
the authors.
128 2.1. Geometric and dimensional defects
Figure 2: An distorted product of LPBF process with stainless steel 316L: (a) CAD model of the designed product; (b)
manufactured product; and (c) the distortion area on the manufactured product. This gure is an original work by the
authors.
129 Geometric deviation is one of the common defects in the LBAM processes. Geometric defects may be caused by
130 machine errors and errors in the generation of machine code that controls laser movement. Laser position error and
131 platform movement error are two machine errors that lead to geometric defects [16]. Layer-based slicing can introduce
132 step-wise marks on surfaces, where their contours may also introduce inaccuracy due to melt pool dimensions.
133 For dimensional inaccuracy, the main influence factor is shrinkage/distortion. In the LBAM processes, there are two
134 kinds of shrinkage: sintering shrinkage and thermal shrinkage. Sintering shrinkage is mainly produced by densification,
135 while thermal shrinkage is caused by cyclic heating and cooling, which leads to significant residual stress, thus, local
136 plastic deformation [17, 18, 19]. An example of a distorted canonical geometry is shown in Fig. 2, where shrinkage
137 was observed near the top.

Figure 3: Product's surface metrology of LPBF process with stainless steel 316L. This gure is an original work by the
authors.
138 Another factor influencing product geometric and dimensional accuracy is surface finish accuracy, also known
139 as surface roughness [20]. During the printing process, partially melted particles and spatters attach to the product
140 surface, causing the final dimension to vary from that designed [21]. Poor surface roughness not only impacts the
141 product’s usefulness but also influences the material’s properties (e.g., fatigue life) [22]. Fig. 3 shows a typical surface
142 morphology of the LPBF process in 316L stainless steel, where melt pool tracks and attached particles can be observed.
143 2.2. Porosity
Figure 4: Optical image of porosities in LBAM printing part with stainless steel 316L. This gure is an original work by
the authors.
144 Porosity is a common defect in the LBAM processes and negatively impacts product density [23]. It is particularly
145 deleterious to mechanical properties, such as fatigue for structural components. The porosity size varies with
146 distinguishing shapes under different forming mechanisms [24]. Typically, pores form due to insufficient powder
147 spreading, lack of fusion (as part of the “incomplete fusion defect” discussed next), keyholing, shrinkage, and gas

148 involvement. The first three types are large in size and are due to improper process parameters. For example, low metal
149 powder packing density (e.g., less than 50%) and spatters may lead to large voids in the powder bed, which cannot
150 be filled during melting. High laser energy density leads to keyholing, and metal vapor pressure results in pores at
151 the bottom of the deep melt pool. Through optimization of LBAM processes parameters, these defects can be fully
152 eliminated [25]. The latter two types, in general, are smaller in size. Pores due to solidification shrinkage are normally
153 located between grains and follow the shape of grain boundaries [26]. Gas-induced pores are spherical in shape and
154 can possibly originate from the gas trapped in the powder feedstock particles due to the gas-atomized process, gas
155 dissolved in metal due to different gas solubility at different temperatures, metal vapor generated during melting, or
156 moisture from the powder surface. An example of porosity is shown in Fig. 4. This type of pores cannot be completely
157 avoided due to the nature of its forming mechanism.
158 2.3. Incomplete fusion
Figure 5: Optical image of the lack of fusion defects in LBAM printing part with stainless steel 316L. This gure is an
original work by the authors.
159 The incomplete fusion hole is mainly produced by a lack of energy input in the LBAM printing processes, also
160 known as lack of fusion defects, as shown in Fig. 5. Lack of fusion defects are produced mainly because the metal
161 powder is not fully melted to deposit a new layer on the previous layer with sufficient overlap between them [27, 28].
162 The lack of fusion defects can be categorized into two types: poor bonding defects formed by insufficient molten metal
163 powder in a solidification process and defects due to unmelted metal powder. It is also one of the reasons for large
164 pores in LBAM.
165 When the scan path of the laser density is low, the molten pool width is small, which produces an insufficient
166 overlap between each track. The insufficient overlap then results in unmelted metal powder between the scan tracks as
167 there is insufficient energy to fully melt the powder for a new layer [29]. As a result, incomplete fusions are produced
168 and unmelted powder remains in the LBAM printed product, especially between the scan tracks and the deposited
169 layers. Moreover, the defect formation area makes the product surface rough, directly impacting the molten metal’s
170 flow to form interlayer defects. Then the interlayer defects may gradually extend and propagate to form more significant
171 multi-layer defects in the continuing printing process [30].

172 2.4. Cracks
Figure 6: Image of crack morphology from an LBAM fabricated part with stainless steel 316L: (a) crack morphology; and
(b) enlarged crack morphology. This gure is an original work by the authors.
173 In the LBAM printing processes, with a high laser energy input, the metal powder undergoes fast-melting and rapid
174 solidification. The cooling rate in the melting pool can reach more than 1.6 × 106 K/s following high-temperature
175 gradients [31]. Cracks initiate during and after solidification based on the materials, processes, and part design. For
176 example, solidification cracking forms due to insufficient liquid feed at the last stage of solidification. Strain-age
177 cracking, a type of ductile cracking, forms due to the formation of precipitates during heat treatment processes. Also
178 common is the tremendous amount of residual stress is developed in the printed product, which may exceed the ultimate
179 strength of the materials and lead to cracks [32, 33]. In Fig. 6(a), solidification cracking was developed along grain
180 boundaries in the LPBF printing process fabricated part. Fig. 6(b) shows the enlarged crack morphology of the red box
181 in Fig. 6(a).

182 3. Machine learning algorithms for defect detection
Figure 7: Categories of dierent machine learning algorithms (all the abbreviations are listed in the nomenclature and can
be found in the paper). This gure is an original work by the authors.
183 ML is a subset of artificial intelligence (AI) that provides systems the capacity to learn and improve from experience
184 without being explicitly programmed. Lots of the ML models have been built and applied by researchers for defect
185 detection in AM [14]. This section reports ML algorithms used explicitly for defect detection during the metal LBAM
186 processes. The field of ML can generally be divided into four domains, as shown in Fig. 7: 1) supervised machine
187 learning; 2) unsupervised machine learning; 3) semi-supervised machine learning, and; 4) reinforcement machine
188 learning. This section presents the state-of-the-art in defect detection and is organized by domain: section 3.1 presents
189 supervised machine learning algorithms; section 3.2 describes unsupervised algorithms; section 3.3 lists the semi-
190 supervised machine learning algorithms; and section 3.4 discusses reinforcement learning algorithms.

Y. Fu et al.: Preprint submitted to Elsevier
Table 1: Supervised ML algorithms utilized in defect detection for LBAM processes.
ML ML AM
Material type Defect type Dataset type Accuracy Reference
category algorithm technology
17-4 precipitation Melt pool
LPBF Anomalous, good quality 99.6% [34]
Logistic stainless steel thermal image
regression Mixing Ti-6Al-4V
DED Pore, crack AE signal 1.72973 [35]
with H13 tool steel
(MSE)
Gaussian process laser power,
LPBF 17-4 stainless steel Porosity No [36]

regression scanning speed
specific
accuracy
Underheating, medium
underheating, normal,
LPBF Stainless steel 304L Acoustic signal 89.13% [37]
medium overheating,
overheating
Melt pool
DED Ti-6Al-4V Porosity 97.97% [38]
thermal image
17-4 precipitation Melt pool
Supervised LPBF Anomalous, good quality 99.5% [34]
stainless steel thermal image
learning
Melt pool
DED Ti-6AL-4V. Healthy layer, unhealthy layer 91.65% [39]
thermal image
Support vector
Desirable, balling,
machine
severe keyholing, Melt pool
LPBF Inconel 718 85.1% [40]
keyholing porosity, morphology
or under-melting
Layerwise
LPBF Stainless steel GP-1 Anomalous, good quality 80% [41]
Optical imaging
Layerwise
LPBF Ti-6Al-4V Porosity 89.36% [42]
Optical imaging
LPBF Stainless steel 316L Track continuity Optical image 90.1% [43]
LPBF Stainless steel GP-1 Anomalous, good quality Optical image 85% [44]
DED Stainless steel 316L Track depositing height Printing parameters 2.89E- [45]
08(MSE)
Bayesian LPBF Inconel 625 Porosity, quality of fusion Visual image 89.5% [46]
Page 9 of 31
classifier AlSi10Mg
LPBF Key hole, lack of fusion Optical image 77% [47]
aluminum powder
ML ML AM
Melt pool
DED Ti-6Al-4V Porosity 98.44% [38]
thermal image
Layerwise
K-Nearest Optical imaging
neighbors 17-4 precipitation Melt pool
LPBF Anomalous, good quality 99.1% [34]
stainless steel thermal image

LPBF Inconel 718 Porosity Photodetector data 90% [48]
No defect, bulge defect,
LPBF Stainless steel 316L Point cloud 93.15% [49]
dent defect, wavy defect
Random Macropores, micropores, Optical microscope
DED Al-5083 powder 94.41% [50]
decision tree elongated pores image
Layerwise
LPBF Stainless steel GP-1 Anomalous, good quality 90% [41]
Optical imaging
Layerwise
Optical imaging
Supervised
Mixing Ti-6Al-4V
learning DED Pore, crack AE signal 1.702703 [35]
with H13 tool steel
(MSE)
Printing process
LPBF Inconel 625 Surface texture Refer to [51]
parameters
Artificial paper
neural network Balling, lack-of-fusion, Pyrometer data,
LPBF Stainless steel 316L Refer to [52]
conduction, key hole video camera data
paper
LPBF Ti-6Al-4V Geometry variation Layerwise image 92.50 ± [53]
1.03%
Titanium alloy,
LPBF STM B348 Grade 23 Printing condition variation Powder bed image 97.14% [54]
Ti-6Al-4V
LPBF Stainless steel 316L Track continuity Optical image 92.7% [43]
Stainless steel 304,
stainless steel 316,
Page 10 of 31
Good quality, crack, Digital

DED Ti-6Al-4V, 92.1% [55]
gas porosity, lack of fusion microscope image
Convolutional AlCoCrFeNi alloys,
neural network Inconel 718
ML ML AM
Thin wall’s thickness, density,
LPBF Ti-6Al-4V Optical image 85% [56]
edge smoothness, discontinuity
Good quality, average quality, QM-meltpool
LPBF CL 31 98.9% [57]
bad quality 3D-generated image
LPBF Stainless steel 316L Track continuity Printing process 93.1% [58]
video

