Robotics and Computer–Integrated Manufacturing 73 (2022) 102222

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing

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

An automatic method for constructing machining process knowledge base

from knowledge graph
Liang Guo a, *, Fu Yan a, Tian Li b, Tao Yang a, Yuqian Lu c
a
School of Mechatronic Engineering, Southwest Petroleum University, Chengdu, China
b
AECC Chengdu Engine Co., Ltd., Chengdu 610503, China
c
Department of Mechanical Engineering, The University of Auckland, Auckland, New Zealand

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

Keywords: The process knowledge base is the key module in intelligent process design, it determines the intelligence degree
Process Knowledge Base of the design system and affects the quality of product design. However, traditional process knowledge base
Knowledge Graph construction is non-automated, time consuming and requires much manual work, which is not sufficient to meet
Fuzzy Evaluation
the demands of the modern manufacturing mode. Moreover, the knowledge base often adopts a single knowledge
NLP
representation, and this may lead to ambiguity in the meaning of some knowledge, which will affect the quality
of the process knowledge base. To overcome the above problems, an automatic construction framework for the
process knowledge base in the field of machining based on knowledge graph (KG) is introduced. First, the
knowledge is classified and annotated based on the function-behavior-states (FBS) design method. Second, a
knowledge extraction framework based on BERT-BiLSTM-CRF is established to perform the automatic knowledge
extraction of process text. Third, a knowledge representation method based on fuzzy comprehensive evaluation is
established, forming three types of knowledge representation with the KG as the main, production rules and two-
dimensional data linked list as a supplement. In addition, to overcome the redundancy in the knowledge fusion
stage, a hybrid algorithm based on an improved edit distance and attribute weighting is built. Finally, a pro­
totype system is developed, and quality analysis is carried out. Compared with the F values of BiLSTM-CRF and
CNN-BiLSTM-CRF, that of the proposed extraction method in the machining domain is increased by 7.35% and
3.87%, respectively.

1. Introduction experience of various interdisciplinary experts. In addition, by sharing,

exchanging, and reusing design knowledge in a distributed resource
The process knowledge base is the core component of intelligent environment, the process knowledge base can make full use of a great
process design, and it determines the degree of intelligence of the design deal of the tacit knowledge of enterprises and can then configure and
system and affects the quality of process reasoning. In the information optimize the human and manufacturing resources of enterprises
age, the development of modern products is facing unprecedented accordingly in order to minimize resource waste and time loss during
challenges. For modern products in the machinery field and other in­ engineering design [4]. Therefore, we think that in the intelligent
dustries, there is enormous pressure from the market in terms of de­ manufacturing environment, the life cycle of manufactured products
mands for shorter lead time and stricter quality requirement [1]. In should not be calculated only in a narrow sense, and the construction
practice, this means that products of increasing complexity have to be time of the knowledge base should also be included in a broad product
developed in a shorter time frame [2]. As an important part of the life cycle. For this reason, how to construct a high-quality process
modern intelligent decision-making system for process, the process knowledge base in a shorter period of time is essential for achieving
knowledge base is also a bridge connecting product design and product shorter cycle development and process design of complex products in
manufacturing in the modern manufacturing industry [3]. It directly modern manufacturing models.
affects the organizational efficiency of production, the product cost and However, the application of a knowledge base in product
the production cycle of an enterprise by integrating the knowledge and manufacturing is still considered a challenging task for the following

* Corresponding author.
E-mail address: [email protected] (L. Guo).

https://doi.org/10.1016/j.rcim.2021.102222
Received 20 November 2020; Received in revised form 21 June 2021; Accepted 19 July 2021
Available online 27 August 2021
0736-5845/© 2021 Elsevier Ltd. All rights reserved.
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

reasons: On one hand, product manufacturing is one of the most time- finance and agriculture and has very broad prospects. In recent years,
consuming and knowledge-intensive tasks across the modern product the automatic construction of a domain knowledge base has become a
development process, which involves a series of decision-making ac­ popular issue and has attracted many scholars to conduct research. This
tivities. Most of these activities are highly dependent upon personal section mainly reviews related work on the automatic construction of a
knowledge accumulated from the previous projects over a long period of knowledge base: natural language processing, knowledge representation
time and enterprise equipment resources [3,5]. Because format diversity and the application of knowledge bases.
and heterogeneity of manufacturing-related knowledge among different
service providers (CATIA/UG/PROE), effective sharing and reuse of 2.1. Natural language processing
these knowledge become quite difficult. As a result, traditional con­
struction of a knowledge base often need to be obtained by manual As an interdisciplinary research field, natural language processing
methods, which is time consuming and labor intensive, resulting in (NLP) involves language and speech processing, human language tech­
difficulty in meeting the requirements of automatic construction of nology, computational linguistics, and speech recognition and synthesis
knowledge base [6]. Studies have shown that to acquire supporting in­ [8].With the advent of the information age and the rapid development of
formation from different sources, engineers spend approximately 40% of computing technology, NLP has played an important role in various
their time searching and reorganizing raw data [7]. Therefore, if the applications, including dialogue agents and dialogue systems [9], ma­
automated construction of the knowledge base cannot be completed, a chine translation [10], knowledge mining and reasoning [11], and
lot of time will be spent in the preparation phase to manually create the search engines [12]. Although NLP has made great progress in the
knowledge base, which greatly increases the product manufacturing general field, its application in the machining field still faces great
cycle. On the other hand, the knowledge applied to automatic process challenges. Due to the complexity of mechanical proprietary entities and
decision-making system is more complicated than other applications the difficulty of defining feature boundaries, the traditional knowledge
such as fault detection, and it is difficult for a single knowledge repre­ base construction methods in the machining field have low accuracy, are
sentation form (ontology, triples, etc.) to properly represent the existing time consuming and require much manual work, which completely fails
design knowledge. The lack of formal representations of the existing to meet the requirements of automated construction [8].
design knowledge makes it extremely difficult for even experienced Early research on NLP tended to focus on word processing (including
design engineers to trace past design routines. As a consequence, inef­ representation and calculation). Regular expressions [13], finite state
ficient retrieval of the design knowledge (e.g., with respect to the automata [14], sensors [15], the N-gram model [16], part-of-speech tags
format, type, and content), as well as incorrect interpretations (due to [17], the hidden Markov model [18] and maximum entropy model [19]
the specific domain), can lead to cost- and time-intensive iterations in are the most widely used models for dealing with vocabulary-related
product development, rendering the knowledge base ineffective for issues. After better solutions were found for vocabulary processing,
practical use [8]. scholars carried out related research on grammar processing. For
Hence, in order to solve the above problems, the focus in this work is grammar processing, early researchers used common tools such as
developing a framework for the knowledge based engineering system formal grammars and treebanks [20]. With the deepening of research,
that supports the automated construction of a knowledge base and probability models and formalisms based on attribute constraints [21]
related design activities. The main contributions of this research are have become more widely used. The current research on NLP focuses on
threefold. First, we propose an architecture for automatically con­ semantic and pragmatic processing. First-order logic, semantic net­
structing a knowledge base in the machining field. The novelty of this works, concept dependency graphs, etc. are used to represent semantic
architecture is that it allows the automatic extraction of manufacturing information [22].
knowledge from existing simulation models and process documents. Recently, with the development of artificial intelligence technology,
Second, based on the function-behavior-states (FBS) design framework, machine learning technology [23,24] and statistical learning technology
various data and knowledge types in the machining field are summa­ have been introduced into NLP. Through a co-occurrence matrix and
rized and discussed, and the knowledge representation method is judged singular value decomposition, the vocabulary is transformed into a
by the fuzzy comprehensive evaluation algorithm. Finally, a knowledge dense vector. Word2vec, developed by Google [25], was the earliest
extraction system based on the browser/server (B/S) architecture is word vector representation. Through Continuous Bag-of-Words Model
developed, which not only performs feature/part process reasoning in (CBOW) or Continuous Skip-gram Model (Skip-gram), words can be
the application but also supports process engineers in seamlessly inte­ vectorized based on context [26]. However, Word2vec considers only
grating data topics in other Web environments; this system can assist adjacent contexts. Although it can be established quickly and directly,
with other manufacturing stages, technologies and application areas in its ability to differentiate semantically is limited. Because of this defi­
product development. ciency, GloVe [27] was proposed; it uses global vectors to represent
The structure of this paper is as follows: Section 2 reviews the related words and can consider local and global information in vector training
work of knowledge base automatic construction technology. The overall through global co-occurrence matrices. Although GloVe is more
construction methods and framework of a knowledge base for the powerful than Word2vec, it is difficult to adapt it to different environ­
machining field are introduced in Section 3. Section 4 focuses on the ments. To address this problem, researchers have proposed employing
knowledge extraction model in the machining field. The choice of the multiple long short-term memory (LSTM) stacked neural models to learn
knowledge representation scheme based on a fuzzy comprehensive morphological features at different levels (ELMo) [28]. This method
evaluation algorithm is described in Section 5. Section 6 focuses on the synthetically considers the needs of disambiguation, parts of speech and
elimination of redundancy in the knowledge base by improving the grammar. In addition, it can be easily extended to specific applications
similarity judgement of the editing distance in the knowledge fusion through pre-training and fine-tuning. However, due to the characteris­
stage. A prototype system is developed, and the application and tics of LSTM, training the ELMo model is a sequential process. Therefore,
extraction effects of the system are discussed in Section 7. Section 8 the earlier contexts have a higher priority than the later contexts. Sub­
presents the conclusions. sequently, the BERT model was proposed to learn morphemes from
multiple attention nodes through a two-way encoder. The BERT model
2. Related research has been extensively validated and can be directly used to extract
morpheme features from language texts through pre-training.
As the brain of a modern intelligent system, the knowledge base
provides a large knowledge reserve and support for intelligent infor­
mation processing. It has been widely used in the fields of medicine,

