Mechanical Fault Prediction Based on Event Knowledge Graph

Li He

, Liang Zhang

2,∗

and Wei Yan

1,∗

School of Information Science and Engineering, Shandong Normal University, Jinan 250358, Shandong, China

Institute of Frontier and Interdiscriplinary Science and Key Laboratory of Particle Physics and Particle Irradiation (MOE),

Shandong University, Qingdao 266237, Shandong, China

Keywords:

Event Knowledge Graph, Mechanical Fault, Convolutional Neural Networks, Fault Diagnosis, Fault

Prediction.

Abstract:

Currently, the majority of diagnoses in the ﬁeld of mechanical faults are performed by experts or expert

systems, which require domain experts to guide the completion while having subpar and limited portability.

Consequently, we analyzed the current situation of rolling bearing fault, event knowledge graph and convo-

lutional neural network, explored the intelligent fault diagnosis and prediction technology of rolling bearing,

introduced event knowledge graph and convolutional neural network in rolling bearing fault diagnosis, pro-

vided support for fault diagnosis and prediction, enhanced the accuracy of fault diagnosis and prediction, and

enabled the predictive of complex mechanical equations.

1 INTRODUCTION

Despite the fact that fault diagnosis technology is an

applied and marginal ﬁeld, the emphasis is transi-

tioning toward information technology, digitalization,

and artiﬁcial intelligence as a result of technologi-

cal advancements. The sensitivity and complexity of

the large-scale mechanical apparatus have increased.

Moreover, calamitous incidents caused by mechanical

equipment malfunctions have become more prevalent

in recent years as a result of the signiﬁcant expansion

in the complexity of mechanical equipment. There-

fore, the use of fault diagnosis and predictive tech-

niques for predictive maintenance of apparatus and

equipment is crucial for maintaining the equipment’s

operation stability and preventing losses due to out-

ages.

Theoretical foundations for fault diagnosis tech-

nologies include modern control theory, computer en-

gineering, mathematical statistics, fuzzy set theory,

signal processing, pattern recognition, etc. The objec-

tive of fault diagnosis is to determine the fault’s char-

acteristic description and to detect and isolate the fault

based on the various (measurable or unmeasurable)

quantities in the system, or some of them, exhibiting

characteristics distinct from their normal state when a

fault occurs(An, 2008).In the initial stages, manual di-

agnosis is used. However, the traditional technology

*Corresponding authors.

of fault diagnosis has been unable to satisfy the actual

requirements for the operation of complex mechani-

cal equipment. The prediction model is a crucial com-

ponent of the technology for predicting mechanical

faults. Current fault prediction models include curve

ﬁtting models, ﬁltering models, time series models,

grey models, artiﬁcial neural network models, and

fuzzy models.

Knowledge Graph (KG) is a popular form of

knowledge representation, published by Google in

2012. It focuses on entities and their relationships,

thus representing static knowledge. And existing

knowledge graphs in the ﬁeld of mechanical fault

are frequently static graph, where the degree of fault

is inferred by associating semantic information with

the data(Yan et al., 2022). However, the world con-

tains a vast quantity of event information that conveys

dynamic and procedural knowledge, making event-

centric knowledge representations such as the Event

Knowledge Graph (EKG) essential.

In this paper, in contrast to the conventional ap-

proach of using static knowledge graph, we use

Event-Ontology(Han et al., 2007) combined with sig-

nal data from mechanical fault data sets to construct

event knowledge graph. We also utilize the data in the

event knowledge graph to construct a convolutional

neural network model for fault diagnosis and predic-

tion functions, and then update the event knowledge

graph with the results of the diagnosis and prediction.

164

He, L., Zhang, L. and Yan, W.

Mechanical Fault Prediction Based on Event Knowledge Graph.

DOI: 10.5220/0012178200003598

In Proceedings of the 15th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2023) - Volume 2: KEOD, pages 164-173

ISBN: 978-989-758-671-2; ISSN: 2184-3228

This knowledge graph can be utilized for predictive

maintenance of machinery and equipment to avoid or

minimize downtime, thereby ensuring that machinery

and equipment can be utilized for as long as possible.