Under-melt, beautiful-weld,
LPBF ASTM F75 I CoCrMo Micrograph image 100% [59]
over-melt
DED Stainless steel 316L Dilution MWIR image 2.8% [60]
(RMSE)
DED Sponge Titanium Porosity Melt pool image 91.2% [61]
powder
Stainless steel 304,
Gas porosity, crack,
DED stainless steel 316, Optical image 92.1% [55]
lack of fusion
Ti-6Al-4V, AlCoCrFeNi
DED Ti-6Al-4V Porosity Pyrometer image 100% [62]
Supervised Convolutional
DED 0Cr18Ni9 powder Printing parameter variation Thermal image 80% [63]
learning neural network
AlSi10Mg, bronze, Recoater hopping, recoater
Inconel 625, Inconel 718, streaking, porosity,
LBPF stainless steel 316L, swelling, spatter, soot, debris, Powder bed image Refer to [64]
Ti-6Al-4V, Fe-3Si, super-elevation, part damage, paper
stainless steel 17-4 PH incomplete spreading
Short feed defect, warpage, Powder layer
LPBF Alloys, Inconel 718 94%, [65]
part shifting defect optical image
alloys 96%,
94%
AlSi10Mg, bronze,
Recoater hopping, recoater
Inconel 625, Inconel 718,
streaking, debris,
LPBF stainless steel 316L, Powder bed image 97% [66]
super-elevation, part failure,
Ti-6Al-4V, Fe-3Si,
incomplete spreading
stainless steel 17-4 PH
Page 11 of 31
Overheating, normal,
LPBF Inconel 718, stainless Optical image 99.7% [67]
irregularity, balling
ML ML AM
Stainless steel 316L,
DED Distortion Thermal image 24 𝜇𝑚 [68]
Ti-6Al-4V, Fe-3Si,
(RMSE)
Defects from standard energy, Powder layer images,
LPBF Stainless steel 17-4 PH low energy, high energy, part slice images 99.4% [69]
very low energy after laser scanning

Metal fracture
Metal fracture, microscope images,
LBAM Not mentioned 82%, [70]
metallographic defects Metal metallographic
87.5%
images
Acoustic
Porosity,
Convolutional LPBF Ti-6Al-4V, spectroscopy, No [71]
surface imperfections
Supervised neural network CM247-LC optical image specific
learning accuracy
Laser additive manufacturing
LBAM Ti–6Al–4V transvers and longitudinal, Microstructure 90.4% [72]
laser cladding
Delamination, splatters, Thermographic
LPBF H13 tool steel 96.80% [73]
good quality off-axis imaging
LPBF Stainless steel 316L Poor, medium, high quality AE signal 83% - [74]
89%
Long-term
recurrent
DED Ti–6Al–4V Porosity Thermal image 92.07% [75]
convolutional
networks
Page 12 of 31
191 3.1. Supervised learning

192 Supervised learning is the most widely used ML technique. The training dataset needs to be labeled with input
193 values and the corresponding output values. During the training process, the ML algorithm uses this labeled dataset
194 and learns the relationship between input data and output data. Supervised learning is appropriate for classification
195 and regression; therefore, many defect detection systems use supervised learning approaches to detect and classify
196 different defects in LBAM. For clarity, the supervised learning subsection of this paper is divided into three parts:
197 traditional classifier, artificial neural network (without convolutional layers), and convolutional neural network. As
198 most of the reviewed research utilizes several classification algorithms in one paper and compare their performance, this
199 paper lists all traditional classification algorithms together. Traditional classification algorithms include support vector
200 machine, Bayesian classifier, logistic classification, K-nearest neighbors, and decision tree. Artificial neural network
201 (ANN) can also be formulated as a classifier and are discussed here in detail with one subsection. Furthermore, ANNs
202 that utilizes convolutional layers in their architectures are separated into their own subsection due to their prevalence
203 in the literature. Table 1 compiles a list of supervised ML algorithms utilized for detecting defects in components
204 manufactured using the LBAM processes. In Table 1, key features of the reviewed research are listed, including ML
205 domain, dataset type, material type, defect type, and algorithm performance.
206 3.1.1. Traditional regression

207 Regression is a method of modeling target values based on independent predictors. This method is mainly used
208 to uncover cause and effect relationships between variables. Regression techniques differ primarily depending on the
209 number of independent variables and the relationship between the dependent and independent variables. Regression
210 algorithms may be linear as well as non-linear.
211 Different regression algorithms also have been applied to defect detection in the LBAM process. Mahmoudi et al.
212 built an anomaly detection system for the LPBF printing process [34]. This system detects printing process deviations
213 by using the thermal signals obtained from the thermal images of the melt pool. Logistic regression is utilized to
214 decide the printing process quality (in control or out of control) through a step-by-step process. The result shows this
215 framework had a very low error rate for cavity defect detection, which can detect the 750 𝜇𝑚 diameter cylindrical cavity
216 on a 5.5 × 8 × 9 mm rectangular prism printed by 17-4 precipitation hardening stainless steel powder. Gaja et al. used a
217 logistic regression model to detect product defects simulated by mixing Ti-6Al-4V powder with H13 tool steel powder
218 in the LMD printing process [35]. With the AE signal as the input data, this approach can detect two primary defects
219 between crack and porosity. Moreover, Tapia et al. proposed a spatial Gaussian process regression model to learn and
220 predict porosity for the SLM process [36]. Two printing parameters, laser power and scanning speed, which have the
221 most significant effect on porosity formation, are fed into the Gaussian prediction model. The case study validated on
222 10 mm × 10 mm × 10 mm test coupons printed with 17-4 PH stainless steel powder. The result shows that the proposed
223 model provided accurate porosity prediction under any printing parameters.
224 3.1.2. Traditional classifiers

225 Classification is a class of ML that focuses on predicting the label associated with a given piece of data. Sometimes
226 classes are termed as labels, categories, or targets. Classification predictive modeling is the task of approximating a
227 mapping function (𝑓 ) from input variables (𝑋) to discrete output variables (𝑦). In a classification task, the output can
228 be a binary or multi-class classification according to the chosen classifier, such as good or bad quality, defect types.

Figure 8: Demonstration of porosity prediction procedure using supervised machine learning [38].
229 Support vector machine (SVM) is a popular ML tool that offers a solution for both classification and regression.
230 Various researchers have used this algorithm for defect detection; for example, Khanzadeha et al. applied multiple
231 supervised learning methods (i.e., decision tree (DT), K-nearest neighbor (KNN), SVM, linear discriminant analysis
232 (LDA), and quadratic discriminant analysis (QDA)) to predict porosity on the single track thin wall specimens (Ti-6Al-
233 4V) by using melt pool thermal images [38]. In the SVM method, when selecting the polynomial as the kernel function,
234 the porosity prediction accuracy is 97.97%. Fig. 8 shows the overall SVM methodology for porosity prediction. To
235 classify the melt pool morphologies, Scime et al. developed a flaw formation identification method for the LPBF
236 process with a multi-class SVM, which can detect five melt pool types on a 10 mm × 20 mm rectangle made with
237 Inconel 718 powder: under-melting, balling, spatter, porosity, and desirable [40]. Gobert et al. applied SVM to detect
238 defects on a staircase cylinder fabricated with stainless steel GP-1 powder in the LPBF process [44]. By labeling
239 different flaws (i.e., porosity, incomplete fusion, crack, or inclusions) with computer tomography (CT), the properly
240 trained model achieved an over 80% defect detection accuracy. The flaw with a diameter larger than 47 𝜇m can be
241 identified. Petrich et al. developed an approach based on the SVM and neural networks to detect defects during the
242 LPBF process with high-resolution layer-wise images [41]. By utilizing a cross-validation methodology, the proposed
243 algorithm in situ anomaly detection achieved a 90% accuracy for a staircase cylinder made with stainless steel GP-1
244 powder. Imani et al. investigated the printing process to figure out the effects different printing process conditions
245 had on fusion porosity in LPBF [42]. By analyzing the in-process layer-by-layer optical images, the result shows the
246 process conditions significantly impact the porosity generation. This approach satisfies statistical fidelity by linking the
247 features extracted from the layer-by-layer images to the process condition with ML methods. Several ML classifiers
248 (SVM, complex tree, LDA, KNN, bagged trees, and feed-forward neural network) are applied in this research. The SVM
249 classifier achieved an 89.36% porosity defection accuracy with F-score for the cylinders of 10 mm diameter × 25 mm
250 height manufactured with Ti–6Al–4V. The resolution for the pore detection is larger than 65 𝜇m. Mahmoudi et al. built
251 an anomaly detection system for the LPBF printing process [34]. This system detects printing process deviations by
252 using the thermal signals obtained from the thermal images of the melt pool. Classifiers, including logistic regression

253 (LR), KNN, SVM, and random forests (RF), are utilized to decide whether the printing process is in control or out
254 of control through a step-by-step process. The result shows this framework had a very low error rate for cavity defect
255 detection.
256 SVM is also utilized by researchers for product quality prediction and inspection. Seifi et al. proposed a method to
257 detect product anomalies in real-time for LBAM and verified on the DED printing process [39]. After extracting the
258 key layer-wise signature features from the melt pool images with multilinear principal component analysis (MPCA),
259 an SVM classifier is used to extract key layer-wise signature features to predict the product’s quality. The proposed
260 methodology prediction accuracy obtained an F1-score of 91.65% and was validated on a thin wall fabricated with
261 Ti-6AL-4V. Lu et al. utilized the least square support vector machine (LS-SVM) to predict the track’s depositing
262 height for the DED process [45]. With laser scanning speed, laser power, and powder feeding speed as inputs, this
263 model provided knowledge on the whole printing part’s precision. The experimental result on a thin wall printed
264 with AISI316L powder shows that the LS-SVM algorithm prediction value had a high correlation coefficient with the
265 experimental value of 0.971. Ye et al. proposed a defect-recognition approach for the LPBF process [37]. Using SVM
266 and extracted features from recorded acoustic signals, this method achieved accurate diagnosis for five different melted
267 states (i.e., overheating, medium overheating, underheating, medium underheating, and normal) for the single tracks
268 printed by 304L stainless steel powder.
269 The Bayesian classifier is a probabilistic classifier that gives probability information about the examined product
270 layer being defective, which could also be considered a product quality quantitative method. Therefore, the Bayesian
271 classifier is also a useful tool for defect detection in the LBAM processes. A Bayesian classifier was built and trained
272 by Aminzadeh et al. to classify the product quality fabricated by Inconel 625, which can detect the defective regions
273 or abnormal layers in LPBF [46]. The training data is labeled printing process images with detailed layer defects
274 and porosity. The result shows that with appropriate feature selection, the defect identification performance is 89.5%.
275 To predict the product quality (stainless steel 316L powder) printed by LPBF, Hertleina et al. proposed a Bayesian
276 network, which combines the printing process parameters with product quality characteristics [76]. After training, the
277 forecasted mean of a printed product quality characteristic (hardness) is within 0.41 standard deviations of the true
278 value. Bartlett et al. utilized a three-dimensional digital image correlation system to predict microstructural defects
279 (i.e., lack of fusion and keyhole) from AlSi10Mg aluminum powder bed quality in direct metal laser sintering (DMLS)
280 printing process [47]. By feeding the powder anomaly topology images into the Naïve-Bayes classifier, the algorithm
281 can predict microstructural defect formation probability according to the in-process powder bed error. The defect and
282 no defect prediction accuracy for three energy density printing conditions (i.e., low, standard, and high) are 66%, 77%,
283 and 72%, respectively.
284 In addition, other classifiers are also utilized by researchers for detecting defects in the LBAM processes. For
285 example, Montazeri et al. presented an in-process porosity monitoring approach using optical emission spectroscopy
286 [48]. The experimental result shows that using the graph Fourier transform coefficients as the input feature within
287 the KNN model had the best layer porosity-level prediction accuracy for two-level classification (90% F-score) on the
288 disk manufactured with nickel alloy 718 powder. In addition, the computation time only needs less than 0.5 s. Chen
289 et al. proposed a rapid surface defect detection method for the DED technology, which integrates the in situ point
290 cloud processing with ML [49]. The algorithm used in this research combined unsupervised with supervised ML to
291 identify and classify the surface defect. The unsupervised algorithm is used to segregate the potential surface defect
292 region from the point cloud. This approach registered the defect’s existence and typed recognition by inputting the
293 clustering result from an unsupervised algorithm into the supervised algorithm. The surface defects are categorized
294 into four classes: no defect, bulge defect, dent defect, and wavy defect. To compare the classification accuracy, eight
295 algorithms are applied; the most accurate algorithm for classification is the KNN classifier, with an accuracy of 93.15%.
296 Garcia-Moreno et al. presented an image-based porosity classification method by utilizing a random forest algorithm
297 for the LMD printing process [50]. The random decision tree classifier achieved an accuracy of 96.45%, 94.56%, and
298 92.21% for the macropores, micropores, and elongated pores on the dogbone sample fabricated with Al-5083 powder,
299 respectively. The proposed method can segment and classify pores between 5 and 250 𝜇𝑚.
300 3.1.3. Artificial neural network

