Safety Science xxx (xxxx) xxx–xxx
Contents lists available at ScienceDirect
Safety Science journal homepage: www.elsevier.com/locate/safety
A process mining approach to improve emergency rescue processes of fatal gas explosion accidents in Chinese coal mines ⁎
Zhen Hea, Qiong Wua, Lijie Wenb, , Gui Fuc a
College of Management and Economics, Tianjin University, No 92 Weijin Rd., Nankai District, Tianjin 300072, China School of Software, Tsinghua University, Room 405, Region 11, East Main Building, Haidian District, Beijing 100084, China c Faculty of Resource and Safety Engineering, China University of Mining & Technology, Haidian District, Beijing 100083, China b
A R T I C LE I N FO
A B S T R A C T
Keywords: Gas explosion accidents Emergency rescue processes Process mining
Gas explosion has always been one of the major accident types in the coal mines of China, and reliable emergency rescue process is one of the key guarantees to minimize accident losses. As an emerging discipline to discover bottlenecks and position deviations of processes, process mining is widely used to improve various business processes in recent years. Nevertheless, scarcely does any research apply it to the field of coal mine emergency rescue to our knowledge. In this study, we apply the technique of process mining to domain of gas explosion accident emergency rescue in coal mines. 50 considerable and major gas explosion accidents, which occurred from the second half of 2006–2014 in Chinese coal mines, are selected as log data. The aspects of control-flow perspective, case perspective, helicopter view, organizational dimension and performance perspective are analyzed by process mining. In terms of control-flow perspective, the emergency rescue model is extracted by the inductive miner in ProM, correspondingly, conformance checking and comparison with real-life model have shown the validity of our work. Research also shows that the average length of cases is 23.72 events, and 60% of the cases are under the average length; 56% of the gas explosion accident emergency rescue processes last less than two days; emergency rescue headquarters, barmaster and medical staff share the highest centrality in social network of rescue; the rescue performance is more affected by accident grade, compared with the factors of accident site, region of coal mine and coal mine ownership.
1. Introduction
1.2. Emergency rescue situation of coal mines in China
1.1. Prevention situation of gas explosion accidents in Chinese coal mines
The definition of mishap is the results of unintentional death, injury, loss of property and so on (DoD, 2012). On the one hand, safety researchers focus on reducing the accident incidence, on the other hand, they pay attention to the reduction of the mishap severity. Once a gas explosion accident (GEA for short) in underground coal mine occurs, emergency rescue is the vital pathway to minimize casualties and property losses. As of 2013, there are 397 specialized mine rescue teams in China (SAWS, 2014), serving more than 12,000 coal mines nationwide (SCIOPRC, 2014). Though the state has been shutting down small coal mines in a large number of cases annually, the number of coal mines is still in the absolute advantage relative to the number of emergency rescue teams up to present. Many correlative factors in an underground coal mine system are critical to a GEA emergency rescue, such as the workers’ knowledge and response ability to complex and ever-changing underground (KowalskiTrakofler et al., 2010). Furthermore, some rescue processes which have
As of now, China is the world’s largest producer and consumer of coal (Wang et al., 2014). Although the number of fatalities has steadily decreased year by year, compared with the United States under the same safety performance index, the world’s major coal-producing country, there is still a great gap of safety in coal mines between China and the USA (Nieto et al., 2014; Yin et al., 2017). For various types of coal mine accidents, gas explosion accident is undoubtedly one of the biggest threats to underground coal mines (Chen et al., 2013; Yin et al., 2017). Even with the downward trend of death toll during 2004–2014 for considerable and major gas explosion accidents in Chinese coal mines, the proportion of gas explosion accidents in total casualties of considerable and major coal mine accidents seems unstable, and what’s more, has shown an upward trend between the year of 2010 and 2014 (Fig. 1).
⁎
Corresponding author. E-mail addresses:
[email protected] (Z. He),
[email protected] (Q. Wu),
[email protected] (L. Wen).
https://doi.org/10.1016/j.ssci.2018.07.006 Received 17 November 2017; Received in revised form 22 March 2018; Accepted 8 July 2018 0925-7535/ © 2018 Elsevier Ltd. All rights reserved.
