Semantic Web 1 (0) 1–5 1IOS Press

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Combining Chronicle Mining and Semanticsfor Predictive Maintenance in ManufacturingProcessesQiushi Cao a,*, Ahmed Samet b, Cecilia Zanni-Merk a, François de Bertrand de Beuvron b andChristoph Reich c

a Normandie Université, INSA Rouen, LITIS, 76000 Saint-Étienne-du-Rouvray, FranceE-mails: qiushi.cao@insa-rouen.fr, cecilia.zanni-merk@insa-rouen.frb ICUBE/SDC Team (UMR CNRS 7357)-Pole API BP 10413, 67412 Illkirch, FranceE-mails: ahmed.samet@insa-strasbourg.fr, francois.debertranddebeuvron@insa-strasbourg.frc Hochschule Furtwangen University, 78120 Furtwangen, GermanyE-mail: rch@hs-furtwangen.de

Editors: First Editor, University or Company name, Country; Second Editor, University or Company name, CountrySolicited reviews: First Solicited Reviewer, University or Company name, Country; Second Solicited Reviewer, University or Company name,CountryOpen reviews: First Open Reviewer, University or Company name, Country; Second Open Reviewer, University or Company name, Country

Abstract. Within manufacturing processes, faults and failures may cause severe economic loss. With the vision of Industry 4.0,artificial intelligence techniques such as data mining play a crucial role in automatic fault and failure prediction. However, datamining results normally lack both machine and human-understandable representation and interpretation of knowledge. This maycause the semantic gap issue, which stands for the incoherence between the knowledge extracted from industrial data and theinterpretation of the knowledge from a user.

To tackle this issue, in this paper we introduce a novel hybrid approach to facilitate predictive maintenance tasks in manu-facturing processes. The proposed approach is a combination of data mining and semantics, within which chronicle mining isused to predict the future failures of the monitored industrial machinery, and a Manufacturing Predictive Maintenance Ontology(MPMO) with its rule-based extension is used to predict temporal constraints of failures and to represent the predictive resultsformally. As a result, Semantic Web Rule Language (SWRL) rules are constructed for predicting occurrence time of machineryfailures in the future. The proposed rules provide explicit knowledge representation and semantic enrichment of failure predic-tion results, thus easing the understanding of the inferred knowledge. A case study on a semi-conductor manufacturing processis used to demonstrate our approach in detail.

Keywords: Semantics, Chronicle Mining, Predictive Maintenance, Manufacturing Process, Industry 4.0

1. Introduction

Manufacturing processes are sets of structured oper-ations to transform raw material or semi-finished prod-uct parts into further completed products. To ensure

*Corresponding author. E-mail: qiushi.cao@insa-rouen.fr.

high productivity, availability and efficiency of manu-facturing processes, the detection of harmful tenden-cies and conditions of production lines is a crucial is-sue for manufacturers. In general, anomaly detectionon production lines is performed by analyzing data col-lected by sensors, which are located on machine com-ponents and also in production environments. The col-

1570-0844/0-1900/$35.00 c© 0 – IOS Press and the authors. All rights reserved

2 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

lected data record real-time situations and reflect thecorrectness of mechanical system conditions. Whenthe tendency of a mechanical failure emerges, experi-enced operators in factories are able to take appropri-ate operations to prevent the outage situations of pro-duction systems. However, as the collected data be-come more heterogeneous and complex, it is conceiv-able that the operators may fail to respond to mechani-cal failures timely and accurately. In the context of In-dustry 4.0, advanced techniques such as the IndustryInternet of Things (IIoT) and Cloud Computing enablemachines and production systems in smart factories tobe interconnected to exchange data continuously. Thistrend has brought opportunities to manufactures to ef-fectively manage and use the collected big data andhas triggered the demand of methodologies to detectanomalies on production lines automatically.

In the manufacturing domain, the detection ofanomalies such as mechanical faults and failures en-ables the launching of predictive maintenance tasks,which aim to predict future faults, errors, and failuresand also enable maintenance actions. Normally, a pre-dictive maintenance task relies on the monitoring of ameasurable system diagnostic parameter, which iden-tifies the state of a system [2]. In this way, maintenancedecisions, such as calling the intervention of a machineoperator, are proposed based on the severity of anoma-lies, to prevent the halt of the production lines and tominimize economic loss. Several techniques have beenused to detect wear and tear in mechanical units and topredict future machinery conditions, such as machinelearning, data mining, statistics, and information the-ory [46].

1.1. Existing Challenges of the PredictiveMaintenance Tasks in Industry 4.0

With the trend of Industry 4.0, the predictive main-tenance tasks are benefiting from a cyber-physical ap-proach. Within cyber-physical systems (CPS), produc-tion facilities are able to exchange information withautonomy and intelligence, which enable manufactur-ers to optimize the production processes [39]. Fig. 1shows the architecture of a CPS designed for predic-tive maintenance tasks. Within a CPS, predictive main-tenance of manufacturing entities is performed basedon a three-layer collaboration between the cyber spaceand the physical space: 1). The Physical Space, wheremachine operating data is gathered using sensors lo-cated on the machines and machine components. Ad-ditional data is collected from the products, manufac-

turing environments, as well as the machine operators’experience; 2). The Cyber-Physical Interface, wherestatistical techniques such as data mining and machinelearning use the collected data to understand the manu-facturing processes and to learn from operators’ expe-rience; 3). The Cyber Space, where decision-makingabout machine failure prediction and maintenance areproposed. In this layer, machine degradation models,and knowledge base of machine health are employedto predict machine damage, quality loss or mainte-nance demands in the future.

In the second layer of the architecture, data min-ing is normally performed by obtaining and process-ing sensor data that contain measurements of physi-cal signals of machinery, such as temperature, voltage,and vibration. By identifying events and patterns thatare not consistent with the expected behavior, potentialhazards in production systems, such as power outageof the systems, could be detected.

However, sometimes the knowledge extracted fromdata mining is presented in a complex structure, there-fore formal knowledge representation methods are re-quired to facilitate the understanding and exploitationof it [38]. Furthermore, there may exist the seman-tic gap issue, which stands for the incoherence be-tween the knowledge extracted from industrial dataand the interpretation of the knowledge from a user[8]. To overcome these issues, semantic technologieshave been utilized in several research efforts to pro-mote the interpretation and management of knowledge[3, 8, 38, 39]. Also, since semantic technologies ensurethe explicit representations of machine-interpretabledomain semantics, they can support the semantic in-teroperability in a large heterogeneous environment ofloosely coupled systems [24]. In the data mining do-main, several stages can benefit from the involvementof formal semantics, such as data transformation, al-gorithm selection, and post-processing [8]. Moreover,the use of semantic technologies can also integrate thecapitalization of domain experts’ experience. For ex-ample, in a predictive maintenance task of machinecutting tool, when data mining algorithms fail to iden-tify the occurrence time of a future cutter failure, logic-based expert rules which capitalize experience of do-main experts can be applied to propose predictive de-cisions.

In the context of predictive maintenance in smartfactories, pattern mining has been widely used todiscover frequently occurring temporally-constrainedpatterns, through which warning signals can be sentto humans for a timely intervention [4]. Among pat-

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 3

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 1. A predictive maintenance task based on a cyber-physical approach.

tern mining techniques, chronicle mining has beenapplied to industrial data sets for extracting tempo-ral information of events and to predict potential ma-chinery failures [6]. However, even though chroniclemining results are expressive and interpretable repre-sentations of complex temporal information, domainknowledge is required for users to have a compre-hensive understanding of the mined chronicles [19].As the predictive maintenance domain is becomingmore knowledge-intensive, tasks performed in this do-main can often benefit from incorporating domain andcontextual knowledge, by which the semantics of thechronicle mining results can be explicitly representedand clearly interpreted. This helps to reduce the se-mantic gap issue. However, to the best of our knowl-edge, no work has been proposed to combine chron-icle mining, and semantics to facilitate the predictivemaintenance of manufacturing processes. Also, mostof the existing research works about predictive main-tenance in the manufacturing domain merely focus onthe classification of operating conditions of machines(e.g., normal operating condition, breakdown condi-tion...), while lacking the extraction of specific tempo-ral information of failure occurrence. This brings ob-stacles for users to perform maintenance actions withthe consideration of temporal constraints.

1.2. Contributions of This Paper

To address the aforementioned issues, in this pa-per, we propose an ontology-based approach to repre-sent chronicle mining results in a semantic rich format,which enhances the representation and reuse of knowl-edge. By specifying domain semantics and annotatingindustrial data with rich and formal semantics, ontolo-gies with their rule-based extensions help to addressthe issues described in Section 1.1. In more detail, thecontributions of this paper are as follows:

– We present a domain ontology named Manufac-turing Predictive Maintenance Ontology (MPMO),which is a Web Ontology Language (OWL) [9]based ontology designed to model the knowl-edge related to condition-based maintenance. TheMPMO ontology provides the foundation to for-mally represent chronicles with their numericaltime constraints, for the purpose of predictivemaintenance.

– We propose an algorithm to transform chroni-cles into Semantic Web Rule Language (SWRL)based logic rules, by which the predictive resultsare formalized, thus interpretable for both humanand machines. The proposed transformation en-ables the automatic generation of SWRL rulesfrom chronicle mining results, thus allowing an

4 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

automatic semantic approach for machinery fail-ure prediction.

