CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57

Contents lists available at ScienceDirect

CIRP Journal of Manufacturing Science and Technology

journal homepage: www.elsevier.com/locate/cirpj

Manufacturing feature recognition with a 2D convolutional neural

network
Yang Shia , Yicha Zhangb , Ramy Harika,*
a
Department of Mechanical Engineering, College of Engineering and Computing, University of South Carolina, 1000 Catawba Street, Columbia, SC 29201,
United States
b
Mechanical Engineering and Design Department, Université de Bourgogne Franche-Comté, Université de Technologie de Belfort-Montbéliard, ICB UMR CNRS
6303, 90010 Belfort Cedex, France

A R T I C L E I N F O A B S T R A C T

Article history: Feature recognition is critical to connect CAX tools in automation via the extract of significant geometric
Available online 27 June 2020 information from CAD models. However, to extract meaningful geometric information is not easy. There
are still a couple of problems, such as lack of robustness, inability to learn, limited feature types, difficult
to deal with interacting features, etc. To fix these problems, a new feature recognition method based on
Keywords: 2D convolutional neural networks (CNNs) is proposed in this paper. Firstly, a novel feature representation
Manufacturing feature recognition
scheme based on heat kernel signature is developed. Then, the feature recognition problem is transferred
Artificial neural network
into a graph learning problem by using a percentage similarity clustering and node embedding technique.
Shape signature
Feature interaction
After that, CNN models for feature recognition are trained via the use of a large dataset of manufacturing
CAX feature models. The dataset includes ten different types of isolated features and fifteen pairs of
interacting features. Finally, a set of tests for method validation are conducted. The experimental results
indicate that the proposed approach not only performs well on recognizing isolated features, but also is
effective in handling interacting features. The state-of-the-art performance of interacting features
recognition has been improved.
© 2020 CIRP.

Introduction with inherent attributes related to design intent and manufactur-

Feature is an important notion in CAM. It is a higher-level
description of geometric shapes. It characterizes the engineering
significance of designs in term of mathematical descriptions of
surface or volume, which facilitate the automation of CAX
engineering activities. Feature recognition is a process that
captures and refines expert knowledge of features for downstream
engineering activities. It can be considered as a necessary and
fundamental component to integrate design and downstream
applications such as engineering analysis [1–3] reverse engineer-
ing [4], optimization [5–7] design validation [8,9] manufacturing
planning [10–15] or DFX consideration ([16,17]; [18]), etc. It is
inevitable in CAX systems due to the fact that CAD model cannot be
used directly for these engineering analysis and decision-making
systems. CAD models lack high-level description of geometrical
and topological entities [19]. Therefore, the key actions occurring
in feature recognition are translating the low-level geometric
entities from the CAD models into a set of appropriate “features”
* Corresponding author.
E-mail address: HARIK@mailbox.sc.edu (R. Harik).
ing functions ([20,21].
The development of feature recognition methods has been
active for more than two decades in both industry and academic
communities [22–24]. Various feature recognition systems had
been developed, such as rule-based, graph-based, volume decom-
position, hint-based, artificial neural network (ANN) and hybrid
approaches. However, there are still drawbacks of the existing
methods that hinder its applications, such as lack of robustness,
inability to learn, limited class of features, and computational
complexity. The most critical one is the difficulty of recognizing
interacting features, where boundaries of predefined features
changed by feature interactions. To solve these problems, a novel
feature recognition method based on 2D CNN is proposed. Instead
of defining feature as a set of connected faces or volumes, a shape
signature, which can describe both the topology and geometry
characteristics of 3D shapes [25], is introduced. It goes beyond
individual geometric entities, and transforms the whole part into a
heat persistence map (HPM). After clustering, the HPM can be
simplified into attributed graph to reduce the discreteness and
computational complexity. It is not only informative and
comparable, but also has other advantages, e.g. generalization,
robustness, transformation and scale invariance. To be fed into 2D