2 RELATED WORK

Currently, mechanical fault prediction (MFP) and

diagnosis are usually implemented independently

by knowledge graphs or convolutional neural net-

works. Laibin Zhang et al.(Zhang et al., 2007) pro-

posed a multidimensional fault characteristic parame-

ter model for rotating parts for the purpose of mechan-

ical fault diagnosis technology. Speciﬁcally, a fuzzy

gray optimal prediction method is proposed to predict

short-term faults. The method is able to build a pre-

diction model with a minimum of four data and can

effectively handle the nonlinearity of the prediction

data. Shi Yu et al.(Yu Shi, 2022) proposed a model

for fault diagnosis using a knowledge graph and fuzzy

inference. By constructing a knowledge graph of the

mechanical fault domain, a fuzzy ontology, and com-

municating the knowledge graph and fuzzy ontology

in a mapping manner, the equipment status can be di-

agnosed using fuzzy reasoning to derive fault fuzzy

values. Although the fuzzy inference model can also

accomplish the diagnosis and prediction of mechani-

cal faults, its implementation requires the manual or

semi-automatic construction of a fuzzy inference rule

base and the setting of thresholds for the addition

of inference labels. The rules in the rule repository

should not have an excessive number of parameters,

as this would reduce the efﬁcacy of inference.

Although Google introduced the Knowledge

Graph concept formally in 2012, it has been under

development since the Semantic Web was ﬁrst pro-

posed in 1960. Early knowledge bases, open knowl-

edge graphs, Chinese common knowledge graphs,

and domain knowledge graphs are the various types

of knowledge graphs(Tian et al., 2021). The event

knowledge graph is a dynamic knowledge graph, as

opposed to a static knowledge graph, and the data is

frequently updated and has a certain timeliness(Zhao

et al., 2022); thus, it is difﬁcult to achieve improved

results when using inference methods such as fuzzy

inference for prediction or diagnosis. Typically, event

knowledge graphs employ an ontology-based con-

struction process for knowledge modelling, wherein

the top-level representation model of event knowl-

edge graphs is developed ﬁrst, followed by the re-

ﬁnement of supporting examples. This method is

comparable to that of static knowledge graphs. And

the method proposed by Zhao Xiaojun et al.(Zhao

et al., 2022) for constructing a dynamic military do-

main graph by processing dynamic data according to

data update frequency and usage frequency can better

satisfy the requirements of a domain dynamic knowl-

edge graph for high quality and timeliness.

Knowledge graph is a signiﬁcant infrastructure for

AI technology and is essential for computers to ac-

complish human-like inference and prediction capa-

bilities. Compared with shallow models, deep learn-

ing has more obvious advantages in both feature ex-

traction and modeling. Deep learning is better at min-

ing abstract feature representations from original data,

and these feature representations usually have bet-

ter generalization ability. From AlexNet(Krizhevsky

et al., 2012), VGGnet(Simonyan and Zisserman,

2014), and GoogLeNet(Szegedy et al., 2015), which

participated in the ILSVRC competition and won, it is

clear that convolutional neural networks are the undis-

puted leaders in computer vision and have excellent

generalization. Therefore, it is reasonable to assume

that convolutional neural networks can contribute new

concepts and techniques to the ﬁeld of mechanical

fault diagnosis.

By taking advantage of convolutional neural net-

works and building CNN models that act directly

on time-domain signals, features useful for diagnosis

can be learned automatically without manual extrac-

tion(Zhang, 2017). Using the upper channel CNN to

extract feature information from the vibration signal

and the lower channel Bidirectional Gated Recurrent

Unit(BiGRU) network to extract the temporal features

of the vibration signal from both positive and negative

directions enables the diagnosis of faults by combin-

ing the beneﬁts of both network models.

Most implementations of the fault prediction func-

tion in the aforementioned papers use CNNs for pre-

diction or only KGs for inference, whereas by com-

bining the prediction capability of CNNs with the

event attributes of EKGs, we are able to more effec-

tively predict mechanical faults.

3 METHODOLOGY

The methodology framework for the mechanical fault

prediction model based on event knowledge graph

and deep convolutional neural network is depicted in

Figure 1.