2
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

2.2. Knowledge representation construction of the knowledge base in the knowledge based engineering
(KBE) system by scholars. In the field of machining process, Shen et al.
Expert knowledge extracted from documents by text mining or [42] adopted EXPRESS language to design the knowledge base and
recorded in expert interviews with experienced simulation engineers has application framework of welding data task model in detail. The
to be formalized so that it can be provided in a knowledge base [29]. The knowledge model includes information on accurate geometry, dimen­
purpose of exploring knowledge representations is to store knowledge in sion and welding requirement, which is enough to support intelligent
a computer-readable structured representation and express its meaning process planning in robot welding field. However, due to the lack of
most appropriately. In addition, establishing an effective knowledge semantic description and consideration in EXPRESS model language,
representation model to achieve sharing and integration between en­ knowledge sharing and data exchange are still hindered to a certain
terprises has always been the focus of knowledge management research extent. Wan et al. [43] uses ontology to build domain knowledge base,
in engineering design. The latest knowledge representation models can and proposesd a knowledge-based maintenance planning system. The
be classified as described below. system adopted Adaptation-Guided Retrieval (AGR), Case-based
The research on knowledge representation models can be traced Reasoning (CBR) and semantic description as retrieval methods, which
back to the 1990s. Early research mainly focused on function-based improved the effectiveness in managing structured and unstructured
knowledge representation models. The most representative model is knowledge over traditional engineering data management systems.
Gero’s functional behavior structure (FBS) model, which describes the Nevertheless, it does not take into account the knowledge that are not or
main components involved in engineering design activities [30]. The cannot be represented in this format, so it has limitations on the singular
FBS model combines the three main components of design in order to mapping in case retrieval. In addition, Zhang et al. [44] proposed an
effectively combine the design model with the knowledge of design ontology-based machining process knowledge representation model for
experts. Similar models include the functional representation (FR) a feature-based machining operation (such as turning or milling),
model and structural behavior function (SBF) model [31,32], which thereby constructing a machining knowledge base. Eum et al. [45]
have also attracted the application of many scholars [33–35]. Later, proposed an ontology-based process planning knowledge modeling
influenced by process-oriented thinking, scholars discovered that design method. This method utilizes relevant attributes to match the knowl­
knowledge can be described using components of the design process in edge of features, machining methods, and machining capabilities, and it
the form of linked process models with associated data and information. can perform knowledge-based selection of corresponding machining
For instance, design resources can be defined at different design stages methods for multi-axis machining features. However, these studies are
according to different manufacturing targets and then associated with more focused on using limited methods to build knowledge bases, which
various types of production planning knowledge [36,37]. However, for lack a framework for automated construction of knowledge bases. With
complex fields, especially the machining field, a process-oriented the development of ontology technology, Helgoson and Kalhori [46]
knowledge representation is difficult to expand and reuse, and its ac­ proposed a conceptual knowledge model which includes an adoption of
curacy is often low, which makes it difficult to use in practical appli­ the methodology of knowledge acquisition (MOKA) and contains the
cations. Because of this deficiency, scholars have tried using activities as to identify, justify, capture, formalize, package, and activate
object-oriented knowledge representations. A typical example is Mod­ for sharing and integration of knowledge in process planning to increase
elica, which uses classes to define the data and behavior of objects. Then, the level of efficiency, reliability, and productivity in process planning
the classes in Modelica use domains and equations to represent the so­ process. Liang [34] proposed a new AM OntoProc knowledge model
lution result and the solution process, respectively. In this way, the rules, based on MOKA, which could construct a process planning knowledge
equations and documents in the database can be converted into base in assisted additive manufacturing (AM). But, the system does not
multi-domain knowledge. Poorkiany et al. [38] introduced an object fully consider the manufacturing data exchange and semantic analysis in
diagram that integrates SolidWorks, Word, Excel and Wiki pages to the knowledge model. In recent research based on the knowledge base in
capture, construct and access design principles during product and tool the KBE system, Sormaz [47] used a novel foundation ontologies
design. In addition, Grabowik et al. [39] used object-oriented methods (upper-level) to construct a manufacturing process planning knowledge
to express knowledge. base model, which can realize the interoperability with other
Recently, in Paulheim [40], ontology was defined as a clear speci­ upper-layer manufacturing ontology and the process planning design
fication of (shared) conceptualizations. It defines a general under­ case of prism parts in the cloud manufacturing environment. Never­
standing of a domain in a machine-readable way. In addition to its theless, the knowledge type of the system is mainly based on the upper
machine-readable form, the benefit of ontology is a unique under­ ontology and semantic rules, there is still room for research on the
standing of an explicit representation of knowledge, which is achieved representation and complete semantic integration of process knowledge
through formal and semantic definitions of terms and axioms. Beside, which is not suitable for ontology representation. In addition, with the
due to its linguistic nature, ontology is particularly suitable for mapping deep research and successful application of knowledge graph technology
knowledge obtained from natural language documents. On the basis of in recent years, many scholars have constructed knowledge bases by
ontology, the knowledge graph proposed by Google becomes more using knowledge graphs. Zhang et al. [48] proposed an extraction
promising, and it can encode the semantic information of knowledge method based on the semantic distance of subgraphs and constructed a
entities in consideration of their physical connection strengths, de­ knowledge base in the field of metal materials by using knowledge
pendencies and causality [40]. This idea is used in the construction of graphs. Zhou et al. [49] proposed a knowledge graph-driven assembly
various knowledge bases, such as DBpedia, YAGO, Freebase, Wikidata, process generation and evaluation method for complex components,
Spark, Google’s Knowledge Vault, Microsoft’s Satori and Facebook’s which provides assembly decision-making support for the evaluation
entity graphs [41]. Overall, the knowledge graph provides a new way to method of interference detection of assembly sequence based on point
represent knowledge. cloud assembly feature recognition. Li et al. [50] aimed to address the
difficulty that process knowledge is difficult to reuse and share, and a
2.3. Construction of knowledge base in the machining field knowledge base in the field of machining was constructed by using
knowledge graphs to represent process knowledge in a unified way.
The above research has introduced two key technologies in the Kertkidkachorn et al. [51] proposed a rule-based and similarity-based
automatic construction of knowledge base: natural language processing method by using knowledge graphs in the T2KG framework to com­
and formal representation of knowledge. In order to more accurately plete the construction of the knowledge base. Zhou et al. [52] presented
realize the automatic construction framework of the knowledge base in a unified knowledge graph-driven production resource allocation
the machining field, it is also necessary to explore and review the approach, which fully integrated the complex data of workshop

3
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Fig. 1. Label classification based on the FBS framework.

resources and realized the rapid decision-making for resource allocation and automatically construct the knowledge base in the machining field.
in the manufacturing field.
In summary, although these KBE systems have made many attempts 3. Automatic construction framework for the machining
to capture process planning knowledge in various domains, very few knowledge base
researches actually focused on developing an automated model for
capturing process planning knowledge, which is independent of do­ 3.1. Classification of process knowledge
mains. Most of the literature utilizes limited and manually created
ontology or knowledge graph structures to obtain, capture and formalize The classification of process knowledge directly affects the quality of
simulation knowledge, and an automatic framework for constructing a the knowledge base. To ensure the rapid and accurate acquisition of
knowledge base to obtain target-oriented process knowledge in me­ different types of process knowledge in the process of automatic
chanical product design is lacking. In short, these developed systems do knowledge base construction, the process knowledge is divided into the
not automatically acquire knowledge from text-based documents. In following three categories according to the knowledge classification
addition, the knowledge representation form adopted by most systems is method in [50]:
often single; that is, a knowledge representation form is used to express
all process knowledge. However, regardless of whether knowledge (1) Rule-based process knowledge
graph technology based on triples or ontology technology, it cannot
accurately and appropriately express every piece of knowledge. For From the perspective of rule reasoning, rule-based knowledge is the
example, considering “When turning the step shaft, in order to ensure basis of process planning. It is mainly derived from process design
rigid turning, the larger-diameter part should be turned first, and then manuals, various engineering standards, expert experience and knowl­
the smaller-diameter part”. In this textual knowledge resource, it is edge, and it is the set of correct knowledge that has been verified. At
evident that the triple representation of the knowledge graph cannot present, it mainly relies on manual summary. Examples include the se­
accurately express this knowledge. Therefore, because of the de­ lection rules of machining methods (face before the hole, first rough and
ficiencies in the abovementioned research, this paper proposes a method then refined, etc.). This knowledge has important guiding significance
of automatically constructing a knowledge base for the machining field. for machining process decision-making.
The innovation of this method is to allow the automatic extraction of
manufacturing knowledge from existing process documents as well as to (2) Same-property process template knowledge
modify and expand text and data mining methods to automate knowl­
edge acquisition. In addition, three knowledge representation methods The same-property process template knowledge is based on the
are used to comprehensively judge the knowledge representation similarity of the set and process content, and it is generated by gathering
through the fuzzy comprehensive evaluation algorithm to realize the and classifying the processing technology of typical parts, features and
sharing and integration between enterprises in knowledge management local structures. For example, for the end face of grey cast iron, if the