https://doi.org/10.1016/j.cirpj.2020.04.001
1755-5817/© 2020 CIRP.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
CNN, the graphs are converted into learnable 2D tensors by graph
node embedding technique. Finally, a 2D CNN is built to classify
unknown manufacturing features by using a collection of training
samples. Taking advantage of the robustness of both heat
persistence signature and machine learning, the state-of-the-art
performance for recognizing interacting features has been
improved.
The proposed feature recognition method mainly consists of
three steps: feature representation, graph learning and feature
classification. Before presenting the details in Section 3, a
comprehensive literature review on feature recognition methods
is presented in Section 2 to show the objective and position of this
work. Section 4 presents the method demonstration on solving
feature interaction problem. Section 5 summarizes the paper with
profound discussion and perspective on future work.
Related works
The main purpose of this section is to review how existing
feature recognition methods solve the problems and what
limitations still exist. This article focuses on five group of
approaches that attracted the most extensive research interest
and the hybrid methods or systems, combinations derived from
the fives methods.
Rule-based approach
The rule-based approach was among the earliest to be
investigated due to the granted advantages of expert system, such
as [26],[27], [28], [29] and [30]. Features are generalized as
templates consisting of characteristic patterns, but no explicit
representation scheme was defined. The recognizing process is
performed by using inference rules represented as If-Then manner.
If the predefined conditions are satisfied, then the corresponding
region on a part is recognized as a feature. Nonetheless, the feature
representation is ambiguous and the rule based reasoning is
difficult to deal with new features. Moreover, it is not realistic to
define rules for complex and interacting features.
Hint-based approach
To deal with feature interactions and be more flexible with the
feature searching, hint-based methods were developed based on
the idea that incomplete representation can be searched for, so as
to indicate the existence of certain features. Hence, a hint is
defined as a minimal indispensable portion of a feature's
boundary which must be present in the part even when features
intersect. Hint-based approaches use a two-step procedure for
feature recognition: in the first step, hints are generated by
extraction rules based on different ways, such as geometric and
topologic reasoning [19,31–33], feature taxonomies ([16,34],
Abouel [35]), and combined probabilities of ranking potential
feature hints [36,37]. In the second step, these hints are processed
and directly matched by applying rules ([16], [18], [31]). However,
since the hints are predifined for each feature class, the
recognition algorithm is hardcoded, so the system extensibility
is sacrificed. In addition, it is difficult to determine the
characteristics of a feature for hint definition in case of complex
and interacting features.
Graph-based approach
The Graph-based approach is among the most-researched
methods due to the inherent advantage of a graph’s structural
similarity with B-Rep based solid models. The approach devel-
oped by [38] can be considered as the first formal graph-based
37
feature recognition method, introducing the concept of an
attributed adjacency graph (AAG). AAG captures the concave/
convex relationships of the part’s adjacent faces, and analyze the
adjacency graph in order to decompose it into subgraphs as
features. The feature searching can be realized by subgraph
isomorphic matching.
[39] extended AAG by adding curved surface node and
developed surface-based attributed adjacency graph (SAAG).
[40–42] developed the Multi-Attributed Adjacency Graph/Matrix
(MAAG/MAAM) to overcome the limitations of AAG. To better
support interacting features, [43] extended AAG with more
attributes of edge and face, naming Extended Attributed Adjacency
Graph (EAAG). [44] improved EAAG by adding edge node type and
quantitative attributes such as face normal vector, face angle, and
edge length, and named it as holistic attribute adjacency graph
(HAAG). [45] extended the graph method to freeform features by
dividing the shape into different regions based on curvature and
using these regions as graph nodes, and named it as Region
Adjacency Graph (RAG).
Numerous different graph representation schemes were
developed to enhance the representation capability of graphical
models for more complex features. The information stored in
graphs gradually became more complete and less ambiguous,
nevertheless, at the cost of robustness and computation complex-
ity. It can be stated that graph-based approaches are quite effective
in the domain of primitive features. However, the most significant
shortcoming of all graph-based methods is the incapability of
handling topology variations, because variations change the arcs
and nodes that are indispensable for defining the feature.
Moreover, due to the combinatorial difficulties and exponential
time complexity, the complex graph with a large number of nodes
takes expensive computational costs for graph matching process.
Consequently, graph-based approaches have difficulties to handle
real industrial parts with small-scale variations such as fillets and
chamfers.
Volume decomposition approach
The volume decomposition approach identifies the removal
(machining) volume of stock material from the solid model and
decomposes the volume into intermediate volumes first, and then
the features are generated by combining the intermediate volumes
based on some rules. According to the way of decomposing,
volume decomposition methods can be generally divided into two
sub-groups: convex-hull decomposition and cell-based volume
decomposition.
The convex-hull decomposition method was introduced by [46]
to decompose non-convex objects into a sequence of convex
volumes with arbitrary shapes called Alternating Sum of Volumes
(ASV) Decomposition. To solve the non-convergence problem, [47]
proposed the method of Alternative Sum of Volumes with
Partitioning (ASVP). The convex-hull decomposition method is
effective in finding delta volumes for polyhedral parts, but has
difficulty in curved surface.
The essential methodology of cell-based decomposition ap-
proach is to decompose the volume or delta volume of an object
into minimal cells with a simple shape. The cells are then
recombined as a larger volume that can be removed by a single
machining operation. Finally, the combined volume is checked for
topological and geometrical characteristics and recognized as a
machining feature. Sakurai and Chin [48] proposed a representa-
tive method to generate all the possible combination sets of the
minimal convex cells decomposed by face extension. [49]
introduced a new maximal volume decomposition method, which
decomposes the delta volume by recursively bisecting sub-
volumes into two smaller volumes.
38 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Cell-based decomposition approach is suitable for manufactur-
ing planning and NC coding. However, the volume feature
representation is deficient with geometric and topological feature
information, so it is not able to recognize the types of functional
feature and not suitable for other engineering applications. There
are also some other disadvantages such as an expensive
computational load and the inability to guarantee the generation
of the features of interest.
Artificial neural network approach
The major characteristic of ANN that makes it one of the most
promising feature recognition methods is its ability to derive
implicit patterns through training with examples. The other
advantage that attracts interest for employing ANN in feature
recognition is the robustness for tolerating exceptions or incom-
plete input patterns, so it can be possible to recognize the non-
orthogonal interacting features.
Building a perfect input representation scheme is the most
difficult part of ANN-based feature recognition method.
Manufacturing features are characterized by both topological
and geometrical information derived from the CAD model, while
neural networks typically use numerical values as input. Therefore,
this raises the problem of how to convert a solid model to a suitable
input representation for neural network, since simple numerical
representations are not always sufficient to represent solid models.
The reviewed input representation schemes can be generalized
Fig. 1. Overview of the proposed approach.
into three types: 2D projections based ([50], Chuang 1999, [51]),
graph-based [52–54], and face score vector [55–58].
Feed-forward neural network with back-propagation is the
most often used ANN topology for feature recognition. Because of
simple structure, good learning and generalization capabilities for
pattern association and pattern recognition problems, three-layer
feed-forward neural network has been used as a classifier by many
researchers to prove the feasibility of their methods, such as
([21,50–52,57,59]. To deal with large number of types of feature
and complex interactions, [60] and [58] adopted a standard four-
layer FF neural network with two hidden layers due to its better
convergence results. [61] and [62] presented a network structure
by cascading a number of typical three-layer feed-forward back-
propagation neural networks to be more expandable for new
features.
Although ANN-based feature recognition systems have certain
advantages over conventional methods, the limitations are
obvious: their feature representation schemes still have a degree
of ambiguity and limited applicability, such that, only a limited
range of features and feature intersections can be recognized; most
of ANN-based systems use supervised learning, which requires a
retraining for including new feature class.
Hybrid approach
Hint-based, graph-based, volume decomposition and neural
networks are four basic and promising approaches in feature
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
recognition field. They all have their advantages and limitations.
Therefore, it is evident that a hybrid system adopting selective
characteristics of these approaches could provide a constructive
and practical solution to overcome the shortcomings of existing
feature recognition systems.
To amend the graph-based method’s incapability of interacting
features, [43] combined graph-based and hint-based methods.
[63] and [64] developed a similar graph-hint hybrid system to
extract subgraphs as feature hints. [65] directly connected a graph-
based and a hint-based system to get a hybrid system, in which the
isolated and interacting features are processed respectively. To
take advantage of the strong ability of generalization and
robustness, ([53,54], Ding 2004, tried to integrate ANN as feature
classifier with inputs generated from existing graph-based
representation schemes. [66] used the ANN to convert the non-
orthogonal feature into a number of triangular blocks for
combining into a virtual orthogonal feature, which can be
recognized by graph-based method in the next step.
Due to the remarkable volumetric representation and exhaus-
tive nature of decomposition, some researchers attempted to
integrate graph-based and volume decomposition approach for
better solution of interacting features. The most common type of
volume-graph hybrid approach uses maximal volume decomposi-
tion to extract machining features and graph matching approach to
recognize features, such as [67–69]. The other example is ([21]b),
where the graph-based approach was used to extract features from
solid models and the volume decomposition approach was
incorporated to generate delta volumes and multiple interpreta-
tions of machining operation sequences.
Although hybrid feature recognition systems eliminate some
drawbacks of original approaches, there still remains limitations
like difficulties to extend feature class, limited range of feature
interaction, and high computational cost. As discussed above, most
of the existing feature recognition approaches were based on
identification of the geometrical entities composing a feature,
together with their interrelationships, against certain predefined
set of rules or templates. Their target features were usually
classified by some high-level characteristics such as function,
usage, and manufacturing methods. Each type of methods has
more or less limitations in feature recognition, especially for
complex interacting feature recognition. To make some progress in
this direction, in this article, a new method to use a shape signature
to represent features and recognize them by 2D convolutional
neural network is proposed. The next section presents details of
the proposed method.
Proposed 2D CNN method
In comparison to knowledge-based systems, machine learning
algorithms do not implement any logical operations explicitly. By
Fig. 2. Example of heat persistence map and similarity clustering.
39
performing only simple arithmetic operations, they can derive
many kinds of knowledge or discover regularities through training
with input patterns that are difficult to describe adequately with
knowledge-based systems. Therefore, the proposed feature
recognition algorithm is independent of the features library. It is
expandable to any secondary features without any need to rebuild
the algorithm. The general flowchart of the proposed approach is
depicted in Fig. 1. At first, the input CAD models are converted into
HPM. Then, the HPM is simplified into attributed graph by
similarity clustering, and 2D histograms later by node embedding.
The histograms are finally sent to a 2D CNN for the classification.
The details of our feature representation scheme, the node
embedding strategy, and the CNN model implementation will be
introduced in following subsections.
Feature representation
The first and most important step of employing ANN is to
convert a solid model into a suitable input representation for the
neural network, since simple numerical representation is not
always sufficient to represent geometry and topology informa-
tion stored in CAD models. A satisfactory input representation
should at least have the following basic requirements: (1) it
contains all the necessary information for identifying the
patterns; (2) it must be unambiguous, each class having unique
representation; (3) it is implementable for computation. In this
proposed method, Heat Kernel Signature (HKS) is adopted for
pattern representation.
HKS is a pointwise feature descriptor for representing the
point's local and global geometric properties. It is based on heat