– We evaluate the feasibility and effectiveness ofour approach by conducting experimentation ona real industrial data set. The performance ofSWRL rule construction and the quality of fail-ure prediction is evaluated against the aforemen-tioned data set.

The rest of this paper is structured as follows. Sec-tion 2 provides a review of existing ontology-basedmodels and systems developed for predictive mainte-nance. Section 3 introduces the foundations and ba-sic notions of chronicle mining and semantics that arenecessary for describing our approach. It contains for-mal definitions of chronicles and the Semantic WebRule Language (SWRL). Sections 4 presents a hy-brid semantic approach for automatic failure predic-tion. The approach includes the use of the MPMO on-tology, which models necessary and principle knowl-edge related to chronicles. We introduce a real-worldexample scenario and use it to describe our approach indetail. Also, one algorithm for transforming chroniclesto SWRL-based predictive rules is introduced. Section5 evaluates our approach through a real industrial dataset. Section 6 gives concluding remarks and outlinesfuture research directions.

2. Related Work

There are plentiful of works that deal with the pre-dictive maintenance tasks in Industry 4.0. In this sec-tion, we analyze the related works according to twoperspectives: 1). The common approaches for machin-ery failure prediction within manufacturing processes;2). Existing ontologies and ontological models that areused by knowledge-based predictive maintenance sys-tems.

2.1. Common Approaches to Predictive Maintenancein Industry 4.0

The main objective of machinery failure predictionis to estimate the time of the occurrence of future fail-ures. The prediction is normally based on the exami-nation of the current condition of the machinery andthe past operation profile. The estimation of remain-ing useful life (RUL) is the main approach that is usedin predictive maintenance. The commonly used meth-ods for RUL estimation can be classified into four cat-egories [21, 22]:

– Knowledge-based Models. This type of modelsassess the similarity between an observed situ-ation and a databank of previously defined fail-ures and deduce the life expectancy from previousevents [21]. Normally, a knowledge-based modelcontains a rule base that consists of a set of rules.The rules are formulated as IF-THEN statements,and they are often proposed based on heuristicfacts acquired by domain experts. Knowledge-based models can be further classified into ExpertSystems and Fuzzy Systems.

– Life Expectancy Models. They determine theRUL of individual machine components with re-spect to the expected risk of deterioration un-der known operating conditions [21]. This typeof models can be further grouped into StochasticModels and Statistical Models.

– Artificial Neural Networks. They compute an es-timated output for the RUL of a piece of machin-ery, directly or indirectly, from a mathematicalrepresentation of the machinery derived from ob-servation data rather than a physical understand-ing of the failure processes [21]. This type ofmodels can be used for direct RUL forecasting orparametric estimation for other models.

– Physical Models. They compute an estimated out-put for the RUL of a piece of machinery from amathematical representation of the physical be-haviour of the degradation processes [21]. Byusing physical models, users can obtain a thor-ough understanding of the system behaviour in re-sponse to different levels of stress and burden, atboth macroscopic and microscopic levels.

In this work, we focus on the use of Knowledge-based Models in predictive maintenance. In the nextsubsection, we present the existing solutions, espe-cially the ontology-based approaches that are appliedto failure prediction tasks.

2.2. Existing Knowledge-based Models to PredictiveMaintenance

In recent years, several knowledge-based modelshave been proposed to facilitate the failure predictiontasks in the predictive maintenance domain. Amongthem, the ontology-based approach is an effective andnotable method that has drawn considerable attentionfrom researchers. Ontologies are explicit specificationsof conceptualizations, and they are comprehensive andreusable knowledge repositories in various domains

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 5

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

[44]. In general, this type of approach uses ontolo-gies to formally define the semantics of knowledge anddata, and utilizes sets of logic rules to enable onto-logical reasoning, for inferring new knowledge. Theavailable research works related to this approach canbe categorized into two major fields, according to dif-ferent purposes: i) using ontology-based approach torepresent data mining results in a formal and struc-tured way, to further enrich knowledge bases; ii) usingontology-based approach to facilitate knowledge for-malization,sharing and reuse in the predictive mainte-nance domain.

To formalize the data mining results and to facili-tate the interpretation of them, many researchers triedto incorporate explicit domain knowledge with usingontologies. The DAMON ontology [26] is developedas a data mining ontology to simplify the developmentof distributed knowledge discovery systems. The on-tology is used as a knowledge reference model to helpdomain experts solve tasks. Also, the ontology enablesusers to search for data mining resources and soft-ware when they want to find solutions for a specificproblem. The EXPO ontology [23] formalizes con-cepts about experimental design, methodologies andresults representation in a general way. The ontologypromotes the sharing of experimental results withinand among different subjects, and it can reduce the in-formation duplication and loss in the sharing process.The OntoDM-core ontology [37] is developed to for-mally describe core data mining entities. The ontol-ogy provides a framework to represent essential andbasic data mining concepts, such as data sets, datamining tasks, algorithms, and constraints. The advan-tage of this ontology is its powerful representation ofconstraint-based data mining activities.

The use of an ontology-based approach can alsofacilitate knowledge formalization, sharing and reusein the predictive maintenance domain. In the con-text of predictive maintenance, several ontologies andontology-based intelligent systems are developed toachieve this goal. To enhance the expressiveness ofthese ontologies, several rule-based extensions wereproposed to perform ontological reasoning, in order tofacilitate maintenance decisions of users. We reviewexisting ontologies according to two aspects: ontolo-gies that model manufacturing processes and ontolo-gies that model preventive maintenance tasks.

As indicated in the introduction, manufacturing pro-cesses are structured sets of operations that transformraw materials or semi-finished product segments intofurther completed product parts. Over the last decades,

several ontologies have been developed to representknowledge about manufacturing processes. The Pro-cess Specification Language (PSL) ontology [27] isone of the early-stage contributions. This ontology ax-iomatizes a set of semantic primitives that are essen-tial for describing a wide range of manufacturing pro-cesses. The axioms defined in this ontology model thekey elements of manufacturing processes, such as pro-cess scheduling, process modeling, production plan-ning, and project management [27]. Another contribu-tion in this subdomain is the Manufacturing ServiceDescription Language (MSDL) ontology, which de-fines a well-defined framework for formal represen-tation of manufacturing services [14]. This ontologyformalizes manufacturing capabilities of manufactur-ing resources in different levels of abstraction, basedon which a rule-based extension of the ontology isproposed to enable automatic supplier discovery. Atlast we mention the Manufacturing Reference Ontol-ogy (MRO) [50] that is developed to formalize a setof core concepts about the manufacturing in a highabstraction level. The ontology categorizes the manu-facturing domain into eight general concepts: RealizedPart, Part Version, Manufacturing Facility, Manufac-turing Resource, Manufacturing Method, Manufactur-ing Process, Feature and Part Family. This categoriza-tion enables further development of more specific on-tologies in the production domain.

Compared to ontologies that model manufacturingprocesses, ontologies for predictive maintenance aremuch less numerous. These type of ontologies nor-mally focus on the issues of fault or failure prognosticsand machine health monitoring. Among these ontolo-gies, the OntoProg Ontology [10] addresses the failureprediction of machines in smart factories. The ontol-ogy is developed based on a set of international stan-dards, and a classification for severity criteria, detec-tion, diagnostics and prognostics of failure modes isprovided. The ontology standardizes the concepts thatare necessary for tackling machinery failure analysistasks. As another most recent contribution, the SensingSystem Ontology [12] is proposed to define the em-bedded sensing systems for industrial Product-ServiceSystems (PSSs). This ontology is used as the back-bone of the PSS knowledge-based framework and itdescribes the sensors that are embedded on PSSs, forthe aim of providing customized services for users.

We summarize the domain coverage of existingontologies in Table 1. A comparison among exist-ing ontologies with respect to the MPMO ontologyis also presented. We evaluate the domain coverage

6 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Table 1A comparison between the proposed MPMO ontology and the existing ontologies with respect to their domain coverage.

Ontologies Context Condition Monitoring Manufacturing

Identity Activity Time Location Anomaly Fault Failure Severity Prognostics Diagnostics Alarm Alert Product Process Resource

MASON [42] 3 3 7 7 7 7 7 7 7 7 7 7 3 3 3

MSDL [14] 3 3 3 3 7 7 7 7 7 7 7 7 3 3 3

MRO [50] 3 3 3 3 7 7 7 7 7 7 7 7 3 3 3

ONTO-PDM [15] 3 3 3 3 7 7 7 7 7 7 7 7 3 3 3

MCCO [49] 3 3 3 7 3 7 7 7 7 7 3 7 3 3 3

MaRCO [11] 3 3 7 3 7 7 7 7 7 7 7 7 3 3 3

Sensing System Ontology [12] 3 3 3 7 7 7 3 7 3 7 3 3 7 7 7

OntoProg Ontology [10] 3 3 7 3 3 3 3 3 3 3 3 3 7 7 3

MPMO 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3

and scopes of these ontologies by examining whetherthe key concepts required for describing the predic-tive maintenance domain are covered and formally de-scribed in existing ontologies. These key concepts canbe categorized into three subdomains: Manufacturing,Context, and Condition Monitoring. For the Manufac-turing subdomain, the key concepts are Product, Pro-cess and Resource. For the Context subdomain, thekey concepts are Identity, Activity, Time, and Loca-tion. While for the Condition Monitoring subdomain,Anomaly, Fault, Failure, Severity, Prognostics, Diag-nostics, Alarm, and Alert are the key concepts. Theseconcepts form the columns of the Table 1, and the on-tologies are enumerated by rows. If the concept is cov-ered by the ontology, a check mark is placed in the ta-ble. Otherwise, a cross mark is assigned.