Four major components comprise the model: the

input layer, the EKG layer, the Deep Convolu-

tion Neural Networks with Wide First-Layer Kernels

(WDCNN) layer, and the output layer. The model

input consists of as much data as feasible, including

fan end and drive end vibration data, motor revolu-

Mechanical Fault Prediction Based on Event Knowledge Graph

165

Figure 1: Framework of the prediction system.

tions per minute (RPM), and time series data. The

EKG layer then extracts relevant information from the

data to establish entities and their relationships. By

reading the EKG layer’s data, the WDCNN layer per-

forms fault prediction using the pre-trained WDCNN

model. When a fault is predicted to be possible, a

new entity is added to the EKG, and its relationships

with other entities, such as the (entity-fault type) re-

lationship, are established; if no fault is predicted, the

corresponding entity’s attributes and relationships are

merely updated. The output layer obtains the output

by reading the attributes and inter-entity relationships

of the altered or added entities in the EKG layer.

3.1 EKG Layer

In this paper, we constructed the event knowledge

graph based on the seven-step method(Noy and

Mcguiness, 1995) and the 6H method(Han et al.,

2007). At present, the more widely used knowledge

graph construction methods are: IDEF-5 method,

Skeletal Methontolody, TOVE method, seven-step

method(Noy and Mcguiness, 1995) and Methontol-

ogy method(Fern

andez-L

opez et al., 1997). In the

construction of domain ontologies, the seven-step

method and the NeOn Methontology method are now

utilized more frequently. Below are brief explanations

of the two aforementioned methods: Currently, the

most popular methods for building knowledge graphs

are the IDEF-5 method, Skeletal Methontolody,

TOVE method, seven-step method(Noy and Mcgui-

ness, 1995), and Methontology method(Fern

andez-

opez et al., 1997). More frequently, the seven-step

method and the NeOn Methontology method are em-

ployed in the construction of domain ontologies to-

day. Detailed explanations of the two aforementioned

methodologies are provided below.

(i) Seven-Step Method (Noy and Mcguiness, 1995):

The Stanford University School of Medicine devised

this method, which is primarily used for the construc-

tion of domain ontologies. The seven stages of the

method are: (1) identifying the domain of expertise

and scope of the ontology; (2) investigating the pos-

sibility of utilizing existing ontologies; and (3) devel-

oping the ontology. (3) listing the key terms in the

ontology; (4) deﬁning classes and class hierarchies;

(5) deﬁning the attributes of classes; (6) deﬁning the

constraints of the attributes; (7) generating instances.

(ii) NeOn Methontology Method (Gomez-Perez

and Su

arez-Figueroa, 2009): The core idea of

NeOn methodology is to integrate ontologies from

different domains into a uniﬁed knowledge graph.

It is constructed using a ”divide-and-conquer” ap-

proach, in which the overall problem to be addressed

is decomposed into subproblems, and solutions to the

overall problem (i.e., the development of the ontology

network) are obtained by combining the solutions to

the subproblems.

Event-Ontology is deﬁned based on fundamental

factors which describe an event. That means the 6H of

’who’, ’what’, ’where’, and ’how’ are represented as

properties of an event. The property value of ”who”

is the subject of an event, while the property value

of ”how” is the subject’s action or the event’s con-

tent. ”What” refers to the substance of an event, while

”where” and ”when” indicate the event’s location and

time, respectively.(Han et al., 2007).

EKG gave to construct entities and relationships

between entities by extracting as much data infor-

mation as possible in the dataset, such as vibration

signal data, fault occurrence locations, fault types,

etc. This paper refered to the knowledge graph de-

veloped by Liu Xin(Liu, 2019) of Beijing University

of Posts and Telecommunications using the improved

seven-step methodology. Because mechanical fault

event knowledge graph is a type of domain knowledge

graph, and in order to share a common understand-

ing of the information structure and make the domain

knowledge reusable, it is typically necessary to deter-

mine whether there are existing ontologies that can be

reused before developing the ontology. The structure

Liu proposed is enhanced on the basis of the ontology

class hierarchy diagram of this knowledge graph to

emphasize the features of the event knowledge graph.

By deleting class structures that are not relevant to this

paper and adding classes corresponding to the infor-

mation extracted above, an ontology class hierarchy

that better highlights the characteristics of the event

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

166

knowledge graph is constructed.