4
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Table 1 In this paper, we adopt the following definitions of the key compo­
Labeling classification and examples. nents of the FBS ontology. Some of these definitions can be found in
Knowledge Category Symbolic Concrete Examples original works by Umeda et al. and by Gero; others have been derived
Representation from works on qualitative physics [53] and from related works [54-56]
Method Condition MC Requirements that should be met; that complete the FBS framework. According to the above research
(Functions) requirements; demand foundation and the processing of patent documents in Fantoni [57],
Purpose (Functions) P To; thus; purpose machining domain knowledge documents can be classified into entity
Calculation Formula CF F=m*a tags, as shown in Fig. 1.
(Behaviors)
Theorem (Behaviors) T Coulomb’s law of friction
Functions are interpretations of physical behaviors according to the
Feature (States) F Shaft; hole; end face; groove user’s goals. While Umeda et al. [58] define functions as ‘‘descriptions of
Equipment (States) E Lathes; drilling machines, milling behavior recognized by a human through abstraction in order to utilize
machines it’’, Gero [59] describes them as the motivation for the product existence
Method (States) M Rough machining; rough turning
or, more generally, ascribes them to teleology (what the object is for).
Cutting Elements CE Feed amount; back cutting depth
(States) Therefore, the purposes and motivations of the machining field included
Material (States) MT Carbon steel; white steel in the top-level function framework in Fig. 1 are mainly divided into
Feature Attributes FA Accuracy grade, surface roughness, three entity labels: goals, needs and functional requirements.
(States) etc. Behaviors are the “physical phenomena” that cause a change in the
Equipment Attributes EA Spindle speed; maximum cutting
(States) length
‘‘states’’ of a system. Both Hirschberg and Manning [22] and Russo and
Material Attributes MTA Heat transfer; gloss, stiffness; Montecchi [60] introduced Physical Effects into this view. In our view,
(States) behavior is an equation described in natural language (belonging to
Processing PP Feed amount: 2 mm physics, chemistry, mechanics, etc.) to describe the evolution of a system
Parameters (States)
(machining field). Therefore, under the framework of behavior, the
entity tags in the machining field are divided into three types: physical
machining accuracy is level 6 and the surface roughness is 6.3, then the events, behaviors and interactions. To facilitate understanding, a few
recommended machining method is rough machining. arrows are added from the entity tag class to the abstract instance in
Fig. 1, and the FBS framework can be better understood through an
(3) Personalized process instance knowledge explanation of this abstract instance. For example, the abstract examples
of physical event labels refer to axioms and theorems. Specific examples
Personalized process instance knowledge is mainly based on objec­ are the Joule effect, Coulomb friction law, Hertz contact, and other
tive facts and descriptive process knowledge generated by describing terms that imply structure and behavioral information.
mechanical entities or enterprise equipment. For example, the The state level is relatively easy to understand, and its abstract ex­
maximum turning diameter of the C630 ordinary lathe is 630 mm, the amples are consistent with common objects in the machining field. It is
maximum workpiece length is 3000*4000 mm, the main motor power is worth noting that to allow the computer to attain a better extraction
11 kw, the machine width is 1380 mm, and the height is 1710 mm. effect, this paper distinguishes the Attributes entity tag and Properties
entity tag. On the one hand, we prefer the term “Attributes” to be
defined as ‘‘the specific characteristics of a single part of the product”,
3.2. Knowledge labeling based on FBS such as the length, width and maximum cutting length. On the other
hand, we define “Properties” as the chemical, physical, mechanical,
For the field of machining, the key information of knowledge biological, etc. properties of the material the product is composed of (e.
extraction is more on the functional level, since the key elements are g. the Young’s modulus, resistance to acid or flame, transparency to
assumed to be the activities (functions) that the artefact must perform to certain light frequencies, thermal or electrical conductivity and
achieve its desired outputs. Studying functions as independent entities porosity). According to the label classification in Fig. 1, the specific la­
allows abstraction from particular technical solutions and therefore beling method is as shown in Table 1.
constitutes a compelling design strategy. Moreover, functions capture According to the above definition of machining process classification
both the physical action that produces the output and the desired goal.

Fig. 2. Automatic construction framework of the knowledge base in the machining field.

5
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Table 2 evaluation algorithm to select a process knowledge representation

Notations of the knowledge extraction framework. scheme to determine the best representation of each type of process
Notation Definition Notation Definition knowledge and finally obtain a preliminary process knowledge base.
Step 3: Knowledge fusion. Due to the wide range of knowledge
x sentence vector rn The n reverse LSTM
sources, the knowledge base constructed according to the two steps
etn Characteristics of the n yn Label sequence of the n
word sentence above has high redundancy, so it is necessary to use latent semantic
esn Characteristics of the n A Parameters of the CRF analysis, similarity calculations and attribute weighting to eliminate
sentence layer redundancy in the knowledge. First, the entity triples in the preliminary
epn Characteristics of the n Aij The transfer score from the knowledge base are mapped with the Protégé ontology library, and then
place i label to the j label, the semantic web rule language (SWRL) is used to represent the
Ck Input parameters of the k score(x, y) The model score for empirical rule knowledge. Finally, the data layer is instantiated to
word vector based on Bert sentence x with y label construct the final knowledge base.
sequences
In order to allow readers to better understand the model, the math­
ci The hidden layer state Y The set of all label
ematical representation of this manuscript is presented. The definition of
vector of i position
the notations used in the knowledge extraction framework is given in
pi ci is mapped to the k yj
′
The value of j true label in
dimension Y Table 2.
eXk The k word vector p(x, y) Training samples with y
determined by the BERT labels in sentence x 4. Knowledge extraction model in the machining domain
vocabulary provided by
Google This paper’s topic is oriented toward the field of machining, and
En The n feature vector log(p(y|x)) Log-likelihood function of realizes the knowledge extraction in the machining field through a
sequence p(x, y)
multistage neural network collaboration model, which is composed of
ln The n forward LSTM y∗ The optimal label sequence
obtained by maximizing
the vector expression, bidirectional LSTM (BiLSTM) layer and condi­
the log(p(y|x)) tional random field (CRF) layer. At the same time, in view of the problem
U The evaluation factor set rPR Membership function of PR that current scholars cannot accurately express process knowledge, the
of knowledge fuzzy comprehensive evaluation algorithm is used to determine the
representation scheme appropriate expression form of the extracted knowledge, which better
me The number of rKG Membership function of
addresses the problem that triples cannot properly represent some
mechanical entity tags KG
knowledge in the machining fields.
pm The number of processing m Reduction coefficient
method tags
pc The number of processing λ Correction coefficient 4.1. Preprocessing of data text
condition tags
pp The number of processing σ The control variable of the The natural text in the machining field is arranged in a scattered,
purpose tags knowledge graph
V Judgment set in fuzzy μ The control variable of the
complex and irregular pattern. If the information extraction method is
comprehensive evaluation knowledge graph used directly, it is difficult to obtain satisfactory knowledge extraction
processing label results. To this end, consistent with the NLP processing in the general
KG Knowledge graph Aw The weight of influencing field, the author has carried out preprocessing on the data text. The
representation scheme factors in the fuzzy
specific steps are as follows:
comprehensive evaluation
PR Production rule Simp (Txi , Similarity value of Step 1: It is necessary to check whether the craft text is a meaningful
representation scheme Tyj ) attributes inXand Y language before text processing. If it is, different languages should be
w1 Threshold for Semantic EDa,b (i, j) The edit distance between treated in different ways. So the natural language recognition tool called
similarity calculation the i characters in sentence Langid is used. The accuracy of Langid is higher but the speed is slightly
a and the j characters in b slower. Moreover, ISO 639-1 was set in the system as the language
w2 Threshold for attribute Simc (i, j) Similarity value based on
coding standard. For more accurate identification, the team also added a
weighting calculation improved edit distance
algorithm programming language test, whose reference sample is Linguist in
w3 The similarity threshold in function(i, The cost of changing a Github.
the edit distance j) String to make i equalj Step 2: Sentence segmentation/Part-of-speech tagging. The Jieba
algorithm
word segmentation tool was used in the system, which supports three
Xi The number of attributes Ki×j The incidence matrix of
in sentence X entityiand entityj word segmentation modes, supports traditional Chinese, and supports
Yi The number of attributes
custom dictionaries. In addition, the system uses SnowNLP and PkuSeg
in sentence Y as auxiliary tools to make the tagging more accurate.
Step 3: The general situation in the process text is still considered,
that is, it cannot confirm whether there is sensitive information in the
and annotation, this paper proposes a framework for the automatic data set. Therefore, the Cherry classifier was used to screen the data
construction of a machining domain knowledge base. Fig. 2 shows the through secondary screening to prevent the model from deteriorating.
overall process of the automatic construction of the machining domain Among them, thresholds and logical structures are added to ensure the
knowledge base, which includes three core steps: accuracy of the model.
Step 1: Knowledge extraction. First, in a large number of process text In addition, the system has made reasoning judgments about labels in
documents, a method based on multi-neural collaboration is used to the text of the craft document whose paragraph is within 128 words.
extract knowledge, and the extracted knowledge is classified accord­ From a technical perspective, regularization tool in the system was used
ingly through tags, which are divided into three categories: rule-based to process and split the sentence, whose split identifier is the Chinese
process knowledge, same-property process template knowledge and period. For the second case, if the knowledge text paragraph exceeds
personalized process instance knowledge. 128 words, the period will be used for regular segmentation, and the
Step 2: Knowledge representation based on fuzzy comprehensive system will recognize the redundant sentence as the next sentence.
evaluation: Since the knowledge graph cannot accurately represent each Moreover, for paragraphs with character length of less than 128 words,
type of process knowledge, this paper uses the fuzzy comprehensive the system will adopt a conventional processing method: Filling the

6
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

unfilled matrix with 0. For specific details, please refer to this article score sentence x with y label sequences:
[61].
∑
n ∑
n+1
score(x, y) = Pi,yi + Ayi− 1, yi (2)
4.2. Characteristics of the model input i=1 i=1

