Graph_Convolutional_Neural_Network_for_Intelligent_Fault_Diagnosis_of_Machines_via_Knowledge_Graph.pdf

观海

9页

1次

2024-11-27

免费下载

7862 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 20, NO. 5, MAY 2024

Graph Convolutional Neural Network for

Intelligent Fault Diagnosis of Machines

via Knowledge Graph

Zehui Mao , Senior Member, IEEE, Huan Wang , Bin Jiang , Fellow, IEEE, Juan Xu , Member, IEEE,

and Huifeng Guo

Abstract—Considering the challenge of deep mining of

root causes in machine failures, a knowledge aggrega-

tion fault diagnosis (KAFD) model is proposed, in which

the graph convolutional network (GCN) GraphSAGE is im-

proved and introduced into the knowledge graph (KG)-

based fault diagnosis. Historical maintenance data of ma-

chines is used to construct a fault phenomenon-FBG,

which is then combined with the fault diagnosis knowledge

graph (FDKG) to form a collaborative FDKG. A single-layer

knowledge aggregation network (KAN) that incorporates

sensitivity factors and conﬁgures different types of GCN

aggregators is constructed in the proposed KAFD. Based

on deep neighbor aggregation operations on collaborative

FDKG, KAFD obtained by stacking multiple KANs, can cap-

ture the higher order structural information and semantic

information, which results in the multihop reasoning, im-

provement of the rationality and diversity of fault cause

tracing. The KAFD is experimentally validated through two

fault diagnosis datasets, which are constructed by the

maintenance data of an industrial enterprise, and the re-

sults demonstrate the excellent performance.

Index Terms—Fault diagnosis, graph neural networks,

industrial machines, knowledge graph (KG).

I. INTRODUCTION

NDUSTRIAL machines have been becoming more complex

and expensive with the high performance, as the advanced

intelligent devices and monitoring technologies are introduced

Manuscript received 23 October 2023; revised 10 December 2023;

accepted 6 February 2024. Date of publication 29 February 2024; date

of current version 6 May 2024. This work was supported in part by the

National Key Research and Development Program of China under Grant

2021YFB3301300 and in part by ZTE Industry-University-Institute Co-

operation Funds. Paper no. TII-23-3628. (Corresponding author: Juan

Xu.)

Zehui Mao, Huan Wang, and Bin Jiang are with the College of

Automation Engineering, Nanjing University of Aeronautics and As-

tronautics, Nanjing 210016, China (e-mail: zehuimao@nuaa.edu.cn;

wanghuan233@nuaa.edu.cn; binjiang@nuaa.edu.cn).

Juan Xu is with the College of Computer Science and Technology,

Nanjing University of Aeronautics and Astronautics, Nanjing 210016,

China (e-mail: juanxu@nuaa.edu.cn).

Huifeng Guo is with the State Key Laboratory of Mobile Network and

Mobile Multimedia Technology, Shenzhen 518000, China, and also with

the ZTE Corporation, Shenzhen 518000, China (e-mail: guo.huifeng2

@zte.com.cn).

Color versions of one or more ﬁgures in this article are available at

https://doi.org/10.1109/TII.2024.3367010.

Digital Object Identiﬁer 10.1109/TII.2024.3367010

into them. However, this leads to an increasing requirement on

reliability and safety of industrial machines subjected to faults

and failures [1]. Fault diagnosis that detects the occurrence of a

fault as early as possible and identiﬁes the location and type of

the fault as accurately as possible, is a key mean to ensure the

safety of industrial machines [2], [3].

With the development of technology, the fault diagnosis

methods are receiving more and more attention, including the

model-based methods [4], [5], data driven-based methods [6],

[7], [8], [9], and knowledge-based methods [10]. As a signiﬁcant

amount of maintenance records can be accumulated during the

long-term maintenance and repair processes of industrial ma-

chines, which contain the valuable knowledge and information

related to machine fault diagnosis.