3.2 WDCNN Layer

The WDCNN layer determines the structure of the

model and the parameters of each layer by extract-

ing the vibration signal data from the established

knowledge map of mechanical failure events and sev-

eral preliminary experiments, thus establishing a WD-

CNN model with the function of diagnosis and pre-

diction of mechanical faults, whose diagnostic func-

tion can recognize ten types of failures. As a nor-

mal Convolution Neural Network(CNN), WDCNN

also has a ﬁve-layer structure , which are input layer,

convolution layer, pooling layer(or sampling layer),

full connection layer and output layer. Among them,

the convolution layer and pooling layer are set alter-

nately. In contrast to conventional convolution neural

networks, however, WDCNN have a large convolu-

tion kernel in the ﬁrst layer and smaller convolution

kernels in subsequent layers. Due to the small num-

ber of convolution kernel parameters, WDCNN has

a deeper network and can suppress overﬁtting, so it

has a better expressive ability compared with ordinary

CNNs. We can also ascertain the applicability of the

model’s ﬁndings by examining the displayed model

assessment curves. It implements its prediction func-

tion by determining whether mechanical equipment is

likely to malfunction based on a threshold value of the

difference between predicted and actual values.

3.3 Output

In the output layer, mechanical fault diagnosis and

prediction results are obtained by retrieving the up-

dated EKG output results, which can be done by

clicking on the visualization interface or using query

statements.

We implemented the query function of this me-

chanical fault event knowledge graph in Neo4j. Be-

cause of its ability to visualize the entities we create

and the relationships between them compared to other

databases. In addition, Neo4j can be queried with

Cypher statements or simple clicks, which makes it

accessible to people who do not have knowledge of

database queries. Cypher’s fundamental syntax com-

prises of four distinct sections, each with a unique

rule: (1) Start: identify the starting node in the graph;

(2) Match: match the graph pattern to locate sub-

graphs of the desired data; and (3) Where: ﬁlter the

data based on certain criteria. data; (4) Return: de-

liver the desired results.

4 EXPERIMENTAL SETUP

In this section, we discuss all the experimental envi-

ronments we considered to perform the Event Knowl-

edge Graph of Mechanical Fault Prediction. In the

following subsections we discuss: (i) the data set used

to build and test; (ii) the metrics used to assess perfor-

mances; and (iii) the implementation aspects of the

proposed EKG of MFP model.

4.1 Data Sets Description

The experimental dataset used for this paper is the

rolling bearing dataset published by the Rolling Bear-

ing Data Center at Case Western Reserve Univer-

sity(CWRU, 2023). This dataset is presently recog-

nized as the standard dataset for fault diagnosis of

rolling bearings worldwide. The investigations in this

paper therefore make use of the Normal Baseline, 12k

Drive End Bearing Fault, and 12k Fan End Bearing

Fault data sets from the CWRU dataset. We used the

data in Normal Baseline and 12k Drive End Bearing

Fault to construct the event knowledge graph, then

train and construct WDCNN using the event knowl-

edge graph’s data, and ﬁnally use the data in 12k Fan

End Bearing Fault to test the generalizability of the

established EKG of MFP.

4.2 Data Sets Pre-Processing

The initial data set is stored in Matlab (*.mat) ﬁles,

each ﬁle contains fan and drive end vibration data in

addition to motor speed.

The experimental object of this paper is a deep

groove ball bearing SKF6205 model drive end bear-

ing. The faulty bearing is machined by EDM with a

system sampling frequency of 12kHz, and it is sepa-

rated into three categories based on its damaged parts:

inner race, outer race, and ball, and four categories

based on its fault diameters: 0.007, 0.014, 0.021, and

0.028 inches. However, due to the lack of outer race

damage data for the 0.028 inch diameter portion, the

0.007, 0.014, and 0.021 damage diameters were se-

lected, for a total of 3×3 fault states. As outer race

damage is a stationary damage, the location of the

fault in relation to the bearing load zone has a direct

impact on the vibratory response of the motor/bearing

system. The outer race fault location chosen for this

research is in the 6 o’clock direction (directly within

the load zone). The data sets used are shown in the

Table 1 and Table 2.

Then, the data set is randomly partitioned into

train, valid, and test, and the partitioning ratio is set

to 0.7: 0.1: 0.2 so that training and test set samples

Mechanical Fault Prediction Based on Event Knowledge Graph

167

Table 1: Normal Baseline Data.

