ArticlePDF Available

Machine Learning for Object Recognition in Manufacturing Applications


Abstract and Figures

Feature recognition and manufacturability analysis from computer-aided design (CAD) models are indispensable technologies for better decision making in manufacturing processes. It is important to transform the knowledge embedded within a CAD model to manufacturing instructions for companies to remain competitive as experienced baby-boomer experts are going to retire. Automatic feature recognition and computer-aided process planning have a long history in research, and recent developments regarding algorithms and computing power are bringing machine learning (ML) capability within reach of manufacturers. Feature recognition using ML has emerged as an alternative to conventional methods. This study reviews ML techniques to recognize objects, features, and construct process plans. It describes the potential for ML in object or feature recognition and offers insight into its implementation in various smart manufacturing applications. The study describes ML methods frequently used in manufacturing, with a brief introduction of underlying principles. After a review of conventional object recognition methods, the study discusses recent studies and outlooks on feature recognition and manufacturability analysis using ML.
Content may be subject to copyright.
International Journal of Precision Engineering and Manufacturing
1 3
Online ISSN 2005-4602
Print ISSN 2234-7593
Machine Learning forObject Recognition inManufacturing
HuitaekYun1· EunseobKim2· DongMinKim3 · HyungWookPark4· MartinByung‑GukJun1,2
Received: 28 March 2021 / Revised: 16 December 2022 / Accepted: 19 December 2022
© The Author(s) 2022
Feature recognition and manufacturability analysis from computer-aided design (CAD) models are indispensable technologies
for better decision making in manufacturing processes. It is important to transform the knowledge embedded within a CAD
model to manufacturing instructions for companies to remain competitive as experienced baby-boomer experts are going
to retire. Automatic feature recognition and computer-aided process planning have a long history in research, and recent
developments regarding algorithms and computing power are bringing machine learning (ML) capability within reach of
manufacturers. Feature recognition using ML has emerged as an alternative to conventional methods. This study reviews ML
techniques to recognize objects, features, and construct process plans. It describes the potential for ML in object or feature
recognition and offers insight into its implementation in various smart manufacturing applications. The study describes ML
methods frequently used in manufacturing, with a brief introduction of underlying principles. After a review of conventional
object recognition methods, the study discusses recent studies and outlooks on feature recognition and manufacturability
analysis using ML.
Keywords Machine learning (ML)· Manufacturability· Automated feature recognition (AFR)· Object recognition
1 Introduction
Cyber manufacturing is a new strategy for future manufac-
turing systems, which draws upon such recent technologies
as cloud computing, low-cost sensors, wireless communica-
tion, cyber-physical systems, machine learning (ML), and
mechanistic simulation and modeling [13]. The concept of
cyber manufacturing enables us to share information rapidly
among a manufacturer, suppliers, customers, and govern-
ments. Given this importance, several nations and compa-
nies have globally developed new manufacturing concepts
such as “Industry 4.0” by Germany, “Monozukuri” by Japan,
“Factories of the Future” by Europe, and “Industrial Inter-
net” by General Electric [4].
Due to the improved capability of big data in cyber man-
ufacturing, finding meaningful information from the data
(data mining) has drawn attention recently [57]. Accord-
ingly, applications of ML combined with big data have gen-
erated more profit in many industries [8, 9]. Thus, many
case-studies about ML applications in manufacturing fields
have emerged [10, 11]. For example, the tool wear predic-
tion model can be established by ML containing relation-
ships of complex parameters, which is difficult via model- or
physics-based predictive models [12]. Such predictive main-
tenance in ML improves machine intelligence. Moreover,
the capability of ML can be extended to automate conven-
tional decision-making procedures through artificial intelli-
gence, subject to the acceptance of manufacturers. Notably,
a candidate is planning a manufacturing process based on a
designer’s computer-aided design (CAD) model.
* Dong Min Kim
* Martin Byung-Guk Jun
1 Indiana Manufacturing Competitiveness Center (IN-MaC),
Purdue University, 1105 Endeavour Drive, WestLafayette,
IN47906, USA
2 School ofMechanical Engineering, Purdue University, 585
Purdue Mall, WestLafayette, IN47907, USA
3 Dongnam Regional Division, Korea Institute ofIndustrial
Technology, Jinju-si, Gyeongsangnam-do, RepublicofKorea
4 Department ofMechanical Engineering, Ulsan National
Institute ofScience andTechnology, UNIST-gil 50,
Eonyang-eup, Ulju-gun, Ulsan689-798, RepublicofKorea
International Journal of Precision Engineering and Manufacturing
1 3
The typical iterative process for production planning
is as follows. Designers ascertain the mechanical draw-
ings to meet the engineering specifications of the prod-
ucts. Manufacturers then verify the manufacturability of
the product design. Process planners draw flowcharts and
enlist required machines to minimize costs and maximize
productivity and quality while satisfying the specifica-
tions. If the plan is not satisfactory, the design or spec-
ifications is altered. Iterations of the feedback flow are
time-consuming, and the costs are high [13]. Furthermore,
the experience or skill of manufacturing personnel, espe-
cially those from the “baby boomer” generation, has been
indispensable regarding making manufacturing-related
decisions. However, such individuals will be retiring over
the next several decades, and their knowledge, know-how,
and experience will be lost from the workforce [4]. Thus,
strategies are required to replace this knowledge in the
cyber manufacturing framework. In cyber manufacturing,
cloud-based databases and big data may be accessed by
companies from across the design and manufacturing sup-
ply chain [14]. When the designer develops a new product
concept, cyber manufacturing may be used to determine
manufacturing strategies, production and process plans
[15], and logistics chains [16].
Among the mentioned steps, estimating manufacturability
from the drawings relies on human experience and know-
how. Several decades have gone into automating the pro-
cess of automated feature recognition (AFR). However, there
are numerous ways to recognize features and assign suitable
manufacturing processes. Moreover, model complexity by
interacting features hinders accurate estimation of manufac-
turability. Other than AFR, several tools have been proposed
to reduce the losses. Technical data package (TDP) [17]
is a technical description providing information from the
design to the production. However, dimensions, tolerances,
and product quality of a new conceptual design remain sub-
ject to substantial uncertainty [18]. Alternatively, design for
manufacturing (DFM) predicts the manufacturability before
accepting the production plans of the newly designed prod-
ucts. 80% of the avoidable costs in traditional production are
generated during the initial design stages DFM is a useful
tool to achieve lower costs for manufacturing new designs.
Design for additive manufacturing (DfAM) provide the
guide-line for product design for the additive manufactur-
ing process [19]. Furthermore, simulation method is intro-
duced to predict the surface accuracy of the manufacturing
process [20]. Another considering factor, the tolerance are
a a significant factor in deciding product quality, and that is
influenced by manufacturing process. Therefore, the know-
ing and tolerance information of manufacturing process is
import [21]. Therefore, designer still require manufacturing
knowledge considering which manufacturing process will be
used in their design. At the same time, AFR for manufactur-
ing becomes challenging as the model becomes complex
according to diversified demands from customers.
Thus, this study reviews the object recognition techniques
for the manufacturing of a CAD model via the utilization
of ML techniques. It covers the steps of feature recognition
techniques from the CAD model and estimating manufac-
turability before computer-aided process planning (CAPP).
Section2 briefly describes the theoretical background of
ML. Section3 shows the research opportunities for manu-
facturability analysis against the backdrop of ML techniques.
Section4 mentions traditional feature extraction techniques
from CAD data for manufacturability. Section5 describes
feature extraction methods from the CAD model that have
the high potential to be applied in manufacturability recogni-
tion via ML techniques. Section6 shows recent case studies.
Figure1 shows the research scope and brief history of the
feature extraction process for manufacturability.
2 A Brief Theoretical Background
ofMachine Learning Techniques
2.1 Introduction toMachine Learning
ML has a characteristic of self-improving performance
through learning progress. ML techniques have been
applied in manufacturing fields and various interdiscipli-
nary fields such as human pose estimation, object classifica-
tion, multiple object detection, and model segmentation and
The representative techniques of ML are supervised ML,
unsupervised-ML, and reinforcement ML. The supervised
neural-net defines the classification for each data [22]. For
instance, weight factors and thresholds are updated through
the neural-net when the pre-classified or labeled images
are fed to the neural network (NN). The trained NN then
classifies the new undefined images. Unsupervised ML is
the model where input data are fed without corresponding
output labels. The goal of the unsupervised ML is to find
meaningful relationships and hidden structures among the
data [22]. Some of the unsupervised learning techniques are
self-organizing maps, singular value decomposition, nearest-
neighbor mapping, and k-mean clustering. The reinforce-
ment model is a learning algorithm that obtains experiences
through action and reward. The representative reinforcement
learnings are Q-learning and Deep-Q-Network (DQN) [10].
The following section describes core ML techniques used in
object recognition for manufacturing.
International Journal of Precision Engineering and Manufacturing
1 3
2.2 Support‑Vector Machine (SVM)
A support-vector machine (SVM) is a traditional and
widely-used algorithm. SVM provides answers for distin-
guishing different status of interests by dividing a feature
space with decision boundaries. Vapnik first proposed
the linear classifier algorithm in 1963. Boser etal. [23]
improved the classifier for driving the decision boundaries
(known as the hyperplane) using the kernel trick, which
enables non-linear classification. Figure2 and Eq.(1)
describes a training dataset
points in bilinear
classification problems with two classes as
is kth input and
is the label. Equation(2)
describes the decision function
is the predefined function of
is a vector
orthogonal to the hyperplane, and
is a visa of the decision
function. From Eq.(1), the distance between the hyperplane
and the kth data point
is given Eq.(3) for margin
Therefore, the maximizing margin
yields the cor-
responding finding sector
. Further, this statement
results in the minimax problem, which is equivalent to a
quadratic problem [23]. Equation(4) is constrained with
Lagrangian induced the optimal solution without a
local-minimum problem [25]. As mentioned above, SVM
was initially designed for the linear classification problem.
However, mapping input data into a higher-dimensional
space can be applied to non-linear classifications using a
kernel trick, as shown in Fig.3.
X= {(x1,y1),(x2,y2),,(xn,yn)}
yk=1if xk
=−1if x
1980 ‛83‛84 ‛88‛92 ‛932010‛11 ‛15‛18
Fig. 1 The research scope for manufacturability recognition
Fig. 2 Hyperplane, samples, and a margin in 2D space within the lin-
ear case
International Journal of Precision Engineering and Manufacturing
1 3
2.3 Decision Tree
A decision tree is the concatenation of multiple classifi-
ers known as leaves and internal nodes. [26, 27] defined
the leaves, terminal nodes, or decision nodes without any
descendants. Each node divides feature space into multiple
subspaces by certain conditions in the decision tree algo-
rithm. Figure4 shows an example of the decision tree clas-
sifier and partitioned 2D space [27].
Furthermore, it is crucial to specify structural parameters
to improve the performance of the decision tree. The depth of
the tree, the order of features, or the number of nodes domi-
nate the calculation load and accuracy of the classification.
Several researchers proposed the optimization of the deci-
sion tree with variant parameters. The main target of those
optimizations is the structure of a tree. The iterative dichot-
omiser 3 (ID3) algorithm was emerged with this concept,
thus implementing optimization by changing structural
attributes (e.g., depth of the tree and number of nodes). This
optimization that changes the inner structure of the tree is
also called “greedy algorithm.” To enhance the performance
of the greedy algorithm, Olaru and Wehenkel [28] devel-
oped the soft decision tree (SDT) method using fuzzy logic.
The fuzzy logic-based method shows higher accuracy than
the ID3-based algorithm due to adaptively assigned fuzzi-
ness. However, the greedy algorithm suffers from overfitting
Fig. 3 A schematics of kernel trick for a polynomial classifier and b circular classifier
Fig. 4 A schematic of the
decision tree; a The decision
tree process; b The partitioned
feature space
International Journal of Precision Engineering and Manufacturing
1 3
and updating. Thus, to update the decision tree based on the
greedy algorithm with unexperienced data, the tree needs
to be optimized regarding structural parameters from the
beginning. However, it costs a load that is as heavy as the
first-time construction. Hence, Bennet [29] improved the
single-decision optimization method using global tree opti-
mization (GTO). It is a non-greedy algorithm that considers
overall decisions simultaneously. Basically, GTO starts with
an existing decision tree, and it minimizes the error rate only
by changing decisions, not the structural parameters of the
tree. In this aspect of leaving the structure of tree unchanged,
the benefit of GTO against the greedy algorithm is easy to
update when it faces unprecedented information. As another
approach of the non-greedy algorithm, Guo and Gelfand [30]
introduced an NN-based decision tree optimization. They
replaced the leaves with a multi-layer perceptron having the
structure of NN. The NN-based method showed better perfor-
mances with the decision tree by reducing the total number
of nodes, which termed as called pruning.
2.4 Artificial Neural Network (ANN)
An artificial neural network (ANN) works like a human
brain. Moreover, it has been applied to feature recogni-
tion since the 1990s. ANN is a large-scale interconnected
network of neurons, which have simple elements such as
an input layer, interconnected-neuron layers, and an out-
put layer (Fig.5a). The input layer obtains signals from
external sources. These external signals are passed through
the connected links between neurons; they then flow to
other neuron branches through the output layer (Fig.5b).
Each node is obtained from arithmetic operations, which
determine weights factors and numerical calculations dur-
ing the signals flow [31]. The ANN model updates them
via training from a dataset, and the model (after training)
predicts the output from the test inputs reasonably. Logical
rules are not used; only simple calculations are employed.
Therefore, it is faster than other NN methods. The math-
ematical function among the neuron networks can be
expressed as Eq.(5).
is the result through the neuron network,
is the
number of inputs,
is the weight factor attributed from ith
is the input information,
is the ANN’s parameters,
is the bias.
2.5 Convolutional Neural Network (CNN)
In 1998, LeCun etal. [32] proposed the CNN, which is
called LeNet-5. A modern CNN has progressed through
two steps called feature extraction and classification. Fig-
ure6 shows a schematic of CNN. Feature extraction lay-
ers recognize the features from input images and generate
“Feature map” in convolution layers and pooling layers.
A convolution layer (or kernel) is like an image filter that
extracts features from the imported input matrix. Arrays
of the 2D images are imported to CNN, and it is convolved
by filters to generate features maps. Equation(6) [33] rep-
resents the convolution below.
is an imported two-dimensional array,
is a two-
dimensional kernel array, and
is a feature map through
Fig. 5 A schematic of ANN; a The neuron representation in computation comparing to a human brain; b The conceptual configuration of artifi-
cial neuron networks (ANN)
International Journal of Precision Engineering and Manufacturing
1 3
According to the literature [34, 35], the use of convolution
has three main advantages. First, the feature map shares the
weight to reduce the variables. Second, the kernel extracts
correlations between the localized features. Third, the sig-
moid function as the activation function achieves scale invari-
ance. From the advantages, CNN is faster and more accurate
than other fully connected NN models [34, 36, 37].
The following is a pooling layer which reduces the dimen-
sions of feature maps. The pooling layer transforms images
invariantly and compresses the information. Max pool-
ing consists of the grid or pyramid pooling structure with
smoothing operation. The pooling layers provide several
estimates of the sample groups at the detail levels. The max
pooling method is widely used in CNN to improve perfor-
mance [38]. Max pooling is given in Eq.(7) as follows.
is the vector in the pooling dimensions, and
is a
pooling operation which translates a rectangular array to a
single scalar
. The pooling process obtains the maxi-
mum values in the rectangular dimension. For example, the
max pooling layer compresses 16 × 16 features maps to 8 × 8
dimensional arrays with strides of two.