Knowledge-based fault diagnosis methods commonly include

expert systems, fault tree analysis, and knowledge graphs (KGs).

By automating the acquisition, organization, and analysis of

various machine information, including maintenance history and

expert experience, KG-based fault diagnosis methods can iden-

tify fault root causes and provide solutions. Existing research

mainly utilizes KG reasoning techniques [11], [12], [13], which

can make inferences about potential causes or solutions based

on entities and relations. During the machine operating, the

new data and information reﬂecting new faults often emerge

often generates. Incorporating new data and information into

existing KGs and reasoning requires ofﬂine updates, difﬁcult to

be updated and improved dynamically.

Recommendation systems, by continuously collecting feed-

back from maintenance personnel during the maintenance pro-

cess, can achieve dynamic updates and continuously improve

the accuracy of fault cause recommendations, addressing the

issue of dynamic updates in KG-based fault diagnosis. Fusing

the KGs and recommendation algorithms can address the issue

of dynamic updates in KG-based fault diagnosis, which can be

primarily achieved through the propagation-based methods [14],

[15]. Propagation-based methods can expand the depth of infor-

mation reception by following the deep aggregation, thereby

fully utilizing the information in the KG to better predict fault

causes and locate potential fault causes. Deep aggregation is

the core of graph convolutional neural networks (GCN) [16],in

which GraphSAGE [17] is a representative model. But GCNs

like GraphSAGE is not feasible for weighted graphs, as they

can only perform equally weighted aggregations of neighboring

See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: ZTE CORPORATION. Downloaded on November 26,2024 at 06:05:55 UTC from IEEE Xplore. Restrictions apply.

MAO et al.: GRAPH CONVOLUTIONAL NEURAL NETWORK FOR INTELLIGENT FAULT DIAGNOSIS OF MACHINES VIA KNOWLEDGE GRAPH 7863

nodes. In addition, GraphSAGE imposes limitations on the

number of sampled neighbors, leading to the loss of important

local information for some nodes and making it unsuitable

for representation learning tasks in KGs and requiring further

improvements.

This article aims to address the challenges of difﬁcult fault

knowledge mining and low real-time performance by combining

KGs with recommendation systems. To tackle these challenges,

a novel knowledge aggregation fault diagnosis (KAFD) model

is proposed. The main contributions of this article are as follows.

1) As KGs and recommendation algorithms are combined

and introduced into the industrial machine fault diagnosis,

a new knowledge aggregation network (KAN) is de-

signed, which incorporates sensitivity factor to make the

GCN suitable for the structure of fault diagnosis knowl-

edge graph (FDKG). This design allows for weighted

operations on different nodes of the FDKG, considering

the importance of different neighboring entities.

2) The deep aggregation idea is applied, and the KAFD

model is constructed. By stacking KAN, KAFD expands

the depth of information reception, facilitating multihop

reasoning in the KG and further exploration of potential

fault causes.

The rest of this article are organized as follows. Section II

describes the fault diagnosis problem based on KGs. Section III

introduces the structure of the KAFD model. Section IV presents

and analyzes the performance of the KAFD model. Finally,

Section V concludes this article.

II. P

ROBLEM FORMULATION

This section analyzes the task of KG-based fault diagnosis

for industrial machines, and provides a formulation to describe

the fault diagnosis problem achieved by recommending the fault

root cause from the fault phenomenon.

A. Fault Phenomenon-Fault Root Cause Bipartite Graph

(FBG)

For fault diagnosis, recommendation system determines the

root cause of the fault from the fault phenomenon. In the

recommendation system, associations between fault phenom-

ena and causes are represented as an FBG deﬁned as U =

|f ∈F,s∈S}, where F is the set of fault phenomena,

S is the set of fault root causes, and y

correlation between

fault root causes and fault phenomena as



1,fand s have connection in history

0, otherwise.