301 Artificial neural networks (ANN) are a class of ML algorithms modeled loosely after the human brain and
302 designed to recognize patterns. A typical ANN usually has three parts: the input layer (first layer), one or more hidden
303 layers (middle layers), and the output layer (last layer). Each layer consists of a collection of neurons, which are
304 nodes connected with the other nodes via links (also termed synapse). Each link has a weight, which decides one

305 node’s impacting strength to another node. During the training process, the ANN model forms probability-weighted
306 associations between the input and target. The target output is given; thus, training is performed by reducing the
307 difference between the network’s processed output and the target output. Backpropagation is used to adjust the model’s
308 weights and biases, given the change of error at each step (also termed the gradient) and a predefined set of learning
309 rules. After successive and sufficient adjustments, the network’s output is increasingly similar to the target output,
310 which means the network has been fully trained and can be used for future prediction.
311 When fed a labeled dataset for training, the ANN is helpful for clustering and classification. Petrich et al. developed
312 an approach based on the ANN to detect defects during LPBF with high-resolution layer-wise images [41]. By utilizing
313 cross-validation strategies for training and testing, close to 90% accuracy is achieved when using the ANN model to
314 detect the defect for a staircase cylinder fabricated by stainless steel GP-1 powder. Gaja et al. used an ANN model
315 to detect product defects in the LMD printing process [35]. With the AE signal as the input data, this approach can
316 detect crack and porosity simulated by mixing Ti-6Al-4V powder with H13 tool steel powder. To predict the product
317 surface texture in LPBF, Ozel et al. applied several ML algorithms to establish the relationship between printing process
318 parameters (i.e., laser energy density and scan strategy) and measured surface texture parameters [51]. For the ANN
319 algorithm, it is able to reflect the actual conditions when the measurement data is sufficient; for the genetic programming
320 (GP), it is good at selecting the best LPBF process parameters for the cubes fabricated with nickel alloy 625 with the
321 dimension of 16 mm × 16 mm × 15 mm. Gaikwad et al. developed and evaluated an ML-based quality assessment
322 model for a single track printed with stainless steel 316L powder by LPBF process [52]. The study researched various
323 printing parameters (i.e., laser power and laser velocity) effects on single-track’s quality and analyzed the quality
324 measurement with height-map by using the extracted mean and standard deviation of width and percent continuity.
325 The experimental result shows that the proposed sequential decision analysis neural network (SeDANN), which uses
326 the sensor data-derived features, performs better than the other ML models, such as long short-term memory recurrent
327 neural network (LSTM-RNN), in speed and accuracy for balling, lack-of-fusion, conduction, and keyhole detection.
328 Fig. 9 shows the schematic of the SeDANN proposed by Gaikwad et al.
329 A deep neural network (DNN) is an ANN with multiple layers between the input and output layers. The DNN
330 algorithm finds suitable mathematical manipulation to exploit the input information into the output, whether it is a
331 non-linear or linear relationship. DNNs have also been applied to defect detection in the LBAM processes. For example,
332 Imani et al. designed a DNN model for real-time incipient geometry defects detection from the spatial characterization
333 images in the LPBF process [53]. The experimental result shows that the proposed DNN model effectively achieved
334 geometry flaw detection on drag link joint object (23.7 mm × 13.3 mm × 27.3 mm). The accuracy achieved is 92.5%
335 for the 750 𝜇𝑚 cylinder and cube defect. As laser power has a huge impact on the pores and cracks formation, which
336 directly determines the printing product quality, Kwon et al. applied a DNN in LPBF to find the link between melt-pool
337 images and laser power [77]. The cuboid product with a dimension of 8.5 mm × 8.5 mm × 4 mm is fabricated with
338 stainless steel 316L. The proposed DNN model’s classification accuracy on this product is 98.9%, which is helpful
339 for understanding the product microstructure formation by abnormal laser power. Mohammadi et al. presented a novel
340 approach for defect detection in the LPBF process by utilizing the elastic waves from the AE sensor [78]. The proposed
341 approach can detect three types of defect (i.e., parts with minimum defects, parts that had only intentional cracks, and
342 parts that had both intentional cracks and porosities) on the cylindrical parts with a dimension of 20 mm diameter and
343 10 mm height fabricated with H13 tool steel powder. Three ML algorithms are utilized to analyze and interpret the
344 data. Firstly, a K-means clustering is applied for data labeling, then followed by a DNN to match the AE signal with
345 the correct defect type. Secondly, a PCA is employed to reduce the data’s dimensionality. A Gaussian Mixture Model
346 (GMM) is utilized to accelerate the defect detection speed. Thirdly, a variational auto-encoder method is applied to get
347 a general feature of the AE signal that could be an input for the classifier. The unique contribution of this work is that
348 the proposed approach can generalize the classifier to be used for different materials without the need for training.

Figure 9: A schematic of the sequential decision analysis neural network (SeDANN). The sensor data and height map
shown above belong to a single-track deposited at linear energy density (EL) of 0.33 (i.e., balling regime). The statistical
probability distribution features extracted from the pyrometer are used in the rst echelon ANN to predict the laser process
parameters (P and V) followed by melt pool features derived from the high-speed video camera to predict the mean width
and standard deviation and single-track continuity at higher echelons [52].
349 3.1.4. Convolutional neural network

350 A convolutional neural network (CNN) is a class of ANN that incorporates one or more convolutional layers into
351 the architecture of the neural network for feature extracting, one or more pooling layers for subsampling, and then with
352 one or more fully connected layers for the output. CNN is the most popular deep learning (DL) algorithm for image
353 recognition, image classification, and object detection due to its outstanding image processing and pattern recognition
354 performance.
355 The CNN algorithm has been used for defect detection in the LBAM process. For example, Scime et al. utilized
356 a CNN algorithm for autonomous anomaly detection and classification in LPBF [66]. The proposed model achieved
357 six types of flaws detection (i.e., recoater hopping, recoater streaking, debris, super-elevation, part failure, incomplete
358 spreading) with an overall accuracy of 97%. A case study validated the proposed algorithm’s ability by printing a
359 heat exchanger model with Inconel 718 powder. Fig. 10 shows the flowchart of the ML technology, which is utilized
360 by Scime et al. for defect detection. Baumgartl et al. developed a printing defect detection system by combining
361 thermographic off-axis images with a DL-based CNN [73]. By using the depthwise-separable convolutions to reduce
362 the dimension of channels into the neural network architecture, this technique achieved an accuracy of 96.80%
363 for delamination and splatter defects. Furthermore, the model’s architecture is small and therefore has reduced

364 computational cost, which shows great potential to apply the proposed algorithms for online defect detection. Cui
365 et al. utilized a CNN algorithm to inspect the defect for the product produced by LMD [55]. The materials in this
366 research include AISI 304 stainless steel, AISI 316 stainless steel, Ti-6Al-4V, AlCoCrFeNi alloys, and Inconel 718
367 alloys. The proposed algorithm had an accuracy of 92.1% for defects, such as gas porosity, crack, and lack of fusion.
368 Zhang et al. described a porosity prediction CNN model for the LBAM process [61]. With the melt pool image from
369 the single track manufactured by sponge titanium powder, the algorithm had a 91.2% porosity detection accuracy and is
370 able to predict the micro-pores below 100 𝜇𝑚. Moreover, Guo et al. presented a physical-driven CNN model to predict
371 porosity on thin-wall structure from Ti-6Al-4V powder by using metal pool thermal images obtained from a pyrometer
372 for the DED process [62]. When combining the data-driven feature from the pyrometer image with the physical feature
373 from finite element analysis, the proposed algorithm had a reported 100% porosity prediction accuracy. Zhang et al.
374 proposed an online monitoring method for LBAM products, which relies on a CNN algorithm [59]. The experiment
375 results show that when using a small local image block, the classification accuracy for bonding quality (i.e., under-melt,
376 normal, and over-melt) is 82%; when utilizing the full images, the model accuracy was reported as 100%.
Figure 10: Flowchart of the implementation of the multi-scale CNN ML technique [66].
377 Product quality guarantee can also be achieved with CNN algorithms for the LBAM processes. Yuan et al. built a
378 CNN model to monitor the LPBF printing process by analyzing the single track quality made with 316L stainless steel
379 [58]. After training with the frames obtained from the melt pool video, this model is able to predict the track width,
380 width standard deviations, and track continuity. Zhang et al. utilized CNN for product quality level identification in
381 the LPBF process [43]. The input data is melt pool, plume, and spatter images obtained by a high-speed camera from
382 the 8 mm single track printing process with stainless steel 316L powder. The proposed CNN model’s quality level
383 identification accuracy is 92.7%. Furthermore, one image identification only takes 8.01 milliseconds. A CNN-based
384 approach was proposed by Gaikwad et al. to monitor and predict the thin-wall build quality with Ti-6Al-4V in LPBF
385 [56]. After training with the layer-wise thin-wall optical images captured by an optical camera in the LBPF printer,
386 the model can predict the X-ray computed tomography (XCT) derived statistical quality features, such as thickness
387 and edge consistency, with an accuracy exceeding 85%. Kunkel et al. proposed a product quality assurance method
388 for LBPF by utilizing a CNN-based image classification algorithm [57]. This CNN method had a promising quality
389 identification accuracy of 98.9% for the product made with AlSi10Mg powder. Li et al. proposed a DL-based process
390 monitoring for the DED technology [63]. The proposed CNN model used thermal images from the straight stick part
391 manufactured with 0Cr18Ni9 powder as input, which had an accuracy of over 80% for different printing condition
392 recognition (i.e., normal, lower power, low and high speed).