The following approaches are well-known pooling layers:
stochastic, spatial pyramid, and Def. Stochastic pooling layer
arbitrarily selects the activations within each pool of neurons by
a multinomial distribution [39]. Max pooling is susceptible to
overfitting of the training data. However, it approves the slight
local deformation to avoid the overfitting issue. Spatial pyramid
pooling [40] excerpts the information with restrained-lengths
from the images or regions. It enables a flexible performance
regardless of various scales, sizes, and ratios of input data.
Therefore, the spatial pyramid pooling layer is applied to most
CNN frames for better operations. Ouyang etal. [41] proposed
the Def-pooling method, which is useful in handling deforma-
tion problems, such as the object recognition task or learning
the deformed geometric model. The common methods (i.e.,
max pooling or average pooling) cannot learn object deforma-
tion patterns. Thus, the pooling layers should be purposefully
selected for object learning and better performance of CNN.
The structure of fully connected layers is similar to the
structure of conventional NNs that transform the 2D struc-
ture to a vector layer. The adjusted information through a
fully connected layer is fed into a SoftMax function, which is
placed at the end of CNN. SoftMax is the activation function
that consists of real numbers between 0 and 1. Equation(8)
[42] expresses SoftMax function as follows.
is the kth outcome,
is the number of neurons in
the output layer and
is a vector of the inputs.
Moreover, the loss functions evaluate the predicted values
of the trained models. For the loss functions, there are two
representative functions, the minimum square error (MSE)
and the cross-entropy. Stochastic gradient descent (SGD) is
usually used to update the weight parameters for minimizing
loss functions. In summary, CNN has serial structures, such as
the convolution layer, pooling layer, and fully connected layer,
to provide a model of classification with high performance.
3 Research Challenges forManufacturability
Using Machine Learning
The storage capacity of computers has been increased
enough to store big data for engineering. Among the types
of digital data, those regarding manufacturing engineering
Fig. 6 The architectures of CNN
International Journal of Precision Engineering and Manufacturing
1 3
are categorized into structured and unstructured data.
Structured data stores their information as rows and col-
umns. CSV files, enterprise resource planning (ERP),
and computer logs correspond to the structured data. In
contrast, unconstructed data has no restrictions from cer-
tain structures. They include videos, pictures, 3D scans,
reports, and CAD models that contain the information of
geometries without any descriptions [43]. Artificial intel-
ligence (AI) can handle such unstructured data. Moreover,
it is successful in its applications in manufacturing indus-
tries such as operation monitoring [4446], optimization
[47], inspection [4852], maintenance [53, 54], schedul-
ing [5557], logistic [58], and decision support [59, 60].
In Table1, the listed papers are explained in detail which
datasets and ML methods are utilized. Table2 recatego-
rizes the studies in Table1 with extra case studies and
explains which input, output, and feature extraction meth-
ods are used. More examples of ML in the industries can
also be found in [61], which are categorized as products
(vehicle, battery, robotics, and renewable energy) and pro-
cesses (steel and semiconductor) showing how classifica-
tion or regression techniques with sensory input data are
used to improve manufacturing. Especially, human–robot
collaboration requires environmental perception and
object localization in various applications [62] in which
ML plays a vital role.
Several researchers have studied the design for man-
ufacturability (DFM) techniques combined with ML
to improve productivity. Ding etal. [65] proposed the
detection process of critical features, such as a bounded
rectangle, T-shape, and L-shape, in the hot spot point of
the lithography process. The hot-spot influences contour
precision in the process. Moreover, 5-dimensional vectors
are width (W), length (L), coordinates in the upper-left
corner (X, Y), and direction (D). The information defines
the bounded rectangular features. The gray-shaded zones
encircling the bounded rectangular features are repre-
sented as T-shape (T-f) and L-shape (L-f) features. Criti-
cal features are then derived in the form of (W, L, X, Y,
D, T-f, L-f) at each selected target metal area. The ANN
was implemented to detect hotspots resulting in over
90% of the prediction accuracy. Yu etal. [66] proposed
an ML-based hotspot detection framework by combining
topological classification with critical feature extractions.
They formulated topological patterns by a string- and den-
sity-based method. It classified hotspot features with over
98.2% accuracy. Raviwongse and Allada [67] introduced
a complexity index of the injection molding process using
ANN. They defined 14 features for each molding design
and searched the features in the model, which resulted in
a complexity index from 1 to 10. Jeong etal. [68] used
SVM to decide optimal lengths of pipes in an air-condi-
tioner with the constraints of vibration fatigue life, natural
frequency, and maximum stress. The studies mentioned
above show that ML can be applied to various DFM prob-
lems beyond machinability.
Designers draw the mechanical drawings of products
in various industry fields thinking which CAD design
increases productivity and quality. CAD is indispensa-
ble to portrait detailed mechanical or other engineering
information. However, when designers are not familiar
with the knowledge of manufacturing, information can be
misunderstood or missing from the perspective of expert
engineers. Therefore, the “feature extraction process”
has been used to analyze machinability, which finds suit-
able manufacturing processes from the CAD model. The
expert can decide which manufacturing process is required
for each feature in the CAD model. This process is dif-
ficult for a computer to perform automatically without
expert-designed rules. As an alternative to full and com-
plex implementation of the rules, ML techniques show
the potential to apply the distinction of manufacturability
from the CAD model. The hierarchical learning function
in the deep learning technique, convolutional neural net-
work (CNN) model, for example, enables the recognition
of machinable features from several steps of using convo-
lution kernels that are made of interesting units of basic
features. In this case, it is necessary to design convolution
kernels, pooling layers, and classifiers that can enhance
the performance of feature extraction from CAD models.
However, it is less complex than rule-based techniques.
Searching for patterns in engineering data is challeng-
ing as indicated by its long history [69]. The pattern rec-
ognition method automatically obtains the regularities of
data by computer algorithms, which, in turn, accompa-
nies classification or categorization. Dekhtiar etal. [43]
mention that the five tasks of “Object Recognition” are
object classification, object localization, object detection
or segmentation, object identification, and shape retrieval.
Pre-processing of the information or optimization of the
procedures improves the speed and accuracy of “Object
Recognition”. Further, ML-based feature recognition can
solve the problems of “Object Recognition” without strict
rules. In this context, the ML-based approaches have the
potential to recognize features of DFM and manufactur-
ability due to their simplicity, scalability, and adjustability.
Figure7 shows a summary of the research opportunities.
4 Conventional Feature Recognition
Techniques forManufacturability
Research about automatic feature recognition (AFR) for
CAPP has been conducted for a few decades [70]. In this
chapter, a brief history and ideas of previous research are
introduced. The most recent studies are then reviewed.
International Journal of Precision Engineering and Manufacturing
1 3
Table 1 Relevant utilization of artificial intelligence in the manufacturing industry
Objective Manufacturing application Data AI method Results References
Monitoring Milling operation Time–frequency domain signal Gaussian process regression (GPR),
Bayesian ridge regression, k-near-
est neighbors regression (KNN),
support vector regression (SVR),
decision trees regression
Tool wear estimation Aghazadeh etal. [44]
Vibration signal Artificial neural network (ANN) Surface roughness estimation Khorasani and Yazdi [45]
Acoustic emission signal LSTM-Autoencoder (LAM) Tool breakage Nam etal. [46]
Optimization MEMS Printed features Firefly algorithm, grey relation
coefficient, genetic algorithm
(GA), particle swarm optimization,
response surface methodology
Printing quality Amit etal. [47]
Inspection Cold rolling Images NN Defect detection Yazdchi etal. [48]
Plastic mold Images Principal component analysis
(PCA), multi-layer perceptron
(MLP), artificial neural network
Damage classification Librantz etal. [49]
Hot rolling Images (greyscale) SVM
Defect detection in real-time Jia etal. [50]
Cover glass Images CNN
defects Yuan etal. [51]
Machine vision Images FRR-CNN, YOLO Defects Choi etal. [52]
Maintenance Semiconductor manufacturing Physical/electrical variable and
Filament health and duration Susto etal. [53]
Rotating machinery Vibration PCA
Shaft misalignment Lee etal. [54]
Scheduling Job shop scheduling Job sequencing Genetic algorithm (GA) Machine assigning Lei [55]
Job shop scheduling Bill of materials (BOM) Genetic algorithm (GA) Operation assigning Chen etal. [56]
Disassembly planning Fuzzy scores Genetic algorithm (GA) Design for disassembly Lee etal. [57]
Logistic Supply chain Cost (materials, supplier, manufac-
turing, distribution)
Particle swarm approach Optimization of the supply chain
Shankar etal. [58]
Decision support Non-standard manufacturing
(ex. Large bath production)
All potential risks Multi-layer perceptron (MLP), artifi-
cial neural network (ANN)
Total costs of risk estimation Kłosowski and Gola [59]
Machining process (electri-
cal discharge machining,
Dielectric fluid database grinding
wheel specification
Decision tree Classification of dielectric fluids,
Selections of grinding tools
Filipič and Junkar [60]
International Journal of Precision Engineering and Manufacturing
1 3
Feature recognition methods are divided into rule-based,
graph-based, volume-decomposition, hint-based, hybrid,
and NN methods.
4.1 Rule‑Based Approach
Rule-based approaches compare model representations
with patterns in the knowledge base, which consist of
if–then rules. The rule-based approaches are the earliest
forms of feature recognition processes. However, they lack
unified criteria, leaving different interpretations for a sin-
gle CAD model in addition to the concern of the process-
ing time [71].
Henderson and Anderson [72] proposed a procedure to
recognize features, as in Fig.8a. The method is extracted
from the features from a B-rep model using predefined rules
between entities and features (e.g., swept and non-swept fea-
tures as in Fig.8b). Chan and Case [73] proposed a process
planning tool for 2.5D machined parts by defining rules
for each feature. The rules can be extended from learning
shapes and their machining information. Xu and Hinduja
[74] found cut-volumes from concave and convex entities in
the finished model, and a feature taxonomy recognized the
volumes. Sadaiah etal. [75] also developed process planning
of prismatic components. Owodunni and Hinduja [76, 77]
developed a method to detect six types of features according
to its presence of cavity, single or multiple loop, concavity,
and visibility. Abouel Nasr and Kamrani etal. [78] estab-
lished a rule-based model to find features from the B-rep
model, which is an object-oriented structure from different
types of CAD files.
In addition to boundary representation (B-rep) model
uses, Sheen and You [79] generated a machining tool
Table 2 Machine learning techniques for manufacturability
ML technique Paper Application Input Feature extraction Output
Support-vector machine
(SVM) or regression
[44] Tool condition monitor-
Cutting force, spindle
acceleration, CNC
Wavelet Tool wear
[53] Semiconductor mainte-
nance (ion implanter)
Current, pressure, volt-
age, etc
Min, max, average, etc Maintenance interval
[54] Shaft misalignment
Vibration FFT and PCA Abnormal behavior
[50] Surface defect monitoring Image Gray scale contrast,
disparity, average, vari-
ance, etc
Defect detection
Decision tree [60] Dielectric fluids for EDM
and grinding wheel
Set of learning exam-
ples in attribute-based
Process attributes Decision procedure
Artificial neural network
[45] Surface roughness moni-
Cutting speed, feed rate,
depth of cut, material
type, vibration
RMS Roughness
[48] Surface defect monitoring Image Entropy, variances, aver-
age, power of correla-
tion, etc
Defect and type of defect
[63] Grinding process Cutting parameters Synthetic minority over-
sampling techniques
(SMOTE) functions
Convolutional neural
[51] Defect detection Image No extraction Defect
[64] Defect detection Image RoD transform Defect
Fig. 7 Research challenges for manufacturability in the CAD model
International Journal of Precision Engineering and Manufacturing
1 3
path from slicing models. Ismail etal. [80] defined rules
to find cylindrical and conical features from boundary
edges. Furthermore, the rule-based approaches have ana-
lyzed features from sheet metal parts. Gupta and Guru-
moorthy [81] found freeform surfaces such as protrusion
and saddle from B-rep CAD models. In a further study,
they developed a method to find features such as com-
ponents, dents, beads, and flanges. Sunil and Pande [82]
proposed a rule-based AFR system for sheet metal parts.
Recently, Zehtaban and Roller [83] developed an Opitz
code, a rule to discern features from a STEP file. The pre-
defined rule assigned a code for each component/feature
are recognized via the codes. Moreover, Wang and Yu [84]
proposed ontology-based AFR, as shown in Fig.9. The
model compared B-rep data from the STEP file with a
predefined ontology model, which is a hierarchical struc-
ture with entities and their relations to recognize features.
4.2 Graph‑Based Approach
B-Rep information determines model shapes by faces sur-
rounded by line entities. Graphs of B-Rep is one of the
model description methods that can represent it by multi-
ple details with its level, which enables inexact matching
by checking similarity. Moreover, regarding B-Rep, graphs
represent other information such as height, curvature, geo-
desic distances, and the skeleton of 2D or 3D models [85].
Fig. 8 a feature extraction
procedures, b categorizations of
swept and non-swept features
in the rule-based approach
(Adapted from [72] with per-
HolesSlots Pockets (Extruded)
Divergent /
Convergent Facing Operations Pockets (Non-extruded)
1. Swept
2. Non-swept
International Journal of Precision Engineering and Manufacturing
1 3
However, this study focuses on graph-based methods regard-
ing manufacturability.
Joshi and Chang [86] firstly introduced a graph-based
approach with the attributed adjacency graph (AAG) of
B-Rep polyhedral parts. A graph G = (N, A, T) (N, the set of
nodes; A, set of arcs; T, set of attributes to arcs in A) defines
the relationship between lines, arcs, and boundary faces.
Figure10 illustrates the example of the AAG representation.
The method successfully expresses the meaningful informa-
tion to recognize the features from set arcs or nodes of solid
parts. However, researchers [87, 88] highlighted problems
in the graph-based representations, which are characterized
by the difficulty in recognizing intersections, not considering
tool access, and increased data size for model complexity.
For its completeness, the algorithm should define every sub-
graph pattern; otherwise, it leaves ambiguous representation.
The approaches are an easy way to obtain boundary informa-
tion but are not suitable for volumetric representation [89].
Previous research has endeavored to solve the highlighted
problems. Trika and Kashyap [90] proved that if differences
between a stock and a final part are not recognized by a
union of all volumetric features from the algorithm, it can-
not be machined. Moreover, they developed an algorithm
to generate virtual links for cavity features such as steps,
slots, holes, and pockets in CAD models to be recognized.
Gavankar and Henderson [91] developed a method to sep-
arate protruded or depressed parts from a solid model as
biconnected components in the edge. Marefat and Kashyap
[92, 93] added virtual links to solve interacting features and
compared the subgraphs with predefined machining features.
Thus, a manufacturing plan was established automatically.
Qamhiyah etal. [94] proposed a concept of “Form Features,
which are basic sets of changes from the initial shape. The
Form Features are classified from the graph-based represen-
tation of boundaries. Yuen etal. [95, 96] introduced a similar
concept called the primitive features (PTF) and variation of
PTFs as VPTFs representing information of boundary inter-
acting types. Ibrhim and McCormack [97] defined a new tax-
onomy for vertical milling processes such as depression and
profusion to reduce attempts to find sub-graphs. Huang and
Yip-Hoi [98] used the feature relation graph (FRG) to extract
high-level features such as stepped holes for gas injector
Fig. 9 An example of the subclass features using ontology (Adapted from [84] with permission)
International Journal of Precision Engineering and Manufacturing
1 3
head from low-level features. Figure11 illustrates the pro-
cedure. Verma and Rajotia [99] introduced “Feature Vec-
tor” to represent parts containing curved faces. It represents
subgraphs of AAG into a single vector, which is advanta-
geous to reduce computational time in graph-based methods.