The score of the entire tag sequence is equal to the sum of the scores
At present, the mainstream pre-training models for generating word of each position, and the score of each position consists of two parts: one
vectors and term vectors include Word2vec and GloVe. Based on the part is directly determined by the output of the LSTM, and the other part
BERT pre-training model, the transformer encoder is used as the lan­ is determined by the transition matrix of the CRF. Define all the label
guage model [62–64], which is more efficient than the recurrent neural sets as Y, where yj ∈ Y represents the j-th true label value in Y. We use
′

network (RNN) model and can capture longer-distance dependencies. the softmax function to obtain the normalized probability:
This model can solve the unidirectionality problem of most current word
vector generation models. To obtain higher-quality input vectors, this p(y|x) = ∑
exp(socre(x, y))
(3)
paper constructs a word vector expression layer based on BERT ac­
′
y exp(socre(x, y ))
′

cording to the fine-turning mechanism provided by BERT and the par­

For a given training samplep(x, y), the log-likelihood function
ticularity of the knowledge extraction problem in the machining domain
islog(p(y|x)). During training, the optimal label
[65–67].
sequencey∗ = argmaxp(x, y) for the input sequence x = (x1 , x2 , ⋅⋅⋅, xn ) is
We define each word in the sentence vector x = (x1 , x2 , ⋅⋅⋅, xn ) to y∈Y
calculate 3 characteristics: word characteristics, sentence characteris­ obtained by maximizing the log-likelihood function.
tics, and location characteristics. The characteristics are correspond­
ingly expressed as (et1 ,et2 ,⋅⋅⋅,etn ), (es1 ,es2 ,⋅⋅⋅,esn ), and(ep1 ,ep2 ,⋅⋅⋅,epn ). Since the 5. Knowledge representation based on fuzzy comprehensive
smallest-input unit of the process text in the machining field is a sen­ evaluation
tence, the recognition unit is also a sentence, and the sentence itself is
invalid for knowledge extraction, so this article sets the sentence feature At present, most scholars use knowledge graph triples to formally
to 0. Therefore, it can be concluded that the input of the word vector represent knowledge, but knowledge graphs often have difficulty
representation layer is Ck = esk + etk + epk , where Ck ∈ C, C = (C1 , C2 , ⋅⋅⋅, accurately expressing meaning when elaborating general knowledge,
Cn ) and eXk are the corresponding vectors determined by the BERT vo­ such as “face before the hole”, “first rough and then refined”. In addi­
cabulary provided by Google and they represent the corresponding tion, in the process of automatically constructing a knowledge base, the
characteristics of the k − th word. We calculate C = (C1 , C2 , ⋅⋅⋅, Cn ) source of the knowledge is complex process text documents. These
through a multi-level Transformer and finally obtain the character-level documents not only contain rule-based process knowledge but also
feature vector sequence E = (E1 ,E2 ,⋅⋅⋅,En ), which is used as the input of contain process template knowledge and process instance knowledge.
the BiLSTM neural network layer. Therefore, relying on only the knowledge graphs to express the extrac­
ted knowledge clearly does not enable accurate explanations of the
meaning of knowledge. For this reason, a knowledge expression form
4.3. BiLSTM neural network layer
based on a fuzzy comprehensive evaluation is proposed. The core ideas
are as follows: ① Extract the corresponding knowledge from the process
There is a long-distance dependence problem in knowledge extrac­
text document. ② Judge the optimal representation method for the
tion in the machining field. This paper uses BiLSTM to solve it. BiLSTM
knowledge through a fuzzy comprehensive evaluation (knowledge
[68,69] is used to automatically extract sentence features. We use each
graphs, IF-THEN production rules and two-dimensional data linked
word of a sentence as the input of a time step of BiLSTM and denote the
lists). ③ If the result of evaluation is a knowledge graph, then construct
forward LSTM as l = {l1 , l2 , ⋅⋅⋅, ln } and the reverse LSTM as r = {r1 ,r2 ,⋅⋅⋅,
the node of the pattern layer; if the result of evaluation is an IF-THEN
rn }. The inputs l1 and r1 are the vector inputs E1 based on BERT. At the
production rule, then convert it into a SWRL representation; and if it
same time, meeting the conditions starting from l2 and r2 , the input is
is a two-dimensional data table, read and extract the entity tag infor­
Ei ⊕ li− 1 ,Ei ⊕ ri− 1 . Then, splicing l and r, the hidden layer state vector of
mation for data-driven. In addition, it should be noted that the carrier of
each position is obtained as ci = [li ; ri ]. Finally, the result of BiLSTM is
entity information identified by BERT+BiLSTM+CRF is a tag. In addi­
connected to a linear layer, and the hidden state vector ci is mapped to
tion, we can also understand the difference in knowledge types is mainly
the k-th dimension to obtain pi so that the automatically extracted
due to the variations in name tags. For example, considering the
sentence features are obtained, which are recorded as a matrix:
knowledge of the turning rule “To ensure the rigidity of turning, the
⎡ ⎤
p11 ... p1m larger-diameter part should be turned first, and the smaller-diameter
p = (p1 , p2 , ⋅⋅⋅, pn ) = ⎣ ⋮ ... ⋮ ⎦ (1) part afterward” and the process template knowledge “For hardened
pn1 ... pnm steel holes, the precision grade is 8, the RA is 3.2, and the machining
method is drilling-expanding-hinging”; The rule knowledge has two
Each dimension pij of pi can be regarded as the scoring value for
main tags: the processing purpose (to ensure rigidity during turning) and
classifying xi under the j-th label. According to the traditional method,
the processing conditions (the larger-diameter part should be turned
softmax can be directly connected for classification, but by this method,
first, and then the smaller-diameter part) in the Function layer. For
the information that has already been labeled cannot be used when la­
personalized process instance knowledge such as “The maximum
beling each position; therefore, we will next access a CRF layer for
machining diameter of a TC20/32 lathe is 260 mm, and the maximum
labeling.
turning diameter is 380 mm”, in contrast to the previous two types of
knowledge, there are no “processing conditions” or “processing pur­
4.4. CRF layer pose” tags.
Therefore, based on the FBS annotation framework and the type and
The parameter of the CRF layer is a (k + 2) × (k + 2) matrix A, where number of the tags identified by named entities, the evaluation factor set
Aij represents the transfer score from the i-th label to the j-th label, and a of the knowledge representation scheme is defined as follows:
label that has been marked previously can be used when marking a
position [70,71]. k + 2 is used because it is necessary to add a starting U = {u1 , u2 , u3 , u4 } = {me, pm, pc, pp}; (4)
state and an ending state to the sentence. We define a sentence label me represents the number of mechanical entity tags in knowledge
sequence y = (y1 , y2 , ⋅⋅⋅, yn ) with length equal to n, then the model will

7
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

extraction; pm represents the number of processing method tags in After the fuzzy comprehensive evaluation is completed, it is still
knowledge extraction; pc represents the number of processing condition necessary to make a supplementary explanation for the case where the
tags in knowledge extraction; and pp represents the number of pro­ evaluation result is knowledge graph. As we know, the knowledge graph
cessing purpose tags. In addition, to increase the calculation speed and is composed of the triples (entity-relationship-entity), and it seems that
facilitate calculation, this paper makes the following simplifications: the using BERT-BiLSTM-CRF can only be able to identify the entity infor­
entity tags (the mechanical features tag, mechanical equipment tag, mation in the process document. But what we need to pay attention to is
materials tag, etc.) are unified into the category of mechanical entity that the carrier of entity information are labels. Due to the particularity
tags, and the secondary attribute tags are linked by the child-parent of the machining field, it does not contain many tags such as place names
relationship. In addition, because a two-dimensional data linked list and people in the general field. Therefore, in view of the characteristics
cannot express rule knowledge containing processing conditions and of the machining field, the author has summarized the relationships in
processing purposes, the two-dimensional data linked list is removed the machining field through the FBS design framework in the Section 3.
from the judgement set V, and its best expression is set as instance To put it simply, commonly used relationships in the machining field
knowledge. Through the above simplifications, a judgement set V = have been set as labels in FBS. When the knowledge expression form is
{KG, PR} is constructed as shown, where KG represents the knowledge judged to be knowledge graphs, the BiLSTM-CRF framework can still be
graph scheme and PR represents the production rule scheme. Regarding used to automate the construction of triples (entity-relationship-entity).
the classification of knowledge, the representation advantage of PR over From the perspective of the relationship, the States level tends to be
KG is that it can explicitly represent general knowledge such as causal more internal to the entity (the relationship between the child and the
judgements but is not very good at representing mechanical entities or parent class), and the Function level tends to be the external relationship
processing methods. Therefore, for the production rule, the membership (the connection between nodes and nodes).
function of the mechanical entity factor and the processing method More details are implemented through the system code. For example,
factor can be set to the minor deflection type; the processing condition there is a same property process template knowledge: For the end face of
factor and the processing purpose factor can be set to the large type. gray cast iron, the machining accuracy is level 6, and the surface
Through fuzzy statistics, the corresponding membership function is roughness is 6.3, then the recommended machining method is rough
determined as: machining.
⎧ ( x )2 Through the extraction framework, it can label the text knowledge.
( )
⎪
⎪
⎪
⎪ , ume The label of processing method is rough processing; The material label is
⎧ 4− x ⎪ 8
gray cast iron; The mechanical characteristics label is the end face and
m
⎪
⎪
⎪
⎪ ; ume ⎪
⎪ x − 0.6
⎪
⎪
⎪
⎪
4 ⎪
⎪
⎪
⎪ , 0 < x ≤ 1; upm the characteristic attribute is IT=6, RA=6.3. Input these information
⎪ ⎪ 4+λ
⎪
⎪
⎪
( ) ⎪
⎪
⎪ into the system, and the algorithm flow steps of the system operation are
⎪ ⎪ [ ]
⎪ ⎪ as follows:
− 2 x− 21 2
⎪
⎪ ⎪
⎪ x − 0.6
⎪ ⎪
⎪
⎨ e ; u pm ⎨ 4 + λ , 1 < x < 6; upm
⎪
Step 1: The system obtains relevant tags and judges the type of
rPR = 1 rKG = (5) knowledge (Same property process template knowledge). The system
⎪
⎪ x− ⎪
⎪ 1, x ≥ 6; upm
⎪ ⎪
⎪
⎪
⎪ 2 ; upc ⎪
⎪
⎪ implements a fuzzy comprehensive evaluation algorithm, selects the
⎪ ⎪
⎪
⎪
⎪
⎪
8 ⎪
⎪
⎪
⎪ 4− λ− x best knowledge representation algorithm, and the result is knowledge
⎪ ⎪ ; upc
⎪
⎪
⎪ x −
1 ⎪
⎪
⎪ (4 − λ)σ graph. Then the system perform knowledge redundancy judgment.
⎪ ⎪
⎩ 2; u
pp
⎪
⎪
⎪
⎪ Step 2: Start Neo4jUtils Jar and call related methods to realize the
4 ⎪
⎪
⎪
⎩
4− x
; upp connection with Neo4j.
16μ Step 3: It is very important for the system to search for the Function
For the membership function of the production rule program rPR , mis sub-tab, but the results shows that it is not found. Because the system
the reduction coefficient, and the author set m = 2; regarding the does not find the Function level sub-tag responsible for external rela­
membership function rKG for the knowledge graph, we set λ as the tionship connections, its priority is lowered. Then, the system converts
correction coefficient, which is taken as 1. Since the knowledge graph is the connection of the external relationship to search for the States label,
very sensitive to the effects of processing conditions and processing and the Processing Method label will be found by the system, whose text
purposes, the control amount σ, μ of the knowledge graph processing content is rough machining. Therefore, the system will create a rough
label is set to 3, 2. machining accuracy node. (If there is a Function sub-tag, the system will
Setting the weights of the above factors in the decision scheme as Aw first create a Function sub-tag node, and will not grant Process_Method
= (a1 ,a2 ,a3 ,a4 ), they are incorporated into the membership function to priority to connect to external nodes).
complete the overall evaluation of the scheme. For example: for the Step 4: Searching for entity tags. The system creates entity nodes
knowledge text “To ensure the rigidity of turning, the part with the (end faces) and fills in attribute information (this can be achieved
larger outer diameter should be turned first, and the part with the through sub-parent relationship). After this step, two knowledge nodes
smaller outer diameter should be turned afterwards”, in fuzzy reasoning, have appeared, and the main content of Step5 is to establish related
the number of mechanical entities is 3, the number of processing relationships.
methods is 1, the number of processing purposes is 1, and the number of Step 5: Building related relationships. According to the knowledge
processing conditions is 2. The evaluation matrix R can be obtained by type judged by step1 and the system operation of step3, the relationship
calculating the degree of membership, and the comprehensive evalua­ between the two nodes can be judged as Process Method. Fig. 5 in the
tion matrixB = Aw ⋅R = (0.2, 0.2, 0.3, 0.3)⋅R is: manuscript described this example.
⎡ ⎤
0.0625 0.1406 6. Eliminating redundancy based on knowledge fusion
⎢ 0.3680 0.0800 ⎥
B = Aw ⋅⎢ ⎥
⎣ 0.1875 0.0370 ⎦ = [0.1801, 0.0832] (6)
The algorithm based on multistage neural network collaboration
0.1250 0.0930
performs knowledge extraction for the machining field. However, due to
From the judgement result, it can be seen that the membership de­ the wide variety of data sources in this field, the extraction results often
gree of the production rule representation scheme is greater than that of contain a large amount of repetitive, confusing, and even conflicting
the knowledge graph representation scheme, so this knowledge is rep­ knowledge. Therefore, to improve the quality of the knowledge base, it
resented by production rules. is necessary to address the redundancy in the extracted knowledge. In
this paper, for the extraction of entities from the semantic, attribute and