Motor Load (HP) Approx. Motor Speed (rpm) Normal Baseline Data

0 1797 97.mat

1 1772 98.mat

2 1750 99.mat

3 1730 100.mat

do not overlap. In each experiment, 1024 data are uti-

lized for fault diagnosis.

In addition, there are 10 fault categories that have

been artiﬁcially speciﬁed, corresponding to 1 fault-

free status and 9 fault statuses with three fault loca-

tions (inner race, ball, outer race) and three fault di-

ameters (0.007, 0.014, 0.021) in two combinations.

Table 3 lists the types corresponding to these fault sta-

tus.

Due to the fact that the subsequent query includes

bearing location, motor load horsepower (HP), mo-

tor speed, bearing operation status, fault diameter (in

fault status), fault occurrence location (in fault status),

and fault name, etc., the corresponding data columns

are added to the data set and output as excel (.csv)

ﬁles.

Before conducting fault prediction, it is necessary

to obtain new data. In this paper, we utilize the data in

the 12k DriveEndBearingFaultData ﬁle with the split

ratio set to [train: valid: test]=[0.1: 0.1: 0.80], pre-

sume the data in the test section to be recently col-

lected data, and transform the imported data into a

format acceptable by the model input layer.

4.3 Metrics

The purpose of this paper is to accomplish supervised

diagnosis and prediction of mechanical faults. Ac-

curacy and loss curves are frequently used as crucial

metrics for evaluating the efﬁcacy of machine learn-

ing models, and loss curves can reﬂect the dynamic

trend of network training. To evaluate the diagnostic

and predictive quality of the WDCNN model devel-

oped in this paper for bearing faults, we observe the

loss curve to determine whether the model converges

and ﬁts, and the acc curve to determine the model’s

applicability for bearing fault prediction.

4.4 Implementation Details

The construction of EKG and the establishment of

WDCNN are carried out using the methodologies de-

scribed above.

EKG Construction: First, we analyzed the data in

the CWRU bearing fault dataset and identify the key

terms in the ontology; then, we deﬁned the classes and

the class hierarchy based on these terms. We used a

combination of top-down and bottom-up approaches

and referred to the ontology class hierarchy of the

knowledge graph created by Xin Liu of Beijing Uni-

versity of Posts and Telecommunications(Liu, 2019)

to deﬁne classes and class hierarchies in the EKG.

Analysis of the CWRU rolling bearing dataset from

the bottom-up layer provides access to the important

terms in the ﬁeld of rolling bearing faults: bearing

operating part, bearing physical component, bearing

fault type, bearing operating status, bearing operating

conditions, bearing protection methods, etc., which

can all be used as concepts extracted from the top-

down layer to construct the ontology. The bottom

stratum of data contains bearing fault diameters, time

series data, fault varieties, etc. Therefore, the bearing

operation status can be divided into two categories:

normal and abnormal, with the fault status subdivided

further into fault vibration signal, fault diameter, etc.

In summary, the class hierarchy can be deduced as fol-

lows: A rolling bearing includes several categories of

operating part, physical component, operating status,

faults, and bearing protection methods. The operating

status contains subcategories for normal state and ab-

normal state. The normal state subcategory contains

normal state signal data, while the abnormal state sub-

category contains fault signal data and speciﬁc fault

type. The operating condition also includes motor

load horsepower and motor speed.

Then, to facilitate subsequent query functions and

data updates, we specify the class’s data attributes and

relational attributes, upon which we deﬁne the con-

straints for each attribute. Instances are subsequently

created in the Neo4j graph database to conclude the

construction of the bearing fault event knowledge

graph.

WDCNN Establishment: The purpose of establish-

ing a convolutional neural network in this paper is

to realize the diagnosis and prediction of mechani-

cal faults, using tensorﬂow in python to establish the

initial CNN and then selecting the WDCNN with su-

perior performance based on a series of preliminary

experiments. The WDCNN consists of ﬁve convolu-

tional layers, ﬁve pooling layers, one full connection

and one Softmax layer. Following are the parame-

ters of the convolutional neural network: The size of

the ﬁrst convolutional kernel is 64×11 with a 16×11

stride size, the size of the remaining convolutional

layers is 3×11 with a 2×11 stride size, and the size of

each of the ﬁve pooling layers is 2×11 with a 2×11

stride size. In addition, the Softmax layer is conﬁg-

ured to generate 10 outputs for 10 bearing fault states.