kernel, which is a fundamental solution to the heat equation. The
eigen-decomposition of heat kernel is expressed as
X
1
Ht ðx; yÞ ¼ eli t fi ðxÞfi ðyÞ ð3:1Þ
i¼1
where li and fi are the eigenvalues and the corresponding
eigenfunctions of the Laplace-Beltrami operator D over M. It
should be noted that heat kernel can be successfully approximated
by only a few of the eigenfunctions fi .
Using the heat kernel, one can derive the quantity of heat
received by a local area Dx  M around a point x at a given time
instance t. If assume that the initial heat distribution was a Dirac
delta distribution dx ðyÞ, i.e. only a source point x has an infinite
heat value and the other points have zero heat, and that there is no
loss of the net heat, the heat value Rx received by the local area Dx
can be expressed as:
Z
Rx ðtÞ ¼ 1  Ht ðx; yÞdy ð3:2Þ
MDx
40 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
It is defined that the heat persistence value as the incremental
value of Rx , until the time where heat drops lower than a threshold
value m. It can be computed as the integral of the heat value Rx :
Zti
Rv ðxÞ ¼ Rx ðtÞdt ð3:3Þ
0 In the computation experiments, the mesh size was uniform for
The threshold m is determined by the mean of persistence
values of the points set at the time ti , which is the reference time
when the relative standard deviation of the collection of heat
persistence value RX reaches a peak value.
After getting the heat persistence value for every vertex in a
mesh model, a heat persistence map can be obtained as the feature
representation. However, there are normally thousands of nodes in
the map, so it is too large to serve as the input of ANN. Hence, the
HPM is divided into subsets by grouping adjacent vertices with
similar heat persistence values. The clustering algorithm starts
from the node having the highest heat persistence value. Then, the
nodes connected to this start node will be clustered with it if their
heat persistence values are higher than the similarity limit.
Otherwise, the nodes will be used as new start node for another
clustering process. Fig. 2 shows an example of applying this
similarity clustering strategy. One can clearly note from Fig. 3 that,
as similarity percentage limit increases, the shape pattern is
clearer, but at the cost of more clusters.
Feature selection
As mentioned in section 2, there are numerous feature
classification schemes based on the specific needs and interest
of the researchers, but manufacturing feature is the most common
objective. In this article, a set of ten primitive machining features
by reference to ISO STEP-AP 224 for university are selected.
However, as a matter of fact, the selection of feature classes is
independent of building the recognition framework. Under the
same framework, the feature family can be easily extended to some
secondary features, such as through pocket, T-slot, tapered hole
and so on.
Ideally, the training dataset should be collected from real
industrial parts, but the machine learning method usually needs a
large number of training samples to ensure the generality and
robustness, so these features are generated by scripts. As shown in
Fig. 4, the negative features lay in a cubic raw stock with 100 mm
side length, and the positive features are placed on a plate with 10
mm thick, 100 mm length and width. The size and position of each
feature are randomly set in a specific range. It should be noted that
the heat kernel signature is invariant to translation, rotation and
scale. Therefore, the substrate is fixed to origin and all faces are
aligned with the principal axes, and the features are only located
on the top face of the substrate.
The solid models used for feature representation are first
created by a CAD modeling software called FreeCAD [70]. And the
Fig. 3. Examples of similarity clustering using different percentages.
neutral output format of FreeCAD files selected for the recognizer is
*.step (IOS STEP 10303). Then these B-Rep models need to be
tessellated for HPM computation. In order to ensure the
correspondence between the CAD models’ B-Rep representation
and meshes points, the mesh models are created by a mesh
generator named Triangle [71] instead of using CAD software.
every model to make sure the heat persistence values are most
comparable. The computational time cost of HPM and storage cost
of models depend on the number of vertices in the mesh model. In
the initial experiments, the mesh size was set at 1 mm, the number
of vertices in a single model roughly range from 20,000 to 50,000.
The size of 1 mm gives a good mesh resolution and ensures that
features with small dimension have enough number of clusters to
represent the shape, and that the computational cost is acceptable.
Technically, when the trained model is applied to parts with large
dimensions, the mesh size can be proportionally increased to
guarantee a fast recognition.
Feature extraction
Manufacturing features are normally represented by a set of
connected faces, which can be generated by material removing
process. Therefore, before getting the cluster graph as the input to
the neural network, the features need to be extracted from the
substrate first. Although we introduced a novel feature extraction
method based on HKS in [25], its disadvantage of computational
complexity hinders the application to a large amount of feature
database. Considering all of the target features are primitive
features, a hint-based feature extraction method is implemented
on the basis of Fu’s work [16].
Hint-based feature extraction method is an effective way to
identify surface features that are of interest. The concept of hint
can be defined as a pattern in the part boundary that provides a
trace for searching the existence of certain features. Here, the
characteristics of edge loops are utilized as hint to identify form
features. An edge loop is a set of connected edges that forms the
boundary of a surface. As shown in Fig. 5, an external loop is the
exterior boundary, while the internal loop is inside of the face sand
surrounding a depression or protrusion.
As shown in Fig. 6, the features are extracted by the following
steps:
1 Read the geometry and topology entities from the generated
STEP file, including vertices, edges, boundary loops and faces of
the solid model.
2 Identify the hierarchical and adjacent relationship between
geometry entities.
3 Identify the face type (internal or external) and the loop type
(inner, outer and hybrid).
4 Search the feature by using the inner loop or hybrid loop in
external face as hints.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Fig. 4. A list of 10 isolated features selected in this research, and the range of their dimensions.
Fig. 5. Internal loop and external loop.
Fig. 6. Pseudocode of Feature Extraction.
41
5 Extract the feature by collecting the faces that are directly or
indirectly connected to the hint loop.
6 Classify the feature into negative or positive by identifying the
convexity (concave, convex and hybrid) of its entrance loop.
After feature extraction, the boundary information of the
manufacturing features is obtained. By combining with the heat
persistence clusters, the manufacturing features are further
represented by only clusters and their corresponding heat
persistence values. To be processed subsequent algorithms, the
feature representation is formatted by graph structure. Let the
graph be G = (V, E), where V is a set of persistence clusters, E is the
corresponding connectivity between clusters. The cluster set V can
be described by N  F attributes vector X, where N is the number of
clusters and F is number of cluster attributes. As shown in Fig. 7, the
edges E are encoded by an N  N adjacency matrix A.
Graph learning
Graphs are a powerful data structure that extensively used to
model structured data within computer science and related fields.
In this article, the basic idea is to frame 3D shape feature
recognition as graph learning problem (Fig. 8). Specifically, it is to
build a machine learning framework that can learn a function to be
used for classifying unknown shape features using a collection of
sample graphs.
The key of graph learning is how to convert the discrete graph
structures into informative and learnable features that can support
effective classification. However, CNNs are typically designed for
input data structure with meaningful or consistent order. For
example, images are composed of pixels that are aligned by the
spatial coordinates. Nevertheless, graphs do not have such
underlying spatial structure and the spatial dependence features.
They and their adjacency matrices are irregular, no-Euclidean, not
associated with spatial position and the notion of Euclidean
distance. Hence, generalizing neural network to graphs is a
challenging problem.
In order to transform the graphs into ordered tensor
representations for CNN, the first step is determining the node
42 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
sequences to make sure that nodes with similar structural roles in
the graphs are positioned similarly in their tensor representations.
Then the tensors are assembled and normalized so that their
structural information can be comparable. At last, the normalized
tensors can serve as receptive fields of a convolutional neural
network.
In the proposed graph learning framework, the graph nodes are
arranged in order of descending heat persistence value. Heat
persistence value is an efficient node attribute that can be used for
the arrangement of node sequence. It is directly based on the
Gaussian curvature of a surface, and is also closely related to
diffusion maps and diffusion distances [72], which means it
represents both the shape and relative position of a vertex.
Therefore, it provides a consistent pattern for the node sequences.
Graph node embedding
The graph learning problem is formulated as: let G = (V, E) be a
given cluster graph, it is undirected, unweighted without self-
loops. Each graph can be represented by a n  n adjacency matrix
A, where n is the number of nodes (clusters). It is binary and
symmetric. Moreover, a n2 column vector X is defined to denote
the node attributes: heat persistence value and quantity propor-
tion of vertices. X will be used as extra feature channel to be
normalized and compressed with the node embedding. Finally,
each graph G in the training dataset has a corresponding label that
can be used for learning a model to predict the class of an unseen
graph.
The goal of node embedding is to project A and X into low-
dimensional vectors that summarize the structure and feature
information contained in them. In the node embedding space, the
Euclidean distance between two points should be proportional to
the similarity of graph structure role of the corresponding graph
nodes [73].
The problem of node embedding can be taken as a machine
learning task independently. Fig. 9 illustrates the process of node
embedding visually. The nodes are colored with respect to heat
persistence value. The node embedding vectors are learned by
Novd2vec [74], which is a neutral representation learning
framework. It can learn a mapping of nodes as points to a low-
dimensional vector space. The basic idea of learning is to optimize
this mapping by minimizing the information loss between
geometric relationships in the learned space and the structure
of the original graph [75]. After obtaining the embedding vectors,
in order to get fixed length input vectors for downstream graph
learning task, principal component analysis (PCA) is used to reduce
the dimensions of the learned embedding.
The last step is to transform PCA node embedding into an
image-like histogram. To mimic the schema of CNNs on images,
every two columns of the PCA node embedding to 2D plane are
projected and the projections are turned into a square and number
Fig. 7. Example of feature extraction and graph representation.
fixed 2D grids, where each grid is associated with a number
indicating how many nodes locate in the grid. In Fig. 10, there are
40 2D histograms of dome feature to visually demonstrate that the
features in the same class have similar layout of 2D histograms.
2D CNN configuration
Technically, the 2D histograms can be input to any kind of
CNNs that are suitable for images. In this research, the
convolution architecture in [76] is adopted as an example,
because it is known to be able to provide good prediction results.
As illustrated in Fig. 11, it is basically a LeNet-5 [77] with four
parallel convolutional-pooling blocks followed by a fully con-
nected layer with 128 units. The filter sizes of convolutional layers
are 3  3, 4  4, 5  5, and 6  6 respectively. The first layer
convolutional-pooling has 64 filters, and the second layer has 96
filters. The max pooling filter is 2  2. A softmax layer was added
at the end to output the prediction of feature classes. The ReLU
function was used for activation, and regularization is achieved by
a dropout rate of 0.3.
A standard 10-fold cross-validation was used and repeated each
fold 3 times in all of the experiments. The dataset was shuffled into
a 90-10 split of the training-validation subsets for each fold. Batch
size was set to 32, and the learning rate and number of training
epochs was optimized by Adam on each fold. The learning model
was implemented in Python3.6 using Keras Library with tensor-
flow backend. All experiments were run on the same machine.
Parameters
First of all, experiments were conducted on datasets with
different number of samples, and the validation accuracy was
improved as the number increased. Finally, the number 5,000 is
picked for each feature class to balance the good results and
computational expense. Therefore, the entire dataset consists of 10
features and 50,000 models.
Before giving the final results of feature recognition, the
parameters of input pre-processing were tuned to find out their
impacts on the classification results. When a target parameter was
being tested, all other parameters shared the same setting. The
performance comparison between different values of parameter
were based on the validation accuracy of trained CNN model. In
addition, the tuning experiments are only employed on negative
features for simplicity.
Similarity of clustering: the percentage similarity is used to
control the number of clusters in a single feature. The ideal
condition is that the models in the same feature class have similar
number of clusters, but the clustering process is stochastic and
the number of clusters varies. Therefore, a minimum similarity
and minimum number of clusters for the clustering process is set.
The minimum similarity ensures the features with large dimen-
sion can have a good resolution of clusters to represent the shape,
Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57 43