8
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

core tag perspectives, the similarity evaluation of the entities is deter­ mass production types and whose locations are very close to each other.
mined, and the results of the similarity evaluation are organically If the precision attributes are considered alone, without considering the
combined to yield the final similarity value. location attributes or batch attributes, this entity is often mapped to the
general hole in the knowledge base whose attribute are IT6 and RA3.2. If
6.1. Redundancy in knowledge extraction the entity in the knowledge base is used in the downstream process
reasoning, it will be identified as a general method: drilling-expanding-
For a given entity reference item, there are often multiple entities reaming. However, this process method easily causes the thin wall be­
with the same semantics and attributes in the knowledge base; that is, tween the two holes to break in actual processing, which leads to the
the same ontology has different term descriptions. For example, for two scrapping of the workpiece. Therefore, after the semantic similarity
knowledge triples A =< s1 , p1 , o1 > and B =< s2 , p2 , o2 > in the knowl­ calculation, the weighted calculation of the attributes is also required.
edge base, if the content of this pair of triples is the same or there is a The threshold w2 is introduced, and we define that if the score after
case in which the subject s1 ⇔ s2 , predicate p1 ⇔ p2 , and object o1 ⇔ o2 , it attribute mapping is higher thanw2 , the distance function will be set to 0.
is called knowledge redundancy. Therefore, as the number of knowledge This means that the semantics and attributes of these two entities are the
extraction entities increases, if similar knowledge triples are not elimi­ same, and there is no distance. Otherwise, the two entities have the same
nated, the redundancy of the knowledge base will become increasingly semantics but different attribute values, and the distance function is set
large, which will lead to a reduction in the quality of the knowledge base to 1. If the core entities are different, the distance function is set to 2. The
and further affect the accuracy of process reasoning in downstream corresponding calculation formulas for Steps 1 and 2 are as follows:
applications. ∑
n
xi × yj
( ) ∑
n
(8)
i=1,j=1
Simp Txi , Tyj = Cos(xi ,yi ) = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
6.2. Similarity judgement based on improved edit distance ∑n ∑ n
i=1
xi 2 × yi 2
i=1,j=1 i=1,j=1
The traditional edit distance (ED) is used as a function to measure the
distance of a string. It can only judge whether the degree of literal
( ) ∑n
matching of strings is the same but cannot judge whether the degree of SimT Xi , Yj = wp ∗ Simp (Txi , Tyj ) (9)
semantics is the same. Therefore, we introduce the Word2vec tool into p=1

the traditional edit distance function. It is defined as follows:

⎧ where Xi Yj represent the number of entities in the sentences, Simp (Txi ,
⎨0 Tyj ) represents the similarity of the attributes in concepts Xi and Yj , and
function(i, j) = 1 (7) wp represents the weight of the p-th attribute.
⎩
2
Step 3: Semantics-based improved edit distance calculation. Through
In this formula, we use the following definition: If the i-th entity in the above two steps, the word vector after the semantic calculation and
sentence a and the j-th entity in sentence b in the knowledge base have attribute weighting can be obtained. At this time, the improved Lev­
the same semantics and attributes, we set function(i, j) = 0; if the i-th enshtein calculation is used to calculate the final similarity. The final
entity in sentence a and the j-th entity in sentence b in the knowledge similarity threshold w3 is introduced. If the similarity result is greater
base have different semantics and attributes, we set function(i, j) = 1; than w3 , then the extracted sentences are redundant with respect to the
and if the i-th core entity in sentence a and the j-th core entity in sen­ knowledge triples in the knowledge base and need to be deduplicated;
tence b in the knowledge base have different semantics and attributes, otherwise, the extracted sentences are input into the knowledge base.
we set function(i, j) = 2. ⎧ max(i, j)
The main steps of knowledge fusion are: ⎪
⎨ ⎧
⎨ EDa,b (i − 1, j) + 1
Step 1: Semantic similarity calculation. First, the Word2vec tool is ED(i, j) = (10)
⎩ min⎩
⎪ EDa,b (i, j − 1) + 1
used for unsupervised learning to obtain a low-dimensional vector EDa,b (i − 1, j − 1) + function(i, j)
representation of the semantics of the words in the sentence. At the same
time, the cosine formula is used to calculate the vector angle between ( ) ∑n

the two word embeddings in the semantic space. Generally, the more SimT Xi , Yj = wp ∗ Simp (Txi , Tyj ) (11)
p=1
similar the semantics are, the greater the cosine score. Then, the se­
mantic similarity threshold is set as w1 , and the knowledge type is where EDa,b (i, j) is the edit distance between the first i characters in
judged through a fuzzy comprehensive evaluation: if it is general sentence a and the first j characters in b. function(i, j) is defined as the
knowledge, the core entities are the “processing purpose” and “pro­ cost of changing (inserting, deleting, replacing) a string.
cessing conditions”; If it is general process template knowledge, the core For example, a= “For the end face with a surface roughness of 6.3,
entity is “mechanical characteristics”; and if it is personalized knowl­ the processing strategy is generally rough machining” is a process
edge, the core entity is a “mechanical entity”. Finally, the cosine score template knowledge text. There is knowledge of b= “The main pro­
values are compared. If the cosine score value is higher than w1 , indi­ cessing method for the outer cylindrical surface of RA 6.3 is rough
cating that the semantics are the same between the entities, the distance turning” in the knowledge base. According to the semantic calculation
function is set to 0; otherwise, the entity is different from its semantics, and attribute weighting of the above formula, the semantic and attribute
the distance function is set to 1, and the core entity distance function is similarity between the entities is:
set to 2.
Step 2: Attribute mapping. In knowledge fusion, the entity attributes
also affect the quality of the knowledge base. For example, consider the According to multiple sets of experimental data, we take w1 as 0.75,
case of extracting two hole entities whose attributes are IT6, RA3.2,

( ) ( )
Sims surface roughness, Ra = 0.89; Sim end face, outer cylindrical = 0.24;
Sim(processing strategy, processing method) = 0.78; Sim(rough machining, rough turning) = 0.76.

9
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Fig. 3. Process of constructing the knowledge base.