Stefano etal. [100] introduced the “Semanteme,” which are
features that have engineering importance such as concave
parts, axial symmetric parts, and linear sweep parts. The
graph can represent those Semantemes with neighbor attrib-
utes such as parallelism, coaxially, and perpendicularity.
In a recent study, Zhu etal. [101] found machining fea-
tures from a graph-based method to optimize machining
processes in a multitasking machine such as a turn-mill.
After establishing AAG of the model from a STEP file, the
method searched machinable volumes such as slots, bosses,
and blind holes by comparing the analyzed subgraphs with
predefined ones. The model categorized interacting features
into four features—isolation, division, inner loop connecting,
and outer loop connection. In the machining cost optimizing
step, rules of process priority and turning proceeds before
milling, for example, are set to reduce computational loads.
Fig. 10 An example of AAG representation a A 3D CAD model; b The model’s AAG (Adapted from [86] with permission)
Fig. 11 An example of a high-level feature recognition (Adapted from [98], open access)
International Journal of Precision Engineering and Manufacturing
1 3
4.3 Volume Decomposition Approach
The volume decomposition approach decomposes a vol-
ume into small-scaled volumes and analyzes them to extract
meaningful features. It is more advantageous to interpret
intersecting features than previous methods with fewer scal-
ability issues. However, the result may diverge due to dif-
ferent representations [102]. The approach consists of cell
decomposition and the convex hull method.
The cell decomposition method decomposes volumes
into small cells, and a combination of the cells is classified
as one of the machinable features. Sakurai and Dave [103]
introduced a concept of the maximal volume, which consists
of minimal cells with concave edges from an object with a
planar or curved surface. Shah etal. [104] also used the cell
decomposition method. However, they classified volumes
to possible swept volumes from a 3-axis machining center.
Tseng and Joshi [105] extracted machining volumes from
B-Rep data. They then divided the volumes to smaller ones
and reconnected them to obtain features. Figure12 illustrates
the principle that a face and two slots are recognized as fea-
tures after combining sub-volumes.
Recently, Wu etal. [106] decomposed cutting volumes of
milling and turning into cells to optimize the processes. For the
turning volume, edges on 2-D cross-section divided the volume
into cells with variable sizes, and the edges similarly divided
milling volumes but as 3-D segmentations. These cells were
optimized to reduce machining time showing better results than
the hint-based or the convex hull decomposition method.
The convex hull method finds the maximum convex vol-
umes and subtracts them from the original model, and its
difference is iteratively analyzed until there is no convex vol-
ume. Researchers have developed the method since 1980 to
apply it to manufacturing process plans [107110]. Woo and
Sakurai [111] proposed the concept of the maximal feature,
the maximum size of the volume that is machinable with
a single tool. With recursive decomposition, the maximal
feature enabled the improvement of calculation time and
reduced multiple feature interpretation problems.
As one of the recent studies, Bok and Mansor [112] devel-
oped algorithms to recognize regular and freeform surfaces.
The method divided material removal volume (MRR) for
roughing and finishing into sub-volumes such as the overall
delta volume (ODV) to be machined, sub-delta volume for
roughing (SDVR), and finishing (SDVF). Figure13 illustrates
the classification of the CAD model to each sub-volume. In
the following research, Kataraki and Mansor [113] calculated
ODV without any material removal volume discontinuity or
overlaps. Thus, to achieve the goal, the ODV was classified
into SDVR, SDVF, arbitrary volume to be filled (SDVF filled
region) to preserve the continuity of SDVF, and volumetric
features (SDV-VF) to obtain the net shape. The method divided
the sub-volumes stepwise using contours and vectors. The study
validated the method by comparing the calculated ODV to the
manual one, and the difference was within 0.003%. Similarly,
Zubair and Mansor [114] used the method for AFR of symmet-
rical and non-symmetrical cylinder parts for turning and milling
operations. External features are analyzed from faces and edges
to derive roughing and finishing volumes for turning operations.
Asymmetric but turntable internal features are also detected by
comparing the center of the axis. Algorithms for detecting gaps,
fillets, and conical shapes are also established. The validation
shows a 0.01% error level of the ODV difference.
4.4 Hint‑Based Approach
The hint-based approach utilizes information in the CAD
model. For example, a tab hole should have a base drill oper-
ation. The algorithm then finds a cylinder volume for the
drilling. Researchers have studied the method since Vanden-
brande and Requicha’s research [115]. Regli etal. [116, 117]
established the concept of “trace,” a hint to find manufactur-
ing features. For example, a trace of a cylindrical volume is
an indication of the drill hole. Kang etal. [118] proposed a
framework to use tolerance information such as geometry,
dimension, and surface roughness to generate machining
features from the STEP file format. As in Fig.14, Han and
Requicha [119] used hint ranks for the analysis to be much
desirable. Meeran etal. [120] extracted manufacturing fea-
tures from hints in 2D CAD drawings without hidden lines.
Verma and Rajotia [121] established a complete algorithm
for 3-axis vertical milling stages by finding hints from inter-
acting features and repeatedly testing manufacturability and
repairing them.
Fig. 12 An example of the
cell decomposition method
(Adapted from [105] with
Face Slot Slot
International Journal of Precision Engineering and Manufacturing
1 3
Fig. 13 The 3D geometric model in a Isometric top view of CAD and b isometric bottom view of CAD model (Adapted from [112] with permission)
International Journal of Precision Engineering and Manufacturing
1 3
Hints are dependent on specific manufacturing features
such as drill holes, slots, and channels. Thus, it is hard to
find manufacturing features with new tools or new designs.
However, once rules to treat hints are established, the cal-
culation is less exhaustive than rule- and graph-based
approaches [121].
4.5 Hybrid Approach
Real CAD models are complex with Boolean operations,
thus leaving interactive parts. Therefore, time for feature
recognition is also increased as well [122]. Several studies
develop hybrid methods to find the most optimal representa-
tion of features with less time consumption. They used the
NN with other methods to avoid complexity in calculating
interacting features. This section illustrates the combinations
of methods mentioned above. The next section describes the
hybrid methods using the NNs.
First, the hint-based method can clarify interacting
features as a graph representation. Gao and Shah [123]
extracted isolated features from AAG but used the hint-
based approach for interacting features. The hints are defined
by the extended AAG with virtual links. Rahmani and Are-
zoo [124] combined the graph- and hint-based method. For
milling parts, they analyzed milling traces by hints and rep-
resented them as graphs; thus, whole graphs consisted of
known sub-graphs. Ye etal. [125] developed an extended
graph of AAG to discern undercut parts from its subset,
while face properties and parting lines are used as hints to
find undercut features. Sunil etal. [126] used hint-based
graph representation for multiple-sided features without
virtual links. As shown in Fig.15, faces sharing the same
axis are bundled with their adjacencies, thus helping to find
multiple sided interacting features.
Moreover, researchers combined the volume decomposi-
tion method with other methods. Kim and Wang [127] used
both the face pattern-based feature recognition and volume
decomposition. Thus, to calculate stock volumes for cast-
then-machined parts, the method initially searched for face
patterns from predefined atomic features such pockets, holes,
slots, and steps. Subrahmanyam [128] developed “heuristic
slicing,” volume decomposition, and recompositing using
the type of lumps. Woo etal. [129] merged graph, cell-based,
and convex hull decomposition. The graph-based method
filters out non-interconnecting features like holes. Maximal
volume decomposition also filters out conical, spherical, and
toroidal parts. Negative feature decomposition then changed
negative removal volumes to machining features generating
hierarchical structure of the features.
4.6 Conventional Neural Network (NN)‑Based
NN has the advantage of learning from examples. NN is
an excellent tool for pattern recognition if there are enough
datasets [130]. Prabhakar and Handerson [131] showed the
potential of NN-based techniques in feature recognition.
They developed an input format of the neural-net, which
is a combination of the face description and face to face
relationship of the 3D solid model. However, it is necessary
to prepare the input strictly with the rules to construct the
adjacency matrix. Nezis and Vosniakos [132] demonstrated
the feature recognition of topological information such as
planar and straightforward curve faces. This information
was in the form of an attributed adjacency graph (AAG)
that was fed to NN. The neural-net recognized the pocket,
hole, passage, slot, step, protrusion, blind slot, and corner
pocket, showing faster speed than the rule-based recognizer.
Kumara etal. [133] proposed the super relation graph (SRG)
method to identify machined features from solid models.
SRG defines super-concavity and face-to-face relationships,
which became the input data of the NN.
Hwang [134] described the feature recognition method
from a B-rep solid model by using the “perceptron neural
net.” The method used eight-element face score vectors as
input data in the neural-net that enabled the recognition
of partial features. The descriptor recognized simple fea-
tures such as slots, pockets, blind holes, through holes, and
steps. Lankalapalli etal. [135] proposed a self-organizing
NN, which was based on the adaptive resonance theory
(ART). The theory was applied to feature recognition
from B-rep solid models. The continuous-valued vector
measured the face complexity score based on convexity or
concavity and assessed nine classified features. ART-NN
was the unsupervised recognition methods. Moreover, it
Fig. 14 An illustration of the hint ranks; a A 2D geometry with four slot hints (f1–f4); b The calculation of ranks among the hints; c The
obtained design features (DF). (Adapted from [119] with permission)
International Journal of Precision Engineering and Manufacturing
1 3
consumes less memory space. Onwubolu [136] employed
a backpropagation neural network (BPN). The face com-
plexity codes described the topological relationships. BPN
recognized the nine features such as tabs, slots, protru-
sions, pockets, through-holes, bosses, steps, cross-slots,
and blind-holes. Sunil and Pande [137] used the multi-
layer feed-forward back-propagation network (BPNN).
The research showed that the 12-node vector scheme could
represent features such as pockets, passages, blind slots,
through slots, blind steps, and through steps. Öztürk and
Öztürk [138] extracted the face-score values of the com-
plex relationships from B-rep geometries. The NN was
trained from the constructed face-scores and recognized
non-standard complex shapes.
Zulkifli and Meeran [139] developed a cross-sectional
layer technique to search feature volumes from the solid
model. This method defined the feature patterns for edges
and vertices. The detected features were used as the input
to the NN model, which recognized both interacting and
non-interacting features. Chen and Lee [140] described the
feature recognition of a sheet metal part by using an NN. The
NN model classified the model into six features, including
rectangles, slots, trapezoids, parallelograms, V-slots, and
Figure16 shows the feature recognition procedures of
the NN. The solid models are converted into topological
information, such as graphs. They are then used to train
the NN. The input model recognizes the machined fea-
tures. NN-based feature recognition for machinability has
been improved, and the calculation is faster than graph- or
rule-based methods. However, NN needs to preprocess
the input data as adjacency graphs, matrices, codes, and
vectors, which describe the relationship among entities
of a model.
5 Deep Learning‑Based Feature Recognition
As previously mentioned, ML techniques can be applied
to various manufacturing fields. For example, NN based
methods can identify features from a complex CAD
design. B-rep expresses 3D CAD models as boundary
Fig. 15 An illustration of the
hints for circular holes com-
bined with the face adjacency
graph (FAG) method (Adapted
from [126] with permission)
10 11
12 13
H1 –Set of coaxial faces having axis A1
H2 –Set of coaxial faces having axis A2
A1 -Axis 1
A2 -Axis 2
International Journal of Precision Engineering and Manufacturing
1 3
entities such as faces and lines, processes data as graphs
or matrix, and trains the NN model. However, when the
model becomes complex, the amount of input data is
increased, and solving them to several manufacturable fea-
tures also becomes more difficult. Therefore, researchers
have proposed several feature recognition techniques other
than using the B-rep entities highlighted thus far. This
section introduces the methods based on deep-learning
techniques that have the potential to enhance the decision
making of manufacturability in complex 3D CAD models.
5.1 View‑Based Method
In the computer vision research field, researchers have stud-
ied the utilization of 2D images from the 3D CAD models for
feature recognition. In recent years, it has been studied as the
view-based method combined with CNN. Su etal. [141] pro-
posed a multi-view image method for 3D shape recognition.
Multi-view convolution neural network (MVCNN) extracted
the features from 2D images of 12 different views. The CNN
results in the images are pooled and passed to a unified CNN
model. It then produces a single compact descriptor for the
3D shape. Thus, MVCNN achieved better accuracy than the
standard CNN for the classification of 3D shapes. Xie etal.
[142] also studied the feature learnings with multi-view depth
images from the 3D model. Figure17 shows how to obtain
depth images from the projected views. Cao etal. [143] devel-
oped the spherical projected view method, which used cap-
tured images from 12-vertical stripe projection. It is similar
to the multi-view method. There were two sub-captures meth-
ods, the depth-based projection, and the image-based projec-
tion. The depth-based projection determined the depth values,
the distances calculated between the 3D model located in the
center, and each point on the sphere. The image-based projec-
tion captured the image set on 36 spherical viewpoints, which
is then used to train the CNN. The spherical representation
can classify 3D models, and it showed similar performance
compared to other methods. Papadakis etal. [144] proposed
PANORAMA to handle large-scale 3D shape models. They
obtained a set of panoramic projection images from the 3D
model. Then, 2D discrete wavelet transformation and 2D dis-
crete Fourier transformation converted the projection images
to the feature images. PANORAMA provided a significant
reduction of memory storage and calculation time. Shi etal.
[145] introduced deep panoramic representations (DeepPano)
for 3D shape recognition. Panoramic views, as a cylinder pro-
jection, detected 2D images from 3D geometry datasets. The
technique showed higher accuracy than 3D ShapesNets [144],
spherical harmonic (SPH) [146], and light field descriptor
(LFD) [147].
Fig. 16 The feature recognition procedure of the NN-based approach
Fig. 17 The projection plane of the 3D model and captured depth images (Adapted from [142] with permission)
International Journal of Precision Engineering and Manufacturing
1 3
Johns etal. [148] suggested the pairwise decomposi-
tion method with depth images, greyscale images, or
both arrangements. The image-sets were captured over
unconstrained camera trajectories. This method has the
advantage of training for any trajectories. It decomposed
a sequence of images into a set of view pairs. Feng etal.
[149] proposed a hierarchical view-group-shape archi-
tecture for content-based discrimination. The architec-
ture is called a group-view convolutional neural network
(GVCNN). Initially, an expanded CNN extracted the
descriptor in a view level of the 3D shape. The proposed
group module then described the content discrimination of
each view. The module distinguished the view images as
different groups. The architecture merged each group level
descriptor with the shape level descriptor and the sub-
sequent discriminative weight. GVCNN achieved higher
accuracy for the 3D shape classification compared to SPH
[146], LFD [147], MVCNN [141].
The view-based ML is acceptable in recognizing fea-
tures from the 3D model using the CNN architecture.
Moreover, 2D images can be retrieved from the projections
of the 3D model with unstrained directions. The method
can reduce the size of data while preserving the full infor-
mation of the 3D model.
5.2 oud‑Based methodPoint cl
The point cloud was introduced in 2011 [150]; it can rep-
resent the information of 3D shapes, effectively. A point
cloud contains a set of 3D points
i=1, ,n
, where
the vector set at the ith point (
) is each point
Figure18a shows an example of the point cloud contain-
ing coordinates information. Qi etal. [151] designed a
deep learning architecture called PointNet. They only used
the three axial-coordinate information from a point cloud.