The parameters of the WDCNN model used in the ex-

periments are shown in Table 4.

Moreover, the Adadelta optimization algo-

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

168

Table 2: 12k Drive End Bearing Fault Data.

Fault Diameter Motor Load (HP) Approx. Motor Speed (rpm) Inner Race Ball

Outer Race

Position Relative to Load Zone (Load Zone Centered at 6:00)

Centered

@6:00

Centered

@3:00

Centered

@12:00

0.007” 0 1797 105.mat 118.mat 130.mat 144.mat 156.mat

1 1772 106.mat 119.mat 131.mat 145.mat 158.mat

2 1750 107.mat 120.mat 132.mat 146.mat 159.mat

3 1730 108.mat 121.mat 133.mat 147.mat 160.mat

0.014” 0 1797 169.mat 185.mat 197.mat * *

1 1772 170.mat 186.mat 198.mat * *

2 1750 171.mat 187.mat 199.mat * *

3 1730 172.mat 188.mat 200.mat * *

0.021” 0 1797 209.mat 222.mat 234.mat 246.mat 258.mat

1 1772 210.mat 223.mat 235.mat 247.mat 259.mat

2 1750 211.mat 224.mat 236.mat 248.mat 260.mat

3 1730 212.mat 225.mat 237.mat 249.mat 261.mat

0.028” 0 1797 3001.mat 3005.mat * * *

1 1772 3002.mat 3006.mat * * *

2 1750 3003.mat 3007.mat * * *

3 1730 3004.mat 3008.mat * * *

Table 3: 10 types of fault names.

diameter\position inner race ball outer race

0.007 1 2 3

0.014 4 5 6

0.021 7 8 9

normal (no fault) 0

Table 4: Parameters of WDCNN.

No. Network layers

kernel size

/stride size

Number of kernels

output

(width × depth)

1 Convolutional layer1 64×1/16×1 16 128×116

2 Pooling layer 2×11/2×11 16 64×116

3 Convolutional layer 3×11/1×11 32 64×132

4 Pooling layer 2×11/2×11 32 32×132

5 Convolutional layer 3×11/1×11 64 32×164

6 Pooling layer 2×11/2×11 64 16×164

7 Convolutional layer 3×11/1×11 64 16×164

8 Pooling layer 2×11/2×11 64 8×164

9 Convolutional layer 3×11/1×11 64 6×164

10 Pooling layer 2×11/2×11 64 3×164

11 full connection 100 1 100×11

12 Softmax 10 1 10

rithm(Zeiler, 2012) has been used to revise the

weights in this paper. The Adadelta optimization

algorithm is an improvement of the Adagrad op-

timization algorithm, in which the exponentially

weighted average of the squares of the gradients is

used instead of the sum of squares of all gradients

for each dimension, thereby avoiding the problem

that the update magnitude gradually tends to zero in

the late training period; the exponentially weighted

average of the squares of the updates dynamically

replaces the global scalar learning rate, thereby

avoiding truncation.

5 RESULT AND DISCUSSION

In this section, we discuss the query function imple-

mented by the bearing fault EKG and the performance

of WDCNN in terms of loss and accuracy of fault di-

agnosis and prediction.

5.1 Query Functions of Bearing Fault

EKG

In this paper, we present four kinds of most prevalent

query functions based on the Cypher query pattern

and the visual interface provided by Neo4j. Analysis

of the constructed event knowledge graph of bearing

fault can identify the subclass or subclasses of entities

that the entity being queried falls under. By pressing

on nodes, the data attributes associated with nodes

and the relationships between nodes can be queried

based on the visual interface.

Rolling Bearing Operation Status and Physical

Components and Other Parameters: This query

searches for entities and their vibration signal data/

fault vibration signal data data attributes that belong

to the Normal or Abnormal subclasses of the oper-

ating status class. Figure 2 depicts the query result

when the operation status is normal, while Figure 3

depicts the query result when the operation status is

abnormal.

Rolling Bearing Fault Type and Fault Position:

The query is for the ”where” attribute in the event

Mechanical Fault Prediction Based on Event Knowledge Graph

169

Figure 2: Query result of normal status.