⎡ ⎤
$ surface roughness end face processing strategy rough machining
⎢ Ra 0 1 2 3 ⎥
⎢ ⎥
Ki×j =⎢
⎢ outer cylindrical 1 1 2 3 ⎥
⎥
⎣ processing strategy 2 2 1 3 ⎦
rough turning 3 3 3 1

10
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Table 3 Table 4
Description of the relevant parameters in the experiment. Comparison of the effectiveness with that of other models.
Parameter name Parameter value Model P% R% F%

Batch processing parameters 32 BiLSTM-CRF 77.8% 73.22% 75.44%

Gradient 0.5 CNN-BiLSTM-CRF 78.62% 79.24% 78.92%
Inactivation rate 0.5 BERT-BiLSTM-CRF 81.29% 84.36% 82.79%
Learning rate 10− 5
Maximum sentence length 128
Length of LSTM 128
Optimizer algorithm SGD Table 5
Extraction effect based on the BERT-BISTM-CRF model.
Knowledge Category P% R% F%
w2 as 0.7, and w3 as 0.7. If the influence of the core label is not
considered, the maximum distance of function is set to 1, and the inci­ Method Condition (MC) 63.93% 71.86% 67.66%
Processing Purpose (P) 64.39% 73.42% 68.60%
dence matrix between two sentence entities is:
Calculation Formula (CF) 92.89% 94.56% 93.71%
Theorem (T) 90.51% 91.39% 90.94%
Feature (F) 87.35% 88.23% 87.78%
It can be calculated that ED(i, j) = 1 and Simc (i, j) = 1 − 14 = 0.75 are Equipment (E) 90.23% 92.15% 91.17%
Method (M) 78.36% 82.48% 80.36%
greater than the final similarity threshold w3 , and these two sentences Cutting Elements (CE) 80.24% 84.59% 82.36%
are redundant and need to be deduplicated. However, this result is Material (MT) 91.41% 92.15% 91.78%
obviously not in line with actual application because the core label “end Feature Attributes (FA) 74.36% 78.69% 76.46%
surface-outer cylindrical” has different semantics. Therefore, the influ­ Equipment Attributes (EA) 72.36% 79.63% 75.82%
Material Attributes (MTA) 79.83% 84.67% 82.18%
ence of the core label must be considered. Then, ED(i, j) = 2 and Simc (i, j)
Processing Parameters (PP) 70.25% 76.32% 73.16%
= 1 − 24 = 0.5 are less than the final similarity w3 , there is no redun­
dancy, and a is entered into the knowledge base without eliminating
redundancy. ② is the extraction result page. When the extracted text is “In order to
meet the accuracy requirements, the material is grey cast iron, the sur­
7. Prototype system and analysis face roughness is 1.6, and the end face with accuracy grade 6 should
undergo rough turning–semi-finishing turning–finishing turning”, the
To verify the effectiveness of the method proposed in this paper, we extraction result is as shown in ②.
develop a prototype system of the knowledge base for the machining Second, the fuzzy comprehensive evaluation judgement is carried
field. The system uses Microsoft Visual Studio 2003, the Jena and OWL out, and the appropriate representation method for the knowledge is
API functions and Myeclipse2018 as development tools. Protégé soft­ calculated according to the membership degree. ③, ④, and ⑤ in Fig. 3
ware is used to build the ontology, and the SWRL language is used to show examples of knowledge graphs, two-dimensional data tables and
express the knowledge of rules. Three.js constructs the knowledge graph production rules in the knowledge base, respectively. It should be noted
triples and performs visualization. MySQL is used to store personalized that to address the problem that IF-THEN rules are difficult to display
process instance knowledge, process knowledge graphs and system and represent, the system uses the SWRL language to integrate pro­
ontology instantiation data. Two experiments are set up to evaluate the duction rules into Protégé to obtain their semantic expressions, as shown
effect of the prototype system by studying: (1) the effect of knowledge in ⑤. Based on the process data and knowledge extraction, the process
extraction in the machining domain and (2) the process application of knowledge graph visualization interface of the knowledge base is
the knowledge base in the machining domain. developed in the Java environment, as shown in ③ and ⑥.
Finally, redundant processing is carried out in the knowledge fusion
stage. Through the semantic similarity, attribute mapping and editing
7.1. Prototype system distance, the redundant knowledge in the previous extraction results is
removed, and the conflicting or ambiguous knowledge is processed at
The process knowledge and data used in the construction of the the same time. The above steps can improve the quality of the knowl­
system in this article mainly come from the international industry edge base and provide an intelligent reasoning basis for the application
standard data model, the US National Design Repository (NDR) data­ of downstream process knowledge.
base, China’s Baidu Encyclopedia, the US Western Reserve University,
the official NASA website, the process manual, and the process data
accumulated by the research team over many years. At the same time, 7.2. Quality analysis of knowledge extraction
the standard data division method is used to divide the training data set
and the test data set. Fig. 3 shows the knowledge base prototype system To verify the extraction quality of the prototype system, for the entity
oriented toward the machining field that is constructed in this paper, recognition models based on BiLSTM-CRF and CNN-BILSTM-CRF,
which mainly includes text extraction, display of the extraction results, Word2vec and BiLSTM are used to train the word vectors. The feature
three types of knowledge representation, and knowledge base visuali­ matrix trained by the above steps is used to obtain the global optimal tag
zation. Fig. 3 shows that the construction of the knowledge base pro­ sequence of the target sentence from the CRF level. In addition, due to
ceeds through three main steps: the different sentence lengths, the experiment is based on the longest
First, build the BiLSTM-CRF extraction environment and obtain the sentence, and the left and right ends of the sentences are supplemented
results for the entity tags by extracting the text. The main page of with zeroes to obtain a uniform extraction matrix. The parameters set in
knowledge extraction for the prototype system is shown in Figure . The the experiment are shown in Table 3:
corresponding interfaces of the knowledge base are the feature On the same training set and test set, the effectiveness of the above
reasoning in the knowledge base application module, process reasoning three named entity recognition models and the model proposed in this
and visualization of the knowledge base. This system can perform the paper are compared. The experimental results are shown in Table 4.
functions of feature/part process reasoning and process knowledge The experimental results show that the BERT-BiLSTM-CRF used in
visualization and can provide service and support for the application of this paper is better than the other two extraction models for the
process knowledge and the intelligent decisions of the process. In Fig. 3, extraction of knowledge in the machining domain. Since the

11
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Fig. 4. Data of the aero-engine cone gear axis.

Fig. 5. Examples of Process Knowledge Graphs.

combination of word-level feature vectors considers only the charac­ vectors and uses Transformer to train the word vectors, which takes full
teristics of the words and ignores the shortcomings of entity recognition account of the impact of contextual information on the entities. Table 5
in combination with the context, the model used in this paper achieves a shows the experimental results for identifying entities of various cate­
better entity recognition effect. This is because the model used combines gories based on the BERT-BiLSTM-CRF model.
word features, sentence features and position features to generate word Of course, mistakes still occur during the analysis; however, they

12
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Fig. 6. Technological process design of the aero-engine cone gear axis.

have only a few possible causes that can be addressed in the future. to reasoning about the most likely meaning and attributes of each word.
These causes can be grouped into the following families:

• Pronouns 7.3. Case study

Pronouns (that, which, etc.) are in general very difficult to resolve This system uses an aero-engine cone gear axis as an example to
properly; moreover, the subordinate sentences from which the pronouns verify the application of the knowledge base system. The aero-engine
originate may not be correctly linked to their real subjects. cone gear axis page in Fig. 4 shows the many characteristics of the
cone gear axis, which mainly include typical features such as conical
• Missing words surfaces, cylindrical surfaces, end surfaces, chamfers, round grooves,
helical tooth surfaces and thread surfaces. In addition, according to the
Many elements are still missing in all three FBS classes of the accuracy requirements, the machining process includes rough
knowledge base. machining and finishing machining. For the reasoning process, first, the
parsed part information model is written to the knowledge graph in­
• Word structure formation model. Then, the processing strategy, processing equipment,
processing data and other parameters of the processing operation are
In the process of knowledge fusion, there are still limitations based mapped through the geometric information, accuracy information, and
on the improved editing distance algorithm, and there will be certain attribute information. For example, consider the feature machining
extraction obstacles for words with multiple attribution structures. In knowledge Rule 1 in the knowledge base: For non-hardened cavity
the future, we may consider applying statistical learning for algorithms feature machining, the machining accuracy level is 12 and the surface
roughness is 6.3. It can be inferred from the following rules that the