393 Deep convolutional neural network (DCNN) algorithm is a subclass of CNN, and therefore ANN, that utilizes
394 more layers in the architecture. A DCNN is defined as one that utilizes at least 20 layers in its formulation in this work.
395 There are several specific models for DCNNs with distinct structures, such as GoogleNet (also called InceptionNet),
396 VGGNet, and ResNet. Generally, DCNN is better suited at detecting non-linear relationship and feature extraction than
397 traditional CNN but require more data and longer training time than their more-traditional counterparts. Due to their
398 capability to detect non-linear patterns, DCNN algorithms have been widely applied to detect defects in LBAM. Han
399 et al. exploited defect detection in LBAM by utilizing a DCNN (Inception-v4) to analyze and classify images from the
400 manufacturing process [70]. In this research, the DCNN model was applied to two datasets, metal fracture microscopic
401 images and product parts of metal metallographic images. For the metal fracture microscopic image dataset, the model
402 achieved an accuracy of 82%, recall of 93.8%, and precision of 96.8%; the model trained on the metal metallographic
403 images, the accuracy, recall, and precision is 87.5%, 96.3%, and 97.6%. Li et al. built a novel CNN model structure
404 based on a dense convolutional network to recognize the microstructure in LBAM [72]. The result shows the accuracy is
405 90.4% and one microstructure image processing time only needs 0.1 s. Gonzalez-Val et al. proposed a novel ConvLBM
406 (CNN for laser-based manufacturing) method to estimate dilution in the LMD process [60]. Dilution is an important
407 quality indicator: a low dilution may produce insufficient bound and generate warping, but a high value means a large
408 heat-affected zone, which has a high defect probability because of the thermal expansion [79]. ConvLBM is a modified
409 CNN model based on ResNet [80] to extract key features from the raw medium wavelength infrared coaxial images.
410 The trained ConvLBM model estimated dilution with a root-mean-square error (RMSE) of 2.8%.
411 Ensembles of CNNs have also been developed to provide better fault detection capabilities than the traditional
412 stand-alone CNNs. Shevchik et al. investigated combining spectral CNN with acoustic emission for quality monitoring
413 in LBAM [74, 81]. The classification confidence varies between 83% and 89% on a cuboid with a dimension of 10
414 mm × 10 mm × 20 mm manufactured with stainless steel 316L. Xiao et al. developed a two-stage convolutional neural
415 network (TS-CNN) to predict different kinds of defects in LBAM [65]. The research demonstrated that this approach
416 had high accuracy and efficiency in coping with geometrical distortion manufactured with pure polyamide (PA) and
417 PA/TiO2 composite. The respective prediction accuracy for the warpage, part shifting, and short feed defects is 94%,
418 96%, and 94%. Williams et al. developed a CNN model named Densely connected convolutional block architecture
419 for multimodal image regression (DCB-MIR) to detect product defects in metal LBAM [71]. This proposed model
420 utilized the printing product’s SRAS(spatially resolved acoustic spectroscopy)-derived acoustic velocity maps as input
421 data and decoded them to a resembling optical micrograph as output. The model had high defects detection accuracy
422 (i.e., porosity and surface imperfections) for the titanium alloy and nickel alloy samples. The defects were not distinctly
423 recognizable in the as-measured SRAS acoustic map and blurred in the optical image. Zhang et al. proposed a hybrid
424 CNN to monitor the printing process in LPBF [67]. This hybrid CNN architecture consists of two CNN models; the
425 first one is utilized to study the spatial features from a single printing process image and the second one is applied to
426 product quality classification. The model’s overall detection accuracy on the 10 mm single track manufactured with
427 stainless steel 316L powder is up to 99.7% for the overheating, normal, irregularity, and balling defects. Scime et
428 al. built a new dynamic segmentation CNN (DSCNN) model for real-time pixel-wise semantic segmentation with
429 layer-wise powder bed images as input [64]. The proposed DSCNN model had a wide application range of powder-
430 bed-based AM machines, spanning from LPBF, to electron beam PBF, to binder jetting. The model segmented and
431 located the defect, include recoater hopping, recoater streaking, incomplete spreading, swelling, spatter, soot, debris,
432 super-elevation, part damage, and porosity, with good accuracy. Yazdi et al. proposed a hybrid DL model to monitor
433 the printing process parameters, which potentially affect the porosity defect production [54]. The input data is the
434 statistical features extracted from powder bed images by wavelet transform and texture analysis. The proposed model
435 had a 97.14% F-score porosity prediction accuracy on a cylinder (a height of 25 mm and a diameter of 20 mm) fabricated
436 with titanium alloy and Ti-6Al-4V powder. By applying a bi-stream DCNN model, Caggiano et al. built an online defect
437 recognition system for LBAM [69]. The pre-alloyed Inconel718 powder layer image and laser scanning layer image
438 are input into an ML algorithm as training data to manifest the defect produced by improper process conditions. The
439 study result shows that a mean defective condition-related image pattern recognition accuracy of 99.4% was achieved
440 on the final disc with a 40 mm diameter and 20 mm height.

Figure 11: Displaying the CAMP-BD model [68]: (a) the tensor structure of thermal history is visualized; (b) various
process/design parameters are listed as examples of additional inputs to CAMP-BD; and (c) an example prediction of
CAMP-BD.
441 Novel ensembles of ML techniques integrate different algorithms together to enable the new one to take full
442 advantage of the strengths from every integrated algorithm for detecting defects in LBAM, which researchers also
443 utilize. For example, Francis et al. developed a novel DL approach named CAMP-BD (Convolutional and Artificial
444 Neural Network for Additive Manufacturing Prediction using Big Data), which integrates CNN with an ANN algorithm
445 for analyzing the thermal images. It used the relevant process/design parameters as input data to predict fabricated
446 product distortion [68]. The whole CAMP-BD model is displayed in Fig. 11. The distortion prediction result on a
447 disk (5 mm thick and 45 mm diameter) made with Ti-6Al-4V powder, shows that most of the predictions are within
448 the metal LBAM machines’ tolerance limits. The distortion prediction RMSE on training data is 24 𝜇𝑚. Tian et al.
449 developed a DL-based in situ porosity detection method for LBAM, which uses the melt pool thermal images to
450 monitor the melt pool and predict porosity [75]. A PyroNet based on CNN and an IRNet based on long-term RNN two
451 different algorithms are utilized to correct layer-wise porosity. For the PyroNet, the input is in-process images obtained
452 from pyrometry, and for the IRNet, the input data is sequential thermal images captured from a thermal camera. To
453 have a higher porosity prediction accuracy, two algorithms are fused together when making the final decision. The
454 experimental result on a Ti–6Al–4V thin-wall structure shows that the average porosity prediction accuracy on six-
455 folds is 98.93%.

Table 2: Unsupervised ML algorithms utilized in defect detection for LBAM processes.
ML ML AM
AlSi10Mg, bronze,
Recoater hopping, recoater
Inconel 625, Inconel 718,
streaking, debris,
LPBF stainless steel 316L, Powder bed image 98% [82]
super-elevation, part failure,
Ti-6Al-4V, Fe-3Si,
incomplete spreading
stainless steel 17-4 PH
X-ray computed

Ti-6Al-4V, Inconel 718,
LPBF Key hole, lack of fusion 3D pore tomography, No [83]
Ti-5553, Haynes 282
2D pore micrograph specific
accuracy
K-means clustering
LPBF Stainless steel 316L Undesired overheating defect Optical image No [84]
specific
accuracy
Optical
LPBF AlSi10Mg Drift, no drift No [85]
tomography image
specific
accuracy
Mixing Ti-6A1-4V
Unsupervised DED Crack, porosity AE data No [86]
with H13 tool steel
learning specific
accuracy
DED Ti-6Al-4V Process condition variation AE signal 87% [87]
DED 7075-Al alloy powder Surface roughness Spectral Refer to [88]
porosity paper
Balling, slight balling,
LPBF Stainless steel 304 normal, overheating, Acoustic signal 95.93% [89]
slight overheating
Over-melted, middle
Deep belief network
over-melted,
LPBF Stainless steel 304L normal-melted, NIR image 83.40% [90]
middle under-melted,
under-melted
Page 21 of 31
Super-micropores, cracks,
Optical
DED Al-5083 hybrid pores, 95% [91]
Self-organized maps microscope image
inter-micropores defect
DED Ti-6Al-4V Porosity Melt pool image 96% [92]
456 3.2. Unsupervised learning

457 Unsupervised learning is a category of ML that looks for formerly undetected patterns in a dataset without pre-
458 existing targets or labels and only involves minimal human supervision. Compared with supervised learning, which
459 uses training data with labels, unsupervised learning, also termed self-organization, can model probability densities
460 based on the input data. The advantage of unsupervised learning is that no labeled data is needed. Table 3.1.4 shows
461 a list of unsupervised ML algorithms utilized for detecting defects in components manufactured using the LBAM
462 processes, which contains key features of the reviewed research, including dataset type, material type, defect type, and
463 algorithm performance.
464 3.2.1. K-means clustering algorithm

465 The K-means clustering (KMC) algorithm is one of the simplest yet most powerful unsupervised learning
466 algorithms for solving clustering problems. The procedure follows a simple method to classify a given dataset through
467 a pre-defined number of clusters (assume k clusters) fixed apriori. The main idea is to define k centers, one for each
468 cluster. Scime et al. presented an approach for in situ monitoring and analysis of powder bed images, which shows
469 promising potential to be used as a real-time control system in the LPBF process [82]. A computer vision algorithm
470 is used in this approach for automatic defect detection and classification during the powder spreading process. Defect
471 detection and classification are implemented by using a standard k-means unsupervised clustering algorithm to process
472 the printing image. The final algorithm achieved six types of flaw detection (i.e., recoater hopping, recoater streaking,
473 debris, super-elevation, part failure, and incomplete spreading). Snell et al. trialed an unsupervised ML method (KMC)
474 for rapid pore classification in metal additive manufacturing [83]. In this research, the pore data was collected by two
475 different techniques: 3D pore data from the XCT and 2D pore data from microscopy, with multiple alloy (Ti-6Al-4V,
476 Inconel 718, Ti-5553, and Haynes 282). The result shows that KMC is suitable for 3D pores, it performs well both on
477 lack-of-fusion pores and keyholes pores, but it can not be used to analyze the 2D pore data. Grasso et al. developed an
478 in-process defect spatial detection method by utilizing image data analysis for the LPBF process during the layer-wise
479 printing process [84]. The proposed image KMC algorithm achieved defect detection automatically by detecting and
480 positioning the potential defect in each layer. Taheri et al. developed an in situ process condition monitoring method for
481 the DED process [87]. By utilizing the KMC to process the acoustic signatures from a single-layer (Ti-6Al-4V) printing
482 process, the algorithm identified different process conditions (i.e., normal, low powder, and powder spray) with an
483 accuracy of over 87%. In addition, Gaja et al. presented a novel real-time defect detection and classification system for
484 the LMD process, which uses an acoustic emission sensor and a KMC unsupervised ML algorithm [86]. By analyzing
485 the AE signal, the experimental result shows that the KMC is able to recognize crack and porosity simulated by mixing
486 Ti-6Al-4V powder with H13 tool steel powder, two different defects effectively. Ren et al. built an unsupervised model
487 for product quality recognition manufactured by DED [88]. The proposed model has an LSTM-based autoencoder for
488 extracting encoded features which contain information to express the original spectra. Then the extracted features are
489 utilized to K-means clustering for product quality classification. Four 40 mm tracks with Al7075 alloy are printed in
490 the research, and the result shows that the classification matches with the real track quality.
491 3.2.2. Deep belief network

492 A deep belief network (DBN) is a generative graphical model, or alternatively a class of deep neural networks,
493 which is composed of multiple latent variable layers, and each layer is connected, but the units within each layer are
494 not related. DBN is another useful ML algorithm, which has been utilized for defect detection in LBAM. Ye et al.
495 demonstrated that acoustic signals are feasible for product quality monitoring, and DBN algorithms achieved a high
496 defect detection rate among five melted states on single tracks manufactured with 304 stainless steel powder (i.e.,
497 balling, slight balling, normal, slight overheating, and overheating) without signal preprocessing [89]. Fig. 12 displays
498 a generic classification-based DBN architecture that can be used for defect recognition with the acoustic signals from
499 the LBAM printing process. As the plume and spatter signatures have a very close relationship with the melted state
500 and laser energy density. Ye et al. applied DBN to the plume and spatter NIR images obtained from the LBAM printing
501 process, and this approach achieved an 83.4% classification rate for over-melted, middle over-melted, normal, middle
502 under-melted, and under-melted melted states on single tracks produced with 304 stainless steel powder [90].