PointNet has two networks, the classification network, and
the segmentation network, which provides the capability
of classifying 3D shapes and part segmentation. Their NN
model demonstrated the high performance of 3D recogni-
tion. Fan etal. [152] showed that the point cloud is adequate
for transforming and reforming 3D features. Their training
data sets were formed by recovering the point cloud of a 3D
structure, which is obtained from the rendering of 2D views
of CAD models. Their NN model has a strong performance
in reconstructing various 3D point clouds. Additionally, a
mathematical method has been introduced for recovering
a 3D point cloud models from 2D webcam images [153].
Klokov and Lempitsky [154] proposed a deep learning
architecture (Kd-Net) to recognize 3D point cloud data.
They used the kd-tree, which has good performance for
training–testing times to classify and segment parts. Wang
etal. [155] suggested an NN model called EdgeConv. Each
point contains coordinates with additional information such
as color and surface normal. The κ-nearest neighbor (κ-NN)
graph defined the edge features. The CNN-based model has
two EdgeConv layers, implemented by pooling layers and
three fully-connected layers for classification and segmenta-
tion using point clouds. It achieved a high prediction accu-
racy compared to PointNet or Kd-Net.
Point datasets usually consist of unconstructed information
with additional noises. The expression of surfaces is mostly
arbitrary with sharp geometry due to the noise. There is no
representation of statistical distribution for the patterns of
point cloud data. However, the approach has a less complex
structure than B-rep and constructive solid geometry (CSG).
Therefore, it is suitable to be applied to ML algorithms.
5.3 Volumetric‑Based Methods
3D ShapeNets [156] represented 3D shapes by a probability
distribution of binary variables.
Figure18b shows an example of voxelized 3D shapes.
When binary values are 1 and 0, the voxel is inside and out-
side of the mesh surface, respectively. The 3D shape is sliced
as 30 × 30 × 30 voxels. Each voxel indicates free space, sur-
face, or occluded in the depth map. Free space and surface
voxels represented 3D objects. Further, the occluded voxels
indicated the missing data of the object. This representa-
tion technique is beneficial for learning large-scale 3D CAD
models. Maturana and Scherer [157] developed VoxNet to
Fig. 18 Examples of a point
clouds and b voxelizations for
3D CAD models
(a) (b)
Point coordinates (X, Y, Z) Voxelizaons
International Journal of Precision Engineering and Manufacturing
1 3
recognize a real-time object by a 3D convolutional neural
network algorithm. VoxNet represented 3D shapes by using
occupancy grids corresponding to each voxel. It scales to fit
30 × 30 × 30 voxelization of the 3D CAD dataset. VoxNet
provided the high accuracy for the real-time feature recogni-
tion, thereby classifying hundreds of instances per second.
Qi etal. [158] developed a 3D CAD recognition tech-
nique using the combinations of voxelization and multi-view
images. Multi-orientation volumetric CNN (MO-VCNN)
used the captured images of the voxelated model in the vari-
ous orientations, and CNN architecture extracted the fea-
tures from them. However, low the resolution of as much
as 30 × 30 × 30 confined the performance due to the raised
bottleneck. Hegde and Zadeh [159] proposed FusionNet by
combining volumetric representation with pixel representa-
tion. The 3D object representations of FusionNet are simi-
lar to MO-VCNN [158]. There are three different networks:
V-CNN I, V-CNN II, and MV-CNN. The neural models were
merged at the score layers to classify the 3D CAD model. The
combination of representations shows a better performance
in comparison to individual representation. Sedaghat etal.
[160] proposed the orientation-boosted voxel nets, which is
comparable to MO-VCNN. The voxel grid transformed the
3D CAD model to the volumetric voxel. CNN had two sepa-
rate output layers for N-th class labels and N-th class orien-
tations. Moreover, it attained better classification accuracy.
Riegler etal. [161] proposed OctNet, where the convolutional
network partitions the space of the 3D CAD model. It is a
concept of the unbalanced octree, which is flexible according
to the density of 3D structure. Therefore, OctNet allocates
the smaller storage to represent the 3D model, which, in turn,
improves the calculation speed than the octree.
Moreover, meshes represent the volumes of 3D CAD
models. The meshes have advantages where they can
describe deformations or transformed shapes for finite ele-
ment analysis [162165]. Kalogerakis etal. [166] studied the
segmentation and labeling problem of 3D mesh data. The
pairwise feature algorithm segments the mesh data of 3D
models. The mesh representation outperformed the segmen-
tation of the 3D CAD. Moreover, Tan etal. [167] developed
the extraction algorithm for localized deformation. They
used the mesh-based autoencoders and predicted large-scale
deformations of the 3D models, such as the human pose.
6 Machine Learning‑Based
Feature Recognition Techniques
6.1 A Large‑Set ofComplex Feature Recognition
Only a limited number of studies have explored deep learn-
ing-based techniques for manufacturability. Zhang etal. [168]
proposed a deep-learning-based feature recognition method,
called as FeatureNet, for a large set of complex feature recog-
nition. A set of 24 machining features (common geometries
used in the industry) was selected. Figure19 shows a set of
the selected machining features. A thousand CAD models
were created from 24 features using CAD software. Whole
CAD models have cubic blocks with 10cm lengths. The vol-
ume was removed from the blocks, then specific machining
features were generated. The random values of feature param-
eters within specific ranges determined the models. Further-
more, total datasets had 144,000 models due to placing fea-
tures on six faces of each block. The models were voxelized
with 64 × 64 × 64 grids to feed them into the CNN network.
FeatureNet consists of eight layers as follows: an input
layer, four convolution layers, a max-polling layer, a fully
connected layer, and a classification output layer. Figure20
depicts the CNN architecture of FeatureNet. Each convolution
layer had convolutional calculations with filters to generate
feature maps. Simultaneously, ReLU, as an activation func-
tion, normalized the feature maps after the convolution layers.
In the subsequent fourth-convolution layer, the max-pooling
layer produced down-sized feature maps. A fully connected
layer classified 24 features using a Softmax activation func-
tion. FeatureNet used three optimizers, such as the stochastic
gradient descent (SGD) algorithm, stochastic gradient descent
with learning rate decay (SGDLR) algorithm, and Adam algo-
rithm. The cross-entropy as an objective function was used to
minimize differences between predictions and supervised lev-
els. The total dataset of 144,000 CAD models was separated
into a training set (70%), validation set (15%), and testing set
(15%), respectively. The batch size and initial learning rate
were 40 sets and 0.001 during the training, respectively.
The FeatureNet selected the Adam optimizer due to its
faster convergence than SGD and SGD with learning rate
decay (SDGLR). The test accuracy of the Adam optimizer
was 96.70%. The 16 × 16 × 16 voxel resolution had a train-
ing time of 7.5min while the 64 × 64 × 64 voxel resolu-
tion took 390min. However, the classification accuracy
of 64 × 64 × 64 voxel resolution was 97.4%, which was
higher than others due to increased discretization. Moreo-
ver, FeatureNet recognized multiple machining features in
the CAD models. Practical industry components have high
complexity due to a combination of 24 features, as shown
in Fig.21. FeatureNet used the watershed segmentation
algorithm to subdivide into single features. Figure21
shows the prediction results for the high complexity exam-
ples. This CNN architecture classified 179 of 190 features
and showed 94.21% of prediction accuracy.
6.2 The Recognition ofManufacturable Drilled Hole
Conventional feature recognition methods in Sect.4 are
being examined for the full recognition of complex shapes in
International Journal of Precision Engineering and Manufacturing
1 3
multiple manufacturing processes. FeatureNet can recognize
machining features. However, it does not estimate manufac-
turability. Alternatively, Ghadai etal. [169] proposed the
deep learning-based tool for the identification of difficult-to-
manufacture drilled holes. Deep learning-based design for
manufacturing (DLDFM) framework decided the manufac-
turable drilled holes with DFM rules: (1) depth-to-diameter
ratio, (2) through-holes, (3) holes close to the edges, and (4)
thin sections in the direction of the holes. Figure22 depicts
the rules. The first rule describes the manufacturability of
drilled holes, where the depth-to-diameter ratio is fewer than
5. The second rule is the manufacturability, where the ratio
for a “through-hole” is less than 10. The third rule is that
the drilling process is not manufacturable while the hole
is adjacent to the wall of the stock material. The last rule
describes the situation where flexible materials should have
greater dimensions than hole diameters.
They prepared solid models for manufacturable or non-
manufacturable drilled holes according to the DFM rule.
The solid model had a single drilled hole in a block with
5.0 inches. The diameters, depths, and positions of the drill
holes were randomly determined on six faces of the block.
This case study used the voxel-based occupancy grid to train
3D CNN with the solid model. According to Sect.5, the vox-
elized geometry is an efficient method to represent a solid
model. However, boundary information of the 3D model
is missing in the voxel-based representation. Therefore, the
surface normal using the intersection with each axis-aligned
bounding box (AABB) and B-Rep model were used to pre-
vent missing data. The voxelization with the surface normal
Fig. 19 A set of 24 machining features of FeatureNet (Adapted from [168] with permission)
Fig. 20 The proposed architecture of the CNN network trained to recognize machining features on 3D CAD models (Adapted from [168] with
International Journal of Precision Engineering and Manufacturing
1 3
showed excellent performance for the classification of the
manufacturable drilled holes. Moreover, they considered
multiple holes, L-shaped blocks with drill holes, and cylin-
ders with the drilled holes.
The CNN architecture in DLDFM consists of convolu-
tion layers, max pooling layers, and fully connected lay-
ers. ReLU activation and sigmoid activation works in the
convolution layer and fully connected layer, respectively.
3D CNN learned 75% of the generated 9531 CAD models.
The DLDFM then validated 3D CNN with 25% of the CAD
models. Drawing the class-specific feature maps can help
them to interpret the predictions. Thus, they used a gradient
weighted class activation map (3D-GradCAM) for 3D object
recognition to consider the feature localization map for the
manufacturability. For the least validation loss, it is neces-
sary to fine-tune the hyperparameters. The selected hyperpa-
rameters were as follows: 64 batch size, Adadelta optimizer,
and cross-entropy loss function. These parameters guarantee
optimized learning of the CNN architecture.
Figure23 shows examples of both manufacturable and
non-manufacturable models. 3D-GradCAM predicted manu-
facturability and showed it with color codes. Figure23a–d
show blocks with various types of drilled holes. For instance,
manufacturable drill holes are indicated as blue color code,
as shown in Fig.23a. Furthermore, Fig.23e–h show the
3D-GradCAM for L-shapes with a single hole, cylinder
shape with a single hole, and multi drilled holes, respec-
tively. After that, the DLDFM method was compared with a
hole-ratio based feature detection system. The system had a
training accuracy of 0.7504–0.8136. However, the DLDFM
method had a training accuracy of 0.9310–0.9340. The
DLDFM method outperformed a hole-ratio based feature
detection system for recognizing manufacturable geometries.
Thus, this case study shows the potential of deep-learning
techniques to improve communication between designers
and manufacturers.
Fig. 21 Feature recognition results of the FeatureNet (Adapted from [168] with permission)
Fig. 22 Different DFM rules-based hole examples in classifying manufacturable and non-manufacturable geometries (Adapted from [169] with
International Journal of Precision Engineering and Manufacturing
1 3
7 Research Outlook
Ongoing studies of feature recognition and manufacturabil-
ity analysis will mainly focus on one of key issues: how
to overcome complexity, calculation burden, and ambigu-
ity. Recognizing features and the following machinability
started from the analysis of B-rep or CSG while recent deep
learning techniques converts model to points, voxels, and
planes. It could handle complex models by reducing the
size of a model, however, the conversion also degrades the
resolution of the original model by sacrificing the details.
As one of the solution, Yeo etal. [170] emphasized tight
integration of 3D CAD model into NN by introducing a fea-
ture descriptor. The method recognized 17 types from 75
Fig. 23 Illustrative examples of manufacturability prediction and interpretation using the DLDFM framework (Adapted from [169] with permis-
International Journal of Precision Engineering and Manufacturing
1 3
test models. Panda etal. [171] considered volumetric error
at layer-by-layer calculation during transition from CAD
model to additive manufacturing. Furthermore, to access
manufacturability of 3D mesh or point clouds, converting
the datasets into CAD model as reverse engineering is also
possible. This is about an issue how to find detailed informa-
tion from rough measurement data. The dimension of data is
compressed as space vector and decoded into input to match
between reference CAD models. Kim etal. [172] found pip-
ing contents from 3D point cloud model of a plant using
MVCNN. Including such recent efforts, future studies will
improve the accuracy and flexibility of the feature recogni-
tion by introducing novel machine learning and information
processing techniques.
In the future, feature recognition can be extended to the
study of assembly planning as another field of manufactur-
ability analysis. As one of the studies, a liaison graph was
used to filter out impossible sequences from the assembly of
reference CAD models [173]. Recently, reinforced learning
was used to plan assembly automatically from the feasibility
analysis of module connection [174]. In addition, to handle
the complexity of assembly with various parts, a machine
learning model provided optimized decision making which
is built from previous knowledge [175]. It is expected that
integrating machine learning techniques into feature recog-
nition will provide assessment of assembly directly from
complex CAD assembly models or measured 3D point
clouds. The assembly planning is also expected to be further
improved by converting human skills to building artificial
intelligence. Surface fitting of 3D measurements to CAD
models [153] will recognize subassembly parts and assist
the smart assembly planning.
Technologies such as cyber-physical systems (CPS) and
cloud networks are key technologies of smart manufacturing
[176]. Due to the advantages of ML models and big data-
sets, the feature recognition and manufacturability analysis
will be advanced with the current technological develop-
ment. The smart manufacturing framework of the design
and manufacturing chain combined with developed object
recognition models gives further scope for future research.
Moreover, developing related applications of the machine
learning techniques such as finding a suitable machine shop
for the customer’s CAD model is anticipated as a future
research topic, which is related to smart logistics and dis-
tributed manufacturing.
8 Conclusions
This study reviews ML-based object recognition for analyz-
ing manufacturability. Here is a list of conclusions.
1. In Sects.2 and 3, frequently used ML techniques are
briefly explained and applications for manufacturability
using ML are introduced. From the list of examples,
the scope is narrowed down to feature recognition and
manufacturability assessment from part models.
2. In Sect.4, conventional studies of feature recognition
from CAD model are reviewed. Over a few decades,
researchers in the field mainly dealt with information
regarding B-rep or CSG. The section reviewed research
elements such as graphs, volume decomposition, NN,
hints, and hybrid methods for feature recognition. The
rule-based approach was improved by introducing an
ontology-based technique. Since AAG was proposed,
many works used a graph-based approach in its modi-
fication, given its clear data representation and scal-
ability. The volume decomposition method discretized
the 3D CAD model into sub-cells or maximal features
for enhanced scalability and less calculation; however,
issues of multiple representations remain. Although the
hint-based approach was specific to certain manufactur-
ing processes, it utilized intuitive information to find
machinable volumes, thus resulting in less calculation
load. NN methods using the CAD data was proposed
for less model complexity. A combination of these
approaches, hybrid methods, was studied to enhance
the feature recognition algorithms.