Figure 3: Query result of abnormal status.

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

170

Figure 4: Query result of fault type and position.

knowledge graph, and there are two querying meth-

ods: (i) directly querying the fault type and returning

the fault type node directly; (ii) transforming it into a

query for the ”happened at” relationship query, which

returns the type of fault that occurred on the physi-

cal component. The query result for the fault type

”fault 7” is illustrated in Figure 4.

Rolling Bearing Fault Type and Solution: The

query aims to identify a solution based on the type

of defect, which can be translated into a query for the

”solution is” relationship for the starting node speci-

ﬁed. Take the fault type ”fault 5” as an example, the

result of its fault solution query is shown in Figure 5.

Figure 5: Query result of fault type and solution.

Rolling Bearing Physical Components and Main-

tenance Methods: The query intends to identify

the corresponding maintenance solution based on the

physical components of the rolling bearing, i.e., spec-

ify the starting node as a physical component entity

and locate the end node with the relationship ”main-

tenance method is”. Figure 6 depicts the result of the

maintenance method inquiry regarding the inner race

malfunction.

Figure 6: Query result of physical components and mainte-

nance methods.

5.2 WDCNN Evaluation

The dataset is trained using a deep convolutional neu-

ral network constructed according to Section 4.5, with

the learning rate of Adadelta set to 1.0 and the number

of training epochs set to 500. Figure 7 and Figure 8

depict the loss and accuracy curves for each training

and test, respectively.

With a decent ﬁt, the recognition rate of the con-

volutional neural network exceeds 99.8% on the train-

ing dataset and 97% on the test dataset, as depicted

in the ﬁgures. The experimental results demonstrate

that WDCNN has a strong ability to recognize the test

sample data.

The WDCNN model established in the preceding

section is applied to assess the test data and calculate

the average loss and accuracy values for the newly

Mechanical Fault Prediction Based on Event Knowledge Graph

171

Figure 7: Loss curve of training.

Figure 8: Accuracy curve of training.

collected data. If the average loss in the test history

is greater than 0.2 or the accuracy is less than 0.5,

then the operating status of the bearing is ”Abnormal”

if the ﬁnal average loss is greater than 0.2 or the ac-

curacy is less than 0.5 based on the test history, and

”Normal” otherwise. The prediction effect is illus-

trated in Figures 9 and 10.

Figure 9: Loss curve of prediction.

Then, we evaluate the function of the constructed

Figure 10: Accuracy curve of prediction.

model to create new nodes based on the prediction re-

sults, using ”normal” as an example of a prediction

result. We constructed a new node with the name

”new normal” under the class ”normal type” in the

diagram. Establish the ”type of” relationship between

this node and the ”Normal” entity within the ”operat-

ing status” class. Establish the ”fault is” relationship

between this node and the ”fault 0” entity within the

”fault details” class. The newly constructed node is

queried in the manner depicted in Figure 11. Simi-

lar to the status of ”Normal”, when the running state

is ”Abnormal”, create a new node with the name

”new abnormal” under the class ”abnormal type” in

the diagram. Establish the ”type of” relationship be-

tween this node and the ”Abnormal” entity within the

class ”operating status”.

Figure 11: New node of predicted results.

6 CONCLUSION

In this paper, we focused on the analysis of faults in

rolling bearings, event knowledge graphs, and the cur-

rent state of convolutional neural networks, as well as

fault diagnosis and prediction techniques for rolling

bearings. Based on the characteristics of the mechan-

KEOD 2023 - 15th International Conference on Knowledge Engineering and Ontology Development

172

ical fault ﬁeld and the data collected from the CWRU

rolling bearing data center, we extract the relation-

ship between entities and entities using a seven-step

process, store the obtained entities and inter-entity re-

lationships in the Neo4j graph database to complete

the construction of the bearing fault event knowledge

graph, and implement the four types of visual query

functions based on the constructed event kdb. Then,

we construct a deep convolutional neural network

with a wide convolutional kernel on the ﬁrst layer

in order to diagnose the operational status of rolling

bearings. The experimentally established WDCNN

model has an accuracy rate of greater than 99.8% on

the training dataset and greater than 97.3% on the test

dataset. Finally, we predict bearing faults by com-

bining an event knowledge graph with a deep con-

volutional neural network. By acquiring newly col-

lected data during rolling bearing operation, setting

fault thresholds, and applying WDCNN to predict and

analyze them, new nodes and their relationships are

added to the event knowledge graph.