Please cite this article as: He, Z., Safety Science (2018), https://doi.org/10.1016/j.ssci.2018.07.006
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
reliability analysis; Park et al. (2016) did a feasibility study in a nuclear power plant scene to assess the quality of work process; for the sake of identifying appropriate process redesign, van Beest and Maruster (2007) used PM and simulation techniques in a Dutch company belonging to gas industry; PM was applied to analyze the workload and delay in manufacturing process by Park et al. (2015); der Weerdt et al. (2013) verified the value of PM to position organization inefficiency in insurance company; Gupta (2007) found the association rules that the complications C_-VKF, atrium-flutter and C_Oligurie ( < 5 ml/kg/24 u) always occur together for the patients in the ICU and the patients receiving treatment B_Thoraxdine always received the treatment B_Beademing as well with the technique of PM; Rojas et al. (2016) summarized the PM algorithms, techniques, tools, methodologies and approaches in the domain of healthcare improvement; Rebuge and Ferreira (2012) utilized the technique of PM to identify regular behavior, process variants, and exceptional medical cases for a hospital emergency service; Perimal-Lewis et al. (2012) presented a synergy of PM technique and statistical data analysis to discover inpatient process patterns, ward types, waiting time and length of stay; Poelmans et al. (2010) combined PM with data mining techniques to discover the cause of the patients who refused the key intervention “revalidation”; Mans et al. (2015) summarized the implementation of the PM in the field of healthcare, and the data quality issue was elaborated synchronously; Bogarín et al. (2018) did a survey on educational process mining, who outlined the application of PM to discover and analyze the educational process; Intayoad and Becker (2018) propose a methodology to elicit actual business process model in the scene of manufacturing and logistic for large transaction data. However, scarcely does any researcher apply it to the field of coal mine emergency rescue process. Different from empirical emergency rescue process models in the contingency plans mentioned previously, PM can discover emergency rescue process models from GEA reports which have recorded the emergency rescue processes systematically, meanwhile, bottlenecks and deviations will be positioned when analyzing the mining results. It is important to point out that the models in this article are conceptual models (Friedrich, 2010). Correspondingly, the structure of this paper is as follows. Above all, in Section 2, materials and methods of this research are provided. Afterwards, in Sections 3 and 4, the case study is addressed. Discussions are presented in Section 5. Finally, Section 6 provides some conclusions and anticipation of future work.
Fig. 1. Death toll of considerable and major GEAs, and proportion of GEAs in total death toll of considerable and major coal mine accidents (Data for death toll of gas explosion accidents and the proportion is from State Administration of Coal Mine Safety (SACMS, 2015)).
led to the expansion of the accidents are found as well (SACMS, 2015). This is directly related to decisions of emergency rescue commander. When searching the key words of improper coal mine rescue in the SAWS website, a large number of directly relevant messages will be represented. Normally, the guidance of coal mine emergency rescue is mine rescue regulations (by SAWS) and different levels of contingency plans based on the laws and regulations of governments and industry norms. For instance, law of the People’s Republic of China on safety production clearly stipulates that the enterprises shall formulate contingency plans for accidents (NPC, 2014). To make procedural information comprehended by users of every hue, it has become a quintessential activity to apply textual process descriptions with conceptual process models in the enterprises (ver der Aa et al., 2017). With similar principle, at the end of the contingency plan, a flow-chart of emergency rescue is presented in the ordinary course of events (real-life model). Nevertheless, coal mine emergency rescue has not been paid enough attention up to now, and it is a weaker link in the safety work system of coal mine in China (Niu et al., 2012). As a matter of fact, both governments and enterprises at different levels have varieties of emergency plans. In the case of ignoring the individual contingency plans, only integrated plans are often more than 100 pages, and even if the professional safety staff can hardly dig out useful information in good season (Guo et al., 2013). According to our interviews with several safety supervisors in two provinces of China and investigating emergency plans of two large coal mines in the other two provinces in China, for one thing, the contingency plans of coal mines follow the laws and regulations to a certain extent, for another, largely based on experience. Furthermore, excessive reliance on paper plans will weaken the efficiency of emergency management in coal mines (Cui, 2015).
2. Materials and methods Fig. 2 is the framework of this study, which outlines the main procedure. The first part addresses data acquisition, followed by data preprocessing. Then the technique of PM and data reprocessing is presented, which is followed by PM and data reprocessing until the mining results accord with the needs of business staff to the utmost. Final part of the framework is the analysis of mining results and the advices on the improvement of GEA emergency rescue processes in underground coal mines.