Figure 12: An illustration of a generic classication-based deep belief network (DBN) with stacked RBMs (restricted
Boltzmann machines): (a) deep belief network (DBN); and (b) restricted Boltzmann machine (RBMs). This gure is an
original work by the authors.
503 3.2.3. Self-organized maps algorithm

504 Self-organized maps are a type of ANN trained by unsupervised learning to output a low-dimensional (generally
505 two-dimensional), discretized representation of the input space of the training samples. As an effective dimensionality
506 reduction method, self-organized maps are also applied to defect detection. Garcia-Moreno proposed an automatic
507 porosity quantification method by utilizing an unsupervised ML classifier for the LMD printing process [91]. The
508 final classifier based on self-organized maps obtained a 95% accuracy for the super-micropores, cracks, hybrid pores,
509 and inter-micropores defect for the product manufactured with Al-5083 powders. The proposed algorithm is sensitive
510 detecting the average size of 6.33 𝜇𝑚 inter-micropores. Moreover, Khanzadeh et al. proposed an in situ porosity
511 prediction method for the DED process based on the printing melt pool images [92]. According to the melt pool’s
512 temperature distribution, the proposed self-organized maps model predicted the porosity position on a Ti-6Al-4V
513 thin-wall specimen with an accuracy of 96%.
514 3.3. Semi-supervised learning
Table 3
Semi-supervised ML algorithms utilized in defect detection for LBAM processes.
ML ML AM Material Defect Dataset
Accuracy Reference
categories algorithm technology type type type
K-nearest neighbors No
Optical tomography
and K-means clustering DED AlSi10Mg Drift, no-drift specic [85]
image
combination accuracy
Gaussian Mixture
LPBF Inconel 718 Faulty, acceptable Photodiode data 77% [93]
Semi-supervised model
machine learning Track average No
Stainless steel
Semi-supervised DED width, Video data specic [94]
316L
convolutional neural track continuity accuracy
network Under-melt,
ASTM F75 I Micrograph
LBAM beautiful-weld, 90% [95]
CoCrMo image
over-melt
515 As discussed before, supervised learning uses a dataset that includes labels to train the algorithm to understand
516 which features are important to the problem at hand. However, the generation of labeled datasets is a costly process,
517 especially when dealing with large volumes of data. On the other hand, unsupervised learning is trained on unlabeled
518 data and must identify features and determine their importance by themselves based on inherent data patterns. The

519 disadvantage of any unsupervised learning is the limited application range. To counter these disadvantages, the concept
520 of semi-supervised learning was introduced. In semi-supervised learning, the algorithm is trained upon a combination
521 of labeled and unlabeled data. Typically, this combination contains a limited amount of labeled data and a larger
522 amount of unlabeled data. This is helpful for a few reasons. First, the process of labeling massive amounts of data
523 for supervised learning is often prohibitively time-consuming and expensive, however, a small amount of labeled data
524 can significantly improve the learning accuracy when paired with a large volume of unlabeled data. Furthermore,
525 too much labeling can impose human biases on the model. That means including lots of unlabeled data during the
526 training process tends to improve the final model’s accuracy while reducing the time and cost spent building it. In such
527 conditions, semi-supervised learning can be of great practical value. In the research situation, as shown in Table 3,
528 semi-supervised learning is also helpful as there is no need to label all data for training, which saves time and effort.
529 Semi-supervised ML has been studied by researchers for defect detection. Okaroa et al. introduced a semi-
530 supervised ML algorithm for automatic defect detection in the LBAM printing process [93]. The result on the tensile
531 test bars manufactured with Inconel 718 shows that the semi-supervised ML algorithm is a promising method for
532 automatically identifying the LBAM product defects, which achieved a comparable result to a benchmark where all
533 training data are labeled. To detect the inline drift in the LPBF process, Yadav et al. proposed a semi-supervised method,
534 which uses certified product computer tomography as input data [85]. The supervised ML-based KNN algorithm is
535 trained with the labeled input data from the unsupervised ML K-means clustering algorithm. Based on the case studies
536 validated on cylindrical samples (diameter of 10 mm and height of 15 mm) with AlSi10Mg powder, the proposed
537 semi-supervised ML algorithm had a reported 100% accuracy in predicting the exact drift layers.
Figure 13: Semi-supervised CNN architecture [94].
538 Semi-supervised CNNs eliminate the challenges of collecting large labeling datasets, making them more efficient
539 for detecting defects in the LBAM processes. Yuan et al. developed a semi-supervised CNN, which only uses a limited
540 amount of labeled data and a large amount of unlabeled data to monitor the LPBF printing process [94]. This approach
541 used the videos from the single-track printing process with 316L stainless steel powder as input data. The result
542 shows that the semi-supervised approach performs better than the fully supervised approach, no matter whether it is a
543 regression or classification problem. The semi-supervised CNN model’s architecture used in this research is shown in
544 Fig. 13. Li et al. proposed a DCNN model to analyze the product quality for the metal LBAM printing process [95]. In
545 this method, the semi-supervised training data was used to mitigate the demand for a large amount of labeled image
546 data. The algorithm proposed in this research had a 100% accuracy identification performance for under-melt, normal,
547 and over-melt on the product made with ASTM F75 I CoCrMo powder.
548 3.4. Reinforcement learning
Table 4
Reinforcement learning algorithms utilized in defect detection for LBAM processes.
ML ML AM Material Defect Dataset
Accuracy Reference
categories algorithm technology type type type
74%,
Markov Stainless steel Poor, medium,
LBAM AE signal 79%, [96]
decision process 316L high quality
Reinforcement 82%
learning Constrained No
Markov LBAM Not mentioned Defect states Optical image specic [97]
decision process accuracy

549 Reinforcement learning (RL) is one of the three basic ML paradigms, alongside supervised learning and
550 unsupervised learning. It is an area of ML, which is concerned with how agents ought to take actions in a specific
551 environment to get the maximum cumulative reward. Therefore, RL also has been used as an approach for defect
552 detection in LBAM, as shown in Table 4. For example, Wasmer et al. integrated RL with acoustic data obtained
553 from the printing process by acoustic emission to build quality monitoring in LBAM [96]. The classification accuracy
554 obtained from a cuboid shape manufactured with stainless steel 316L shows that this RL approach had a high potential
555 to be conducted in situ and in real-time. Furthermore, Yao et al. proposed a sensor-based product quality control model
556 for the LBAM process through the constrained Markov decision process [97]. The experimental result shows that the
557 proposed constrained Markov decision process provided an efficient policy for taking the right actions to remedy and
558 mitigate incipient defects before the whole part is completed. Fig. 14 shows the flow diagram of the in situ control
559 sequential decision-making framework for the LBAM process through the constrained Markov decision process.
Figure 14: Flow diagram of the sequential framework of constrained Markov decision process to control the quality of
LBAM builds [97].
560 4. Outlook
561 An appropriate and effective defect detection system is a key driver for the development of next-generation
562 technologies in LBAM. Their implementation will continue to increase the quality, efficiency, consistency, and
563 sustainability of metal components manufactured using the LBAM processes [98, 99]. Future development for LBAM
564 defect detection will likely be in the following four areas.
565 1) Integration of defect detection and product quality: The ultimate goal is to ensure the final product quality
566 regardless of which algorithm is utilized for defect detection. In other words, the defect-to-property is the core for
567 defect detection in LBAM. The defect detection should focus on detecting and evaluating the defect impact rather than
568 just finding the defect. The decision-making system is also a critical part after defect impacting evaluation. Based on
569 the evaluation result, the impacting defect level can be divided into negligible, marginal, and impactful. The printing
570 process needs to be canceled to save energy, material, and time when an impactful defect occurs. For the decision-
571 making system, the decision boundary from the machine learning algorithm has promising potential when combined
572 with the defect impacting together.
573 2) Data fusion and advanced algorithms: Signals from the different processes vary significantly in terms of temporal
574 and spatial resolution. Sensors with different accuracies and resolutions need to be synchronized to handle different
575 characteristics of the signal features, which provide a holistic understanding of the printing process. The improvement
576 for the future advanced algorithms has two sides. (1) Fusion algorithms include in-process signal processing and
577 defect detection. Most of the existing machine learning algorithms can only handle one type of input data. However,