13
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

cavity feature machining method is rough milling. underlying knowledge base; In addition, it allows seamless inte­
Rule: Micro_ProcessPlanning(?M)∧Procket(?F)∧HasMaterial(?F, gration with other data subjects in other Web environments.
NotHardenedSteel)HasMethod(?F,?M)∧HasPrecisionGrade(?F,?IT)∧
HasSurfaceRoughness(?F,?R)∧greaterThan(?IT,9)∧greaterThan(? In future work, we will consider using machine learning technology
Ra,6.3)lessThanOrEqual(IT,11) ∧lessThanOrEqual(?Ra,12.5)⇒HasPro­ to automatically expand the capacity of the knowledge base. Finally, we
cessMethod(?M,Rough_Milling) aim to apply the model described in the discussion section to perform
In addition, when the attribute values (batch attributes, location intelligent process reasoning for a knowledge base system under real-
attributes, etc.) and feature names are normal, the processing method time dynamic information.
can be determined from the knowledge graph in the knowledge base.
The applied knowledge graph is shown in Fig. 5. Declaration of Competing Interest
Based on the knowledge-driven prototype system, the process design
of the aero-engine bevel gear shaft is shown in Fig. 6. As seen from the We declare that we do not have any commercial or associative in­
reasoning results page in Fig. 6, the process flow design includes not terest that represents a conflict of interest in connection with the work
only the reasoning of the processing program but also the parameter submitted.
data of the work steps, processing equipment (machine tools, tools,
fixtures and measuring tools), cutting parameters, feed times and
Acknowledgments
working hours. Through the knowledge extracted from the process text
in the knowledge base, process reasoning for the features and parts can
We would like to thank the reviewers and the editor for their
be completed, and automatic process design can be carried out.
constructive comments and suggestions regarding this paper. We also
acknowledge the Youth Science Foundation of the National Natural
8. Conclusions
Science Foundation (No. 51705438), the Sichuan Science and Tech­
nology Project (No. 2018JY0366), the Young Science and Technology
As the foundation of KBE technology and intelligent manufacturing,
Innovation Team of SWPU (No. 2019CXTD02), the Independent Inno­
the process knowledge base plays an extremely important role in
vation Special Fund Project, AECC (No. ZZCX-2017-039) and Chengdu
advanced manufacturing systems, which directly affects the process
International Science and Technology Cooperation Project (No:2020-
decision, product cost and production cycle of an enterprises. Therefore,
GH02-00040-HZ).
we believe that in the intelligent manufacturing environment, the life
cycle of manufactured products should not be calculated only in a nar­
References
row sense, and the construction time of the process knowledge base
should also be included in the extensive product life cycle. [1] G. Peng, H. Wang, H. Zhang, K. Huang, A hypernetwork-based approach to
However, previous researches tend to use limited and manually collaborative retrieval and reasoning of engineering design knowledge, Adv. Eng.
created ways to construct a knowledge base, which lacks the automatic Inform. 42 (2019) 100956.1–100956.10, https://doi.org/10.1016/j.
aei.2019.100956.
construction process of knowledge base for obtaining and target- [2] Y. Lu, X. Xun, L. Wang, Smart manufacturing process and system automation–a
oriented simulation knowledge in mechanical product design. There­ critical review of the standards and envisioned scenarios, J. Manuf. Syst. 56 (2020)
fore, this method is time-consuming and labor-intensive, which has 312–325, https://doi.org/10.1016/j.jmsy.2020.06.010.
[3] J.A. Girodon, et al., An organizational approach to designing an intelligent
obvious limitations in the distributed environment. knowledge-based system: application to the decision-making process in design
To this end, the focus in this work is developing a framework for projects, Adv. Eng. Inform. 29 (3) (2015) 696–713, https://doi.org/10.1016/j.
automatically constructing a knowledge base in the field of machining, aei.2015.07.001.
[4] S.A.S.H. Wan, D.A. Li, J.A. Gao, et al., A knowledge based machine tool
and providing a case showing that this framework can assist the intel­
maintenance planning system using case-based reasoning techniques, Rob.
ligent system in the automated process decision-making during Comput. Integr. Manuf. 58 (1) (2019) 80–96, https://doi.org/10.1016/j.
machining. We think this is a very interesting topic, which transcends rcim.2019.01.012.
[5] L.A Qi, et al., A self-learning and self-optimizing framework for the fault diagnosis
the traditional way of creating knowledge bases based on expert inter­
knowledge base in a workshop, Rob. Comput. Integr. Manuf. 65 (65) (2020)
view or manual way, especially with the restrictions posed by global 101975, https://doi.org/10.1016/j.rcim.2020.101975.
occurrences such as Covid-19 whereby traditional in-person activities [6] H.A. Chen, X.A.X.C. Luo, An automatic literature knowledge graph and reasoning
may be altered significantly. network modeling framework based on ontology and natural language processing,
Adv. Eng. Inform. 42 (2019) 100959, https://doi.org/10.1016/j.aei.2019.100959.
In general, the method has the following advantages: [7] M.A. Robinson, How design engineers spend their time: job content and task
satisfaction, Design Stud. 33 (4) (2012) 391–425, https://doi.org/10.1016/j.
(1) A knowledge extraction model for the machining domain based destud.2012.03.002.
[8] J. Camacho-Collados, M.T. Pilehvar, On the role of text preprocessing in neural
on BERT-BiLSTM-CRF is applied. This architecture allows the network architectures: an evaluation study on text categorization and sentiment
capture, representation and reuse of manufacturing domain analysis, in: Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing
knowledge. In addition, by adding tags to the machining domain and Interpreting Neural Networks for NLP., 2018, pp. 40–46.
[9] M. Gašić, D. Hakkani-Tür, A. Celikyilmaz, Spoken language understanding and
knowledge, the problems of unclear boundaries and complex interaction: machine learning for human-like conversational systems, Comput.
types of domain entities have been solved. Speech Lang. 46 (2017) 249–251, https://doi.org/10.1016/j.csl.2017.05.006.
(2) A fuzzy comprehensive evaluation algorithm is proposed to [10] M. Koponen, Is machine translation post-editing worth the effort? A survey of
research into post-editing and effort, J. Spec. Trans. (25) (2016) 131–148.
determine the best expression of complex knowledge, integrating
[11] M.F. Moens, Argumentation mining: how can a machine acquire common sense
three expression schemes (knowledge graphs, SWRL, and two- and world knowledge? Argum. Comput. 9 (1) (2018) 1–14, https://doi.org/
dimensional data tables) to effectively characterize the techno­ 10.3233/AAC-170025.
[12] M. Zhou, What will search engines be changed by nlp advancements, ICTIR ’18, in:
logical meaning of most knowledge, which can better support the
Proceedings of the 2018 ACM SIGIR International Conference on Theory of
application of comprehensive consideration rules and case Information Retrieval, Tianjin, China, 2018.
knowledge in subsequent actual process design decisions. [13] S.B.S. Biswas, D.S.D. Sengupta, R.B.R. Bhattacharjee, M.H.M. Handique, Text
(3) A similarity judgement based on an improved edit distance is manipulation using regular expressions, in: Proceedings of the IEEE 6th
International Conference on Advanced Computing(IACC), 2016, pp. 62–67.
proposed, which reduces the redundancy of knowledge and im­ [14] A.M.I.U. Maletti, Survey: finite-state technology in natural language processing,
proves the quality of the knowledge base. Theor. Comput. Sci. 679 (2017) 2–17, https://doi.org/10.1016/j.tcs.2016.05.030.
(4) The system adopts the B/S design architecture, which can easily [15] J.A.J.T. Tejedor, D.T.A.D. Toledano, P.A.P.G. Lopez-Otero, L.A.L.G. Docio-
Fernandez, M.A.M.P. Peñagarikano, L.J.A.L. Rodriguez-Fuentes, A.A.A.M. Moreno-
be expanded to update the system and continuously upgrade the Sandoval, Search on speech from spoken queries: the multi-domain international
albayzin 2018 query-by-example spoken term detection evaluation, Eurasip J.