. (1)

Due to the sparsity problem of the practical fault data from

the industrial machines, numerous node in the generated fault

bipartite graph only have a few correlative connections, as shown

in Fig. 1, where blue circles are the fault phenomenon and red

circle are the fault root causes. The sparse correlative connec-

tions cause some information missed in the recommendation

system, and the associations between fault phenomena and fault

root causes cannot be accurately expressed, which effect the

Fig. 1. Example of an FBG sparsity problem.

accuracy of fault diagnosis. However, the FDKG, established us-

ing the triples obtained by the knowledge extraction technology,

contains the substantial knowledge about faults. An extra KG can

help supplement the information and improve the representation

of the associations in FBG [18].

B. Fault Diagnosis Knowledge Graph

The method of taking FDKGs as auxiliary information of the

recommendation system contributes to enrich the representation

of fault phenomena and root causes, relieves the sparsity problem

of fault data, and enhances the accuracy and interpretability of

the model.

Similar to the generic KG, FDKG consists of a large number of

triples (h, r, t) as the basic units, denoted as G = {(h, r, t)|h, t ∈

E,r ∈R}, where E is the set of all entities, and R represents

the set of all relations. In the KG, facts are expressed as triples

(head entity, relation, tail entity), which are interconnected to

express real-world knowledge and facilitate computer under-

standing. In the FDKG, facts are expressed as triples (head

entity, relation, tail entity), which interconnect each other to

express knowledge for computer understanding. In this study, the

FDKG is constructed by unstructured data of industrial machines

accumulated during the production process.

C. Collaborative Knowledge Graph (CKG)

As the FDKG and the FBG are established using the different

and independent information of fault data, a CKG which can

associates these two graphs into a uniﬁed relational network to

unify the two parts of information, is necessary. The associations

between the fault phenomenon and the fault root cause are

expressed in the form of triples (phenomenon, connect, cause)

in the CKG, where y

= 1 denotes as an additional relation

connect between f and s. Deﬁne the alignment set as A, which

denotes that the root cause s of the fault in the FBG can be

aligned to the entity e in the FDKG. Based on the alignment

set A, the FBG and the FDKG are combined into a uniﬁed

CKG G



= {(h, r, t)|h, t ∈E



,r ∈R



}, where E



= E∪Fand



= R∪{connect}.

D. High-Order Relation

In a uniﬁed CKG, high-order relations between fault phe-

nomena and root causes could be established by bridging two

entities that not directly connected. Deﬁne the L-order connec-

tion between nodes as a high-order relation path e

−→ e

−→

···

−→ e

, where r

∈E



, r

∈R



, l ∈ L. More correlation

information between entities could be obtained. As illustrated

Authorized licensed use limited to: ZTE CORPORATION. Downloaded on November 26,2024 at 06:05:55 UTC from IEEE Xplore. Restrictions apply.

7864 IEEE TRANSACTIONS ON INDUSTRIAL INFORMATICS, VOL. 20, NO. 5, MAY 2024

Fig. 2. Example of CKG.

in Fig. 2, the path f

− s−

− e

− s

reveals the

potential connection between f

and s

, which is difﬁcult to

reﬂect in a single FBG. The high-order relation from one entity to

another entity requires the connection of multiple relationships,

which is the so-called multihop.

E. Fault Diagnosis Task

For the fault diagnosis model that recommends root causes

by the fault phenomena, the input is the historical records of the

FBG and the FDKG generated by fault records, while the output

is a sequence of fault root causes corresponding to the observed

fault phenomena.

Given the FBG U and the FDKG G of industrial machine fault,

the probability ˆy

of predicting the candidate fault root cause s

that is the potential cause of the observed fault phenomenon f,

can be expressed as

ˆy

= Func(f,s|Θ, U, G) (2)

where Θ is the parameters of the prediction function. Then, the

fault root causes is sorted according to the prediction probability.