1.3. State-of-the-art for process mining application
2.1. Data acquisition
Extracting knowledge from event logs to discover, monitor and improve real workflows is the basic idea of process mining (PM for short) (van der Aalst, 2016). PM has been widely applied in plenty of fields, e.g., human reliability analysis, workflow redesign of gas industry manufacturing process, healthcare process improvement and so forth (Kelly, 2011; Park et al., 2016; van Beest and Maruster, 2007; Park et al., 2015; der Weerdt et al., 2013; Gupta, 2007; Rojas et al., 2016; Rebuge and Ferreira, 2012; Perimal-Lewis et al., 2012; Poelmans et al., 2010; Mans et al., 2015; Bogarín et al., 2018; Intayoad and Becker, 2018). In more concrete terms, Kelly (2011) incorporated PM into human
As the principle of accident prevention depicted by Heinrich Pyramid Law, in a workplace, of the 330 accidents, 300 accidents result in no injury, 29 accidents cause minor injuries and 1 accident results in serious injury (Heinrich et al., 1980). Accordingly, for the target of accident prevention, all of the no injury mishaps, minor injury mishaps and serious injury mishaps of gas explosion in coal mines should be our source data. However, the central issue of this study is emergency rescue process, and data quality is one of the most important challenge the same as other PM practices (van der Aalst, 2016). As the case stands, coal mine accidents can be classified into four grades in China (considerable accident, major accident, serious and ordinary accident) 2
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
Fig. 2. Framework of this study.
(SCPRC, 2007; Gao et al., 2016). But few of serious and ordinary gas explosion mishaps have systematic and detailed records, let alone emergency rescue processes. Therefore, in this study, the data set of the GEA cases in coal mines are all considerable and major accidents, from the perspective of the death toll, the least death toll of considerable accidents is 30, while the death toll range of major accidents is 10–29 (SCPRC, 2007; Gao et al., 2016). In order to get relatively complete data sets of considerable and major GEAs of coal mines in China, on the basis of our existing data (coal mine accident investigation reports) (SACMS, 2015), a query on the accident inquiry system of State Administration of Work Safety of China with the time interval from 2000 to 2014 are made (SAWS, 2017) accompanied by a web search with the key words of considerable and major GEAs in coal mines with the same time interval. It turns out that 230 considerable and major GEAs are found (Fig. 3). When combing emergency rescue processes of cases from 2000 to 2014, a considerable number of the accident investigation reports before 2006 do not include the emergency rescue processes in fact, or the descriptions are so simple that are not available. This may have to do with the promulgation of mine accident contingency plans in 2006
Fig. 3. Number of considerable and major GEAs in coal mine of China during 2000–2014.
3
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
No Report No Rescue Process Too Sketchy of Rescue Process
problems in the whole process of our research. The carding results are stored in Excel. Table 1 is part of the carding results from GEA investigation reports. In Table 1, each row corresponds to an event, each event relates to a case ID, one timestamp, an activity and a resource. It should be noted that the timestamps in this study refer to the ending time of the activities for the limitation of source data. For various reasons, inaccuracy of timestamps is prevalent (Barga et al., 2007). In the investigation reports of GEAs, all sorts of inaccurate timestamps exist, such as morning, morning shift, the next day, direct missing timestamps and so on. On account of minimum change principle (Bohannon et al., 2005), the timestamp repair needs to satisfy the timing constraint and is closest to the original value (Song et al., 2016). Repairing the imprecise timestamps of the event log is the following task. Specifically, the repairing principles are as follows, (a) if one activity is followed by another activity during emergency rescue processes, its timestamp is earlier; (b) if the emergency rescue process of the case is relatively sketchy, we might supplement the related information by searching the internet (for example, we could get the information of emergency rescue process from the description of some news); (c) if affiliation exists in two institutions, subordinate’s timestamp is earlier than its upper level (for example, county government gets the accident message earlier than municipal government which might make a response earlier generally); (d) if a certain number of missing timestamps are continuous, the description order of the text would be the reference principle; (e) satisfying the timing constraint in the reports and closest to the original value are the most important basis all the time. It turns out to be an event log with 50 cases, 1186 events, 117 types of activities and 122 kinds of resources with the format of XLSX.
35.29% 23.53%
41.18% Fig. 4. Causes of the unavailability of 17 cases.