After reviewing the ontologies mentioned above, werecognize that none of them provides a satisfactoryknowledge representation of the three subdomains.Some of these ontologies focus on a narrow field,such as manufacturing resource planning, and theydo not formalize predictive maintenance-related con-cepts, e.g., machinery Failure and Fault. Also, noneof the existing ontologies standardize the concepts re-lated to chronicle mining. To jointly use chronicle min-ing with semantic technologies for a predictive mainte-nance task, the knowledge-based model should incor-porate not only the machine-interpretable knowledgeof manufacturing entities such as product and processbut also the knowledge about chronicles within whichthe machinery failures are described in a structuredway. To this end, an ontology that formally describesall the concepts in Table 1 is needed. This motivatesus to propose the MPMO ontology. The MPMO ontol-ogy aims to formalize all the predictive maintenance-related concepts as well as relationships.

3. Foundations and Basic Notions

In this section, we introduce the foundations andbasic notions of chronicle mining and semantics thatare necessary for describing our approach. The foun-dations include a formal description of Sequential Pat-tern Mining (SPM) and chronicles, as well as an intro-duction to Semantic Web Rule Language (SWRL).

3.1. Foundations of Sequential Pattern Mining

In industry, data collected for preventive mainte-nance tasks are normally represented as sets of se-quences with time stamps [5]. To cope with this typeof data sets, SPM is one important technique to ex-tract frequently occurring patterns. SPM was first stud-ied by [40], to analyze customer purchase behavior se-quences. One SPM task could be described as follows:Given a data set containing a number of sequences, thegoal of SPM is to find sequential patterns whose sup-port exceed a predefined numeric support threshold.

This support threshold indicates the minimal num-ber of occurrences of the sequential patterns, and thefound patterns are called frequent sequential patterns.For the output of SPM algorithms, each frequent se-quential pattern is a sequence which consists of a setof items in a certain order.

To give a formal description of sequential patterns,in this subsection we review the definitions of key con-cepts. A sequence S is a set of ordered itemsets, de-noted by S =< S ID, < I1 I2 I3 ... In >>, with SIDstanding for the index of the sequence with I j repre-senting a non-empty set of items. Given two sequencesS a =< S ID, < a1 a2 a3 ...am >> and S b =< S ID, <b1 b2 b3 ...bn >>, the sequence S a is considered to bethe subsequence of S b, denoted as S a ⊆ S b, if thereexists integers 1 6 k1 < k2 < ... < km 6 n such thata1 ⊆ bk1, a2 ⊆ bk2, ..., am ⊆ bkm [45]. One exam-ple of sequence data set is shown in Table 2. In the ta-

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 7

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Table 2An example sequence data set.

SID Sequences

10 < c(abe)(ac f ) >

20 < (bcd)(ac)(bd)(ad f ) f >

30 < (cd)(ab)(bc f )e >

40 < b(d f )(bd f )c(ab) >

50 < (ab)(be f )de >

60 < (abe)(cd)(ce) >

ble, each row is a sequence of elements. The elementsare presented with a certain order, showing the prece-dence relationships among them. For example, regard-ing the definitions we recalled before, the sequence< ce(ac) > is the subsequence of < c(abe)(ac f ) >. Ifwe set the minimum support to 3, we can validate that< (ab)c > is a sequential pattern with the support of 3.

Over the last decades, considerable contributionshave been settled in the research field of SPM [34]. Asa result, various SPM algorithms have been proposedto mine frequent sequential patterns. Based on theseproposed SPM algorithms, a variety of approaches andexperiments have been launched to improve the perfor-mance and efficiency of SPM tasks.

3.2. Sequential Pattern Mining with Time Intervals

Even though sequential patterns contain informationabout the orders of items, the algorithms introducedin the previous section can not specify the time in-tervals between elements and items. In real-world sit-uations, the occurrences of events are often recordedwith temporal information, such as time points andtime intervals between events. Thus, several contri-butions have been proposed to obtain the time inter-vals between successive items in sequences. The no-tion of the time-interval sequential pattern is first pre-sented by Yoshida et al. [33]. The authors name thiskind of patterns as “delta patterns". A delta pattern isan ordered list of itemsets with the time intervals be-tween two neighboring itemsets. It can be representedas A [0,3]−−→ B [2,5]−−→ C, where A −→ B −→ C is a fre-quent sequential pattern. The time intervals [0, 3] and[2, 5] are bounding intervals, which means the transi-tion time of A → B is contained in the time interval[0, 3], and the transition time of B→ C is placed in thetime interval [2, 5].

With the introduction of delta patterns, a groupof algorithms were proposed to facilitate the min-ing process in temporal sequence data sets. One sig-nificant contribution is the work by Hirate et al.

[48]. In this work, the authors propose the Hirate-Yamana algorithm to mine all frequent time-extendedsequences. To do this, the authors generalize SPMwith item intervals. In the generalization, they definea set of time-extended sequences , denoted as S t =<S ID, (t1,1, i1), (t1,2, i2), (t1,3, i3), ..., (t1,n, in) >>, wherei j means an item, and tα,β is the item interval betweenitems iα and iβ, tα,β can be interpreted according to twoaspects of conditions [48]:

– If the data sets contain time stamps, which indi-cate the transaction occurrences of items, then tα,βbecomes the time interval and can be computedby the equation tα,β = iβ.time − iα.time, whereiβ.time and iα.time are time stamps of items iα andiβ respectively. For example, one time-extendedsequence could be< (0, c), (1, abe), (3, ac), (5, f )>, which means item c occurs at time point 0,followed by itemset abe occurring at 1 time unitlater. Itemset ac occurs 2 time unites after abe,and the last itemset f occurs 2 time unites afterac.

– If the data sets do not contain time stamps, thentα,β may become the item gap and defined by theequation tα,β = β−α. In this case, the item gap isdefined as the number of items that occur betweentwo items. This type of representation is suitableto be applied to data sets which contain fixed itemintervals, but it is not applicable to data sets whichcontain various length of time intervals.

The study on existing notions and algorithms helpto capture the core concepts in the domain of time-interval SPM. These core concepts form the founda-tions of chronicle mining.

3.3. Foundations of Chronicle Mining

As introduced in the previous section, the temporalpatterns we consider in this paper are chronicles. Togive formal definition of chronicles, we start by intro-ducing the concept of Event, given by [6].

8 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Definition 1 (Event). Let E be a set of event types, andT a time domain such that T ⊆ R. E is assumed totallyordered and is denoted 6E. According to [6], an eventis a couple (e, t) where e ∈ E is the type of the eventand t ∈ T is its time. In SPM, events represent itemsetsof a single sequence.

A sequence contains a set of ordered events, whichare timestamped. The events contained in a sequenceappear according to their time of occurrences.

Definition 2 (Sequence). Let E be a set of eventtypes, and T a time domain such that T ⊆ R. Eis assumed totally ordered and is denoted 6E. Ac-cording to the definition in [6], a sequence is acouple 〈S ID, 〈(e1, t1), (e2, t2), ..., (en, tn)〉〉 such that〈(e1, t1), (e2, t2), ..., (en, tn)〉 is a sequence of events.For all i, j ∈ [1, n], i < j ⇒ ti 6 t j. If ti = t j thenei <E e j.

When the events are time-stamped, how to describethe quantitative time intervals among different eventsis vital important for the prediction of possible futureevents. To achieve this goal, we introduce the notiontemporal constraints in the following definition. Thedefinition of temporal constraints is adopted from theone introduced in [6].

Definition 3 (Temporal constraint). A temporal con-straint is a quadruplet (e1, e2, t−, t+), denoted e1[t−, t+]e2,where e1, e2 ∈ E, e1 6E e2 and t−, t+ ∈ T.

t− and t+ are two integers which are called lowerbound and upper bound of the time interval, such thatt− 6 t+. A couple of events (e1, t1) and (e2, t2) aresaid to satisfy the temporal constraint e1[t−, t+]e2 ifft2 − t1 ∈ [t−, t+].

We say that e1[a, b]e2 ⊆ e′1[a′, b′]e′2 iff [a, b] ⊆

[a′, b′], e1 = e′1, and e2 = e′2With obtaining introducing the events and temporal

constraints among different events within a sequence,we are able to to define the concept of chronicles [6].

Definition 4 (Chronicle). A chronicle is a pair C =(E , T ) such that:

1. E = {e1...en}, where ∀i, ei ∈ E and ei 6E ei+1,2. T = {ti j}16i< j6|E| is a set of temporal con-

straints on E such that for all pairs (i, j) satisfy-ing i < j, ti j is denoted by ei[t−i j , t

+i j ]e j.

E is called the episode of C, according to the defini-tion of episode’s discovery in sequences [6].