3. In Sects.5 and 6, recent feature recognition using
machine learning and the examples on manufacturabil-
ity applications are introduced. Deep learning-based
methods tried to overcome such complexity and ambi-
guity of the model information. Recently, the use of ML
in feature recognition and manufacturability analysis
becomes promising due to the less complex structure,
less pre-processing of input data, reinforcement by self-
learning, improved accuracy, and enlarged hardware
capacity. Although a huge amount of data is required to
improve accuracy for the wide range of CAD models,
ML is worth applying in the manufacturing field due to
its advantages.
4. In Sect.7, current issues and future studies are described.
Several recent studies introduced in Sects.5 and 6 envi-
sions the potential of new methods of object recogni-
tion. However, enhancing accuracy, reducing calculation
load, and removing noise from discretization provide
new scopes for future studies of deep learning-based
techniques. It is also possible that feature recognition
can be extended to the applications of optimization of
assembly planning or decision making for distributed
manufacturing. Furthermore, the methods of combin-
ing subjective knowledge from manufacturing personnel
will also be preserved and implemented to manufactur-
ability analysis.
International Journal of Precision Engineering and Manufacturing
1 3
Acknowledgements This research was supported by the development
of holonic manufacturing system for future industrial environment
funded by the Korea Institute of Industrial Technology (KITECH
EO220001) and this work has supported by the National Research
Foundation of Korea (NRF) Grant funded by the Korea government
(MSIT) (No. 2020R1C1C1008113).
Author contribution Huitaek Yun contributed to the literature review
and the writing of the paper. Eunseob Kim contributed to literature
review. Hyung Wook Park contributed to the advising. Dong Min
Kim contributed to the literature review, proof reading and supervised
the work. Martin Byung-Guk Jun provided supervised the work. All
authors read and approved the fnal manuscript.
Competing interest We wish to confirm that there are no known con-
flicts of interest associated with this publication and there has been no
significant financial support for this work that could have influenced
its outcome.
Open Access This article is licensed under a Creative Commons Attri-
bution 4.0 International License, which permits use, sharing, adapta-
tion, distribution and reproduction in any medium or format, as long
as you give appropriate credit to the original author(s) and the source,
provide a link to the Creative Commons licence, and indicate if changes
were made. The images or other third party material in this article are
included in the article's Creative Commons licence, unless indicated
otherwise in a credit line to the material. If material is not included in
the article's Creative Commons licence and your intended use is not
permitted by statutory regulation or exceeds the permitted use, you will
need to obtain permission directly from the copyright holder. To view a
copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.
1. Ren, L., Zhang, L., Tao, F., Zhao, C., Chai, X., & Zhao, X.
(2015). Cloud manufacturing: From concept to practice. Enter-
prise Information Systems, 9(2), 186–209. https:// doi. org/ 10.
1080/ 17517 575. 2013. 839055
2. Wu, M., Song, Z., & Moon, Y. B. (2017). Detecting cyber-
physical attacks in CyberManufacturing systems with machine
learning methods. Journal of Intelligent Manufacturing. https://
doi. org/ 10. 1007/ s10845- 017- 1315-5
3. Sabkhi, N., Moufki, A., Nouari, M., & Ginting, A. (2020). A
thermomechanical modeling and experimental validation of the
gear finish hobbing process. International Journal of Precision
Engineering and Manufacturing, 21(3), 347–362. https:// doi. org/
10. 1007/ s12541- 019- 00258-y
4. Lee, J., Bagheri, B., & Jin, C. (2016). Introduction to cyber man-
ufacturing. Manufacturing Letters, 8, 11–15. https:// doi. org/ 10.
1016/j. mfglet. 2016. 05. 002
5. Park, K. T., Kang, Y. T., Yang, S. G., Zhao, W. B., Kang,
Y.-S., Im, S. J., Kim, D. H., Choi, S. Y., & Do Noh, S. (2020).
Cyber physical energy system for saving energy of the dyeing
process with industrial internet of things and manufacturing
big data. International Journal of Precision Engineering and
Manufacturing-Green Technology, 7(1), 219–238. https:// doi.
org/ 10. 1007/ s40684- 019- 00084-7
6. Schmetz, A., Lee, T. H., Hoeren, M., Berger, M., Ehret, S.,
Zontar, D., Min, S. H., Ahn, S. H., & Brecher, C. (2020). Eval-
uation of industry 4.0 data formats for digital twin of optical
components. International Journal of Precision Engineering
and Manufacturing-Green Technology, 7(3), 573–584. https://
doi. org/ 10. 1007/ s40684- 020- 00196-5
7. Park, K. T., Lee, D., & Noh, S. D. (2020). Operation proce-
dures of a work-center-level digital twin for sustainable and
smart manufacturing. International Journal of Precision Engi-
neering and Manufacturing-Green Technology, 7(3), 791–814.
https:// doi. org/ 10. 1007/ s40684- 020- 00227-1
8. Syam, N., & Sharma, A. (2018). Waiting for a sales renais-
sance in the fourth industrial revolution: Machine learning and
artificial intelligence in sales research and practice. Industrial
Marketing Management, 69, 135–146. https:// doi. org/ 10.
1016/j. indma rman. 2017. 12. 019
9. Loyer, J.-L., Henriques, E., Fontul, M., & Wiseall, S. (2016).
Comparison of Machine Learning methods applied to the
estimation of manufacturing cost of jet engine components.
International Journal of Production Economics, 178, 109–119.
https:// doi. org/ 10. 1016/j. ijpe. 2016. 05. 006
10. Pham, D., & Afify, A. (2005). Machine-learning techniques and
their applications in manufacturing. Proceedings of the Institu-
tion of Mechanical Engineers, Part B: Journal of Engineering
Manufacture, 219(5), 395–412. https:// doi. org/ 10. 1243/ 09544
0505X 32274
11. Wuest, T., Weimer, D., Irgens, C., & Thoben, K.-D. (2016).
Machine learning in manufacturing: Advantages, challenges,
and applications. Production & Manufacturing Research, 4(1),
23–45. https:// doi. org/ 10. 1080/ 21693 277. 2016. 11925 17
12. Wu, D., Jennings, C., Terpenny, J., Gao, R. X., & Kumara, S.
(2017). A comparative study on machine learning algorithms
for smart manufacturing: Tool wear prediction using random
forests. Journal of Manufacturing Science and Engineering,
139(7), 071018–071018-9. https:// doi. org/ 10. 1115/1. 40363 50
13. Zeng, Y., & Horváth, I. (2012). Fundamentals of next genera-
tion CAD/E systems. Computer-Aided Design, 44(10), 875–
878. https:// doi. org/ 10. 1016/j. cad. 2012. 05. 005
14. Ren, S., Zhang, Y., Sakao, T., Liu, Y., & Cai, R. (2022). An
advanced operation mode with product-service system using
lifecycle big data and deep learning. International Journal of
Precision Engineering and Manufacturing-Green Technology,
9(1), 287–303. https:// doi. org/ 10. 1007/ s40684- 021- 00354-3
15. Aicha, M., Belhadj, I., Hammadi, M., & Aifaoui, N. (2022).
A coupled method for disassembly plans evaluation based
on operating time and quality indexes computing. Interna-
tional Journal of Precision Engineering and Manufacturing-
Green Technology, 9(6), 1493–1510. https:// doi. org/ 10. 1007/
s40684- 021- 00393-w
16. Leiden, A., Thiede, S., & Herrmann, C. (2022). Synergetic
modelling of energy and resource efficiency as well as occupa-
tional safety and health risks of plating process chains. Inter-
national Journal of Precision Engineering and Manufacturing-
Green Technology, 9(3), 795–815. https:// doi. org/ 10. 1007/
s40684- 021- 00402-y
17. Lubell, J., Chen, K., Horst, J., Frechette, S., & Huang, P. (2012).
Model based enterprise/technical data package summit report.
NIST Technical Note.https:// doi. org/ 10. 6028/ NIST. TN. 1753
18. Hoefer, M. J. D. (2017). Automated design for manufacturing and
supply chain using geometric data mining and machine learn-
ing (M.S.). Iowa State University. Retrieved from https:// search.
proqu est. com/ docvi ew/ 19177 41269/ abstr act/ E0D66 2C306 54480
19. Renjith, S. C., Park, K., & Okudan Kremer, G. E. (2020). A
design framework for additive manufacturing: Integration of
additive manufacturing capabilities in the early design pro-
cess. International Journal of Precision Engineering and
Manufacturing, 21(2), 329–345. https:// doi. org/ 10. 1007/
s12541- 019- 00253-3
International Journal of Precision Engineering and Manufacturing
1 3
20. Groch, D., & Poniatowska, M. (2020). Simulation tests of the
accuracy of fitting two freeform surfaces. International Jour-
nal of Precision Engineering and Manufacturing, 21(1), 23–30.
https:// doi. org/ 10. 1007/ s12541- 019- 00252-4
21. Shi, X., Tian, X., & Wang, G. (2020). Screening product toler-
ances considering semantic variation propagation and fusion for
assembly precision analysis. International Journal of Precision
Engineering and Manufacturing, 21(7), 1259–1278. https:// doi.
org/ 10. 1007/ s12541- 020- 00331-x
22. Kashyap, P. (2017). Let’s integrate with machine learning. In P.
Kashyap (Ed.), Machine learning for decision makers: Cognitive
computing fundamentals for better decision making (pp. 1–34).
Apress. https:// doi. org/ 10. 1007/ 978-1- 4842- 2988-0_1
23. Boser, B. E., Guyon, I. M., & Vapnik, V. N. (1992). A training
algorithm for optimal margin classifiers. Presented at the Pro-
ceedings of the fifth annual workshop on Computational learning
theory, ACM (pp. 144–152).
24. Rosenblatt, F. (1961). Principles of neurodynamics: Perceptrons
and the theory of brain mechanisms. Cornell Aeronautical Lab
25. Luenberger, D. G., & Ye, Y. (1984). Linear and nonlinear pro-
gramming (Vol. 2). Springer.
26. Safavian, S. R., & Landgrebe, D. (1991). A survey of decision
tree classifier methodology. IEEE Transactions on Systems, Man,
and Cybernetics, 21(3), 660–674. https:// doi. org/ 10. 1109/ 21.
27. Rokach, L., & Maimon, O. (2005). Top-down induction of deci-
sion trees classifiers-a survey. IEEE Transactions on Systems,
Man, and Cybernetics, Part C (Applications and Reviews), 35(4),
476–487. https:// doi. org/ 10. 1109/ TSMCC. 2004. 843247
28. Olaru, C., & Wehenkel, L. (2003). A complete fuzzy decision
tree technique. Fuzzy Sets and Systems, 138(2), 221–254. https://
doi. org/ 10. 1016/ S0165- 0114(03) 00089-7
29. Bennett, K. P. (1994). Global tree optimization: A non-greedy
decision tree algorithm. Computing Science and Statistics, 26,
30. Guo, H., & Gelfand, S. B. (1992). Classification trees with neu-
ral network feature extraction. IEEE Transactions on Neural
Networks, 3(6), 923–933. https:// doi. org/ 10. 1109/ CVPR. 1992.
31. Henderson, M. R., Srinath, G., Stage, R., Walker, K., & Regli,
W. (1994). Boundary representation-based feature identification.
In Manufacturing research and technology (Vol. 20, pp. 15–38).
32. LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradi-
ent-based learning applied to document recognition. Proceed-
ings of the IEEE, 86(11), 2278–2324. https:// doi. org/ 10. 1109/5.
33. Goodfellow, I., Bengio, Y., & Courville, A. (2016). Deep learn-
ing (Vol. 1). MIT Press.
34. Guo, Y., Liu, Y., Oerlemans, A., Lao, S., Wu, S., & Lew, M.
S. (2016). Deep learning for visual understanding: A review.
Neurocomputing, 187, 27–48. https:// doi. org/ 10. 1016/j. neucom.
2015. 09. 116
35. Zeiler, M. D. (2013). Hierarchical convolutional deep learning
in computer vision. New York University.
36. Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S., Anguelov,
D., Erhan, D., Vanhoucke, V., Rabinovich, A. (2015). Going
deeper with convolutions. Presented at the Proceedings of the
IEEE conference on computer vision and pattern recognition (pp.
37. Oquab, M., Bottou, L., Laptev, I., & Sivic, J. (2015). Is object
localization for free? Weakly-supervised learning with convo-
lutional neural networks. Presented at the Proceedings of the
IEEE conference on computer vision and pattern recognition
(pp. 685–694).
38. Boureau, Y.-L., Ponce, J., & LeCun, Y. (2010). A theoretical
analysis of feature pooling in visual recognition. Presented at
the Proceedings of the 27th international conference on machine
learning (ICML-10) (pp. 111–118).
39. Zeiler, M. D., & Fergus, R. (2013). Stochastic pooling for regu-
larization of deep convolutional neural networks. arXiv preprint.
https:// arxiv. org/ abs/ 1301. 3557. https:// doi. org/ 10. 48550/ arXiv.
1301. 3557
40. He, K., Zhang, X., Ren, S., & Sun, J. (2014). Spatial pyramid
pooling in deep convolutional networks for visual recognition.
Presented at the European conference on computer vision.
Springer (pp. 346–361). https:// doi. org/ 10. 1109/ TPAMI. 2015.
23898 24
41. Ouyang, W., Luo, P., Zeng, X., Qiu, S., Tian, Y., Li, H., Yang,
S., Wang, Z., Xiong, Y., Qian, C., & Zhu, Z. (2014). Deepid-net:
Multi-stage and deformable deep convolutional neural networks
for object detection. arXiv preprint. https:// arxiv. org/ abs/ 1409.
3505. https:// doi. org/ 10. 48550/ arXiv. 1409. 3505
42. Mikolov, T., Kombrink, S., Burget, L., Černocký, J., & Khudan-
pur, S. (2011). Extensions of recurrent neural network language
model. Presented at the IEEE international conference on acous-
tics, speech and signal processing (ICASSP) (pp. 5528–5531).
IEEE. https:// doi. org/ 10. 1109/ ICASSP. 2011. 59476 11
43. Dekhtiar, J., Durupt, A., Bricogne, M., Eynard, B., Rowson, H.,
& Kiritsis, D. (2018). Deep learning for big data applications in
CAD and PLM—Research review, opportunities and case study.
Computers in Industry, 100, 227–243. https:// doi. org/ 10. 1016/j.
compi nd. 2018. 04. 005
44. Aghazadeh, F., Tahan, A., & Thomas, M. (2018). Tool condi-
tion monitoring using spectral subtraction algorithm and artificial
intelligence methods in milling process. International Journal of
Mechanical Engineering and Robotics Research, 7(1), 30–34.
https:// doi. org/ 10. 18178/ ijmerr. 7.1. 30- 34
45. Khorasani, A., & Yazdi, M. R. S. (2017). Development of a
dynamic surface roughness monitoring system based on artifi-
cial neural networks (ANN) in milling operation. The Interna-
tional Journal of Advanced Manufacturing Technology, 93(1),
141–151. https:// doi. org/ 10. 1007/ s00170- 015- 7922-4
46. Nam, J. S., & Kwon, W. T. (2022). A study on tool breakage
detection during milling process using LSTM-autoencoder and
gaussian mixture model. International Journal of Precision
Engineering and Manufacturing, 23(6), 667–675. https:// doi.
org/ 10. 1007/ s12541- 022- 00647-w
47. Ball, A. K., Roy, S. S., Kisku, D. R., & Murmu, N. C. (2020).
A new approach to quantify the uniformity grade of the electro-
hydrodynamic inkjet printed features and optimization of pro-
cess parameters using nature-inspired algorithms. International
Journal of Precision Engineering and Manufacturing, 21(3),
387–402. https:// doi. org/ 10. 1007/ s12541- 019- 00213-x
48. Yazdchi, A. G. Mahyari, & A. Nazeri. (2008). Detection and
classification of surface defects of cold rolling mill steel using
morphology and neural network. In International conference on
computational intelligence for modelling control & automation
(pp. 1071–1076). Presented at the 2008 International conference
on computational intelligence for modelling control & automa-
tion. https:// doi. org/ 10. 1109/ CIMCA. 2008. 130
49. Librantz, A. F., de Araújo, S. A., Alves, W. A., Belan, P. A.,
Mesquita, R. A., & Selvatici, A. H. (2017). Artificial intelligence
based system to improve the inspection of plastic mould surfaces.