Based on the predicted probabilities, the fault root causes are

ranked, which serves as the result of fault diagnosis to enhance

the efﬁciency of troubleshooting for operational personnel.

III. M

ETHOD

In order to improve the accuracy of fault diagnosis, we will

propose a new KG-based fault diagnosis model named KAFD,

to obtain the high-order relations between fault phenomena and

fault root causes, in which the KAN incorporated. Through

the KAN, multihop connections of fault phenomena and root

causes are revealed, which makes the KAFD can learn the fault

diagnosis knowledge from the CKG, to obtain the reason from

observed phenomena to potential root causes.

A. Overall Framework

The KAFD model, shown in Fig. 3, includes multiple KANs

and a prediction module. The KAN module primarily serves

to aggregate t he neighboring entities and enable to the update

of central entity features. Through stacking of h layers KAN,

KAFD can aggregate information from multiple-hop entity,

resulting in more comprehensive fault root cause features. The

prediction module is responsible for calculating the probability

of each pair of fault phenomena f and fault root causes s. KAFD

ultimately outputs the top-k fault root causes with the highest

probabilities as the fault diagnosis results.

B. Knowledge Aggregation Network

KAN is a revised version of GCN GraphSAGE to meet

the needs of KG-based fault diagnosis tasks. There are three

improvements, including the following.

1) Design sensitivity factors to address the issue that differ-

ent fault root causes have diverse importance for the fault

phenomena, achieving weighted aggregation of different

neighbor entities.

2) Propose sampling strategy with controllable depth and

quantity for the scale differences of FDKGs in different

diagnosis tasks, so as to adjust the size of receptive ﬁeld

ﬂexibly for exploring potential fault root causes.

3) Exploit three aggregation functions for various fault di-

agnosis tasks, satisfying the features of different FDKGs.

The structure of KAN is illustrated in Fig. 4.

The speciﬁc measures for the three improvements are shown

as follows.

1) Sensitivity Factors: Considering the varying sensitivity of

fault phenomena to different fault r oot causes, certain fault root

causes are more prone to trigger the fault phenomena, exerting

greater inﬂuence on the fault phenomena features. In order to

prioritize the inﬂuence of fault root causes on fault phenomena

features, sensitivity factors are designed to enable weighted

aggregation of neighboring entities.

For each pair of fault phenomena f and fault root cause s,the

set of neighboring entities adjacent to the fault root cause entity is

deﬁned as N (s), and r

a,b

denotes the relation between entities a

and b. If a fault phenomenon f has been triggered by the s peciﬁc

fault root cause s in fault records, the entities associated with this

fault root cause are searched in the FDKG, and neighbor entity

aggregation is performed. During the aggregation process, the

sensitivity factors are calculated using the inner product function

g to bias the aggregation

= g(f , r) (3)

where f ∈ R

and r ∈ R

are the representations of the fault

phenomenon f and the relation r, respectively, and d is the

dimension of the representations. The parameter ζ

indicates

the importance level of the relation r to the fault phenomenon

f. It can be employed to depict the propensity of a speciﬁc fault

phenomenon being triggered by different types of causes.

To describe the topological neighborhood structure of fault

root cause s in FDKG, it is necessary to compute the linear

combination of sensitivity factors for its neighborhood

N(s)



e∈N(s)



s,e

e (4)

where e is the representation of neighboring entities of the

root cause entity in the FDKG, and



s,e

is the normalized

representation of ζ



s,e

exp



s,e





e∈N(s)

exp



s,e



. (5)

Through the application of sensitivity factors, the neighborhood

representation s

N(s)

of the root cause can be obtained, which

Authorized licensed use limited to: ZTE CORPORATION. Downloaded on November 26,2024 at 06:05:55 UTC from IEEE Xplore. Restrictions apply.

of 9

免费下载

paper goldendb

文档被以下合辑收录

GoldenDB论文（共12篇）

GoldenDB论文

文档被以下合辑收录

相关文档

评论