Fig. 8. Graph Learning Framework.

Fig. 9. Process of node embedding.

44 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57

Fig. 10. Randomly picked 2D histograms of dome feature from the training dataset.

Fig. 11. 2D CNN architecture.

 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
while the minimum cluster number guarantees the features with
small size having sufficient clusters for node embedding vectors.
If the clustering results do not satisfy the requirement of
minimum cluster number, the similarity will be increased by a
small number and redo the clustering. 3 pairs of minimum
similarity and cluster number: [85%, 10], [90%, 15], [95%, 20], are
tried. From the results listed in Fig. 12, it is easy to conclude that
more clusters get better results. However, it is worth to note that,
because of uniform mesh and dimension variations, it needs a high
similarity percentage to get enough clusters for features with small
size. Therefore, if want more than 20 clusters for every model,
there is a need to decrease the mesh size to have more vertices in
the model.
Node2vec: the node2vec algorithm involves a number of
parameters. The conclusion of parameter performance in [74] is
examined. The number of features and the node’s neighborhood
parameters (number of walks r, walk length l, and neighborhood
size k) increasingly improve the performance. Therefore, the
recommended settings: d = 128, r = 10, l = 80, k = 10, are used. For
the in-out parameter p and the return parameter q, which controls
the sampling strategy of random walk, different sets of [p, q]: [4,
0.25], [2, 0.5], [1, 1], [0.5, 2], [0.25, 4], are tested with all same other
parameters. As shown in Fig. 13:, the parameter p and q have
relatively low impacts on the results of validation accuracy. It can
be speculated it was because the structure of cluster graphs is
small, simple and stable. Therefore, p = q = 1 is selected for the rest
of the experiments.
Number of input channels: the 2D histograms are produced by
slicing every 2 columns of the node embedding vectors as a single
Fig. 12. Validation accuracy with different sets of similarity requirements.
45
feature channel, so the maximum number of histograms is limited
by the minimum number of graph nodes in the dataset.
Meanwhile, since the node embedding vectors are sorted by
PCA, the information contained in the histograms is less and less.
Although providing more information to the CNN may help it to
classify the input features, more input channels will increase the
number of parameters in the neural network and make it more
difficult to learn. In addition, more input channels also increase the
computational cost of training process. Therefore, we conduct
experiments to determine the number of input channels according
to the validation accuracy.
It is worthy to be noted that, since the first channel is composed
of heat persistence signature and the vertices number percentage,
it is the most informative one. As shown in Fig. 14, the curve does
not substantially go up, which means the first channel may have
sufficient information for classification purpose. This also proves
that HPM is a great feature representation. For the rest of channels,
which represents the structure of cluster graphs, they are
beneficial to the classification, but also could deteriorate the
learning ability of CNN. In general, experiment of 5 channels gave
the best performance, but 2 channels also worked well and had less
computation cost. Therefore, testing with different channels is
continued in subsequent experiments.
Histogram size: the size of histogram is decided by both of the
size of histogram bins and the numeric range of the node
embedding vectors. Furthermore, the size of bins is determined by
how many intervals the numeric range are divided into. Therefore,
the size of bins can be used to control the size of histogram. For
example, if the vector values range from -2.9 to 2.6, and every
46 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Fig. 13. Validation accuracy with different sets of in-out parameter p and the return parameter q.
single scale 1 is divided into 10 intervals, the histogram size is 55 
55, since j2:6  ð2:9Þj  10 ¼ 55. Intuitively, the smaller the bins,
the higher resolution of the histogram, and the more differentiable
the features are. However, smaller bins produce bigger receptive
fields, such that the CNN has more input nodes and parameters,
and the training could be more difficult and expensive. In order to
select an appropriate bin size, we compare the validation accuracy
of CNN with different bin sizes. Based on the results shown in
Fig. 15, it can be concluded that 1/10 is the best choice of bin size.
Results
In the initial experiments, it is tried to train the CNN model
with a dataset including both of negative and positive features.
However, as listed in Table 1, the result of ten features training
together was not good as separated. By analyzing the distribu-
tion of validation loss shown in Fig. 16, it can find the loss are
mainly between boss and blind hole. Because they are basically
the same shape with different volumetric space, so that their
HPM patterns are very similar. This can be considered as a
limitation of HPM. However, differentiating negative and
positive features can be easily achieved by identifying the
concavity/convexity of their entrance loops in the feature
extraction step. Fig. 18 shows their convergence of loss function
and accuracy with epochs.
Two independent 10-fold trainings of negative and positive
features with the same parameter setting are experimented. As a
preliminary research, the classification accuracy of the proposed
approach is very competitive. The distribution of validation loss is
shown in Fig. 17. For negative features, the biggest loss is between
slot and step. Because the HPM is invariant to rotation, and it
doesn’t take consideration of number of geometry entities.
Therefore, a step feature also can be viewed as a V-slot feature.
For positive features, it can find that the confusions are mainly
among pyramid, cone, dome. They are also similar shapes; all of
them have decreasing cross-section area. The results discussed
above show that the proposed method is capable to recognize
shape features with high accuracy. To further validate the proposed
method, method demonstration is presented in the following
sections via the recognition of interacting shape features, which
are quite difficult for existing feature recognition methods.
Interacting feature recognition
Feature interaction is the most critical issue in the development
of a new feature recognition method. It is the decisive factor that
whether feature recognition technology can be applied to real
industrial components or not. As reviewed in section 2, for those
existing methods, feature interaction is still a tough problem. The
underlying difficulty is that feature interactions destroyed the
indispensable boundaries that characterize a feature, such as edges
and faces. However, for most feature recognition applications,
explicit and unique feature descriptions in topology and geometry
are necessary and intolerant to exceptions. Fig. 19 shows an
example of feature interaction between pocket feature and slot
feature. The figure explains how the topology of their face graph
representations changes by the feature interaction and decompo-
sition. The decomposed slot feature has exactly the same graph as
the original pocket feature.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Fig. 14. Validation accuracy with different number of feature channels.
Cluster graph of interacting feature
As introduced above, the heat persistence method describes a
feature beyond the edges and faces. It is implicit and flexible. The
distribution of clusters follows the shape variation. For example, on
tip like shapes, the cluster wraps all the faces that compose the tip.
On bar shapes, the clusters are like rings that circle these
indispensable faces. When small parts of edges and faces are
missing, the clusters they belong to might be still exist. Moreover,
even when a face is split into multiple faces, they could still share
the same cluster node. As illustrated in Fig. 20, feature interaction
may split a cluster into multiple ones, but the symmetry and
gradualness of the entire cluster graph is constant. These patterns
can be traced to compensate the missing of original boundaries.
Adding the robustness of ANNs to handle exceptions and
incomplete input patterns, it is promising to recognize the
primitive features in the interactions. It has been known that
the difficulties of interacting feature recognition arise from the fact
that feature interaction lead to a vast body of heterogeneous,
uncertain and inherently inconsistent topology and geometry
information, which limits the development of expert systems. For
artificial intelligence, on the contrary, exponential complexity,
uncertainty, inconsistency, and interaction of various kinds of
knowledge are treated as inherent attributes of complex problems.
In the experiments, it assumes that the two interacting feature
share the same entry face. The interacting relationship between
two negative features is classified into four types: touch, intersect,
cross and overlap. These four types of relationship can occur with
47
the same or different depths. Therefore, there could be eight types
of feature interaction. Taking example of slot and pocket feature,
Fig. 21 shows the corresponding cluster graphs of some feature
interactions. To illustrate in an intuitive way, the clustering
similarity is relatively low, so the number of clusters won’t be too
much.
After analyzing the feature interaction scenarios, it can be
noticed that the changes of HPM is mainly caused by the
convexity/concavity switch between new edges and the dis-
appeared ones. For instance, when the interaction destroys part of
a concave edge, replace by a convex one, two corners are created
between the concave and convex edges, and new clusters will be
produced around the corner. In Fig. 22, the interactions of pocket
and slot features are taken as example to illustrate the effects of
interaction on HPM. In the demonstrations, the effects of
interactions based on faces are investigated. They can be
summarized into five typical types of variations: (1) part of the
edge's concavity/convexity of is switched, (2) part of the edge/
face is missing, (3) entire edge’s concavity/convexity of is
switched, (4) face split and (5) hole in the face. As it can be
seen, in the first type, although there is not boundary change in
the face, part of the edge's concavity/convexity of is switched, so
new clusters are created in the face. In the second type, part of the
edges is replaced by new edges with opposite concavity/
convexity. In the third type, entire of an edge is replaced by a
new one with opposite concavity/convexity. For the fourth and
fifth type, four new corners are created in the face. It also should
be note that the clusters are dynamic, the numbers and
48 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57