Journal of Intelligent Manufacturing, 28(1), 181–190. https:// doi.
org/ 10. 1007/ s10845- 014- 0969-5
50. Jia, H., Murphey, Y. L., Shi, J., & Chang, T.-S. (2004). An intel-
ligent real-time vision system for surface defect detection. Pre-
sented at the Proceedings of the 17th international conference
on pattern recognition, ICPR 2004 (Vol. 3, pp. 239–242). IEEE.
https:// doi. org/ 10. 1109/ ICPR. 2004. 13345 12
International Journal of Precision Engineering and Manufacturing
1 3
51. Yuan, Z.-C., Zhang, Z.-T., Su, H., Zhang, L., Shen, F., & Zhang,
F. (2018). Vision-based defect detection for mobile phone cover
glass using deep neural networks. International Journal of Pre-
cision Engineering and Manufacturing, 19(6), 801–810. https://
doi. org/ 10. 1007/ s12541- 018- 0096-x
52. Choi, E., & Kim, J. (2020). Deep learning based defect inspec-
tion using the intersection over minimum between search and
abnormal regions. International Journal of Precision Engineer-
ing and Manufacturing, 21(4), 747–758. https:// doi. org/ 10. 1007/
s12541- 019- 00269-9
53. Susto, G. A., Schirru, A., Pampuri, S., McLoone, S., & Beghi, A.
(2015). Machine learning for predictive maintenance: A multiple
classifier approach. IEEE Transactions on Industrial Informatics,
11(3), 812–820. https:// doi. org/ 10. 1109/ TII. 2014. 23493 59
54. Lee, Y. E., Kim, B.-K., Bae, J.-H., & Kim, K. C. (2021). Mis-
alignment detection of a rotating machine shaft using a support
vector machine learning algorithm. International Journal of Pre-
cision Engineering and Manufacturing, 22(3), 409–416. https://
doi. org/ 10. 1007/ s12541- 020- 00462-1
55. Lei, D. (2012). Co-evolutionary genetic algorithm for fuzzy
flexible job shop scheduling. Applied Soft Computing, 12(8),
2237–2245. https:// doi. org/ 10. 1016/j. asoc. 2012. 03. 025
56. Chen, J. C., Wu, C.-C., Chen, C.-W., & Chen, K.-H. (2012). Flex-
ible job shop scheduling with parallel machines using genetic
algorithm and grouping genetic algorithm. Expert Systems with
Applications, 39(11), 10016–10021. https:// doi. org/ 10. 1016/j.
eswa. 2012. 01. 211
57. Lee, S.-C., Tseng, H.-E., Chang, C.-C., & Huang, Y.-M. (2020).
Applying interactive genetic algorithms to disassembly sequence
planning. International Journal of Precision Engineering and
Manufacturing, 21(4), 663–679. https:// doi. org/ 10. 1007/
s12541- 019- 00276-w
58. Shankar, B. L., Basavarajappa, S., Kadadevaramath, R. S.,
& Chen, J. C. (2013). A bi-objective optimization of supply
chain design and distribution operations using non-dominated
sorting algorithm: A case study. Expert Systems with Applica-
tions, 40(14), 5730–5739. https:// doi. org/ 10. 1016/j. eswa. 2013.
03. 047
59. Kłosowski, G., & Gola, A. (2016). Risk-based estimation of
manufacturing order costs with artificial intelligence. In Feder-
ated conference on computer science and information systems
(FedCSIS). Presented at the 2016 Federated conference on com-
puter science and information systems (FedCSIS) (pp. 729–732).
https:// doi. org/ 10. 15439/ 2016F 323
60. Filipič, B., & Junkar, M. (2000). Using inductive machine learn-
ing to support decision making in machining processes. Com-
puters in Industry, 43(1), 31–41. https:// doi. org/ 10. 1016/ S0166-
3615(00) 00056-7
61. Kim, S. W., Kong, J. H., Lee, S. W., & Lee, S. (2022). Recent
advances of artificial intelligence in manufacturing industrial
sectors: A review. International Journal of Precision Engineer-
ing and Manufacturing, 23(1), 111–129. https:// doi. org/ 10. 1007/
s12541- 021- 00600-3
62. Inkulu, A. K., Bahubalendruni, M. V. A. R., Dara, A., &
SankaranarayanaSamy, K. (2021). Challenges and opportunities
in human robot collaboration context of Industry 4.0—A state
of the art review. Industrial Robot: The International Journal of
Robotics Research and Application, 49(2), 226–239. https:// doi.
org/ 10. 1108/ IR- 04- 2021- 0077
63. Lerra, F., Candido, A., Liverani, E., & Fortunato, A. (2022).
Prediction of micro-scale forces in dry grinding process
through a FEM—ML hybrid approach. International Journal
of Precision Engineering and Manufacturing, 23(1), 15–29.
https:// doi. org/ 10. 1007/ s12541- 021- 00601-2
64. Byun, Y., & Baek, J.-G. (2021). Pattern classification for
small-sized defects using multi-head CNN in semiconductor
manufacturing. International Journal of Precision Engineer-
ing and Manufacturing, 22(10), 1681–1691. https:// doi. org/ 10.
1007/ s12541- 021- 00566-2
65. Ding, D., Wu, X., Ghosh, J., & Pan, D. Z. (2009). Machine
learning based lithographic hotspot detection with critical-fea-
ture extraction and classification. Presented at the IEEE inter-
national conference on IC design and technology, ICICDT’09.
IEEE (pp. 219–222). https:// doi. org/ 10. 1109/ ICICDT. 2009.
51663 00
66. Yu, Y.-T., Lin, G.-H., Jiang, I. H.-R., & Chiang, C. (2013).
Machine-learning-based hotspot detection using topological
classification and critical feature extraction. Presented at the
Proceedings of the 50th annual design automation conference
(p. 67). ACM. https:// doi. org/ 10. 1145/ 24632 09. 24888 16
67. Raviwongse, R., & Allada, V. (1997). Artificial neural network
based model for computation of injection mould complexity. The
International Journal of Advanced Manufacturing Technology,
13(8), 577–586. https:// doi. org/ 10. 1007/ BF011 76302
68. Jeong, S.-H., Choi, D.-H., & Jeong, M. (2012). Feasibility clas-
sification of new design points using support vector machine
trained by reduced dataset. International Journal of Precision
Engineering and Manufacturing, 13(5), 739–746. https:// doi. org/
10. 1007/ s12541- 012- 0096-1
69. Bishop, C. M. (2006). Pattern recognition and machine learning
(information science and statistics). Springer.
70. Xu, X., Wang, L., & Newman, S. T. (2011). Computer-aided
process planning—A critical review of recent developments
and future trends. International Journal of Computer Integrated
Manufacturing, 24(1), 1–31. https:// doi. org/ 10. 1080/ 09511 92X.
2010. 518632
71. Babic, B., Nesic, N., & Miljkovic, Z. (2008). A review of auto-
mated feature recognition with rule-based pattern recognition.
Computers in Industry, 59(4), 321–337. https:// doi. org/ 10. 1016/j.
compi nd. 2007. 09. 001
72. Henderson, M. R., & Anderson, D. C. (1984). Computer rec-
ognition and extraction of form features: A CAD/CAM link.
Computers in Industry, 5(4), 329–339. https:// doi. org/ 10. 1016/
0166- 3615(84) 90056-3
73. Chan, A., & Case, K. (1994). Process planning by recognizing
and learning machining features. International Journal of Com-
puter Integrated Manufacturing, 7(2), 77–99. https:// doi. org/ 10.
1080/ 09511 92940 89445 97
74. Xu, X., & Hinduja, S. (1998). Recognition of rough machin-
ing features in 212D components. Computer-Aided Design,
30(7), 503–516. https:// doi. org/ 10. 1016/ S0010- 4485(97)
75. Sadaiah, M., Yadav, D. R., Mohanram, P. V., & Radhakrishnan, P.
(2002). A generative computer-aided process planning system for
prismatic components. The International Journal of Advanced
Manufacturing Technology, 20(10), 709–719. https:// doi. org/ 10.
1007/ s0017 00200 228
76. Owodunni, O., & Hinduja, S. (2002). Evaluation of existing
and new feature recognition algorithms: Part 1: Theory and
implementation. Proceedings of the Institution of Mechanical
Engineers, Part B: Journal of Engineering Manufacture, 216(6),
839–851. https:// doi. org/ 10. 1243/ 09544 05023 20192 978
77. Owodunni, O., & Hinduja, S. (2005). Systematic development
and evaluation of composite methods for recognition of three-
dimensional subtractive features. Proceedings of the Institution
of Mechanical Engineers, Part B: Journal of Engineering Manu-
facture, 219(12), 871–890. https:// doi. org/ 10. 1243/ 09544 0505X
78. Abouel Nasr, E. S., & Kamrani, A. K. (2006). A new methodol-
ogy for extracting manufacturing features from CAD system.
Computers & Industrial Engineering, 51(3), 389–415. https://
doi. org/ 10. 1016/j. cie. 2006. 08. 004
International Journal of Precision Engineering and Manufacturing
1 3
79. Sheen, B.-T., & You, C.-F. (2006). Machining feature recogni-
tion and tool-path generation for 3-axis CNC milling. Computer-
Aided Design, 38(6), 553–562. https:// doi. org/ 10. 1016/j. cad.
2005. 05. 003
80. Ismail, N., Abu Bakar, N., & Juri, A. H. (2005). Recognition of
cylindrical and conical features using edge boundary classifica-
tion. International Journal of Machine Tools and Manufacture,
45(6), 649–655. https:// doi. org/ 10. 1016/j. ijmac htools. 2004. 10.
81. Gupta, R. K., & Gurumoorthy, B. (2012). Automatic extraction
of free-form surface features (FFSFs). Computer-Aided Design,
44(2), 99–112. https:// doi. org/ 10. 1016/j. cad. 2011. 09. 012
82. Sunil, V. B., & Pande, S. S. (2008). Automatic recognition of
features from freeform surface CAD models. Computer-Aided
Design, 40(4), 502–517. https:// doi. org/ 10. 1016/j. cad. 2008. 01.
83. Zehtaban, L., & Roller, D. (2016). Automated rule-based system
for opitz feature recognition and code generation from STEP.
Computer-Aided Design and Applications, 13(3), 309–319.
https:// doi. org/ 10. 1080/ 16864 360. 2015. 11143 88
84. Wang, Q., & Yu, X. (2014). Ontology based automatic feature
recognition framework. Computers in Industry, 65(7), 1041–
1052. https:// doi. org/ 10. 1016/j. compi nd. 2014. 04. 004
85. Iyer, N., Jayanti, S., Lou, K., Kalyanaraman, Y., & Ramani, K.
(2005). Three-dimensional shape searching: State-of-the-art
review and future trends. Computer-Aided Design, 37(5), 509–
530. https:// doi. org/ 10. 1016/j. cad. 2004. 07. 002
86. Joshi, S., & Chang, T. C. (1988). Graph-based heuristics for
recognition of machined features from a 3D solid model. Com-
puter-Aided Design, 20(2), 58–66. https:// doi. or g/ 10. 1016/ 0010-
4485(88) 90050-4
87. Han, J., Pratt, M., & Regli, W. C. (2000). Manufacturing feature
recognition from solid models: A status report. IEEE Transac-
tions on Robotics and Automation, 16(6), 782–796. https:// doi.
org/ 10. 1109/ 70. 897789
88. Wan, N., Du, K., Zhao, H., & Zhang, S. (2015). Research on
the knowledge recognition and modeling of machining feature
geometric evolution. The International Journal of Advanced
Manufacturing Technology, 79(1–4), 491–501. https:// doi. org/
10. 1007/ s00170- 015- 6814-y
89. Rahmani, K., & Arezoo, B. (2007). A hybrid hint-based and
graph-based framework for recognition of interacting milling fea-
tures. Computers in Industry, 58(4), 304–312. https:// doi. org/ 10.
1016/j. compi nd. 2006. 07. 001
90. Trika, S. N., & Kashyap, R. L. (1994). Geometric reasoning
for extraction of manufacturing features in iso-oriented poly-
hedrons. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 16(11), 1087–1100. https:// doi. org/ 10. 1109/ 34.
91. Gavankar, P., & Henderson, M. R. (1990). Graph-based extrac-
tion of protrusions and depressions from boundary representa-
tions. Computer-Aided Design, 22(7), 442–450. https:// doi. org/
10. 1016/ 0010- 4485(90) 90109-P
92. Marefat, M., & Kashyap, R. L. (1992). Automatic construction
of process plans from solid model representations. IEEE Trans-
actions on Systems, Man, and Cybernetics, 22(5), 1097–1115.
https:// doi. org/ 10. 1109/ 21. 179847
93. Marefat, M., & Kashyap, R. L. (1990). Geometric reasoning for
recognition of three-dimensional object features. IEEE Trans-
actions on Pattern Analysis and Machine Intelligence, 12(10),
949–965. https:// doi. org/ 10. 1109/ 34. 58868
94. Qamhiyah, A. Z., Venter, R. D., & Benhabib, B. (1996). Geo-
metric reasoning for the extraction of form features. Computer-
Aided Design, 28(11), 887–903. https:// doi. org/ 10. 1016/ 0010-
4485(96) 00015-2
95. Yuen, C. F., & Venuvinod, P. (1999). Geometric feature rec-
ognition: Coping with the complexity and infinite variety of
features. International Journal of Computer Integrated Manu-
facturing, 12(5), 439–452. https:// doi. org/ 10. 1080/ 09511 92991
96. Yuen, C. F., Wong, S. Y., & Venuvinod, P. K. (2003). Devel-
opment of a generic computer-aided process planning support
system. Journal of Materials Processing Technology, 139(1),
394–401. https:// doi. org/ 10. 1016/ S0924- 0136(03) 00507-7
97. Ibrhim, R. N., & McCormack, A. D. (2002). Process planning
using adjacency-based feature extraction. The International
Journal of Advanced Manufacturing Technology, 20(11), 817–
823. https:// doi. org/ 10. 1007/ s0017 00200 222
98. Huang, Z., & Yip-Hoi, D. (2002). High-level feature recogni-
tion using feature relationship graphs. Computer-Aided Design,
34(8), 561–582. https:// doi. org/ 10. 1016/ S0010- 4485(01)
99. Verma, A. K., & Rajotia, S. (2004). Feature vector: A graph-
based feature recognition methodology. International Journal
of Production Research, 42(16), 3219–3234. https:// doi. org/ 10.
1080/ 00207 54041 00016 99408
100. Di Stefano, P., Bianconi, F., & Di Angelo, L. (2004). An
approach for feature semantics recognition in geometric models.
Computer-Aided Design, 36(10), 993–1009. https:// doi. org/ 10.