14
L. Guo et al. Robotics and Computer-Integrated Manufacturing 73 (2022) 102222

Audio Speech Music Process. (1) (2019) 1–29, https://doi.org/10.1186/s13636- [45] K. Eum, M. Kang, G. Kim, M.W. Park, J.K. Kim, Ontology-based modeling of
019-0156-x. process selection knowledge for machining feature, Int. J. Precis. Eng. Man Manuf.
[16] J. Graovac, M. Mladenović, I. Tanasijević, Ngramspd: exploring optimal n-gram 14 (10) (2013) 1719–1726.
model for sentiment polarity detection in different languages, Intell. Data. Anal. 23 [46] M. Helgoson, V. Kalhori, A conceptual model for knowledge integration in process
(2) (2019) 279–296, https://doi.org/10.3233/IDA-183879. planning, Procedia CIRP 3 (2012) 573–578, https://doi.org/10.1016/j.
[17] Khan M.R., Ziyadi M., Abdelhady M., MT-BioNER: Multi-task Learning for procir.2012.07.098.
Biomedical Named Entity Recognition using Deep Bidirectional Transformers. [47] D.S.O.E. šormaz, SIMPM–Upper-level ontology for manufacturing process plan
arXiv preprint arXiv:2001.08904, (2020). network generation, Rob. Comput. Integr. Manuf. 55 (B) (2019) 183–198, https://
[18] J.M. Conroy, D.P. O’Leary, Text summarization via hidden markov models, SIGIR doi.org/10.1016/j.rcim.2018.04.002.
’01:Proceedings of the 24th annual international ACM SIGIR conference on [48] X. Zhang, X. Liu, X. Li, D. Pan, MMKG: an approach to generate metallic materials
Research and development in information retrieval, New Orleans, Louisiana, USA. knowledge graph based on dbpedia and wikipedia, Comput. Phys. Commun. 211
[19] Malouf R., A comparison of algorithms for maximum entropy parameter (2017) 98–112, https://doi.org/10.1016/j.cpc.2016.07.005.
estimation, 2008. [49] B. Zhou, J.B.C.C. Aemail, Z. Chen, et al., KGAssembly: Knowledge graph-driven
[20] L.A. Shen, L.B. Champollion, A.K.C. Joshi, Ltag-spinal and the treebank: a new assembly process generation and evaluation for complex components, Int. J.
resource for incremental, dependency and semantic parsing (article), Lang. Resour. Comput. Integr. Manuf. (10) (2021) 1–21, https://doi.org/10.1080/
Eval. 42 (1) (2008) 1–19, https://doi.org/10.1007/s10579-007-9043-7. 0951192X.2021.1891572.
[21] J. Pater, Generative linguistics and neural networks at 60: foundation, friction, and [50] Li Xiuling, Z. Shusheng, H. Rui, H. Bo, Process knowledge graph construction
fusion, Language 95 (1) (2019) c41–c74, https://doi.org/10.1353/lan.2019.0009. method for process reuse, J. Northwest. Polytech. Univ. 37 (6) (2019) 1174–1183,
[22] J. Hirschberg, C.D. Manning, Advances in natural language processing, Science https://doi.org/10.1051/jnwpu/20193761174.
(2015) 261–266, https://doi.org/10.1126/science.aaa8685 (No.6245). [51] N. Kertkeidkachorn, An automatic knowledge graph creation framework from
[23] W. Xiao, J. Xue, Y. Miao, et al., Distributed graph computation meets machine natural language text, IEICE. T. Inf. Syst. 41 (1) (2018), https://doi.org/10.1587/
learning, IEEE. T Parallel Distrib. Syst. 31 (7) (2020) 1588–1604, https://doi.org/ transinf.2017SWP0006.
10.1109/TPDS.2020.2970047. [52] B. Zhou, J.B.C.C. Aemail, J. Li, et al., A novel knowledge graph-based optimization
[24] Agerri R., Rigau G., Robust multilingual named entity recognition with shallow approach for resource allocation in discrete manufacturing workshops, Rob.
semi-supervised features, (2017). Comput. Integr. Manuf. 71 (3) (2021), 102160, https://doi.org/10.1016/j.
[25] B. Chiu, G. Crichton, A. Korhonen, How to train good word embeddings for rcim.2021.102160.
biomedical nlp, in: Proceedings of the 15th Workshop on Biomedical Natural [53] Y. Hinuma, H. Hayashi, Y. Kumagai, Comparison of approximations in density
Language Processing (BioNLP 2016), 2016, p. 9. functional theory calculations: energetics and structure of binary oxides, Phys. Rev.
[26] P. Bojanowski, E. Grave, A. Mikolov, Enriching Word Vectors with Subword 96 (9) (2017) 9, https://doi.org/10.1103/PhysRevB.96.094102.
Information, Trans. Assoc. Comput. Linguist. 5 (1) (2017) 135–146, https://doi. [54] G.G.C.P. Cascini, G.G.F.I. Fantoni, F.F.M.P. Montagna, Situating needs and
org/10.1162/tacl_a_00051. requirements in the FBS framework, Design Stud. 34 (5) (2013) 636–662, https://
[27] C. Lederer, S. Altstadt, S. Andriamonje, J. Andrzejewski, L. Audouin, doi.org/10.1016/j.destud.2012.12.001.
M. Barbagallo, BV. Cares, F.R Becv, et al., Glove: global vectors for word [55] C. Lederer, S. Altstadt, S. Andriamonje, J. Andrzejewski, L. Audouin,
representation, in: Proceedings of the Conference on Empirical Methods in Natural M. Barbagallo, V.C B, F.R Becv, et al., Analysing modifications in the synthesis of
Language Processing, EMNLP 2014, Doha, Qatar, 2014. multiple state mechanical devices using configuration space and topology graphs,
[28] Ilić S., Marrese-Taylor E., Balazs J.A., Matsuo Y., Deep contextualized word in: Proceedings of the 18th International Conference on Engineering Design, ICED
representations for detecting sarcasm and irony, (2018). 11, Copenhagen, Denmark, 2011.
[29] P.K.P.Z. Kestel, Ontology-based approach for the provision of simulation [56] G. Cascini, G. Fantoni, F. Montagna, Reflections on the fbs model: proposal for an
knowledge acquired by data and text mining processes (Article), Adv. Eng. Inform. extension to needs and requirements modeling, in: Proceedings of the 11th
39 (2019) 292–305, https://doi.org/10.1016/j.aei.2019.02.001. International Design Conference (1-3), 2010, pp. 651–660.
[30] J.S. Gero, U. Kannengiesser, A function–behavior–structure ontology of processes, [57] R. Apreda, F. Dell’Orletta, G. Fantoni, M. Monge, Automatic extraction of function-
AI EDAM 21 (4) (2007) 379–391, https://doi.org/10.1017/S0890060407000340. behaviour-state information from patents, Adv. Eng. Inform. 27 (3) (2013)
[31] S. Rugaber, S. Vattam, A.K. Goel, Structure, behavior, and function of complex 317–334, https://doi.org/10.1016/j.aei.2013.04.004.
systems: the structure, behavior, and function modeling language, AI EDAM 23 (1) [58] Y. Umeda, M. Ishii, M. Yoshioka, Y. Shimomura, T. Tomiyamai, Supporting
(2009) 23–35, https://doi.org/10.1017/S0890060409000080. conceptual design based on the function-behavior-state modeler, AI EDAM 10 (4)
[32] D. Baxter, J. Gao, K. Case, J. Harding, B. Young, S. Cochrane, S. Dani, An (1996) 275–288, https://doi.org/10.1017/S0890060400001621.
engineering design knowledge reuse methodology using process modelling, Res. [59] J.S. Gero, U. Kannengiesser, The situated function-behaviour-structure framework,
Eng. Des. 18 (1) (2007) 37–48, https://doi.org/10.1007/s00163-007-0028-8. Artif. Intell. Design. 25 (4) (2004) 373–391, https://doi.org/10.1016/j.
[33] H. Cho, J. Park, Cost-effective concept development using functional modeling destud.2003.10.010.
guidelines, Rob. Comput. Integr. Manuf. 55 (B) (2019) 234–249. [60] Russo D., Montecchi T., A function-behaviour oriented search for patent digging,
[34] JS. Liang, An ontology-oriented knowledge methodology for process planning in Rad. Teach. (2011) 1111-1120.
additive layer manufacturing, Rob. Comput. Integr. Manuf. 53 (2018) 28–44. [61] B.Y. Lin, F. Xu, Z. Luo, et al., Multi-channel BiLSTM-CRF model for emerging
[35] J. Liu, Z.Z.S.E. Zhang, R. Evans, Y. Xie, Web services-based knowledge sharing, named entity recognition in social media, in: Proceedings of the 3rd Workshop on
reuse and integration in the design evaluation of mechanical systems, Rob. Noisy User-generated Text, 2017.
Comput. Integr. Manuf. 57 (2019) 271–281, https://doi.org/10.1016/j. [62] R. Child, S. Gray, A. Radford, et al., Generating long sequences with sparse
rcim.2018.12.010. transformers, Statistics (2019).
[36] Janet.E Burge, Design rationale: researching under uncertainty, AI EDAM 22 (4) [63] Z. Dai, Z. Yang, Y. Yang, et al., Transformer-XL: attentive language models beyond
(2008) 311–324, https://doi.org/10.1017/S0890060408000218. a fixed-length context, in: Proceedings of the 57th Annual Meeting of the
[37] M. Mcharek, M. Hammadi, T. Azib, C. Larouci, J. Choley, Collaborative design Association for Computational Linguistics, 2019.
process and product knowledge methodology for mechatronic systems, Comput. [64] I. Beltagy, M.E. Peters, A. Cohan, Longformer: the long-document transformer, in:
Ind. 105 (2019) 213–228, https://doi.org/10.1016/j.compind.2018.12.008. Proceedings of the International Conference on Advances in Electrical, Computing,
[38] M.M.P.J. Poorkiany, J. Johansson, F. Elgh, Capturing, structuring and accessing Communication and Sustainable Technologies, 2021.
design rationale in integrated product design and manufacturing processes, Adv. [65] W. Wang, B. Bi, M. Yan, et al., StructBERT: incorporating language structures into
Eng. Inform. 30 (3) (2016) 522–536, https://doi.org/10.1016/j.aei.2016.06.004. pre-training for deep language understanding, in: Proceedings of the International
[39] G. Cezary, K. Krzysztof, K. Damian, P. Iwona, K. Wojciech, An attempt of cnc Conference on Learning Representations, 2019, 1908.04577.
machining cycle’s application as a tool of the design feature library elaboration, in: [66] Q.T. Ho, TNNQ, FAD-BERT: Improved prediction of FAD binding sites using pre-
Proceedings of the Matec Web of Conferences, 2017, p. 6019. training of deep bidirectional transformers, Comput. Biol. Med. 131 (2021),
[40] H. Paulheim, Knowledge graph refinement: a survey of approaches and evaluation 104258, https://doi.org/10.1016/j.compbiomed.2021.104258.
methods, Semant. Web Interoperabil. Usabil. Appl. 8 (3) (2017) 489–508, https:// [67] P. Chi, P. Chung, T. Wu, et al., Audio albert: a lite bert for self-supervised learning
doi.org/10.3233/SW-160218. of audio representation, in: Proceedings of the IEEE Spoken Language Technology
[41] M.F.M. Farber, F.B.F. Bartscherer, C.M.C. Menne, A.R.A. Rettinger, Linked data Workshop, 2021.
quality of dbpedia, freebase, opencyc, wikidata, and yago, Semant. Web 9 (1) [68] C. Dong, J. Zhang, C. Zong, et al., Character-based LSTM-CRF with radical-level
(2018) 77–129, https://doi.org/10.3233/SW-170275 (Special SI). features for Chinese named entity recognition, Nat. Lang. Underst. Intell. Appl.
[42] W. Shen, T. Hu, C. Zhang, et al., A welding task data model for intelligent process (2016) 239–250, https://doi.org/10.1007/978-3-319-50496-4_20.
planning of robotic welding, Rob. Comput. Integr. Manuf. 64 (0) (2020) 101934, [69] Z. Zhai, D.Q. Nguyen, K. Verspoor, Comparing CNN and LSTM character-level
https://doi.org/10.1016/j.rcim.2020.101934, 101934. embeddings in BiLSTM-CRF models for chemical and disease named entity
[43] S. Wan, D. Li, J. Gao, et al., A knowledge based machine tool maintenance planning recognition, in: Proceedings of the Ninth International Workshop on Health Text
system using case- based reasoning techniques, Rob. Comput. Integr. Manuf. 58 (1) Mining and Information Analysis, 2018, pp. 38–43.
(2019) 80–96, https://doi.org/10.1016/j.rcim.2019.01.012. [70] S. Jin, H. Jang, W. Kim, Improving bidirectional LSTM-CRF model of sequence
[44] Y. Zhang, X. Luo, H. Zhang, J.W. Sutherland, A knowledge representation for unit tagging by using ontology knowledge based feature, J. Intell. Inf. Syst. (1) (2018)
manufacturing processes, Int. J. Adv. Manuf. Tech. 73 (5–8) (2014) 1011–1031, 253–266, https://doi.org/10.13088/jiis.2018.24.1.253.
https://doi.org/10.1007/s00170-014-5864-x. [71] Z. Huang, X. Wei, Y. Kai, Bidirectional LSTM-CRF models for sequence tagging,
Comput. Sci. (2015). ArXiv abs/1508.01991.

15