In the chronicle discovery process, support is usedas a measure to compute the frequency of a pattern in-side a sequence. It can therefore be formalized by thedefinition below.

Definition 5 (Chronicle support). An occurrence of achronicle C in a sequence S is a set (e1, t1)...(en, tn)of events of the sequence S that satisfies all temporalconstraints defined in C. The support of a chronicle Cin the sequence S is the number of its occurrences inS, or the percentage of its occurrences in the sequenceS [5].

The relevance of a chronicle is essentially based onthe value of its support.

To illustrate these basic definitions, we give an ex-ample including a sequence and a chronicle extractedfrom it. Assuming a sequence S contains three events< A, B,C >, represented as follows:

Fig. 2. A sequence representing three events.

In Fig. 2, time constraints that describe the pattern{A, B, C} are noted by A[2,5]B, B[1,5]C and A[6,7]C.Here [2,5], [1,4] and [6,7] lower and upper bounds ofthe time intervals among events.

After the generation of temporal constraints, theseevents can be represented as a graphical way, as shownin Fig. 3. In the figure, events are represented by thecircles, and temporal constraints are displayed througharrows among events. The values above each arrow arequantitative numerical bounds of temproal constraints.

Fig. 3. Example of a chronicle.

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 9

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 4. The procedure of the semantic approach for predictive maintenance.

In the domain of predictive maintenance, frequentchronicle mining has been used to detect machineanomalies in advance. To combine frequent chroniclemining and semantics for facilitating predictive main-tenance tasks, a special type of chronicles, called fail-ure chronicles is introduced [5].

Definition 6 (Failure chronicle). For a chronicle CF =(E , T ), we say that CF is a failure chronicle if and onlyif the events that describe it are set according to theirorder of occurrence in the sequence, and that the endof the chronicle is the event that represents the failure,i.e. for E = {e1 · · · en|ei 6E ei+1, i ∈ [1, n]}, en is thefailure event.

In [5], a new algorithm called CPM has been intro-duced to mine frequent failure chronicles. Based ontheir work, in this paper, we propose a novel algorithmto automatically generate SWRL rules from frequentfailure chronicles. The generated SWRL rules aim toprovide decision making for predictive maintenance inindustry. The algorithm is introduced in Section 4.

3.4. Semantic Web Rule Language

Semantic Web Rule Language (SWRL) is based ona combination of its sublanguages OWL DL and OWLLite with the RuleMarkup Language. A SWRL rule isin the form of an implication between an antecedent(body) and consequent (head), which can be inter-preted in a way that whenever the conditions speci-fied in the antecedent hold, then the conditions speci-fied in the consequent must also hold [17]. In SWRL, arule has the syntax: Antecedent→ Consequent, whereboth the antecedent (body) and consequent (head) con-tains zero or more atoms. Atoms in SWRL rules can

be the form of C(x), P(x,y), where C(x) is an OWLclass, P is an OWL property, and x,y are either vari-ables, OWL individuals or OWL data values [17].

In this work, the reason we choose SWRL rules istwo-fold. Firstly, SWRL provides model-theoretic se-mantics and has the advantage of its close associationwith OWL ontologies, which enables the definition ofcomplex rules for reasoning about individuals in on-tologies. Secondly, the use of SWRL to write rules isindependent of rule implementation languages withinrule engines, which has the advantage of the flexibleselection of rule engines and inference platform.

To represent data mining results, especially chron-icles, in a formal and structured way, we use ontolo-gies as well as SWRL rules to propose predictive rules.The proposed rules describe events and temporal con-straints within chronicles, and predict a special typeof event (a machinery failure), with corresponding totemporal information.

4. A Novel Hybrid Semantic Approach ForPredictive Maintenance

To propose the novel hybrid semantic approachfor predictive maintenance, we jointly use data min-ing and semantic technologies, within which chroni-cle mining is used to predict the future failures of themonitored industrial machinery, and domain ontolo-gies with their rule-based extension is used to predicttemporal constraints of failures and to represent thepredictive results formally. The procedure of the se-mantic approach is shown in Fig. 4. Firstly, data pre-processing is implemented on raw industry data setsto obtain sequences in the form of pairs (event, time

10 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 5. The global architecture of the MPMO ontology.

stamp), where each sequence finishes with the failureevent. Secondly, frequent chronicle mining algorithmsmine the pre-processed data to discover frequent pat-terns that indicate machinery failures. Thirdly, basedon the mined frequent patterns, semantic technologiesare used to automate the generation of SWRL-basedpredictive rules. These rules enable ontological reason-ing over individuals in ontologies, thus facilitating de-cision making.

4.1. Domain Knowledge

Within an intelligent system, ontologies containthe domain knowledge to operate. In this work, theMPMO ontology is developed to describe the conceptsand relationships within chronicles1. The definitions ofkey concepts and relationships in the MPMO ontologyare formalized based on the basic notions introduced inSection 3. To ensure the reusability of the MPMO on-tology, we adopt the ontology modularization methodduring the development process [35]. As a result, theontology is constructed with three small reusable mod-ules: the Condition Monitoring Module, the Manufac-turing Module, and the Context Module. In this way,other ontology engineers and ontologists can reuse aportion of the MPMO ontology when they need. Thisensures the reusability of the MPMO ontology. Fig.5 shows the global architecture of the ontology. Inthe figure, round rectangles stand for classes, solidlines stand for is-a or subsumption relationships, anddashed lines represent object properties. The classes

1The ontology files can be found at the website:https://sites.google.com/view/combiningchronicleminingandsem/home

with gray background belong to the ManufacturingModule, classes with black background are associatedwith the Context Module, and classes with white back-ground belong to the Condition Monitoring Module.

To describe the main classes in the MPMO ontol-ogy, we use a UML notation where boxes stand for on-tology classes, and arrows represent object properties.Data properties are indicated by class attributes. TheUML diagram for describing the main classes is shownin Fig. 6. For the purpose of clarity, only a subset ofthe whole classes and relationships are presented.

We then give the axioms of the main classes inthe MPMO ontology. The axioms defining the mainclasses are presented below using the description logic(DL) syntax [11].

– ManufacturingResource: This class describes theresources that are used within manufacturing pro-cesses. It consists three subclasses: FinancialRe-source, HumanResource, and PhysicalResource.Among the three subclasses, PhysicalResourcestands for a set of physical entities that the predic-tive maintenance tasks are performed upon, suchas machine tools, workpieces, and final products.The definition of this class is extended from theclass MASON: Resource, in the MASON ontol-ogy [42]. The DL axioms for defining this classand the PhysicalResource class are

Manu f acturingResource ≡ HumanResource t

PhysicalResource t FinancialResource,

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 11

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 6. The main classes in the MPMO ontology.

and

Manu f acturingResource v ∀MakesUse−

O f−1.Manu f acturingProcess.

– ManufacturingProcess: It describes different typesof structured sets of operations that transform rawmaterials or semi-finished product segments intofurther completed product parts [39]. The DL ax-ioms for defining this class are

Manu f acturingProcess ≡ AssemblyProcess t

FinishingProcess t FormingProces t

MachiningProcess t MouldingProcess,

and

Manu f acturingProcess ≡ ∃MakesUseO f .Ma−

nu f acturingResource u ∃hasProcessInput.W−

orkpiece u ∃Produces.RealizedPart.

– Chronicle: Chronicles are a special type of tem-poral patterns, in which temporal orders of eventsare quantified with numerical bounds [6]. To in-troduce this concept in the MPMO ontology, weuse the following axiom.

Chronicle ≡ ∀hasEvent.Event u

(> 1 hasEvent.Event) u ∀hasT imeInterval.Ti−

meInterval u (> 1 hasT imeInterval.TimeInte−

rval) u ∃isLearnedFrom.Manu f acturingPro−

cess.

12 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

– Event: . In predictive maintenance tasks, an Eventis generally associated with a set of Observed-Properties which indicate the correctness of theoperation of a piece of machinery. In this context,the DL axioms for defining this class is

Event ≡ ∀hasObservedProperty.Observed−

Property u (> 1 hasObserved.Property).

– ObservedProperty: This is an attribute which rep-resents some significant measurable characteristicof a monitored ManufacturingProcess, Manufac-turingResource or RealizedPart. The value of anObservedProperty is measured by sensors whichare located at different components of the mon-itored entity. This class is also called Attribute.The DL axioms for defining this class are

ObservedProperty v ∃hasObserved−

Property−1.Event u ∃Observes−1.S ensor.

– Failure: This class represents the Failures that areindicated by Events. A Failure is the inability ofan entity to perform one required function, and itcan be the result of a propagation of a machineryerror [1]. The following axiom is used to definethis class:

Failure v ∀PropagatesInto−1.Error.

– TimeInterval: A temporal entity with an extent orduration. The definition of this class is adoptedfrom the Time Ontology [20]. The axiom for de-scribing this class is

TemporalInterval v ∃hasProceeding−

Event.Event u ∃hasS ubsequentEvent.Event u

∃hasT imeInterval−1.Chronicle.

By defining the common concepts and relationshipsin the predictive maintenance domain, the MPMO on-tology can support semantic interoperability amongdifferent systems and system components. Also, theMPMO ontology provides rich representations ofmachine-interpretable semantics for knowledge-basedpredictive maintenance systems. This ensures the pre-dictive maintenance systems can interoperate withshared semantics and a high level of semantic preci-sion.