1016/j. cad. 2003. 10. 004
101. Zhu, J., Kato, M., Tanaka, T., Yoshioka, H., & Saito, Y. (2015).
Graph based automatic process planning system for multi-tasking
machine. Journal of Advanced Mechanical Design, Systems, and
Manufacturing, 9(3), JAMDSM0034–JAMDSM0034. https://
doi. org/ 10. 1299/ jamdsm. 2015j amdsm 0034
102. Li, H., Huang, Y., Sun, Y., & Chen, L. (2015). Hint-based generic
shape feature recognition from three-dimensional B-rep models.
Advances in Mechanical Engineering, 7(4), 1687814015582082.
https:// doi. org/ 10. 1177/ 16878 14015 582082
103. Sakurai, H., & Dave, P. (1996). Volume decomposition and
feature recognition, part II: Curved objects. Computer-Aided
Design, 28(6), 519–537. https:// doi. org/ 10. 1016/ 0010- 4485(95)
104. Shah, J. J., Shen, Y., & Shirur, A. (1994). Determination of
machining volumes from extensible sets of design features.
Manufacturing Research and Technology, 20, 129–157. https://
doi. org/ 10. 1016/ B978-0- 444- 81600-9. 50012-2
105. Tseng, Y.-J., & Joshi, S. B. (1994). Recognizing multiple inter-
pretations of interacting machining features. Computer-Aided
Design, 26(9), 667–688. https:// doi. org/ 10. 1016/ 0010- 4485(94)
106. Wu, W., Huang, Z., Liu, Q., & Liu, L. (2018). A combinatorial
optimisation approach for recognising interacting machining
features in mill-turn parts. International Journal of Production
Research, 56(11), 1–24. https:// doi. org/ 10. 1080/ 00207 543. 2018.
14250 16
107. Kyprianou, L. K. (1980). Shape classification in computer-aided
design. Ph.D. Thesis. University of Cambridge.
108. Waco, D. L., & Kim, Y. S. (1993). Considerations in positive to
negative conversion for machining features using convex decom-
position. Computers in Engineering, 97645, 35–35. https:// doi.
org/ 10. 1115/ CIE19 93- 0006
109. Kim, Y. S. (1990). Convex decomposition and solid geometric
modeling. Ph.D. Thesis. Stanford University.
110. Kim, Y. S. (1992). Recognition of form features using convex
decomposition. Computer-Aided Design, 24(9), 461–476. https://
doi. org/ 10. 1016/ 0010- 4485(92) 90027-8
111. Woo, Y., & Sakurai, H. (2002). Recognition of maximal fea-
tures by volume decomposition. Computer-Aided Design,
34(3), 195–207. https:// doi. org/ 10. 1016/ S0010- 4485(01)
International Journal of Precision Engineering and Manufacturing
1 3
112. Bok, A. Y., & Mansor, M. S. A. (2013). Generative regular-free-
form surface recognition for generating material removal volume
from stock model. Computers & Industrial Engineering, 64(1),
162–178. https:// doi. org/ 10. 1016/j. cie. 2012. 08. 013
113. Kataraki, P. S., & Mansor, M. S. A. (2017). Auto-recognition
and generation of material removal volume for regular form
surface and its volumetric features using volume decomposition
method. The International Journal of Advanced Manufactur-
ing Technology, 90(5–8), 1479–1506. https:// doi. org/ 10. 1007/
s00170- 016- 9394-6
114. Zubair, A. F., & Mansor, M. S. A. (2018). Automatic feature rec-
ognition of regular features for symmetrical and non-symmetrical
cylinder part using volume decomposition method. Engineer-
ing with Computers, 15, 1269–1285. https:// doi. org/ 10. 1007/
s00366- 018- 0576-8
115. Vandenbrande, J. H., & Requicha, A. A. G. (1993). Spatial rea-
soning for the automatic recognition of machinable features
in solid models. IEEE Transactions on Pattern Analysis and
Machine Intelligence, 15(12), 1269–1285. https:// doi. org/ 10.
1109/ 34. 250845
116. Regli, W. C., Gupta, S. K., & Nau, D. S. (1995). Extracting alter-
native machining features: An algorithmic approach. Research
in Engineering Design, 7(3), 173–192. https:// doi. org/ 10. 1007/
BF016 38098
117. Regli, W. C., Gupta, S. K., & Nau, D. S. (1997). Towards mul-
tiprocessor feature recognition. Computer Aided Design, 29(1),
37–51. https:// doi. org/ 10. 1016/ S0010- 4485(96) 00047-4
118. Kang, M., Han, J., & Moon, J. G. (2003). An approach for inter-
linking design and process planning. Journal of Materials Pro-
cessing Technology, 139(1), 589–595. https:// doi. org/ 10. 1016/
S0924- 0136(03) 00516-8
119. Han, J., & Requicha, A. A. (1997). Integration of feature based
design and feature recognition. Computer-Aided Design, 29(5),
393–403. https:// doi. org/ 10. 1016/ S0010- 4485(96) 00079-6
120. Meeran, S., Taib, J. M., & Afzal, M. T. (2003). Recognizing
features from engineering drawings without using hidden lines:
A framework to link feature recognition and inspection systems.
International Journal of Production Research, 41(3), 465–495.
https:// doi. org/ 10. 1080/ 00207 54021 01488 71
121. Verma, A. K., & Rajotia, S. (2008). A hint-based machining
feature recognition system for 2.5D parts. International Journal
of Production Research, 46(6), 1515–1537. https:// doi. org/ 10.
1080/ 00207 54060 09193 73
122. Li, W. D., Ong, S. K., & Nee, A. Y. C. (2003). A hybrid method
for recognizing interacting machining features. International
Journal of Production Research, 41(9), 1887–1908. https:// doi.
org/ 10. 1080/ 00207 54031 00012 3868
123. Gao, S., & Shah, J. J. (1998). Automatic recognition of interact-
ing machining features based on minimal condition subgraph.
Computer-Aided Design, 30(9), 727–739. https:// doi. org/ 10.
1016/ S0010- 4485(98) 00033-5
124. Rahmani, K., & Arezoo, B. (2006). Boundary analysis and geo-
metric completion for recognition of interacting machining fea-
tures. Computer-Aided Design, 38(8), 845–856. https:// doi. org/
10. 1016/j. cad. 2006. 04. 015
125. Ye, X. G., Fuh, J. Y. H., & Lee, K. S. (2001). A hybrid method
for recognition of undercut features from moulded parts. Com-
puter-Aided Design, 33(14), 1023–1034. https:// doi. org/ 10. 1016/
S0010- 4485(00) 00138-X
126. Sunil, V. B., Agarwal, R., & Pande, S. S. (2010). An approach
to recognize interacting features from B-Rep CAD models of
prismatic machined parts using a hybrid (graph and rule based)
technique. Computers in Industry, 61(7), 686–701. https:// doi.
org/ 10. 1016/j. compi nd. 2010. 03. 011
127. Kim, Y. S., & Wang, E. (2002). Recognition of machining fea-
tures for cast then machined parts. Computer-Aided Design,
34(1), 71–87. https:// doi. org/ 10. 1016/ S0010- 4485(01) 00058-6
128. Subrahmanyam, S. R. (2002). A method for generation of
machining and fixturing features from design features. Comput-
ers in Industry, 47(3), 269–287. https:// doi. org/ 10. 1016/ S0166-
3615(01) 00154-3
129. Woo, Y., Wang, E., Kim, Y. S., & Rho, H. M. (2005). A hybrid
feature recognizer for machining process planning systems. CIRP
Annals-Manufacturing Technology, 54(1), 397–400. https:// doi.
org/ 10. 1016/ S0007- 8506(07) 60131-0
130. Verma, A. K., & Rajotia, S. (2010). A review of machining fea-
ture recognition methodologies. International Journal of Com-
puter Integrated Manufacturing, 23(4), 353–368. https:// doi. org/
10. 1080/ 09511 92100 36421 21
131. Prabhakar, S., & Henderson, M. R. (1992). Automatic form-
feature recognition using neural-network-based techniques on
boundary representations of solid models. Computer-Aided
Design, 24(7), 381–393. https:// doi. org/ 10. 1016/ 0010- 4485(92)
132. Nezis, K., & Vosniakos, G. (1997). Recognizing 212D shape
features using a neural network and heuristics. Computer-Aided
Design, 29(7), 523–539. https:// doi. org/ 10. 1016/ S0010- 4485(97)
133. Kumara, S. R. T., Kao, C.-Y., Gallagher, M. G., & Kasturi, R.
(1994). 3-D interacting manufacturing feature recognition. CIRP
Annals, 43(1), 133–136. https:// doi. org/ 10. 1016/ S0007- 8506(07)
134. Hwang, J.-L. (1991). Applying the perceptron to three-dimen-
sional feature recognition. Arizona State University.
135. Lankalapalli, K., Chatterjee, S., & Chang, T. (1997). Feature rec-
ognition using ART2: A self-organizing neural network. Journal
of Intelligent Manufacturing, 8(3), 203–214. https:// doi. org/ 10.
1023/A: 10185 21207 901
136. Onwubolu, G. C. (1999). Manufacturing features recognition
using backpropagation neural networks. Journal of Intelligent
manufacturing, 10(3–4), 289–299. https:// doi. org/ 10. 1023/A:
10089 04109 029
137. Sunil, V. B., & Pande, S. S. (2009). Automatic recognition of
machining features using artificial neural networks. The Inter-
national Journal of Advanced Manufacturing Technology, 41(9–
10), 932–947. https:// doi. org/ 10. 1007/ s00170- 008- 1536-z
138. Öztürk, N., & Öztürk, F. (2001). Neural network based non-
standard feature recognition to integrate CAD and CAM. Com-
puters in Industry, 45(2), 123–135. https:// doi. org/ 10. 1016/
S0166- 3615(01) 00090-2
139. Zulkifli, A., & Meeran, S. (1999). Feature patterns in recognizing
non-interacting and interacting primitive, circular and slanting
features using a neural network. International Journal of Produc-
tion Research, 37(13), 3063–3100. https:// doi. org/ 10. 1080/ 00207
54991 90428
140. Chen, Y., & Lee, H. (1998). A neural network system feature
recognition for two-dimensional. International Journal of Com-
puter Integrated Manufacturing, 11(2), 111–117. https:// doi. org/
10. 1080/ 09511 92981 30859
141. Su, H., Maji, S., Kalogerakis, E., & Learned-Miller, E. (2015).
Multi-view convolutional neural networks for 3d shape recog-
nition. Presented at the Proceedings of the IEEE international
conference on computer vision (pp. 945–953).
142. Xie, Z., Xu, K., Shan, W., Liu, L., Xiong, Y., & Huang, H.
(2015). Projective feature learning for 3D shapes with multiview
depth images. Presented at the Computer graphics forum, Wiley
Online Library (Vol. 34, pp. 1–11). https:// doi. org/ 10. 1111/ cgf.
143. Cao, Z., Huang, Q., & Karthik, R. (2017). 3d object classifi-
cation via spherical projections. Presented at the International
International Journal of Precision Engineering and Manufacturing
1 3
conference on 3D vision (3DV) (pp. 566–574). IEEE. https:// doi.
org/ 10. 1109/ 3DV. 2017. 00070
144. Papadakis, P., Pratikakis, I., Theoharis, T., & Perantonis, S.
(2010). PANORAMA: A 3D shape descriptor based on pano-
ramic views for unsupervised 3D object retrieval. International
Journal of Computer Vision, 89(2–3), 177–192. https:// doi. org/
10. 1007/ s11263- 009- 0281-6
145. Shi, B., Bai, S., Zhou, Z., & Bai, X. (2015). DeepPano: Deep
panoramic representation for 3-D shape recognition. IEEE Signal
Processing Letters, 22(12), 2339–2343. Presented at the IEEE
signal processing letters. https:// doi. org/ 10. 1109/ LSP. 2015.
24808 02
146. Kazhdan, M., Funkhouser, T., & Rusinkiewicz, S. (2003). Rota-
tion invariant spherical harmonic representation of 3D shape
descriptors. Presented at the Symposium on geometry process-
ing (Vol. 6, pp. 156–164).
147. Chen, D., Tian, X., Shen, Y., & Ouhyoung, M. (2003). On visual
similarity based 3D model retrieval. Presented at the Computer
graphics forum, Wiley Online Library (Vol. 22, pp. 223–232).
https:// doi. org/ 10. 1111/ 1467- 8659. 00669
148. Johns, E., Leutenegger, S., & Davison, A. J. (2016). Pairwise
decomposition of image sequences for active multi-view recog-
nition. Presented at the Proceedings of the IEEE conference on
computer vision and pattern recognition (pp. 3813–3822).
149. Feng, Y., Zhang, Z., Zhao, X., Ji, R., & Gao, Y. (2018). GVCNN:
Group-view convolutional neural networks for 3D shape recog-
nition. Presented at the Proceedings of the IEEE conference on
computer vision and pattern recognition (pp. 264–272).
150. Rusu, R. B., & Cousins, S. (2011). 3D is here: Point Cloud
Library (PCL). In IEEE international conference on robotics
and automation (pp. 1–4). Presented at the IEEE international
conference on robotics and automation. https:// doi. org/ 10. 1109/
ICRA. 2011. 59805 67
151. Qi, C. R., Su, H., Mo, K., & Guibas, L. J. (2017). Pointnet:
Deep learning on point sets for 3d classification and segmenta-
tion. Proceedings of Computer Vision and Pattern Recognition
(CVPR), 1(2), 4.
152. Fan, H., Su, H., & Guibas, L. (2017). A point set generation net-
work for 3d object reconstruction from a single image. Presented
at the Conference on computer vision and pattern recognition
(CVPR) (Vol. 38, p. 1).
153. Abdulqawi, N. I. A., & Abu Mansor, M. S. (2020). Preliminary
study on development of 3D free-form surface reconstruction
system using a webcam imaging technique. International Journal
of Precision Engineering and Manufacturing, 21(3), 437–464.
https:// doi. org/ 10. 1007/ s12541- 019- 00220-y
154. Klokov, R., & Lempitsky, V. (2017). Escape from cells: Deep kd-
networks for the recognition of 3d point cloud models. Presented
at the IEEE international conference on computer vision (ICCV)
(pp. 863–872). IEEE.
155. Wang, Y., Sun, Y., Liu, Z., Sarma, S. E., Bronstein, M. M., &
Solomon, J. M. (2019). Dynamic graph CNN for learning on
point clouds. ACM Transactions on Graphics (ToG), 38(5), 1–12.
https:// doi. org/ 10. 1145/ 33263 62
156. Wu, Z., Song, S., Khosla, A., Yu, F., Zhang, L., Tang, X.,
& Xiao, J. (2015). 3d shapenets: A deep representation for
volumetric shapes. Presented at the Proceedings of the IEEE
conference on computer vision and pattern recognition (pp.
157. Maturana, D., & Scherer, S. (2015). Voxnet: A 3d convolutional
neural network for real-time object recognition. Presented at the
IEEE/RSJ international conference on intelligent robots and sys-
tems (IROS) (pp. 922–928). IEEE. https:// doi. org/ 10. 1109/ IROS.
2015. 73534 81
158. Qi, C. R., Su, H., Niessner, M., Dai, A., Yan, M., & Guibas, L. J.
(2016). Volumetric and multi-view CNNs for object classification
on 3D data. Presented at the Proceedings of the IEEE conference
on computer vision and pattern recognition (pp. 5648–5656).