Fig. 15. Validation accuracy with different bin size.

Table 1
Average validation accuracy of 10-fold 3-repeats training on three different datasets.

Dataset Validation accuracy

Total Slot Step Pocket Through hole Blind hole Boss Pyramid Protrusion Cone Dome
All 98.88  0.18 99.67  0.46 99.48  0.60 99.61  0.28 99.14  0.82 94.11  1.74 98.28  0.59 99.87  0.21 99.32  0.36 99.77  0.21 99.51  0.69
Negative 99.82  0.18 99.98  0.06 99.97  0.07 99.83  0.18 99.70  0.43 99.62  0.29 – – – – -
Positive 99.40  0.14 – – – – – 98.34  0.66 99.82  0.26 99.63  0.38 99.74  0.27 99.48  0.61

Fig. 16. Distribution of validation loss for the 10-fold training with 10 feature classes.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57 49

Fig. 17. Distribution of validation loss for two independent 10-fod trainings on negative and positive features.

Fig. 18. Convergence of loss and accuracy for both learning and validation.

connectivity are decided by the clustering process. But their
overall pattern is always according the shape. The colors are
marked relatively by their heat persistence values. Using these
examples of cluster graphs, it can visually prove that the HPM
characterizes the feature not by predefined boundaries by but by
the whole shape.
Feature selection
In this article, only the interactions between two negative
features are used to validate the proposed idea. On one hand, the
negative features are more common and diverse. All five positive
features have only one entrance face. The interactions of positive
features are simpler than the negative ones. On the other hand,
positive and negative features are similar shapes but with different
volumetric space. In feature interactions, all primitive features
only have three types of defects: edge missing, hole in the face, and
feature split.
In Fig. 23, all the 15 pairs of feature interaction used for
experiments are listed. The size and location of each primitive
feature are selected randomly and independently in same range as
isolated features. The models without interactions are removed to
50 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Fig. 19. Feature interaction and decomposition illustrated by graph-based representations.
Fig. 20. Feature interaction and decomposition illustrated by the cluster graphs.
ensure enough samples of feature interaction. For each pair of
features, there could be four types of relationship: touch, cross,
intersect, and within.
Feature extraction
The hypothesis is that HPM make it possible to recognize
incomplete features by artificial neural network. Therefore, before
the graph learning, it needs to extract and decompose the
interacting feature into primitive features first. In the following
steps, the geometric reasoning to decompose interacting features
is described.
1 The first step is to extract the interacting feature entirely from
the raw stock following the steps in section 3.3.
2 The second step is to identify cylindrical faces from the face
collection of an interacting features. If it has more than one
entrance face, it is a through hole, otherwise it is a blind hole. If it
is a blind hole, the faces connected to the bottom of the
cylindrical face are combined as bottom surface of the blind
hole. The “bottom” of a cylindrical face is determined by the
distance between entrance face and the target face or edge.
3 The next step is to identify bottom surface from the rest faces in
the collection. Then check the adjacent faces of the bottom
surface, if it has two adjacent faces that are entrance face and
connected to each other, this bottom surface belongs to a step
feature. If it has no entrance face in the adjacent faces, it is the
bottom of a pocket feature, otherwise it is the bottom of a slot
feature.
4 Then the bottom surfaces are used as core to reconstruct the
primitive features. Angles between a bottom face and its
adjacent faces are determined by cross product of face normal
directions. If there is a concave edge, the corresponding adjacent
face is added to the face collection of target feature.
5 The final step is to combine split features. Primitive features
might be split if they cross each other. Instead of recognizing
features separately, we combine them together as a single input.
If the number of decomposed features from an interacting
feature is larger than two, the features in same primitive class
are examined for combination. If their bottom surfaces are on
the same plane, and they have one (step feature) or two (slot and
pocket feature) wall face on the same plane, they are combined a
single decomposed feature.
Results
For the dataset of interacting features, a number of 5,000 is
selected for each set of feature interaction. Hence, 150,000 samples
are obtained totally of five decomposed features. Considering the
large variations in the feature interactions, this dataset is not
extraordinary, but sufficient to validate the idea. In future works,
experiments will be conducted on larger dataset with more
complicated interactions.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
Fig. 21. Topology changes of feature interaction and decomposition illustrated by the cluster graph representations.
In the initial experiment, it attempted to classify the decom-
posed features with the CNN model trained with isolated features.
The average accuracy of entire interacting feature dataset is
52.36%. The individual accuracy of five decomposed features is slot
81.09%, step 33.27%, pocket 51.30%, through hole 62.10%, blind hole
34.05%. From these numbers, it can conclude that the CNN model
of isolated features is not able to identify the patterns of
decomposed features. However, the results are still better than
randomly guessing, assumingly because some decomposed
features have the similar HPM pattern as the isolated ones.
To recognize interacting features, there could be two options:
(1) First attempts to classify all extracted features with CNN
models for isolated features. If the probability of classification is
not good enough, it will be classified again by another CNN model,
which is specifically trained by decomposed interacting features.
(2) Using a single CNN model that trained for both interacting
features and isolated features to classify the extracted features.
Ideally, the latter might be better choice, since the isolated features
could serve as an intermedia among different types of feature
interactions. Therefore, two different training experiments are
employed. One is only using the 150,000 decomposed interacting
features, and the other is combined with 50,000 isolated features
(10,000 for each feature). As listed in Table 2, the results of two
datasets are almost the same. Therefore, it may conclude that the
number of decomposed interacting features that have the similar
HPM pattern as isolated features is sufficient for learning
Fig. 24 and Fig. 25 demonstrate the distribution of validation
loss for the corresponding training. Comparing with Fig. 17, it can
be observed that the distributions of wrong predictions for
interacting features are similar to the loss distributions of isolated
features’ training. The confusions are still mostly among similar
51
shapes. Therefore, it is logical to conclude that the feature
representation based on HPM is well capable of handling boundary
loss of interacting features. Furthermore, the performance of
interacting features recognition might be improved by the same
way as recognition of isolated features, such as finer mesh and
better CNN structure. Fig. 26 shows their convergence of loss
function and accuracy with epochs.
Case study
In order to demonstrate the general applicability of the
proposed recognition framework, it is applied on several example
parts from NIST (National Institute of Standards and Technology)
design repository. The example parts are composed of slots, steps,
through holes and blind holes. Most of these features interact with
others and the rest few are isolated. To display the robustness of
the proposed framework, these models are kept as they are, even
though they contain round edges and interactions of multiple
features that are not covered by the training dataset. The STEP files
were downloaded and processed by the same framework as
introduced in previous sections. Then the decomposed features are
classified by the CNN models trained by the dataset including both
interacting and isolated features.
The results are shown in Fig. 27. The prototype system
successfully recognized all isolated features and most of interact-
ing features. The total classification accuracy is 96.88%(62/64). To
be specific, in Fig. 27 (d), a step feature having three holes was not
correctly identified. For Fig. 27 (e), the blind hole interacting with
nine through holes was not successfully recognized. It should be
noted that, both of these two cases are interactions of three or
more features, which are not covered in the training dataset. That
52 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57