4.2. Rules

In the proposed semantic approach, different SWRLrules are used for predicting machinery failures. Thelaunching of these rules allows reasoning over individ-uals contained in the MPMO ontology. In this subsec-tion, we first introduce SWRL rules which are used topredict the time interval between a certain event anda future failure, and then introduce the algorithm de-veloped for transforming chronicles into SWRL rules.The proposed rules and algorithm enable the semanticapproach for automatic failure prediction in the predic-tive maintenance domain.

4.2.1. Failure Time Prediction RulesChronicles provide not only the order of occurrence

of events, but also the intervals of time they occur in.Fig. 7 gives an example failure chronicle within whichthe last event is a failure.

Fig. 7. Example of a failure chronicle.

Inside the chronicle, A, B and C are different events.The three events are identified by their associatedobserved properties and quantitative values. The ob-served properties and quantitative values are obtainedby a feature selection method, that determines the mostrelevant attributes in predicting the future failures. Thelast event C indicates a failure, and the time intervalsamong events A, B with event C gives the temporalinformation of a future failure (event C).

However, even though the chronicle in Fig. 7 is rep-resented in a structured format, it lacks formal seman-tics and domain knowledge to be interpreted by hu-mans and predictive maintenance systems. For exam-ple, the descriptions of events A, B, and C are miss-ing, which may cause the semantic gap between chron-icle mining results and users. To overcome this issue,we use ontologies with their rule-based extensions torepresent chronicles in a semantic rich format, whichhelps the sharing and reuse of chronicle mining results.

As the mining of sequential data sets can generatefrequent failure chronicles, SWRL rules can be pro-

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 13

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 8. Example of a SWRL-based predictive rule, generated from the chronicle introduced in Fig. 7.

posed to reason about temporal information of machin-ery failures. Therefore, when a new sequence of times-tamped events arrive, SWRL rules can be launched topredict the time intervals among different events andfuture failures. As stated in Section 4.1, an event withina chronicle is determined by a set of observed proper-ties (with their associated values). Based on this defini-tion, we construct the antecedent of such a rule by de-scribing quantitative values of observed properties (at-tributes) and the temporal constraints inside a chroni-cle. The consequent of such a rule comprises the lowerand upper bounds of the time intervals among certainevents and the failure.

Based on the chronicle in Fig. 4.2.1, a SWRL rulecan be elicited. Fig. 8 demonstrates how the rule thatdescribes different events and temporal constraints canbe constructed from the chronicle in Fig. 7. Withinthe rule, Chronicle stands for the root class of allthe chronicle individuals in the ontology. hasEvent isthe object property that links individuals of the classChronicle and those under the class Event. hasA1V,hasA2V, hasA3V, and hasA4V are data properties thatassign quantitative values of attributes to the two in-dividuals A and B under the Event class. TimeIn-terval corresponds to the root class of all individu-als of time intervals. There are two object proper-ties that link TimeInterval with Event: hasSubEventand hasProEvent, among which hasSubEvent corre-sponds to the subsequent event of a time interval, andhasProEvent indicates the proceeding event of a timeinterval. In this case, event A is the proceeding eventof the time interval between A and B, and event Bis the subsequent event of this time interval. By de-scribing the numerical values of different attributes andthe time interval with its proceeding and subsequentevents, temporal constraints among events A, B withthe failure C are indicated. The temporal constraintscomprise the minimum time duration between an eventwith the failure, described by the data property has-MinF, and the maximum time duration between an

event with the failure, described by another data prop-erty hasMaxF.

Algorithm 1 Algorithm to transform a chronicle intoa predictive SWRL rule.

Require: CF : A chonicle model within which the lastevent is a failure event, E : the episode of CF whichcontains different types of events in a chronicle.

Ensure: R1: EL← LastNon f ailureEvent(CF , E)2: . Extract the last non-failure event before the

failure within a chronicle.3: R← ∅, A← ∅, C ← ∅, Atoma ← ∅, Atomc ← ∅.4: for each ei[t−i j , t

+i j ]e j ∈ T do

5: pe← ProceedingEvent(ei[t−i j , t+i j ]e j)

6: . Extract the proceeding event of this timeinterval

7: se← S ubsequentEvent(ei[t−i j , t+i j ]e j)

8: . Extract the subsequent event of this timeinterval

9: Atoma ← [t−i j , t+i j ] ∧ pe ∧ se

10: A← Atoma ∧ ([t−i j , t+i j ] ∧ pe ∧ se)

11: end for each12: for each el ∈ EL do13: f tc← FailureT imeConstraint(el,T I)14: . Extract the time constraint between the last

event before the failure and the failure event.15: Atomc ← el ∧ f tc16: C ← Atomc ∧ (el ∧ f tc)17: end for each18: R← (A→ C)19: return R

4.2.2. Automatic Rule Generation Based onChronicles

To enable the automatic generation of a SWRL rule,in this work we propose an algorithm to transformchronicles into predictive SWRL rules. Algorithm 1demonstrates the general idea of our rule transforma-tion method. It runs in four major steps:

14 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

1. The function LastNonfailureEvent extracts thelast non-failure event within a chronicle.

2. For each temporal constraint in a chronicle,the two functions ProceedingEvent and Subse-quentEvent extract the proceeding and subse-quent events of the time interval that is definedin this temporal constraint. Then the two eventsand this time interval forms different atoms inthe antecedent of the rule, and they are treated asconjunctions.

3. For each last non-failure event before the fail-ure (there could be multiple last events before thefailure), extract the temporal constraint betweenthis event and the failure. The extracted temporalconstraint is treated as a conjunction with the lastevent, to form the consequent of the rule.

4. At last, a rule is constructed as an implicationbetween the antecedent and the consequent.

A sequence can be described by one or multiplechronicles. To improve the quality of failure predic-tion, we only keep the most relevant chronicles for therule transformation. In this context, we take featuresof chronicles such as Chronicle Support as a referencemeasure, to select the most relevant chronicles.

5. Experiments

We validate our approach by conducting experimen-tation on the SECOM data set [30], which containsmeasurements of features of semi-conductor produc-tions within a semi-conductor manufacturing process.To evaluate the effectiveness of our approach, a soft-ware prototype is developed based on Java 10.0.2, Pro-tégé 5.5.0 [18], OWL API [28] and SWRL API [29].The reason we choose Protégé and OWL API is theirconvenience of creating, parsing, manipulating, andserializing OWL Ontologies. SWRL API allows usto create and interact with SWRL rules and SQWRLqueries. Also, the graphical tools embedded in SWRLAPI ease the visualization and interpretation of rule-based reasoning and querying results. Among thesetools, OWL API is used to build and manipulate theMPMO ontology. Different types of chronicles are cre-ated as individuals within the MPMO ontology, andSWRL-based predictive rules are proposed using thetransformation algorithm introduced in Section 4.2.2.To enable ontology reasoning, the SWRL API, whichincludes a SWRL Rule Engine API, is used to createthe transformed rules and then execute them. Within

this process, the Drools rule engine [25] is used for ruleexecution. At last, the inferred knowledge is returnedto the OWL API, and stored in the new ontology. Therunning environment of the software prototype is Mi-crosoft Windows 10.

5.1. The SECOM Data Set

In the SECOM data set, 1567 recordings and 590 at-tributes are collected, with each recording being char-acterized by a time stamp referring to the time thatthe data is recorded. Each recording is also associatedwith a label, which is either 1 or -1. The label of everyrecording explains the correctness of the event, with -1corresponding to a non-failure event, and 1 refers to afailure. Timestamps are associated with all the recordsindicating the moment of each specific test point. In to-tal, 104 pieces of records represent the failures of pro-duction. The data is stored in a raw text file, withinwhich each line represents an individual example ofrecording with its timestamp. The features are sepa-rated by spaces.

However, the data contained in SECOM data set donot have the same types of attributes and values, thatsome of the information contained in the data is irrele-vant to the failure prediction task thus is considered asnoise. Moreover, due to the inter-dependency amongindividual features and the complex behavior of com-bined features, it is difficult to extract frequent patternsand rules based on analysis of all the 590 attributes.Thus, in this context, instead of going through the en-tire data set and use all 590 attributes for failure predic-tion, we use feature selection methods [16] to identifyand select the most relevant attributes in predicting thefailures. The selected attributes are subsequently usedto extract the key factors and patterns that lead to ma-chine failures. This reduces the data processing timeand memory consumption.

5.2. The Extraction of Frequent Failure Chronicles

We aim to extract frequent failure chronicles andtest the performance of Algorithm 1 on the SECOMdata set. To obtain frequent failure chronicles, we usethe frequent chronicle mining approach introduced in[5]. In [5], an industrial data pre-processing method isintroduced, including data discretization and sequen-tialization. Fig. 9 shows different steps within the datamining, especially the frequent chronicle mining ap-proach. The steps presented in Fig. 9 elaborates thedata mining procedure which is described in Fig. 4.

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 15

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 9. Different steps used in the frequent failure chronicle mining approach, adapted from [5].

Table 3Extracted failure chronicles that have the highest 10 chronicle support.