578 the defect has many influence factors and performance characteristics. Just detecting defects from only one dimension
579 is not comprehensive and accurate. The advanced algorithm should have the ability to handle multi-type input at
580 the same time. Hybrid machine learning and deep learning will play an increasingly important role in this field.
581 (2) The development of the latest algorithms and their application. Hundreds and thousands of new algorithms are
582 developed nowadays with new features and abilities. This development will significantly broaden the defect detection
583 perspective in LBAM. For example, physics-informed ML will enable the integration of most recent understanding of
584 defect formation mechanisms to reduce the data requirement and increase the detection accuracy [100].
585 3) Real-time feedback control and correction system that couples sensor data, computational models with advanced
586 algorithms: The ways to mitigate the defects impacting can be done by correcting them when the defects are
587 initially detected or when the defects are predicted before they occur. Therefore, real-time feedback control and
588 correction systems and the predictive capability of the algorithms are essential, which should fuse various sensor data,
589 computational models with advanced algorithms together to achieve this goal. Reconstruct 3D objects by the machine
590 learning algorithm based on the different input signals, where physics-based computational models can supplement the
591 experimental data with future status. Then mapping the detected or predicted defects to the printing code. According
592 to the defect information, correct the correspondent code (e.g., laser speed, laser power, and the axis movement) to
593 mitigate the next printing flaws. This may be a feasible way, which has been proven on the other printing technology
594 [101]. In the foreseeable future, the LBAM system will be more widely applied by adding an effective defect detection
595 and feedback system to mitigate printing flaws [102, 103].
596 4) More compatible system: Active work has been performed in defect detection for LBAM; however, the algorithm
597 is mostly specific to one defect and limited to one specific printer. As the development of a new detection system for
598 another printing technology or another printer, even another material is time and effort consuming. Thus, a more unified
599 and compatible framework that flexibly integrates available sensors, algorithms, and computational modeling can be
600 transferred to the other system is important, which certainly will accelerate technology advancement. For example,
601 transfer learning [104], which can boost the speed and accuracy of the training process, will play a significant role in
602 compatible system development.
603 5. Conclusion
604 This paper summarizes recent research focused on the machine learning (ML) algorithms used in defect detection
605 systems for the metal laser-based additive manufacturing (LBAM) processes. The comprehensive and exhaustive
606 information listed in the paper provides a reference to help readers choose a suitable ML algorithm for detecting
607 defects based on different printing technologies, material types, defect types, and data types.
608 Choosing a suitable ML algorithm is critical to achieving the appropriate level of defect detection in any LBAM
609 system. A summary of commonly chosen ML algorithms is presented in this work. Convolutional neural networks
610 (CNN) are the first choice when dealing with image data due to their unique advantages of processing image features.
611 Using powder bed images, product layerwise optical images, or melt pool images, CNNs can be utilized for detecting
612 various defects or product quality levels. Support vector machines using the alternative kernel function are a good
613 choice for classification and are capable of handling both sensor signal and image data to perform both binary and
614 multi-class classification. K-means clustering is a widely utilized algorithm in unsupervised and semi-supervised ML
615 to partition the observations into different defect clusters. The Markov decision process is the most common algorithm
616 for reinforcement learning, which performs well on both sensor signal and image data to detect defects and to evaluate
617 product quality.
618 From the listed algorithms utilized in defect detection for LBAM processes, supervised machine learning
619 algorithms are the most common. However, as supervised machine learning requires that all data be labeled it is
620 time and effort-consuming. Algorithms that leverage unsupervised and semi-supervised learning have started to gain
621 traction in the field and have demonstrated great potential to expand the field of fault detection in LBAM processes.
622 Moreover, reinforcement learning is also being explored for defect detection in the LBAM process and offers the
623 potential to develop accurate and highly efficient fault detection models.
624 While machine learning has its unique advantages; the application prerequisite is rigid. The small databases
625 typically available for manufacturing systems usually results in over fit models that lead to poor fault detection accuracy.
626 A considerable barrier for the real-world adoption of machine learning in defect detection of LBAM is the lack of a
627 comprehensive, precise, and accessible databases for different materials, designs, and printing processes. A solution to
628 this challenge is the development of a standard database with uniform design and fault criteria. Such a database would

629 provide a large volume of accessible data for researchers. By leveraging transfer learning off of a massive database,
630 reliable fault detection methodologies could be developed without the need of generating large specific datasets for
631 each part or process.
632 It is worth pointing out the disproportionality in machine learning approaches for LBAM defect detection. Most
633 of the literature’s machine learning applications focus on directed energy deposition and laser powder bed fusion; no
634 research has been done on laser-based wire-feed systems.
635 Acknowledgements
636 This material is based upon work supported by the South Carolina Research Authority under grant 30228 and the
637 South Carolina Space Grant Consortium under grant 521179-RP-SC007. The support of these agencies is gratefully
638 acknowledged. Any opinions, findings, and conclusions, or recommendations expressed in this material are those of
639 the authors and do not necessarily reflect the views of the South Carolina Research Authority or the South Carolina
640 Space Grant Consortium. Distribution Statement A. Approved for public release: distribution unlimited.

641 References
642 [1] Jingchao Jiang, Yi Xiong, Zhiyuan Zhang, and David W Rosen. Machine learning integrated design for additive manufacturing. Journal of
643 Intelligent Manufacturing, pages 1–14, 2020.
644 [2] Jingchao Jiang, Stephen T Newman, and Ray Y Zhong. A review of multiple degrees of freedom for additive manufacturing machines.
645 International Journal of Computer Integrated Manufacturing, 34(2):195–211, 2021.
646 [3] Sunpreet Singh, Gurminder Singh, Chander Prakash, and Seeram Ramakrishna. Current status and future directions of fused filament
647 fabrication. Journal of Manufacturing Processes, 55:288–306, 2020.
648 [4] Annamaria Gisario, Michele Kazarian, Filomeno Martina, and Mehrshad Mehrpouya. Metal additive manufacturing in the commercial
649 aviation industry: A review. Journal of Manufacturing Systems, 53:124–149, 2019.
650 [5] Wohlers TT. Wohlers report 2021: 3D printing and additive manufacturing state of the industry. Wohlers Associates, 2021.
651 [6] Tarasankar DebRoy, HL Wei, JS Zuback, T Mukherjee, JW Elmer, JO Milewski, Allison Michelle Beese, A de Wilson-Heid, A De, and
652 W Zhang. Additive manufacturing of metallic components–process, structure and properties. Progress in Materials Science, 92:112–224,
653 2018.
654 [7] Sameer Mittal, Muztoba Ahmad Khan, David Romero, and Thorsten Wuest. A critical review of smart manufacturing & industry 4.0 maturity
655 models: Implications for small and medium-sized enterprises (smes). Journal of manufacturing systems, 49:194–214, 2018.
656 [8] Naveed Ur Rahman, David Thomas Allan Matthews, Matthijn De Rooij, Amir Mahyar Khorasani, Ian Gibson, Laura Cordova, and Gert-
657 willem Römer. An overview: Laser-based additive manufacturing for high temperature tribology. Frontiers in Mechanical Engineering,
658 5:16, 2019.
659 [9] Fei Tao, Qinglin Qi, Ang Liu, and Andrew Kusiak. Data-driven smart manufacturing. Journal of Manufacturing Systems, 48:157–169, 2018.
660 [10] Michael Sharp, Ronay Ak, and Thomas Hedberg Jr. A survey of the advancing use and development of machine learning in smart
661 manufacturing. Journal of manufacturing systems, 48:170–179, 2018.
662 [11] Onur Avci, Osama Abdeljaber, Serkan Kiranyaz, Mohammed Hussein, Moncef Gabbouj, and Daniel J Inman. A review of vibration-based
663 damage detection in civil structures: From traditional methods to machine learning and deep learning applications. Mechanical Systems and
664 Signal Processing, 147:107077, 2021.
665 [12] Rakesh Kumar, Manoj Kumar, and Jasgurpreet Singh Chohan. The role of additive manufacturing for biomedical applications: A critical
666 review. Journal of Manufacturing Processes, 64:828–850, 2021.
667 [13] Jinjiang Wang, Yulin Ma, Laibin Zhang, Robert X Gao, and Dazhong Wu. Deep learning for smart manufacturing: Methods and applications.
668 Journal of Manufacturing Systems, 48:144–156, 2018.
669 [14] Lingbin Meng, Brandon McWilliams, William Jarosinski, Hye-Yeong Park, Yeon-Gil Jung, Jehyun Lee, and Jing Zhang. Machine learning
670 in additive manufacturing: A review. JOM, pages 1–15, 2020.
671 [15] Jingchao Jiang, Fei Weng, Shiming Gao, Jonathan Stringer, Xun Xu, and Ping Guo. A support interface method for easy part removal in
672 directed energy deposition. Manufacturing Letters, 20:30–33, 2019.
673 [16] Shaleen Bhatia. Effect of machine positional errors on geometric tolerances in additive manufacturing. PhD thesis, University of Cincinnati,
674 2014.
675 [17] HH Zhu, L Lu, and JYH Fuh. Study on shrinkage behaviour of direct laser sintering metallic powder. Proceedings of the Institution of
676 Mechanical Engineers, Part B: Journal of Engineering Manufacture, 220(2):183–190, 2006.
677 [18] Y Ning, YS Wong, Jerry YH Fuh, and Han Tong Loh. An approach to minimize build errors in direct metal laser sintering. IEEE Transactions
678 on automation science and engineering, 3(1):73–80, 2006.
679 [19] Ratnadeep Paul. Modeling and optimization of powder based additive manufacturing (AM) processes. PhD thesis, University of Cincinnati,
680 2013.
681 [20] Milad Hamidi Nasab, Dario Gastaldi, Nora Francesca Lecis, and Maurizio Vedani. On morphological surface features of the parts printed
682 by selective laser melting (slm). Additive Manufacturing, 24:373–377, 2018.
683 [21] Andrew Townsend, N Senin, Liam Blunt, RK Leach, and JS Taylor. Surface texture metrology for metal additive manufacturing: a review.
684 Precision Engineering, 46:34–47, 2016.
685 [22] Kwai S Chan, Marie Koike, Robert L Mason, and Toru Okabe. Fatigue life of titanium alloys fabricated by additive layer manufacturing
686 techniques for dental implants. Metallurgical and Materials Transactions A, 44(2):1010–1022, 2013.
687 [23] Nesma T Aboulkhair, Nicola M Everitt, Ian Ashcroft, and Chris Tuck. Reducing porosity in alsi10mg parts processed by selective laser
688 melting. Additive Manufacturing, 1:77–86, 2014.
689 [24] Lang Yuan. Solidification defects in additive manufactured materials. JOM, 71(9):3221–3222, 2019.
690 [25] Lore Thijs, Frederik Verhaeghe, Tom Craeghs, Jan Van Humbeeck, and Jean-Pierre Kruth. A study of the microstructural evolution during
691 selective laser melting of ti–6al–4v. Acta materialia, 58(9):3303–3312, 2010.
692 [26] Thomas Vilaro, Christophe Colin, and Jean-Dominique Bartout. As-fabricated and heat-treated microstructures of the ti-6al-4v alloy
693 processed by selective laser melting. Metallurgical and materials transactions A, 42(10):3190–3199, 2011.
694 [27] Haijun Gong, Khalid Rafi, Hengfeng Gu, Thomas Starr, and Brent Stucker. Analysis of defect generation in ti–6al–4v parts made using
695 powder bed fusion additive manufacturing processes. Additive Manufacturing, 1:87–98, 2014.
696 [28] Thomas Vilaro, Christophe Colin, and Jean-Dominique Bartout. As-fabricated and heat-treated microstructures of the ti-6al-4v alloy
697 processed by selective laser melting. Metallurgical and materials transactions A, 42(10):3190–3199, 2011.
698 [29] Qian Chu Liu, Joe Elambasseril, Shou Jin Sun, Martin Leary, Milan Brandt, and Peter Khan Sharp. The effect of manufacturing defects on
699 the fatigue behaviour of ti-6al-4v specimens fabricated using selective laser melting. In Advanced Materials Research, volume 891, pages
700 1519–1524. Trans Tech Publ, 2014.
701 [30] Xin Zhou, Dianzheng Wang, Xihe Liu, DanDan Zhang, Shilian Qu, Jing Ma, Gary London, Zhijian Shen, and Wei Liu. 3d-imaging of
702 selective laser melting defects in a co–cr–mo alloy by synchrotron radiation micro-ct. Acta Materialia, 98:1–16, 2015.