Fig. 22. The effect of feature interaction on HPM pattern.

 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57 53

Fig. 23. 15 pairs of feature interaction between negative features.

Table 2
Average validation accuracy of 10-fold 3-repeats training on dataset of decomposed interacting features and dataset of combining with isolated feature.

Dataset Validation accuracy

Total Slot Step Pocket Through hole Blind hole Boss Pyramid Protrusion Cone Dome
All 98.88  0.18 99.67  0.46 99.48  0.60 99.61  0.28 99.14  0.82 94.11  1.74 98.28  0.59 99.87  0.21 99.32  0.36 99.77  0.21 99.51  0.69
Negative 99.82  0.18 99.98  0.06 99.97  0.07 99.83  0.18 99.70  0.43 99.62  0.29 – – – – –
Positive 99.40  0.14 – – – – – 98.34  0.66 99.82  0.26 99.63  0.38 99.74  0.27 99.48  0.61

Fig. 24. Distribution of validation loss for the dataset of 150,000 decomposed interacting features.

means their classification are not guaranteed in the first place. be seen that the CNN trained by generated samples successfully
However, there are more cases of multi-feature interaction are recognize most features of real industrial components. The
correctly recognized. Although we cannot conclude that the performance of the proposed CNN-based feature recognition
trained CNN is able to recognize multi-feature interactions, it can methods is validated. In addition, these exceptions can be
54 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57

Fig. 25. Distribution of validation loss for the dataset of combining decomposed interacting features and isolated features.

Fig. 26. Convergence of loss and accuracy for both learning and validation.

addressed by extending the range of considered interaction types
of multiple features. The proposed CNN-based feature recognition
framework could easily accommodate new required interaction
types by adding examples to the training dataset.
In this section, the capability of the proposed feature
recognition framework verified to be able handle feature
interactions. It is the first one that successfully recognize
interacting features without recovering the missing geometric
boundaries. The large scale of random sampled dataset verifies that
it has great robustness to incomplete feature representations. The
excellent accuracy of recognizing decomposed features also make
the feature recognition framework outperform any existing
methods. For instance, in most of the existing approaches, the
geometric and topological variations especially the unseen ones
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57 55

Fig. 27. Feature recognition on five complex parts from NIST design repository.