Failure Chronicle Number of Events Number of Time Intervals Attributes Chronicle Support

CF1 3 3 A63, A64, A102, A204, A209, A476 83.65%CF2 3 3 A63, A64, A102, A204, A209, A347, A476 82.69%CF3 3 3 A58, A64, A102, A204, A209, A476 82.69%CF4 3 3 A58, A63, A102, A204, A209, A347 81.73%CF5 3 3 A58, A63, A64, A102, A204, A209, A347, A476 81.73%CF6 3 3 A58, A102, A204, A209, A347, A476 80.77%CF7 3 3 A58, A204, A209, A347, A476 80.77%CF8 4 4 A63, A64, A102, A204, A209, A347, A476 78.84%CF9 4 4 A58, A63, A102, A204, A209, A347 78.84%CF10 4 4 A58, A204, A209, A347, A476 78.84%

The approach starts with the aforementioned featureselection, after which a feature subset of the SECOMdata set is obtained while retaining a suitably high ac-curacy in representing the original data set. As a re-sult, 10 most relevant attributes are selected as the op-timal subset of all 590 attributes. After the feature se-lection, data discretization [43] is employed to dis-cretize continuous values for obtaining nominal ones.Thereafter, data sequentialization is used to transformthe data into the form of pairs (event, time stamp),where each sequence finishes with a failure. With ob-taining sequences that contain failures, CloSpan algo-rithm [47] is applied to the pre-processed data set, toextract frequent sequential patterns. Also, the frequentchronicle mining algorithm introduced in [5] is used toextract the temporal constraints among these sequen-tial patterns. Up to this step, we are able to obtain fre-quent failure chronicles that will be transformed intopredictive rules.

As introduced in Section 4, to improve the qualityof failure prediction, we take Chronicle Support as areference measure, to select the most relevant failurechronicles for failure prediction. As a result, only asubset of all frequent chronicles are used for predictiverule transformation. Table 3 shows the failure chroni-cles that have the 10 highest chronicle support. We usethese chronicles as examples to demonstrate the pre-dictive rule generation approach. In Table 3, each fail-ure chronicle is described by the number of events thatit contains, the number of time intervals among events,all the observed properties (attributes) that character-ize the failure chronicle, and the chronicle support. Forthe ease of demonstration, we label the 590 attributesas A1, A2, A3...A590.

For an event within a failure chronicle, it is not onlyidentified by a set of attributes, but also the quantita-tive values of them. To obtain the corresponding quan-titative attribute values for describing each event, data

16 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Table 4Attributes with their numerical intervals within the failure chronicle CF5.

Event Attribute Numerical Value Interval

A 63 [89.2564, 94.8757)A 204 [4925.1678, 4999.2456)A 209 [20.1884, 23.0750)A 347 [6.4877, 6.9573)A 476 [125.1988, 137.4435)B 58 [4.5537, 4.8994)B 63 [89.3158, 94.8757)B 64 [90.0196, 94.3934)B 102 [-0.1188, 0.5231)B 347 [6.2446, 6.9574)

Fig. 10. The SWRL-based predictive rule transformed from the failure Chronicle CF5, with describing the attributes and their numerical valueintervals.

discretization has been applied to the SECOM data set.After data discretization, the quantitative data has beentranslated into qualitative data. Also, an association be-tween each numerical value and a certain interval hasbeen created. Taking the chronicle that is presented inFig. 7 as an example, Table 4 shows the numericalintervals for describing the events within this failurechronicle. This chronicle is the failure chronicle CF5

introduced in Table 3.

5.3. The Generation of SWRL-based Predictive Rules

Based on the descriptions of the failure chronicleCF5, we use the algorithm introduced in Section 4.2.2

to generate a SWRL-based predictive rule. The resultof this rule generation is shown in Fig. 10. In this rule,hasA58V, hasA63V, hasA64V, hasA102V, hasA204V,hasA209V, hasA347V, hasA476V are data propertiesin the MPMO ontology that link individuals of theEvent class with XML Schema Datatype values. Theycorrespond to the quantitative values of the attributesA58, A63, A64, A102, A204, A209, A347, and A476 inthe SECOM data set. To describe the numerical in-tervals which are obtained by discretization, SWRLBuilt-Ins are used to specify the upper and lower nu-merical boundaries. The consequent of this rule com-prises the temporal constraints among Events A, B andC. The minimum time duration between an event with

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 17

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

the failure is described by the data property hasMinF,and the maximum time duration between an event withthe failure is described by another data property has-MaxF. By this way, the temporal constraints of a futurefailure is inferred by the launching of such a predictiveSWRL rule. This rule is an instantiation of the genericrule introduced in Fig. 8.

5.4. Results Evaluation

To evaluate the usefulness and effectiveness of ourapproach, we conduct results evaluation from two per-spectives: i) the evaluation of the MPMO ontology;and ii) the evaluation of the SWRL rule-based failureprediction results. It should be noted that for evalua-tion we focus on the quality of semantic enrichmentto the chronicle mining results, and the evaluation ofthe performance of the chronicle mining phase is outof the scope of this paper.

5.4.1. Evaluation of the MPMO OntologyOntology evaluation enables users to assess the

quality of ontologies. It is essential for the wide adop-tion of ontologies, since ontologies can be shared andreused by different users, and the quality of ontologiessuch as the consistency, completeness, and concise-ness of taxonomies are key considerations when differ-ent users reuse ontologies in specific contexts. In thispaper, to evaluate the quality of the proposed MPMOontology, we use OOPS!, which is an online ontologyevaluation tool [31]. The reason we choose this tool forontology evaluation is two-fold. Firstly, OOPS! allowsautomatic detection of common pitfalls in ontologies,and the detection of pitfalls can be executed indepen-dently of the ontology development software and plat-forms. Secondly, it enlarges the list of errors that canbe detected by most recent ontology evaluation tools,thus providing a broader scope of anomaly detectionin ontologies [31].

In OOPS!, ontology pitfalls are classified into threecategories: structural, functional, and usability-profiling.Under each category, fine-grained classification crite-ria is provided to cope with specific types of anoma-lies. The MPMO ontology is examined according tothe following three categories [31]:

– Structural dimension: It focuses on anomaly de-tection on syntax and formal semantics. Sincethe MPMO ontology consists of logical axioms,the syntax and logical consistency can be eval-uated and validated through anomaly detectionwithin this category. To be more specific, This

category is composed of five criteria: i) modelingdecisions, which evaluates whether users use theontology implementation language in a correctway; ii) real world modeling or common sense,which evaluates the completeness of the domainknowledge formalized by the MPMO ontology;iii) no inference, which checks whether the de-sired knowledge can be inferred through ontologyreasoning; iv) wrong inference, which refers tothe detection of inference that lead to erroneousor invalid knowledge; and v) ontology language,which assesses the correctness of the ontology de-velopment language of the MPMO ontology.

– Functional dimension: It considers the intendeduse and functionality of the MPMO ontology.Under this category, two specific criteria areused to evaluate the MPMO ontology: i) require-ment completeness, which evaluates coverage ofthe domain knowledge that is formalized by theMPMO ontology; ii) application context, whichevaluates the adequacy of the MPMO ontologyfor a given use case or application.

– Usability-profiling dimension: It evaluates thelevel of ease of communication when differentgroups of users use the same ontology. Withinthis category, two specific criteria are appliedfor ontology evaluation: i) ontology understand-ing, which evaluates the quality of informationor knowledge that is provided to users for eas-ing the understanding of the ontology; ii) ontol-ogy clarity, which assesses the quality of ontologyelements for being easily recognized and under-stood by users. These criteria is commonly usedto check the quality of ontologies when users donot have sufficient domain knowledge.

To evaluate the MPMO ontology according to theaforementioned categories, we uploaded the ontologycode to the OOPS! online tool. After loading the ontol-ogy code, the ontology pitfall scanner is used to checkthe pitfalls that exist in the MPMO ontology. Fig. 11shows the evaluation result. The result shows that ourontology is free of bad practices in the structural, func-tional, and usability-profiling dimensions of evalua-tion. Moreover, the MPMO ontology is developed andformalized using OWL, which is a widely used lan-guage for knowledge representation and ontology de-velopment. This eases the reuse of the MPMO ontol-ogy in other contexts and also simplifies the integrationof the MPMO ontology with other knowledge compo-nents that are developed with the same language.

18 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Fig. 11. Screenshot of the ontology evaluation result using OOPS! online tool.

5.4.2. Evaluation of the SWRL Rule-based FailurePrediction Results

To evaluate the quality of the SWRL rule-based fail-ure prediction results, we apply the SWRL rules on thesequences in the SECOM data set, and three measuresare used to assess the quality of these rules: the TruePositive Rate (TPR), the Precision of failure predic-tion, and the F-measure. The equations for computingthese three measures are shown in Equation 1, 2 and 3.

T PR =T P

T P + FN. (1)

Precision =T P

T P + FP. (2)

F − measure =2T P

2T P + FP + FN. (3)

Among them, the True Positive Rate aims to mea-sure the percentage of positive sequences that havebeen correctly classified. In Equation 1, T P (True Posi-tive) is the true positive results standing for the numberof valid sequences that at least one SWRL rule couldpredict the failures in these sequences, and FN (FalseNegative) is the false negative results which stand forthe number of sequences that no SWRL rule could pre-dict the failures in these sequences.