In rolling bearing fault diagnosis, the introduction

of the event knowledge graph and convolutional neu-

ral network provides support for fault diagnosis and

prediction, enhances the accuracy of fault diagnosis

and prediction, and reduces the need for expertise and

expert system domain knowledge.

ACKNOWLEDGEMENTS

This work is supported by the National Natural Sci-

ence Foundation of China (62002207, 62072290,

62073201), the Shandong Provincial Natural Sci-

ence Foundation (ZR2022YQ05, ZR2020MA102)

and Shandong Provincial Key Laboratory for Novel

Distributed Computer Software Technology.

REFERENCES

An, M. (2008). A survey on fault diagnosis expert systems.

Computer Measurement & Control, 16(9):3.

CWRU (2023). Case western reserve university bearing

data center. accessed on March 11, 2023.

Fern

andez-L

opez, M., Gomez-Perez, A., and Juristo, N.

(1997). Methontology: Fromontological art towards

ontological engineering. Proceedings of International

Conference on Software, Services & Semantic tech-

nologies, pages 33 – 40.

Gomez-Perez, A. and Su

arez-Figueroa, M. C. (2009).

Neon methodology for building ontology networks: a

scenario-based methodology. Proceedings of Interna-

tional Conference on Software, Services & Semantic

technologies.

Han, Y.-J., Park, S.-Y., Park, S.-B., Lee, Y.-H., and Kim,

K.-Y. (2007). Reconstruction of people informa-

tion based on an event ontology. In 2007 Interna-

tional Conference on Natural Language Processing

and Knowledge Engineering, pages 446–451.

Krizhevsky, A., Sutskever, I., and Hinton, G. (2012). Im-

agenet classiﬁcation with deep convolutional neural

networks. Advances in neural information processing

systems, pages 1097–1105.

Liu, X. (2019). Research on Knowledge Graph Construc-

tion Technology for Fault Analysis. PhD thesis, Bei-

jing University of Posts and Telecommunications.

Noy, N. and Mcguiness, D. (1995). A guide to creating your

ﬁrst ontology. Biomedical Informatics Reseach, page

7–25.

Simonyan, K. and Zisserman, A. (2014). Very deep con-

volutional networks for large-scale image recognition.

Computer Science.

Szegedy, C., Liu, W., Jia, Y., Sermanet, P., Reed, S.,

Anguelov, D., Erhan, D., Vanhoucke, V., and Rabi-

novich, A. (2015). Going deeper with convolutions. In

2015 IEEE Conference on Computer Vision and Pat-

tern Recognition (CVPR), pages 1–9, Los Alamitos,

CA, USA. IEEE Computer Society.

Tian, L., Zhang, J., Zhang, J., Zhou, W., and Zhou, X.

(2021). Knowledge graph survey: representation, con-

struction, reasoning and knowledge hypergraph the-

ory. Journal of Computer Applications, 41(8):26.

Yan, W., Shi, Y., Wang, C., Zhang, L., and Wang, J.

(2022). Knowledge graph-based predictive mainte-

nance method and system for mechanical failures.

CN113887751A.

Yu Shi, Y. H. (2022). Mechanical fault diagnosis model

based on knowledge graph and fuzzy inference. Mod-

ern Information Technology, 6:72–76+81.

Zeiler, M. D. (2012). ADADELTA: an adaptive learning

rate method. CoRR, abs/1212.5701.

Zhang, L., Wang, Z., and Zhao, S. (2007). Short-term fault

prediction of mechanical rotating parts on the basis of

fuzzy-grey optimising method. Mechanical Systems

and Signal Processing, 21(2):856–865.

Zhang, W. (2017). Study on Bearing fault Diagnosis Algo-

rithm Based on Convolutional Neural Network. PhD

thesis, Harbin Institute of Technology.

Zhao, X., Wang, W., Li, L., and Xu, Y. (2022). Framework

design and improvement about knowledge graph con-

struction in dynamic military ﬁeld. In Proceedings

of the 10th China Command and Control Conference

(Previous), pages 212–217.

Mechanical Fault Prediction Based on Event Knowledge Graph

173