often lead to unsuccessful recognition of the test cases. Neverthe-
less, feature interaction is still a difficult problem. Since the feature
interactions may involve multiple features, the possible ways of
interaction increase factorially with adding more features to the
combination. Moreover, a large variety of feature interactions may
need huge training cost.
Conclusion and future works
The objective of this research was to use artificial intelligence
technology to develop a feature recognition system that could
overcome the drawbacks in this domain. First, a novel feature
representation scheme based on heat kernel signature is
56 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
developed. Different from conventional representation schemes, it
goes beyond geometric entities, describes both the topology and
geometry characteristics of the shape by a concise HPM. Besides
informative and learnable, HPM also has advantages like robust-
ness of topology and geometry variations, translation, rotation and
scale invariance. Using the technique of percentage similarity
clustering, the HPM is then simplified into cluster graph.
Thereafter, the feature recognition problem is successfully framed
as a graph learning problem. In order to transform the cluster
graphs into suitable input of CNN for effective classification, the
node embedding method is used to project the cluster graphs and
their heat persistence values into low-dimensional tensors. Then
the tensors are assembled and normalized so that their structural
information can be comparable. At last, the 2D histograms of
tensors serve as receptive fields of a convolutional neural network.
A large dataset of CAD models including isolated and interacting
features is generated randomly by scripts for training the CNN
models. The parameters involved in the preprocessing step were
determined by experiments. The details of feature extraction and
classification were presented. The results of recognizing isolated
features have shown that the proposed method has better
performance than any existing ANN based approaches. The feature
recognition framework offers the advantages of learning and
generalization. It is independent of feature selection and could be
extended to various features without any need to redesign the
algorithm. The results of recognizing interacting features indicate
that our feature representation scheme is effective in handling the
boundary loss caused by feature interactions. The state-of-the-art
performance of interacting features recognition has been im-
proved.
Although the proposed feature recognition method shows great
achievements in overcoming the drawbacks of existing methods,
there are several limitations and issues that need further
investigation. Firstly, the heat kernel is invariant to rotation, so
it cannot differentiate the features having the same shape but
defined based on the direction, such as V-slot and step. Similarly,
since HPM only represents surfaces, it also has difficulties to make
a distinction between features with the same shape but defined in
different volume spaces, just like pocket and protrusion. Therefore,
sometimes there is a need of additional mechanisms to ensure
accurate recognition. For instance, the geometric reasoning (inner
loop) is adopted to differentiate negative and positive features.
Secondly, HPM is computed based on the discretized mesh
vertices. Therefore, in order to fully describe the surface, it requires
sufficient number of vertices. Since uniform mesh method is used
to tessellate the solid model, if there is a small surface in the model,
either it can’t be well represented by HPM, or the mesh size has to
be very small. It has been proven that finer meshes has better
accuracy, but it takes more computational cost. The other possible
solution is to use ununiformed mesh and add cluster area to the
node attributes.
Thirdly, the design of convolutional neural networks is
preliminary. It is a classical structure designed for recognizing
hand-writing numbers. Therefore, it is possible to improve the
performance of the feature recognition framework using the state-
of-the-art machine learning algorithms.
Fourthly, there is still large space for improving the solution of
the feature interaction problem. It has been proved that HPM is
capable to handle the boundary loss of feature interactions, but the
CNN model is trained by simple interacting samples between two
negative features. Following this training method, in order to be
more comprehensive, it needs to incorporate more complex
interactions of multiple features, which could make the training
cost unacceptable. However, since the HPM deals with the
interactions at vertex level, it can build the training dataset by
directly removing the vertices in mesh models instead of cutting by
another feature.
In summary, the future research goal is to build a more
complete feature recognition system with a hierarchical structure
to support multiple feature definitions. For instance, at the top
level, the features can be classified as primitive features as defined
in section 3. At the lower level, features in the same primitive class
can be identified based on the geometry variations, such as T-slot,
V-slot, dovetail slot, etc. Also, incorporating the freeform features
is very attractive to this domain. Considering the strengths of HPM
and CNN, it is believed that the proposed method can be applied to
the recognition of 3D real-life object. In addition, to optimize the
algorithms and CNN structure, have more sophisticated training
datasets and more rigorous experiments, to continue to enhance
the implementation to other restrictions of feature recognition,
and to extend the results and application to include more
processing techniques and exploring for other possibilities in
the CAD/CAM domain are future research intension of the authors.
References
[1] Gupta, S.K., 1994, Automated manufacturability analysis of machined parts..
[2] Regli, W.C., Gupta, S.K., Nau, D.S., 1994, Feature Recognition for Manufactur-
ability Analysis..
[3] Shi, Y., Zhang, Y., Baek, S., De Backer, W., Harik, R., 2018, Manufacturability
analysis for additive manufacturing using a novel feature recognition tech-
nique. Computer-Aided Design & Applications, 14:1–14. http://dx.doi.org/
10.1080/16864360.2018.1462574org/10.1080/16864360.2018.1462574.
[4] Xú, S., Anwer, N., Mehdi-Souzani, C., Harik, R., Qiao, L., 2016, STEP-NC based
reverse engineering of in-process model of NC simulation. The International
Journal of Advanced Manufacturing Technology, 86/9–12: 3267–3288.
[5] Allada, V., Anand, S., 1995, Feature-based modelling approaches for integrated
manufacturing: state-of-the-art survey and future research directions. Inter-
national Journal of Computer Integrated Manufacturing, 8/6: 411–440.
[6] Yildiz, A.R., Öztürk, N., Kaya, N., Öztürk, F., 2003, Integrated optimal topology
design and shape optimization using neural networks. Structural and Multi-
disciplinary Optimization, 25/4: 251–260.
[7] Zhang, Y., Bernard, A., Gupta, R.K., Harik, R., 2016, Feature based building
orientation optimization for additive manufacturing. Rapid Prototyping
Journal, .
[8] Chandrasegaran, S.K., Ramani, K., Sriram, R.D., Horváth, I., Bernard, A., Harik, R.
F., Gao, W., 2013, The evolution, challenges, and future of knowledge repre-
sentation in product design systems. Computer-Aided Design, 45/2: 204–228.
[9] Han, J., Kang, M., Choi, H., 2001, STEP-based feature recognition for
manufacturing cost optimization. Computer-Aided Design, 33/9: 671–686.
[10] Harik, R., Capponi, V., Lombard, M., Ris, G., 2006, Enhanced functions support-
ing process planning for aircraft structural parts. Computational Engineering
in Systems Applications, 2:1259–1266.
[11] Harik, R.F., Derigent, W.J.E., Ris, G., 2007, Spécifications de fonctions pour un
système d’aide à la génération automatique de gamme d’usinage: application
aux pièces aéronautiques de structure, prototype logiciel dans le cadre du
projet RNTL USIQUICK [Université Henri Poincaré-Nancy I]. Computer-Aided
Design and Applications, . http://dx.doi.org/10.3722/cadaps.2008.953-962.
(Vol. 5, Issue 6).
[12] Harik, R.F., Derigent, W.J.E., Ris, G., 2008, Computer aided process planning in
aircraft manufacturing. Computer-Aided Design and Applications, . http://dx.
doi.org/10.3722/cadaps.2008.953-962. (Vol. 5, Issue 6, pp. 953–962).
[13] Lin, A.C., Lin, S.-Y., 1998, A volume decomposition approach to process planning
for prismatic parts with depression and protrusion design features. Interna-
tional Journal of Computer Integrated Manufacturing, 11/6: 548–563.
[14] Narang, R.V., 1996, An application-independent methodology of feature rec-
ognition with explicit representation of feature interaction. Journal of Intelli-
gent Manufacturing, 7/6: 479–486.
[15] Wang, M.-T., Chamberlain, M.A., Joneja, A., Chang, T.-C., 1993, Manufacturing
feature extraction and machined volume decomposition in a computer-inte-
grated feature-based design and manufacturing planning environment. Com-
puters in Industry, 23/1–2: 75–86.
[16] Fu, M.W., Ong, S.-K., Lu, W.F., Lee, I.B.H., Nee, A.Y.C., 2003, An approach to
identify design and manufacturing features from a data exchanged part model.
Computer-Aided Design, 35/11: 979–993.
[17] Harik, R.F., Sahmrani, N., 2010, DFMA+, a quantitative DFMA methodology.
Computer-Aided Design and Applications, 7/5: 701–709.
[18] Abouel Nasr, E.S., Kamrani, A.K., 2006, A new methodology for extracting
manufacturing features from CAD system. Computers and Industrial Engineer-
ing, . http://dx.doi.org/10.1016/j.cie.2006.08.004.
[19] Dimov, S.S., Brousseau, E.B., Setchi, R., 2007, A hybrid method for feature
recognition in computer-aided design models. Proceedings of the Institution
of Mechanical Engineers Part B: Journal of Engineering Manufacture, 221/1:
79–96.
 Y. Shi et al. / CIRP Journal of Manufacturing Science and Technology 30 (2020) 36–57
[20] Jones, T.J., Reidsema, C., Smith, A., 2006, Automated Feature Recognition
System for supporting conceptual engineering design. International Journal
of Knowledge-Based and Intelligent Engineering Systems, 10/6: 477–492.
[21] Wong, T.N., Lam, S.M., 2000, Automatic recognition of machining features from
computer aided design part models. Proceedings of the Institution of Mechan-