The Precision of failure prediction measures thepercentage of sequences based on which the SWRL

rules are constructed correctly. For a given sequence,failure chronicles are extracted through chronicle min-ing and SWRL rules are constructed for failure pre-diction. After applying the SWRL rules, if the pre-dicted failure temporal constraints are out of the rangeof the failure occurrence time intervals in the sequence,then it indicates that the SWRL rules could not pre-dict the temporal constraints of the failure in this se-quence. Thus, the failure is classified as False Positive.In Equation 2, T P (True Positive) is the true positiveresults standing for the number of valid sequences thatat least one SWRL rule could predict the failures inthese sequences, and FP (False Positive) is the num-ber of sequences for which the SWRL rules incorrectlypredict the temporal constraints of the future failures.

With obtaining the above two measures, we cancompute the F-measure according to the Equation 3.

Table 5 shows the experimental results of the threemeasures. The three measures are computed accord-ing to different frequency thresholds of sequences inthe data set. We use f tmin to denote the minimum fre-quency threshold of a sequence in the data set.

We can see from Table 5 that all computed valuesfor the three measures are above 80%, which shows theresults are encouraging. As the minimum frequencythreshold f tmin values decreases, the values of threemeasures show an increase tendency. This can be ex-plained as follows: as f tmin increases, the number ofextracted chronicles decreases, which lead to the de-crease of the number of transformed SWRL rules. Forthis reason, each sequence for testing is less likely tobe validated by the transformed SWRL rules.

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 19

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

Table 5True Positive Rate, Precision and F-measure of Failure Prediction Based on SWRL Rules.

f tmin True Positive Rate Precision F-measure

1 83.63% ±6.43% 84.62% ±6.55% 86.55% ±4.89%0.9 85.45% ±4.98% 87.49% ±6.16% 88.54% ±6.06%0.8 87.27% ±7.50% 84.58% ±6.55% 85.71% ±6.98%0.7 89.09% ±6.68% 86.22% ±6.43% 87.52% ±6.51%0.6 90.90% ±7.93% 88.71% ±5.26% 89.21% ±5.43%0.5 90.90% ±7.93% 86.83% ±4.41% 87.88% ±5.77%

Since the SWRL rules are generated from chroni-cle mining results, the quality of their prediction exclu-sively depend on the mined frequent chronicles. In thiscontext, the 10-fold cross validation principle [32] isused to evaluate the quality of failure prediction. To ap-ply the 10-fold cross validation principle, the SECOMdata set is partitioned into two parts: the training setand the test set. Firstly, chronicles are extracted fromthe training sequences in the training set. Then, for thetest set, we check for each sequence, its membership inat least one chronicle among those extracted. The num-ber of sequences validated by the chronicles is com-puted to estimate its percentage with respect to the se-quence set. This procedure is repeated 10 times to val-idate all the sequences of the database.

The launching of such a set of SWRL-based pre-dictive rules enables the prediction of temporal con-straints of future machinery failures. This allows usersto take further maintenance actions, such as the re-placement of the machine tools used on the productionline. The performance of failure prediction could beenhanced by considering a new set of rules that reasonabout the severity levels of failures. We are currentlyapplying machine learning techniques to classify theseverity levels of failures, according to the temporalconstraints among the failures and other events.

6. Conclusion and Future Perspectives

This paper demonstrates a novel hybrid approach forimplementing predictive maintenance in industry. Theproposed hybrid approach is a combination of frequentchronicle mining and semantics, within which chroni-cle mining is used to extract frequent chronicles basedon industrial data sets, and a knowledge-based struc-ture is used to automate the SWRL rule generation pro-cess and to formalize the predictive maintenance re-sults.

The contributions of this paper are three-fold. Firstly,chronicles are formally represented with the use of on-

tologies, by which the main concepts and relationshipsfor describing chronicles are formalized, then easingthe knowledge representation and interpretation of fre-quent chronicle mining results. Secondly, a novel al-gorithm for transforming chronicles into SWRL-basedpredictive rules is introduced. The novel algorithm al-lows the automatic generation of SWRL rules basedon the mined frequent chronicles, thus enabling an au-tomatic semantic approach for predictive maintenance.Thirdly, the reasoning about temporal constraints offuture machinery failures is enabled by the joint useof data mining and semantics, which allows the im-plementation of maintenance actions such as alarmlaunching.

However, there are three major problems that needto be solved. The first problem is the partition methodof numerical values. Since the rules we proposed inSection 5 are based on crisp logic, when the numericvalues of attributes collected by sensors are consider-ably close to partition thresholds, the rules proposedin Section 5 may fail to partition these numeric val-ues into correct categories. To deal with such kind ofuncertainty situations, the use of fuzzy logic shouldbe considered and a fuzzy semantic approach needsto be implemented. This approach will use machinelearning techniques to automatically derive member-ship functions and fuzzy if-then rules from data sets.The fuzzy rules aim to enhance the representation ofimprecise severity level of machinery failures. For ex-ample, an identification of failure will be associatedwith a fuzzy index, indicating the grade of its member-ship to a “low” or “high” level of failure. The fuzzy ap-proach will be applied to tackle the challenge of sym-bol anchoring problem [41].

The second problem is the evolution of the ontol-ogy and the rule base. Since the manufacturing domainis highly-dynamic, the predictive maintenance systemshould be able to adapt itself to dynamic situationsover time, for example, the change of context. Also,when the system fails to provide satisfactory resultsthrough launching the rules, it is required to consult

20 Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

domain experts for decisions about failure predictionand maintenance. In this situation, the domain expertsuse their expertise and experience to assess the currentstate of the system and provide appropriate decisions.For example, when the temperature measured by a sen-sor located at a cutting tool exceeds its threshold andno rule in the rule base is able to warn about his ab-normal condition, domain experts can use their expe-rience and expertise to identify this abnormal condi-tion and provide possible solutions in order to avoidthe production line to produce unqualified products. Inthis way, new rules which capitalize experts’ experi-ence needs to be proposed to update the initial set ofrules in the rule base, in order to facilitate the qual-ity of failure prediction. In this context, when the nexttime a similar situation needs to be addressed, the rulewhich capitalizes domain experts’ experience will belaunched together with the initial rules to identify po-tential failures and to make predictions. This requiresthe ontology and the rule base to be capable of copingwith the dynamic change of knowledge. To deal withthis issue, knowledge base evolution solutions shouldbe proposed: The ontology should be able to adapt it-self efficiently to the changes with using ontology evo-lution techniques, and the rule base should be updatedaccording to the change of context, by implementingcontextual reasoning.

The third problem is the handling of real-time data.Since the manufacturing domain is highly-dynamic,how to process real-time and heterogeneous datastreams is a crucial concern to manufactures. However,the proposed approach uses the classical ontology rea-soning techniques, which can not deal with highly dy-namic data in a timely fashion. To cope with this is-sue, stream reasoning techniques should be adoptedto reason upon a variety of highly dynamic data [7].In stream reasoning, rich query languages are pro-vided by stream reasoners to continuously query datastreams. In this way, predictive maintenance systemsare able to detect and predict machinery failures inreal-time.

Acknowledgements

This work has received funding from INTERREGUpper Rhine (European Regional Development Fund)and the Ministries for Research of Baden-Wrttemberg,Rheinland-Pfalz (Germany) and from the Grand EstFrench Region in the framework of the Science Offen-sive Upper Rhine HALFBACK project.

References

[1] A. Avizienis, J.C. Laprie, B. Randell, and C. Landwehr, Basicconcepts and taxonomy of dependable and secure computing.IEEE transactions on dependable and secure computing, 2004,1(1), pp. 11-33.

[2] A. Grall, L. Dieulle, C. Bérenguer, and M. Roussignol,Continuous-time predictive-maintenance scheduling for a dete-riorating system. IEEE transactions on reliability, 2002, 51(2),pp. 141-150.

[3] B. Vogel-Heuser, and D. Hess, Guest editorial Industry 4.0-prerequisites and visions. IEEE Transactions on AutomationScience and Engineering, 2016, 13(2), pp. 411-413.

[4] C. Dousson, and T.V. Duong, Discovering Chronicles with Nu-merical Time Constraints from Alarm Logs for Monitoring Dy-namic Systems. In Proceedings of the Sixteenth InternationalJoint Conference on Artificial Intelligence, 1999, pp. 620-626.

[5] C. Sellami, C. Miranda, A. Samet, M. Tobji, F.A.B. de Beuvron,On mining frequent chronicles for machine failure prediction.Journal of Intelligent Manufacturing, 2019, pp. 1-17.

[6] D. Cram, B. Mathern, and A. Mille, A complete chronicle dis-covery approach: application to activity analysis. Expert Sys-tems, 2012, 29(4), pp. 321-346.

[7] D. Dell’Aglio, E. Della Valle, F. van Harmelen, and A. Bern-stein, Stream reasoning: A survey and outlook. Data Science,2017, 1(1-2), pp. 59-83.

[8] D. Dou, H. Wang, and H. Liu, Semantic data mining: A sur-vey of ontology-based approaches. In IEEE International Con-ference on Semantic Computing (ICSC), 2015, pp. 244-251.