703 [31] Adrian S Sabau, Lang Yuan, Narendran Raghavan, Matthew Bement, Srdjan Simunovic, John A Turner, and Vipul K Gupta. Fluid dynamics
704 effects on microstructure prediction in single-laser tracks for additive manufacturing of in625. Metallurgical and Materials Transactions B,
705 pages 1–19, 2020.
706 [32] Dongdong Gu, Yves-Christian Hagedorn, Wilhelm Meiners, Guangbin Meng, Rui João Santos Batista, Konrad Wissenbach, and Reinhart
707 Poprawe. Densification behavior, microstructure evolution, and wear performance of selective laser melting processed commercially pure
708 titanium. Acta Materialia, 60(9):3849–3860, 2012.
709 [33] Sheng Zhang, Rui-zhi Gui, QS Wei, and Y Shi. Cracking behavior and formation mechanism of tc4 alloy formed by selective laser melting.
710 Journal of Mechanical Engineering, 49(23):21–27, 2013.
711 [34] Mohamad Mahmoudi, Ahmed Aziz Ezzat, and Alaa Elwany. Layerwise anomaly detection in laser powder-bed fusion metal additive
712 manufacturing. Journal of Manufacturing Science and Engineering, 141(3), 2019.
713 [35] Haythem Gaja and Frank Liou. Defect classification of laser metal deposition using logistic regression and artificial neural networks for
714 pattern recognition. The International Journal of Advanced Manufacturing Technology, 94(1-4):315–326, 2018.
715 [36] G Tapia, AH Elwany, and H Sang. Prediction of porosity in metal-based additive manufacturing using spatial gaussian process models.
716 Additive Manufacturing, 12:282–290, 2016.
717 [37] DS Ye, YHJ Fuh, YJ Zhang, GS Hong, and KP Zhu. Defects recognition in selective laser melting with acoustic signals by svm based on
718 feature reduction. MS&E, 436(1):012020, 2018.
719 [38] Mojtaba Khanzadeh, Sudipta Chowdhury, Mohammad Marufuzzaman, Mark A Tschopp, and Linkan Bian. Porosity prediction: Supervised-
720 learning of thermal history for direct laser deposition. Journal of manufacturing systems, 47:69–82, 2018.
721 [39] Seyyed Hadi Seifi, Wenmeng Tian, Haley Doude, Mark A Tschopp, and Linkan Bian. Layer-wise modeling and anomaly detection for
722 laser-based additive manufacturing. Journal of Manufacturing Science and Engineering, 141(8), 2019.
723 [40] Luke Scime and Jack Beuth. Using machine learning to identify in-situ melt pool signatures indicative of flaw formation in a laser powder
724 bed fusion additive manufacturing process. Additive Manufacturing, 25:151–165, 2019.
725 [41] Jan Petrich, Christian Gobert, Shashi Phoha, Abdalla R Nassar, and Edward W Reutzel. ‘machine learning for defect detection for pbfam using
726 high resolution layerwise imaging coupled with post-build ct scans. In Proceedings of the 27th international Solid Freeform Fabrication
727 Symposium, 2017.
728 [42] Farhad Imani, Aniruddha Gaikwad, Mohammad Montazeri, Prahalada Rao, Hui Yang, and Edward Reutzel. Process mapping and in-process
729 monitoring of porosity in laser powder bed fusion using layerwise optical imaging. Journal of Manufacturing Science and Engineering,
730 140(10), 2018.
731 [43] Yingjie Zhang, Geok Soon Hong, Dongsen Ye, Kunpeng Zhu, and Jerry YH Fuh. Extraction and evaluation of melt pool, plume and spatter
732 information for powder-bed fusion am process monitoring. Materials & Design, 156:458–469, 2018.
733 [44] Christian Gobert, Edward W Reutzel, Jan Petrich, Abdalla R Nassar, and Shashi Phoha. Application of supervised machine learning for defect
734 detection during metallic powder bed fusion additive manufacturing using high resolution imaging. Additive Manufacturing, 21:517–528,
735 2018.
736 [45] ZL Lu, DC Li, BH Lu, AF Zhang, GX Zhu, and G Pi. The prediction of the building precision in the laser engineered net shaping process
737 using advanced networks. Optics and Lasers in Engineering, 48(5):519–525, 2010.
738 [46] Masoumeh Aminzadeh and Thomas R Kurfess. Online quality inspection using bayesian classification in powder-bed additive manufacturing
739 from high-resolution visual camera images. Journal of Intelligent Manufacturing, 30(6):2505–2523, 2019.
740 [47] Jamison L Bartlett, Alex Jarama, Jonaaron Jones, and Xiaodong Li. Prediction of microstructural defects in additive manufacturing from
741 powder bed quality using digital image correlation. Materials Science and Engineering: A, 794:140002, 2020.
742 [48] Mohammad Montazeri, Abdalla R Nassar, Alexander J Dunbar, and Prahalada Rao. In-process monitoring of porosity in additive
743 manufacturing using optical emission spectroscopy. IISE Transactions, 52(5):500–515, 2020.
744 [49] Lequn Chen, Xiling Yao, Peng Xu, Seung Ki Moon, and Guijun Bi. Rapid surface defect identification for additive manufacturing with in-situ
745 point cloud processing and machine learning. Virtual and Physical Prototyping, pages 1–18, 2020.
746 [50] Angel-Iván García-Moreno, Juan-Manuel Alvarado-Orozco, Juansethi Ibarra-Medina, and Enrique Martínez-Franco. Image-based porosity
747 classification in al-alloys by laser metal deposition using random forests. The International Journal of Advanced Manufacturing Technology,
748 110(9):2827–2845, 2020.
749 [51] Tuğrul Özel, Ayça Altay, Bilgin Kaftanoğlu, Richard Leach, Nicola Senin, and Alkan Donmez. Focus variation measurement and prediction
750 of surface texture parameters using machine learning in laser powder bed fusion. Journal of Manufacturing Science and Engineering, 142(1),
751 2020.
752 [52] Aniruddha Gaikwad, Brian Giera, Gabriel M Guss, Jean-Baptiste Forien, Manyalibo J Matthews, and Prahalada Rao. Heterogeneous sensing
753 and scientific machine learning for quality assurance in laser powder bed fusion–a single-track study. Additive Manufacturing, 36:101659,
754 2020.
755 [53] Farhad Imani, Ruimin Chen, Evan Diewald, Edward Reutzel, and Hui Yang. Deep learning of variant geometry in layerwise imaging profiles
756 for additive manufacturing quality control. Journal of Manufacturing Science and Engineering, 141(11), 2019.
757 [54] Reza Mojahed Yazdi, Farhad Imani, and Hui Yang. A hybrid deep learning model of process-build interactions in additive manufacturing.
758 Journal of Manufacturing Systems, 57:460–468, 2020.
759 [55] Wenyuan Cui, Yunlu Zhang, Xinchang Zhang, Lan Li, and Frank Liou. Metal additive manufacturing parts inspection using convolutional
760 neural network. Applied Sciences, 10(2):545, 2020.
761 [56] Aniruddha Gaikwad, Farhad Imani, Hui Yang, Edward Reutzel, and Prahalada Rao. In situ monitoring of thin-wall build quality in laser
762 powder bed fusion using deep learning. 2019.
763 [57] Maximilian Hugo Kunkel, Andreas Gebhardt, Khumbulani Mpofu, and Stephan Kallweit. Quality assurance in metal powder bed fusion via
764 deep-learning-based image classification. Rapid Prototyping Journal, 2019.

765 [58] Bodi Yuan, Gabriel M Guss, Aaron C Wilson, Stefan P Hau-Riege, Phillip J DePond, Sara McMains, Manyalibo J Matthews, and Brian
766 Giera. Machine-learning-based monitoring of laser powder bed fusion. Advanced Materials Technologies, 3(12):1800136, 2018.
767 [59] Binbin Zhang, Prakhar Jaiswal, Rahul Rai, Paul Guerrier, and George Baggs. Convolutional neural network-based inspection of metal additive
768 manufacturing parts. Rapid Prototyping Journal, 2019.
769 [60] Carlos Gonzalez-Val, Adrian Pallas, Veronica Panadeiro, and Alvaro Rodriguez. A convolutional approach to quality monitoring for laser
770 manufacturing. Journal of Intelligent Manufacturing, 31(3):789–795, 2020.
771 [61] Bin Zhang, Shunyu Liu, and Yung C Shin. In-process monitoring of porosity during laser additive manufacturing process. Additive
772 Manufacturing, 28:497–505, 2019.
773 [62] Qi Tian, Shenghan Guo, Yuebin Guo, et al. A physics-driven deep learning model for process-porosity causal relationship and porosity
774 prediction with interpretability in laser metal deposition. CIRP Annals, 2020.
775 [63] Xiang Li, Shahin Siahpour, Jay Lee, Yachao Wang, and Jing Shi. Deep learning-based intelligent process monitoring of directed energy
776 deposition in additive manufacturing with thermal images. Procedia Manufacturing, 48:643–649, 2020.
777 [64] Luke Scime, Derek Siddel, Seth Baird, and Vincent Paquit. Layer-wise anomaly detection and classification for powder bed additive
778 manufacturing processes: A machine-agnostic algorithm for real-time pixel-wise semantic segmentation. Additive Manufacturing, page
779 101453, 2020.
780 [65] Ling Xiao, Mingyuan Lu, and Han Huang. Detection of powder bed defects in selective laser sintering using convolutional neural network.
781 The International Journal of Advanced Manufacturing Technology, pages 1–12, 2020.
782 [66] Luke Scime and Jack Beuth. A multi-scale convolutional neural network for autonomous anomaly detection and classification in a laser
783 powder bed fusion additive manufacturing process. Additive Manufacturing, 24:273–286, 2018.
784 [67] Yingjie Zhang, Hong Geok Soon, Dongsen Ye, Jerry Ying Hsi Fuh, and Kunpeng Zhu. Powder-bed fusion process monitoring by machine
785 vision with hybrid convolutional neural networks. IEEE Transactions on Industrial Informatics, 16(9):5769–5779, 2019.
786 [68] Jack Francis and Linkan Bian. Deep learning for distortion prediction in laser-based additive manufacturing using big data. Manufacturing
787 Letters, 20:10–14, 2019.
788 [69] Alessandra Caggiano, Jianjing Zhang, Vittorio Alfieri, Fabrizia Caiazzo, Robert Gao, and Roberto Teti. Machine learning-based image
789 processing for on-line defect recognition in additive manufacturing. CIRP Annals, 68(1):451–454, 2019.
790 [70] Feng Han, Jingling Zou, Yuan Ai, Chunlin Xu, and Sheng Liu. Image classification and analysis during the additive manufacturing process
791 based on deep convolutional neural networks. In 2019 20th International Conference on Electronic Packaging Technology (ICEPT), pages
792 1–4. IEEE, 2019.
793 [71] Jacob Williams, Paul Dryburgh, Adam Clare, Prahalada Rao, and Ashok Samal. Defect detection and monitoring in metal additive
794 manufactured parts through deep learning of spatially resolved acoustic spectroscopy signals. 2018.
795 [72] Yuemeng Li, Hairong Yan, and Yuefei Zhang. A deep learning method for material performance recognition in laser additive manufacturing.
796 In 2019 IEEE 17th International Conference on Industrial Informatics (INDIN), volume 1, pages 1735–1740. IEEE, 2019.
797 [73] Hermann Baumgartl, Josef Tomas, Ricardo Buettner, and Markus Merkel. A deep learning-based model for defect detection in laser-powder
798 bed fusion using in-situ thermographic monitoring. Progress in Additive Manufacturing, pages 1–9, 2020.
799 [74] Sergey A Shevchik, Christoph Kenel, Christian Leinenbach, and Kilian Wasmer. Acoustic emission for in situ quality monitoring in additive
800 manufacturing using spectral convolutional neural networks. Additive Manufacturing, 21:598–604, 2018.
801 [75] Qi Tian, Shenghan Guo, Erika Melder, Linkan Bian, Weihong Guo, et al. Deep learning-based data fusion method for in situ porosity
802 detection in laser-based additive manufacturing. Journal of Manufacturing Science and Engineering, 143(4), 2021.
803 [76] Nathan Hertlein, Sourabh Deshpande, Vysakh Venugopal, Manish Kumar, and Sam Anand. Prediction of selective laser melting part quality
804 using hybrid bayesian network. Additive Manufacturing, 32:101089, 2020.
805 [77] Ohyung Kwon, Hyung Giun Kim, Min Ji Ham, Wonrae Kim, Gun-Hee Kim, Jae-Hyung Cho, Nam Il Kim, and Kangil Kim. A deep neural
806 network for classification of melt-pool images in metal additive manufacturing. Journal of Intelligent Manufacturing, 31(2):375–386, 2020.
807 [78] Mohammad Ghayoomi Mohammadi, Dalia Mahmoud, and Mohamed Elbestawi. On the application of machine learning for defect detection
808 in l-pbf additive manufacturing. Optics & Laser Technology, 143:107338, 2021.
809 [79] Rasheedat M Mahamood, Esther T Akinlabi, and Moses G Owolabi. Effect of laser power and powder flow rate on dilution rate and
810 surface finish produced during laser metal deposition of titanium alloy. In 2017 8th International Conference on Mechanical and Intelligent
811 Manufacturing Technologies (ICMIMT), pages 6–10. IEEE, 2017.
812 [80] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE
813 conference on computer vision and pattern recognition, pages 770–778, 2016.
814 [81] Sergey A Shevchik, Giulio Masinelli, Christoph Kenel, Christian Leinenbach, and Kilian Wasmer. Deep learning for in situ and real-time
815 quality monitoring in additive manufacturing using acoustic emission. IEEE Transactions on Industrial Informatics, 15(9):5194–5203, 2019.
816 [82] Luke Scime and Jack Beuth. Anomaly detection and classification in a laser powder bed additive manufacturing process using a trained
817 computer vision algorithm. Additive Manufacturing, 19:114–126, 2018.
818 [83] Robert Snell, Sam Tammas-Williams, Lova Chechik, Alistair Lyle, Everth Hernández-Nava, Charlotte Boig, George Panoutsos, and Iain
819 Todd. Methods for rapid pore classification in metal additive manufacturing. JOM, 72(1):101–109, 2020.
820 [84] Marco Grasso, Vittorio Laguzza, Quirico Semeraro, and Bianca Maria Colosimo. In-process monitoring of selective laser melting: spatial
821 detection of defects via image data analysis. Journal of Manufacturing Science and Engineering, 139(5), 2017.
822 [85] Pinku Yadav, Vibhutesh Kumar Singh, Thomas Joffre, Olivier Rigo, Corinne Arvieu, Emilie Le Guen, and Eric Lacoste. Inline drift detection
823 using monitoring systems and machine learning in selective laser melting. Advanced Engineering Materials, page 2000660, 2020.
824 [86] Haythem Gaja and Frank Liou. Defects monitoring of laser metal deposition using acoustic emission sensor. The International Journal of
825 Advanced Manufacturing Technology, 90(1-4):561–574, 2017.
826 [87] Hossein Taheri, Lucas W Koester, Timothy A Bigelow, Eric J Faierson, and Leonard J Bond. In situ additive manufacturing process
827 monitoring with an acoustic technique: clustering performance evaluation using k-means algorithm. Journal of Manufacturing Science