159. Hegde, V., & Zadeh, R. (2016). FusionNet: 3D object classi-
fication using multiple data representations. https:// doi. org/ 10.
48550/ arXiv. 1607. 05695
160. Sedaghat, N., Zolfaghari, M., Amiri, E., & Brox, T. (2017).
Orientation-boosted voxel nets for 3D object recognition. arXiv.
https:// doi. org/ 10. 48550/ arXiv. 1604. 03351
161. Riegler, G., Ulusoy, A. O., & Geiger, A. (2017). Octnet: Learning
deep 3d representations at high resolutions. Presented at the Pro-
ceedings of the IEEE conference on computer vision and pattern
recognition (Vol. 3).
162. Yi, J., Deng, Z., Zhou, W., & Li, S. (2020). Numerical modeling
of transient temperature and stress in WC–10Co4Cr coating
during high-speed grinding. International Journal of Precision
Engineering and Manufacturing, 21(4), 585–598. https:// doi. org/
10. 1007/ s12541- 019- 00285-9
163. Ahmad, A. S., Wu, Y., Gong, H., & Liu, L. (2020). Numerical
simulation of thermal and residual stress field induced by three-
pass TIG welding of Al 2219 considering the effect of inter-
pass cooling. International Journal of Precision Engineering
and Manufacturing, 21(8), 1501–1518. https:// doi. org/ 10. 1007/
s12541- 020- 00357-1
164. Thipprakmas, S., & Sontamino, A. (2021). A novel modified
shaving die design for fabrication with nearly zero die roll
formations. International Journal of Precision Engineering
and Manufacturing, 22(6), 991–1005. https:// doi. org/ 10. 1007/
s12541- 021- 00509-x
165. Ahmed, F., Ko, T. J., Jongmin, L., Kwak, Y., Yoon, I. J.,
& Kumaran, S. T. (2021). Tool geometry optimization of a
ball end mill based on finite element simulation of machin-
ing the tool steel-AISI H13 using grey relational method.
International Journal of Precision Engineering and Man-
ufacturing, 22(7), 1191–1203. https:// doi. org/ 10. 1007/
s12541- 021- 00530-0
166. Kalogerakis, E., Hertzmann, A., & Singh, K. (2010). Learn-
ing 3D mesh segmentation and labeling. ACM Transactions on
Graphics (ToG), 29(4), 102. https:// doi. org/ 10. 1145/ 18333 49.
17788 39
167. Tan, Q., Gao, L., Lai, Y.-K., Yang, J., & Xia, S. (2018). Mesh-
based autoencoders for localized deformation component analy-
sis. Presented at the Proceedings of the AAAI conference on arti-
ficial intelligence (Vol. 32). https:// doi. org/ 10. 1609/ aaai. v32i1.
168. Zhang, Z., Jaiswal, P., & Rai, R. (2018). FeatureNet: Machin-
ing feature recognition based on 3D convolution neural network.
Computer-Aided Design, 101, 12–22. https:// doi. org/ 10. 1016/j.
cad. 2018. 03. 006
169. Ghadai, S., Balu, A., Sarkar, S., & Krishnamurthy, A. (2018).
Learning localized features in 3D CAD models for manufac-
turability analysis of drilled holes. Computer Aided Geometric
Design, 62, 263–275. https:// doi. org/ 10. 1016/j. cagd. 2018. 03. 024
170. Yeo, C., Kim, B. C., Cheon, S., Lee, J., & Mun, D. (2021).
Machining feature recognition based on deep neural net-
works to support tight integration with 3D CAD systems.
Scientific Reports, 11(1), 22147. https:// doi. org/ 10. 1038/
s41598- 021- 01313-3
171. Panda, B. N., Bahubalendruni, R. M., Biswal, B. B., & Leite,
M. (2017). A CAD-based approach for measuring volumetric
error in layered manufacturing. Proceedings of the Institution of
Mechanical Engineers, Part C: Journal of Mechanical Engineer-
ing Science, 231(13), 2398–2406. https:// doi. org/ 10. 1177/ 09544
06216 634746
172. Kim, H., Yeo, C., Lee, I. D., & Mun, D. (2020). Deep-learning-
based retrieval of piping component catalogs for plant 3D CAD
International Journal of Precision Engineering and Manufacturing
1 3
model reconstruction. Computers in Industry, 123, 103320.
https:// doi. org/ 10. 1016/j. compi nd. 2020. 103320
173. Bahubalendruni, M. V. A. R., & Biswal, B. B. (2014). Com-
puter aid for automatic liaisons extraction from cad based
robotic assembly. In IEEE 8th International conference on
intelligent systems and control (ISCO). Presented at the IEEE
8th international conference on intelligent systems and control
(ISCO) (pp. 42–45). https:// doi. org/ 10. 1109/ ISCO. 2014. 71039
174. Zhang, H., Peng, Q., Zhang, J., & Gu, P. (2021). Planning for
automatic product assembly using reinforcement learning. Com-
puters in Industry, 130, 103471. https:// doi. org/ 10. 1016/j. compi
nd. 2021. 103471
175. Zhang, S.-W., Wang, Z., Cheng, D.-J., & Fang, X.-F. (2022).
An intelligent decision-making system for assembly process
planning based on machine learning considering the variety of
assembly unit and assembly process. The International Jour-
nal of Advanced Manufacturing Technology, 121(1), 805–825.
https:// doi. org/ 10. 1007/ s00170- 022- 09350-6
176. Jung, W.-K., Kim, D.-R., Lee, H., Lee, T.-H., Yang, I., Youn,
B. D., Zontar, D., Brockmann, M., Brecher, C., & Ahn, S.-H.
(2021). Appropriate smart factory for SMEs: Concept, applica-
tion and perspective. International Journal of Precision Engi-
neering and Manufacturing, 22(1), 201–215. https:// doi. org/ 10.
1007/ s12541- 020- 00445-2
Publisher's Note Springer Nature remains neutral with regard to
jurisdictional claims in published maps and institutional affiliations.
Huitaek Yun is a senior instru-
ment controls engineer in Indi-
ana Manufacturing Competitive-
ness Center (IN-MaC), Purdue
university, USA. He received his
Ph.D. degree in 2021 from the
School of Mechanical Engineer-
ing, Purdue University, USA. He
is interested in smart manufac-
turing to combine machining
process and systems with infor-
mation technologies: machine
connectivity in cyber-physical
systems (CPS), mixed reality-
based human machine interface,
data analysis from Internet of
Things (IoT), and development of artificial intelligence for self-decision
Eunseob Kim is a Ph.D. student
in the School of Mechanical
Engineering at Purdue Univer-
sity, IN, USA. He received his
B.S. degree in Mechanical Engi-
neering from Gyeongsang
National University, Korea in
2013, and his M.S. degree in
Mechanical and Aerospace Engi-
neering from Seoul National
University, Korea in 2016. His
research interests include smart
monitoring, sound recognition,
and artificial intelligence appli-
cation for manufacturing.
Dong Min Kim earned his B.Sc.
in 2011 from Korea Polytechnic
University. He received his
M.Sc. and Ph.D. in Mechanical
Engineering from Ulsan
National Institute of Science and
Technology (UNIST), 2013 and
2017, respectively. He worked as
a visiting scholar in Purdue Uni-
versity from March to December
2018. He joined Korea Institute
of Industrial Technology
(KITECH) in 2018 and is cur-
rently working as a senior
researcher, Korea. His interest is
in machining technology.
Hyung Wook Park received the
B.S. and M.S. degrees from
Seoul National University in
2000 and 2002, respectively, and
the Ph.D. degree from Georgia
Tech in 2008, all in Mechanical
Engineering. He is currently a
Professor of Mechanical Engi-
neering at Ulsan National Insti-
tute of Science and Technology.
His research interests lie in the
synthesis and fabrication of
multi-functional composite and
the advanced manufacturing
Martin Byung‑Guk Jun is an
Associate Professor of the
School of Mechanical Engineer-
ing at Purdue University, West
Lafayette, IN, USA. Prior to
joining Purdue University, he
was an Associate Professor at the
University of Victoria, Canada.
He received the B.Sc. and
MA.Sc. degrees in Mechanical
Engineering from the University
of British Columbia, Vancouver,
Canada in 1998 and 2000,
respectively. He then received
his Ph.D. degree in 2005 from
the University of Illinois at
Urbana-Champaign in the Department of Mechanical Science and
Engineering. His main research focus is on advanced multi-scale and
smart manufacturing processes and technologies for various applica-
tions. His sound-based smart machine monitoring technology led to a
start-up company on smart sensing. He has authored over 140 peer-
reviewed journal publications. He is an ASME fellow and Associate
Editor of Journal of Manufacturing Processes. He is also the recipient
of the 2011 SME Outstanding Young Manufacturing Engineer Award,
2012 Canadian Society of Mechanical Engineers I.W. Smith Award for
Outstanding Achievements, and 2015 Korean Society of Manufacturing
Technology Engineers Damwoo Award.
Orthopedic metallic screws even post healing causes ‘pinching effect’ and load inflammation to the bone. In this work, the design and development of a novel metallic core and polymeric shell orthopedic screw is discussed along with optimization of manufacturing process. A comparative 2D finite element study for pull-out test was conducted on HA, HB, HC, and HD standards of screws for approximately identical diameters to determine the type of threading for the proposed design of the novel orthopedic screw. The polymeric shell of the orthopedic screws was manufactured by injection molding process. The material used for manufacturing the polymeric shell was poly-l-lactic acid. Injection pressure, injection velocity, packing pressure and packing time of the injection molding process were optimized for the following output responses: maximum axial pull-out strength, and maximal bending strength of the screw. The r-squared and the adjusted r-squared value of the developed regression model for pull-out was found to be 96% and 92%, respectively. For the bending force, the developed regression model had a r-squared and adjusted r-squared value of 87% and 75%, respectively. The regression models were maximised to obtain the desired input parameters. The optimised parameters were validated experimentally at two local maxima for both pull-out and bending force. The developed screw is novel in design and a patent has been filed under US patent office with number US 2022/0000529 A1. The screws can be used by medical practitioners to avoid or minimise revision surgery and inflammation.
Full-text available
Current assembly process planning of complex products depends mainly on existing process templates and the experience of technical personnel, resulting in low design efficiency, poor process pertinence, low intelligence, and difficulty to extract the assembly knowledge. To address these problems, this paper proposes an intelligent decision-making system for complex products assembly process planning based on the machine learning (ML) method, providing a targeted decision-making scheme with an assembly process structure tree. The characteristics, variations, and similarities of the assembly process for complex products were analyzed. A hierarchical model of the product assembly process is established, based on which the assembly process decision-making for overall structure is decomposed into several units. Then, an intelligent decision-making model for new product assembly process planning was constructed through ML by considering the variability of component composition of different product models and existing product assembly data. Based on these, an optimized decision-making model was established through a swarm intelligence algorithm that automatically optimizes the initial weight, threshold, learning rate, and feedback process, improving the decision efficiency and accuracy of system. To validate the system, the intelligent decision-making assembly process was conducted on a cylinder cap as a case study, and the system capability is discussed in terms of ML performance and industrial applicability.
Full-text available
Recently, studies applying deep learning technology to recognize the machining feature of three-dimensional (3D) computer-aided design (CAD) models are increasing. Since the direct utilization of boundary representation (B-rep) models as input data for neural networks in terms of data structure is difficult, B-rep models are generally converted into a voxel, mesh, or point cloud model and used as inputs for neural networks for the application of 3D models to deep learning. However, the model’s resolution decreases during the format conversion of 3D models, causing the loss of some features or difficulties in identifying areas of the converted model corresponding to a specific face of the B-rep model. To solve these problems, this study proposes a method enabling tight integration of a 3D CAD system with a deep neural network using feature descriptors as inputs to neural networks for recognizing machining features. Feature descriptor denotes an explicit representation of the main property items of a face. We constructed 2236 data to train and evaluate the deep neural network. Of these, 1430 were used for training the deep neural network, and 358 were used for validation. And 448 were used to evaluate the performance of the trained deep neural network. In addition, we conducted an experiment to recognize a total of 17 types (16 types of machining features and a non-feature) from the B-rep model, and the types for all 75 test cases were successfully recognized.
Full-text available
The recent advances in artificial intelligence have already begun to penetrate our daily lives. Even though the development is still in its infancy, it has been shown that it can outperform human beings even in terms of intelligence (e.g., AlphaGo by DeepMind), implying a massive potential for its broader application in various industrial sectors. In particular, the growing public interest in industry 4.0, which focuses on revolutionizing the traditional manufacturing scene, has stimulated a deeper investigation of its possible applications in the related industries. Since it has several limitations that hinder its direct usage, research on the convergence of artificial intelligence with other engineering fields, including precision engineering and manufacturing, is ongoing. This overview looks to summarize some of the important achievements made using artificial intelligence in some of the most influential and lucrative manufacturing industries in hopes of transforming the manufacturing sites.
Full-text available
To meet the sustainable development goals of the United Nations, the energy and resource efficiency of industrial processes have to increase, and workplaces have to become decent for the involved workers. Plating process chains are typically associated with high energy and resource demand and the use of hazardous chemicals. For the analysis and improvement of the energy and resource efficiency as well as for modelling the occupational safety and health risks, a variety of separate approaches are available. Combined approaches are not available yet. An agent-based simulation is used as the basis for integrated energy and resource as well as occupational safety and health risk assessment. In particular, an energy and resource flow model provides the life cycle inventory data for an environmental assessment. The integration of a mechanistic inhalation exposure model through a surrogate model approach enables a combined synergetic consideration of environmental and occupational safety and health effects. A simulation case study shows the impact of chrome acid changes in chrome electroplating processes as well as the effect of different rinsing cascade settings and rinsing control strategies.
Full-text available
Every industrial organization, whether it deals with an assembly or a disassembly process, is definitely putting a huge focus on how to optimize its operative mode by reducing: process variation, frequent changes of production tools, quality issues, wastes. That’s why, disassembly plan generation (DP) is a very supportive tool that detects and identifies difficulties and probable blocking assembly/disassembly situations from the early design steps and avoid them. This anticipation is quite beneficial to manufactories point of view cost, time and quality of products as it returns with a remarkable reduction in production stopping time and products’ defects, as a result reduction in cost of production. In order to ameliorate DP algorithm decision generation and make it very similar to real manufacturing circumstances, this article comes with a new approach that combines between two main metrics, which are the Quality index (Qi) and the Timing index (Ti), as criteria while selecting optimal and feasible DP. Qi is calculated based on the Failure Mode, Effects and Criticality Analysis method (FMECA), on the other side, Ti represents the real processing time index with reference to real manufacturing constraints (workplace, lay-out, work-flow, tools, machines…). To highlight the effectiveness of the proposed approach, an industrial gear box example is simulated and compared to a literature study.
Full-text available
Materials with high hardness are usually difficult to machine, and accomplishing precise and economical machining depends on all the cutting conditions. Appropriate tool geometry is one important aspect for the cutting process that can be optimized based on the machining parameters. In this study, the finite element simulation method was applied to analyze the effects of tool geometry on the cutting forces and tool temperature during the ball end milling of tool steel (AISI H13). Multi-objective optimization of the geometrical parameters was performed using the grey relational method, which gave a set of input parameters to obtain the minimum cutting forces and temperature. The findings of this work could be used as a basis for tool design. Experiments were conducted with mono-objective and multi-objective optimal geometries to validate the finite element analysis. The finite element and experimental results were both congruous with an error limit of 5%.