[9] D.L. McGuinness, and F. Van Harmelen, OWL web ontologylanguage overview. W3C recommendation, 2004, 10(10).

[10] D.L. Nuez, and M. Borsato, An ontology-based model forprognostics and health management of machines. Journal of In-dustrial Information Integration. 2017, 6, pp. 33-46.

[11] E. Järvenpää, N. Siltala, O. Hylli, and M. Lanz, The devel-opment of an ontology for describing the capabilities of manu-facturing resources. Journal of Intelligent Manufacturing, 2019,30(2), pp. 959-978.

[12] E. Maleki, F. Belkadi, N. Boli, B.J. van der Zwaag, K. Alex-opoulos, S. Koukas, M. Marin-Perianu, A. Bernard, and D.Mourtzis, Ontology-based framework enabling smart Product-Service Systems: Application of sensing systems for machinehealth monitoring. IEEE Internet of Things Journal, 2018.

[13] F. Ameri, and C. McArthur, Semantic rule modelling for in-telligent supplier discovery. International Journal of ComputerIntegrated Manufacturing, 2014, 27(6), pp. 570–590.

[14] F. Ameri, and D. Dutta, An Upper Ontology for ManufacturingService Description. ASME Conference Proceedings, 2006, pp.651-661.

[15] H. Panetto, M. Dassisti, and A. Tursi, ONTO-PDM: Product-driven ONTOlogy for Product Data Management interoperabil-ity within manufacturing process environment. Advanced Engi-neering Informatics, 2012, 26(2), pp. 334-348.

[16] I. Guyon, and A. Elisseeff, An introduction to variable and fea-ture selection. Journal of machine learning research, 2003, 3,pp. 1157-1182.

[17] I. Horrocks, P.F. Patel-Schneider, H. Boley, S. Tabet, B. Grosof,and M. Dean, SWRL: A semantic web rule language combiningOWL and RuleML. W3C Member submission, 2004, 21, pp. 79.

Qiushi Cao et al. / Combining Chronicle Mining and Semantics for Predictive Maintenance in Manufacturing Processes 21

1 1

2 2

3 3

4 4

5 5

6 6

7 7

8 8

9 9

10 10

11 11

12 12

13 13

14 14

15 15

16 16

17 17

18 18

19 19

20 20

21 21

22 22

23 23

24 24

25 25

26 26

27 27

28 28

29 29

30 30

31 31

32 32

33 33

34 34

35 35

36 36

37 37

38 38

39 39

40 40

41 41

42 42

43 43

44 44

45 45

46 46

47 47

48 48

49 49

50 50

51 51

[18] J.H. Gennari, M.A. Musen, R.W. Fergerson, W.E. Grosso, M.Crubézy, H. Eriksson, N.F. Noy, and S.W. Tu, The evolution ofProtégé 5.5.0: an environment for knowledge-based systems de-velopment. International Journal of Human-computer studies,2003, 58(1), pp. 89-123.

[19] J. Pei, J. Han, and W. Wang, Mining sequential patterns withconstraints in large databases. In Proceedings of the eleventh in-ternational conference on Information and knowledge manage-ment, 2002, pp. 18-25.

[20] J.R. Hobbs, and F. Pan, Time ontology in OWL. W3C workingdraft, 2006, 27, pp. 133.

[21] J.Z. Sikorska, M. Hodkiewicz, and L. Ma, Prognostic mod-elling options for remaining useful life estimation by indus-try. Mechanical systems and signal processing, 2010, 25(5), pp.1803-1836.

[22] K. Efthymiou, N. Papakostas, D. Mourtzis, and G. Chrys-solouris, On a predictive maintenance platform for productionsystems. Procedia CIRP, 2012, 3, pp. 221-226.

[23] L.N. Soldatova, and R.D. King, An ontology of scientific ex-periments. Journal of the Royal Society Interface, 2006, 3(11),pp. 795-803.

[24] L. Obrst, Ontologies for semantically interoperable systems. InProceedings of the twelfth international conference on Informa-tion and knowledge management, 2003, pp. 366-369. ACM.

[25] M. Bali, Drools JBoss Rules 5.0 Developer’s Guide, Packt Pub-lishing Ltd 2009.

[26] M. Cannataro, and C. Comito, A data mining ontology for gridprogramming. In Proceedings of the 1st International Workshopon Semantics in Peer-to-Peer and Grid Computing, 2003, pp.113-134.

[27] M. Grninger, Ontology of the process specification language.Handbook on ontologies, 2004, pp. 575-592. Springer, Berlin,Heidelberg.

[28] M. Horridge, and S. Bechhofer, The owl api: A java api for owlontologies. Semantic Web, 2011, 2(1), pp. 11-21.

[29] M.J. O’Connor, R.D. Shankar, C. Nyulas, A.K. Das, M.A.Musen, The SWRLAPI: A Development Environment forWorking with SWRL Rules. OWL: Experiences and Directions(OWLED), 4th International Workshop, 2008.

[30] M. McCann, and A. Johnston, SECOM data set. 2008, UCIMachine Learning Repository.

[31] M. Poveda-Villalón, A. Gómez-Pérez, and M.C. Suárez-Figueroa, Oops!(ontology pitfall scanner!): An on-line tool forontology evaluation. International Journal on Semantic Weband Information Systems (IJSWIS), 2014, 10(2), pp. 7-34.

[32] M. Stone, Cross-validatory choice and assessment of statisti-cal predictions. Journal of the royal statistical society, Series B(Methodological) 2974, pp. 111-147.

[33] M. Yoshida, T. Iizuka, H. Shiohara, and M. Ishiguro, Miningsequential patterns including time intervals. In Data Mining andKnowledge Discovery: Theory, Tools, and Technology II, 2000,pp. 213-221.

[34] P. Fournier-Viger, J.C-W. Lin, R.U. Kiran, Y.S. Koh, R.Thomas, A survey of sequential pattern mining. Data Scienceand Pattern Recognition, 2017, 1(1), pp. 54-77.

[35] P. Doran, Ontology reuse via ontology modularisation. InKnowledgeWeb PhD Symposium, vol.2006.

[36] P.N. Tan, Introduction to data mining. Pearson Education In-dia, 2007.

[37] P. Panov, L. Soldatova, and S. Džeroski, Ontology of core datamining entities. Data Mining and Knowledge Discovery, 2014,28(5-6), pp. 1222-1265.

[38] P. Ristoski, and H. Paulheim, Semantic Web in data mining andknowledge discovery: A comprehensive survey. Web semantics:science, services and agents on the World Wide Web, 2016, 36,pp. 1-22.

[39] Q. Cao, F. Giustozzi, C. Zanni-Merk, F. de Bertrand de Beu-vron, and C. Reich, Smart Condition Monitoring for Industry 4.0Manufacturing Processes: An Ontology-Based Approach. Cy-bernetics and Systems, 2019, pp. 1-15.

[40] R. Agarwal, and R. Srikant, Fast algorithms for mining associ-ation rules. In Proceeding of the 20th VLDB Conference, 1994,pp. 487-499.

[41] S. Coradeschi, and A. Saffiotti, An introduction to the anchor-ing problem. Robotics and Autonomous Systems, 2003, 43(2-3),pp. 85-96.

[42] S. Lemaignan, A. Siadat, J.Y. Dantan, and A. Semenenko, MA-SON: A proposal for an ontology of manufacturing domain. InIEEE Workshop on Distributed Intelligent Systems, 2006, pp.195-200.

[43] S. Ramírez-Gallego, S. García, H. Mouriño-Talín, D.Martínez-Rego, V. Bolón-Cane., S. Alonso-Betanzo, J.M.Benítez and F. Herrera, Data discretization: taxonomy and bigdata challenge. Wiley Interdisciplinary Reviews: Data Miningand Knowledge Discovery, 2016, 6(1), pp. 5-21.

[44] T.R. Gruber, A translation approach to portable ontology spec-ifications. Knowledge acquisition, 1993, 5(2), pp. 199-220.

[45] T. Slimani, and A. Lazzez, Sequential Mining: Patterns and Al-gorithm Analysis. International Journal of Computer and Elec-tronics Research, 2013.

[46] V. Chandola, A. Banerjee, and V. Kumar, Anomaly detection:A survey. ACM computing surveys (CSUR), 2009, 41(3), pp. 15.

[47] X. Yan, J. Han, and R. Afshar, Clospan: Mining closed sequen-tial patterns in large datasets. In Proceedings of the 2003 SIAMinternational conference on data mining, 2003, pp. 166-177.

[48] Y. Hirate, and H. Yamana, Generalized SPM with Item Inter-vals. Journal of Computers, 2006, 1(3), pp. 51-60.

[49] Z. Usman, R.I.M. Young, N. Chungoora, C. Palmer, K. Case,and J. Harding, A manufacturing core concepts ontology forproduct lifecycle interoperability. In International IFIP Work-ing Conference on Enterprise Interoperability, 2011, pp. 5-18.Springer, Berlin, Heidelberg.

[50] Z. Usman, R.I.M. Young, N. Chungoora, C. Palmer, K.Case,and J.A. Harding, Towards a formal manufacturing referenceontology. International Journal of Production Research, 2013,51(22), pp. 6553-6572.