ical Engineers Part B: Journal of Engineering Manufacture, 214/6: 515–520.
[22] Han, J.H., Pratt, M., Regli, W.C., 2000, Manufacturing feature recognition from
solid models: A status report. IEEE Transactions on Robotics and Automation, .
http://dx.doi.org/10.1109/70.897789.
[23] Shi, Y., Zhang, Y., Xia, K., Harik, R., 2020, A critical review of feature recognition
techniques. Computer-Aided Design and Applications, 17/5: 861–899. http://
dx.doi.org/10.14733/cadaps.2020.861-899.
[24] Verma, Arvind Kumar, Rajotia, S., 2010, A review of machining feature recog-
nition methodologies. International Journal of Computer Integrated
Manufacturing, 23/4: 353–368.
[25] Harik, R., Shi, Y., Baek, S., 2017, Shape Terra: mechanical feature recognition
based on a persistent heat signature. Computer-Aided Design and Applica-
tions, 14/2: 206–218. http://dx.doi.org/10.1080/16864360.2016.1223433.
[26] Hummel, K.E., 1989, Coupling rule-based and object-oriented programming
for the classification of machined features. NASA STI/Recon Technical Report
N, 90.
[27] Donaldson, I.A., Corney, J.R., 1993, Rule-based feature recognition for 2 5D
machined components. International Journal of Computer Integrated
Manufacturing, 6/1–2: 51–64.
[28] Vosniakos, G.C., Davies, B.J., 1993, A shape feature recognition framework and
its application to holes in prismatic parts. The International Journal of Ad-
vanced Manufacturing Technology, 8/6: 345–351.
[29] Nasr, E.S., Abouel Khan, A.A., Alahmari, A.M., Hussein, H.M.A., 2014, A feature
recognition system using geometric reasoning. Procedia CIRP, 18:238–243.
http://dx.doi.org/10.1016/j.procir.2014.06.138.
[30] Chan, A.K.W., Case, K., 1994, Process planning by recognizing and learning
machining features. International Journal of Computer Integrated Manufactur-
ing, 7/2: 77–99. http://dx.doi.org/10.1080/09511929408944597.
[31] Bhandarkar, M.P., Nagi, R., 2000, STEP-based feature extraction from STEP
geometry for Agile Manufacturing. Computers in Industry, 41/1: 3–24. http://
dx.doi.org/10.1016/S0166-3615(99)00040-8.
[32] Gao, J., Zheng, D.T., Gindy, N., 2004, Extraction of machining features for CAD/
CAM integration. The International Journal of Advanced Manufacturing Tech-
nology, 24/7–8: 573–581.
[33] Verma, A.K., Rajotia, S., 2008, A hint-based machining feature recognition
system for 2.5D parts. International Journal of Production Research, . http://dx.
doi.org/10.1080/00207540600919373.
[34] Xu, X., Hinduja, S., 1998, Recognition of rough machining features in 212D
components. Computer-Aided Design, 30/7: 503–516.
[35] Nasr, Emad S. Abouel, Kamrani, A.K., 2006, A new methodology for extracting
manufacturing features from CAD system. Computers & Industrial Engineer-
ing, 51/3: 389–415.
[36] Vandenbrande, J.H., Requicha, A.A.G., 1993, Spatial reasoning for the automatic
recognition of machinable features in solid models. IEEE Transactions on
Pattern Analysis and Machine Intelligence, 15/12: 1269–1285.
[37] Han, J., Requicha, A.A.G., 1997, Integration of feature based design and feature
recognition. Computer-Aided Design, 29/5: 393–403.
[38] Joshi, S., Chang, T.C., 1988, Graph-based heuristics for recognition of machined
features from a 3D solid model. Computer-Aided Design, 20/2: 58–66. http://
dx.doi.org/10.1016/0010-4485(88)90050-4.
[39] Laakko, T., Mäntylä, M., 1993, Feature modelling by incremental feature
recognition. Computer-Aided Design, . http://dx.doi.org/10.1016/0010-4485
(93)90079-4.
[40] Venuvinod, P.K., Yuen, C.F., Merchant, M.E., 1994, Efficient Automated Geo-
metric Feature Recognition through Feature Coding. CIRP Annals - Manufactur-
ing Technology, . http://dx.doi.org/10.1016/S0007-8506(07)62242-2.
[41] Venuvinod, Patri K, Wong, S.Y., 1995, A graph-based expert system approach to
geometric feature recognition. Journal of Intelligent Manufacturing, 6/3: 155–
162.
[42] Yuen, C.F., Venuvinod, P.K., 1999, Geometric feature recognition: Coping with
the complexity and infinite variety of features. International Journal of Com-
puter Integrated Manufacturing, . http://dx.doi.org/10.1080/
095119299130173.
[43] Gao, S., Shah, J.J., 1998, Automatic recognition of interacting machining fea-
tures based on minimal condition subgraph. CAD Computer Aided Design, .
http://dx.doi.org/10.1016/S0010-4485(98)00033-5.
[44] Li, Y.G., Ding, Y.F., Mou, W.P., Guo, H., 2010, Feature recognition technology for
aircraft structural parts based on a holistic attribute adjacency graph. Proceed-
ings of the Institution of Mechanical Engineers Part B: Journal of Engineering
Manufacture, 224/2: 271–278.
[45] Cai, N., Bendjebla, S., Lavernhe, S., Mehdi-Souzani, C., Anwer, N., 2018, Freeform
Machining Feature Recognition with Manufacturability Analysis. Procedia
CIRP, . http://dx.doi.org/10.1016/j.procir.2018.03.261.
[46] Woo, T.C., 1982, Feature extraction by volume decomposition. Conference on
CAD/CAM Technology in Mechanical Engineering, 76–94.
[47] Kim, Y.S., Wilde, D.J., 1992, A convergent convex decomposition of polyhedral
objects. Journal of Mechanical Design, 114/3: 468–476.
57
[48] Sakurai, H., Chin, C.-W., 1994, Definition and recognition of volume features for
process planning.(Vol. 20, pp. 65–80) Elsevier.
[49] Woo, Y., Sakurai, H., 2002, Recognition of maximal features by volume de-
composition. CAD Computer Aided Design, . http://dx.doi.org/10.1016/S0010-
4485(01)00080-X.
[50] Chen, Y.H., Lee, H.M., 1998, A neural network system for two-dimensional
feature recognition. International Journal of Computer Integrated Manufactur-
ing, 11/2: 111–117. http://dx.doi.org/10.1080/095119298130859.
[51] Jun, Y.T., Raja, V.H., 2002, Extracting geometric attributes directly from scanned
data sets for feature recognition. International Journal of Computer Integrated
Manufacturing, 15/1: 50–61.
[52] Ding, L., Yue, Y., 2004, Novel ANN-based feature recognition incorporating
design by features. Computers in Industry, 55/2: 197–222.
[53] Nezis, K., Vosniakos, G., 1997, Recognizing 212D shape features using a neural
network and heuristics. Computer-Aided Design, . http://dx.doi.org/10.1016/
S0010-4485(97)00003-1.
[54] Prabhakar, S., Henderson, M.R., 1992, Automatic form-feature recognition
using neural-network-based techniques on boundary representations of solid
models. Computer-Aided Design, . http://dx.doi.org/10.1016/0010-4485(92)
90064-H.
[55] Hwang, J.-L., 1992, Applying the perceptron to three-dimensional feature
recognition..
[56] Lankalapalli, K., Chatterjee, S., 1997, Feature recognition using ART2: a self-
organizing neural network. Journal of Intelligent Manufacturing, 8/3: 203–214.
[57] Marquez, M., Gill, R., White, A., 1999, Application of neural networks in feature
recognition of mould reinforced plastic parts. Concurrent Engineering-Re-
search and Applications, . http://dx.doi.org/10.1177/1063293X9900700204.
[58] Sunil, V.B., Pande, S.S., 2009, Automatic recognition of machining features
using artificial neural networks. The International Journal of Advanced
Manufacturing Technology, 41/9–10: 932–947. http://dx.doi.org/10.1007/
s00170-008-1536-z.
[59] Li, W.D., Ong, S.K., Nee, A.Y.C., 2000, Recognition of overlapping machining
features based on hybrid artificial intelligent techniques. Journal of Engineer-
ing Manufacture, 214/8: 739–744.
[60] Öztürk, N., Öztürk, F., 2001, Neural network based non-standard feature
recognition to integrate CAD and CAM. Computers in Industry, . http://dx.
doi.org/10.1016/S0166-3615(01)00090-2.
[61] Wu, M.C., Jen, S.R., 1996, A neural network approach to the classification of 3D
prismatic parts. The International Journal of Advanced Manufacturing Tech-
nology, 11/5: 325–335.
[62] Chuang, J., Wang, P., Wu, M., 1999, Automatic classification of block-shaped
parts based on their 2D projections. Computers and Industrial Engineering, .
[63] Rahmani, Keyvan, Arezoo, B., 2006, Boundary analysis and geometric comple-
tion for recognition of interacting machining features. CAD Computer Aided
Design, 38/8: 845–856. http://dx.doi.org/10.1016/j.cad.2006.04.015.
[64] Rahmani, K., Arezoo, B., 2007, A hybrid hint-based and graph-based framework
for recognition of interacting milling features. Computers in Industry, . http://
dx.doi.org/10.1016/j.compind.2006.07.001.
[65] Verma, Arvind Kumar, Rajotia, S., 2009, A hybrid machining Feature Recogni-
tion system. International Journal of Manufacturing Research, 4/3: 343–361.
[66] Lam, S.M., Wong, T.N., 2000, Recognition of machining features-a hybrid
approach. International Journal of Production Research, 38/17: 4301–4316.
[67] Tseng, Y.J., Joshi, S.B., 1994, Recognizing multiple interpretations of interacting
machining features. Computer-Aided Design, 26/9: 667–688. http://dx.doi.
org/10.1016/0010-4485(94)90018-3.
[68] Sakurai, H., Dave, P., 1996, Volume decomposition and feature recognition, part
II: curved objects. Computer-Aided Design, 28/6–7: 519–537. http://dx.doi.org/
10.1016/0010-4485(95)00067-4.
[69] Rameshbabu, V., Shunmugam, M.S., 2009, Hybrid feature recognition method
for setup planning from STEP AP-203. Robotics and Computer-Integrated
Manufacturing, 25/2: 393–408.
[70] Riegel, Juergen, Mayer, Werner, van Havre, Y., 2018, FreeCAD (0.18).https://
www.freecadweb.org.
[71] Shewchuk, J.R., 1996, Triangle: Engineering a 2D quality mesh generator and
Delaunay triangulator.Lin MM.C.C, Manocha DD., (Eds.) Applied Computation-
al Geometry: Towards Geometric Engineering, 1148. Springer-Verlag. p.
pp.203–222.
[72] Sun, J., Ovsjanikov, M., Guibas, L., 2009, A concise and provably informative
multi-scale signature based on heat diffusion. Eurographics Symposium on
Geometry Processing, 28/5: 1383–1392. http://dx.doi.org/10.1111/j.1467-
8659.2009.01515.x.
[73] Tixier, A.J., 2018, Notes on Deep Learning for NLP.1–20.
[74] Grover, A., Leskovec, J., 2016, node2vec: Scalable Feature Learning for Net-
works.. http://dx.doi.org/10.1145/2939672.2939754.
[75] Hamilton, W.L., Ying, R., Leskovec, J., 2017, Representation learning on graphs:
Methods and applications. ArXiv Preprint, . ArXiv:1709.05584.
[76] Tixier, A.J.-P., Nikolentzos, G., Meladianos, P., Vazirgiannis, M., 2018, Graph
Classification with 2D Convolutional Neural Networks..
[77] LeCun, Y., Bottou, L., Bengio, Y., Haffner, P., 1998, Gradient-based learning
applied to document recognition. Proceedings of the IEEE, . http://dx.doi.org/
10.1109/5.726791.