828 and Engineering, 141(4), 2019.

829 [88] Wenjing Ren, Guangrui Wen, Zhifen Zhang, and Jyoti Mazumder. Quality monitoring in additive manufacturing using emission spectroscopy
830 and unsupervised deep learning. Materials and Manufacturing Processes, pages 1–8, 2021.
831 [89] Dongsen Ye, Geok Soon Hong, Yingjie Zhang, Kunpeng Zhu, and Jerry Ying Hsi Fuh. Defect detection in selective laser melting technology
832 by acoustic signals with deep belief networks. The International Journal of Advanced Manufacturing Technology, 96(5-8):2791–2801, 2018.
833 [90] Dongsen Ye, Jerry Ying Hsi Fuh, Yingjie Zhang, Geok Soon Hong, and Kunpeng Zhu. In situ monitoring of selective laser melting using
834 plume and spatter signatures by deep belief networks. ISA transactions, 81:96–104, 2018.
835 [91] Angel-Iván García-Moreno. Automatic quantification of porosity using an intelligent classifier. The International Journal of Advanced
836 Manufacturing Technology, 105(5):1883–1899, 2019.
837 [92] Mojtaba Khanzadeh, Sudipta Chowdhury, Mark A Tschopp, Haley R Doude, Mohammad Marufuzzaman, and Linkan Bian. In-situ
838 monitoring of melt pool images for porosity prediction in directed energy deposition processes. IISE Transactions, 51(5):437–455, 2019.
839 [93] Ikenna A Okaro, Sarini Jayasinghe, Chris Sutcliffe, Kate Black, Paolo Paoletti, and Peter L Green. Automatic fault detection for laser
840 powder-bed fusion using semi-supervised machine learning. Additive Manufacturing, 27:42–53, 2019.
841 [94] Bodi Yuan, Brian Giera, Gabe Guss, Ibo Matthews, and Sara McMains. Semi-supervised convolutional neural networks for in-situ video
842 monitoring of selective laser melting. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 744–753. IEEE,
843 2019.
844 [95] Xiang Li, Xiaodong Jia, Qibo Yang, and Jay Lee. Quality analysis in metal additive manufacturing with deep learning. Journal of Intelligent
845 Manufacturing, pages 1–15, 2020.
846 [96] K Wasmer, T Le-Quang, B Meylan, and SA Shevchik. In situ quality monitoring in am using acoustic emission: A reinforcement learning
847 approach. Journal of Materials Engineering and Performance, 28(2):666–672, 2019.
848 [97] Bing Yao and Hui Yang. Constrained markov decision process modeling for sequential optimization of additive manufacturing build quality.
849 IEEE Access, 6:54786–54794, 2018.
850 [98] Chengcheng Wang, XP Tan, SB Tor, and CS Lim. Machine learning in additive manufacturing: State-of-the-art and perspectives. Additive
851 Manufacturing, page 101538, 2020.
852 [99] Sayyeda Saadia Razvi, Shaw Feng, Anantha Narayanan, Yung-Tsun Tina Lee, and Paul Witherell. A review of machine learning applications
853 in additive manufacturing. In International Design Engineering Technical Conferences and Computers and Information in Engineering
854 Conference, volume 59179, page V001T02A040. American Society of Mechanical Engineers, 2019.
855 [100] George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Paris Perdikaris, Sifan Wang, and Liu Yang. Physics-informed machine learning.
856 Nature Reviews Physics, 3(6):422–440, 2021.
857 [101] Mohammad Abdullah Al Faruque, Sujit Rokka Chhetri, Arquimedes Canedo, and Jiang Wan. Acoustic side-channel attacks on additive
858 manufacturing systems. In 2016 ACM/IEEE 7th international conference on Cyber-Physical Systems (ICCPS), pages 1–10. IEEE, 2016.
859 [102] Michael Molitch-Hou. Overview of additive manufacturing process. In Additive Manufacturing, pages 1–38. Elsevier, 2018.
860 [103] Michael Schmidt, Marion Merklein, David Bourell, Dimitri Dimitrov, Tino Hausotte, Konrad Wegener, Ludger Overmeyer, Frank Vollertsen,
861 and Gideon N Levy. Laser based additive manufacturing in industry and academia. Cirp Annals, 66(2):561–583, 2017.
862 [104] Jason Yosinski, Jeff Clune, Yoshua Bengio, and Hod Lipson. How transferable are features in deep neural networks? arXiv preprint
863 arXiv:1411.1792, 2014.

Machine Learning Algorithms For Defect Detection in Metal Laser-Based Additive Manufacturing: A Review

Uploaded by

Copyright:

Available Formats

Machine Learning Algorithms For Defect Detection in Metal Laser-Based Additive Manufacturing: A Review

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Machine Learning Algorithms For Defect Detection in Metal Laser-Based Additive Manufacturing: A Review

Uploaded by

Copyright:

Available Formats

Version of Record: https://www.sciencedirect.

1 Machine learning algorithms for defect detection in metal

109 ARTICLE INFO ABSTRACT

Y. Fu et al.: Preprint submitted to Elsevier Page 1 of 31

41 KMC K-means clustering

44 LDA linear discriminant analysis

58 RNN recurrent neural network

Y. Fu et al.: Preprint submitted to Elsevier Page 2 of 31

111 2. Defect types in LBAM

Y. Fu et al.: Preprint submitted to Elsevier Page 3 of 31

128 2.1. Geometric and dimensional defects

Y. Fu et al.: Preprint submitted to Elsevier Page 4 of 31

143 2.2. Porosity

Y. Fu et al.: Preprint submitted to Elsevier Page 5 of 31

158 2.3. Incomplete fusion

Y. Fu et al.: Preprint submitted to Elsevier Page 6 of 31

172 2.4. Cracks

Y. Fu et al.: Preprint submitted to Elsevier Page 7 of 31

182 3. Machine learning algorithms for defect detection

Y. Fu et al.: Preprint submitted to Elsevier Page 8 of 31

Machine learning algorithms for defect detection: A review

Machine learning algorithms for defect detection: A review

Good quality, crack, Digital

Machine learning algorithms for defect detection: A review

Machine learning algorithms for defect detection: A review

191 3.1. Supervised learning

206 3.1.1. Traditional regression

224 3.1.2. Traditional classifiers

Y. Fu et al.: Preprint submitted to Elsevier Page 13 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 14 of 31

300 3.1.3. Artificial neural network

Y. Fu et al.: Preprint submitted to Elsevier Page 15 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 16 of 31

349 3.1.4. Convolutional neural network

Y. Fu et al.: Preprint submitted to Elsevier Page 17 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 18 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 19 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 20 of 31

Machine learning algorithms for defect detection: A review

456 3.2. Unsupervised learning

464 3.2.1. K-means clustering algorithm

491 3.2.2. Deep belief network

Y. Fu et al.: Preprint submitted to Elsevier Page 22 of 31

503 3.2.3. Self-organized maps algorithm

514 3.3. Semi-supervised learning

Y. Fu et al.: Preprint submitted to Elsevier Page 23 of 31

Figure 13: Semi-supervised CNN architecture [94].

548 3.4. Reinforcement learning

Y. Fu et al.: Preprint submitted to Elsevier Page 24 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 25 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 26 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 27 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 28 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 29 of 31

Y. Fu et al.: Preprint submitted to Elsevier Page 30 of 31

828 and Engineering, 141(4), 2019.

Y. Fu et al.: Preprint submitted to Elsevier Page 31 of 31

You might also like