(SAWS, 2006) and regulations on reporting and investigation of production safety accidents in 2007 (SCPRC, 2007). The former specifies the hierarchical response in mine accident emergency rescue, the latter clearly stipulates that the accident investigation report needs to describe the accident rescue processes. Hierarchical response is the reason of ordinary and serious GEAs’ exclusion as well, and by contrast, emergency response of ordinary and serious GEAs cannot reflect the whole workflow of emergency rescue (panorama of the GEA rescue processes) previously. Consequently, among the 67 considerable and major GEAs occurred from the second half of 2006 to 2014, 50 cases which have recorded the relatively complete emergency rescue processes are selected as the research objects. In the meantime, the causes for the unavailability of 17 other cases are summed up (Fig. 4). Of the 17 cases, case number of which the accident report hasn’t been found is 7, case number of which there is no rescue process in accident report is 6, case number of which there exists only sketchy emergency rescue process is 4.
2.3. Discussion on completeness and process mining tools Completeness is an important notion when it comes to PM, which relates to noise (van der Aalst, 2016). In the domain of PM, the researchers generally set the threshold to filter the cases for dealing with the noise. Of the two types of completeness, the focus of global completeness is the lack of case (trace) (Pei et al., 2018), and the focus of local completeness is the lack of relationship among the tasks (Yang et al., 2014). Different from lion’s share of process discovery algorithms, inductive miner focuses on the facts of the event log (Leemans et al., 2013, 2014a, 2014b, 2015; Leemans, 2017). Of course, the more logs, the closer our mining results are to the truth and the more instructive to the actual emergency rescue process it is. Furthermore, inductive miner has been integrated into the framework of ProM, and ProM is well known as the most powerful framework for PM currently (van der Aalst, 2016). And this provides us with the possibility of applying inductive miner to model the emergency rescue process (controlflow perspective). Moreover, organizational mining (Song and van der Aalst, 2008) and conformance checking (Munoz-Gama, 2014; MunozGama and Carmona, 2010, 2011) will be implemented with the
2.2. Data preprocessing Event log is crucial for PM, which is the origin of mining valid information from various latitudes (van der Aalst, 2016). Thus, the first task of this study is to extract the emergency rescue processes of the 50 cases in accordance with the original description of the texts. Table 1 is a sample of the process carding results, which can be regarded as a section of an event log. When it comes to the carding of emergency rescue processes, combing principles need to be mentioned, which is the basis of presenting accurate emergency rescue processes. The first principle of all is to keep the original description of accident reports, coal mine emergency rescue experts, coal mine safety experts, safety supervision personnel of government and frontline technical staff of coal mine will be consulted when facing with abstract business Table 1 A section of an event log: each row corresponds to an event. Case id
50 50 50 50 50 50 50 50 50 50
Properties Timestamp
Activity
Resource
14-12-2014:10.23 14-12-2014:10.23 14-12-2014:10.23 14-12-2014:10.23 14-12-2014:10.45 14-12-2014:11.30 14-12-2014:11.42 14-12-2014:11.42 14-12-2014:11.42 14-12-2014:11.57
Find accident Report technical barmaster Organizing miners to rescue themselves Leading miners to rescue Some of the miners go out of well Report group company commissioner Report group company chief engineer Call group company rescue team Start emergency response plan Report county coal administration
Barmaster, safety barmaster Safety barmaster Barmaster, safety barmaster Technical barmaster Working group of coal face, working group of heading face Barmaster Group company commissioner Chief engineer of group company Chief engineer of group company Chief engineer of group company
4
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
parameters of Activities and Paths on the interface of Disco. Of course, the three principles are the reference point of each generated model. The quantitative selection of process model under different number of activities is an interesting and novel topic, which will be involved in our following study. The empirical research method will be used, for example, the approach of questionnaire. On top of that, Disco is only an auxiliary tool to model the process for its high readability in this study, which helps to identify the effects of the activity abstraction. The final process modeling algorithm we choose is the inductive miner in ProM 6.4.1. Meanwhile, both the construct validity (Section 3.2) and the external validity (Section 3.3) will be discussed for the final mined model.
framework of ProM as well. Disco (http://fluxicon.com/) will be used as auxiliary tool to do activity abstraction, case analysis, helicopter view analysis and performance analysis, which is one of the most effective tools among lots of PM tools and an extension of the fuzzy mining algorithm (Günther and van der Aalst, 2007). What we have to mention is the problem of the minimum data quantity can be used to do analysis with Disco. As a matter of fact, there is no real minimum data quantity that is necessary to get a PM result with Disco. The amount of data we should extract relies on the questions that we want to answer for Disco. To be more specific, both Disco 2.0.0 and ProM 6.4.1 are used in our whole study. In particular, the event log with XLSX format will be read and converted to XES (eXtensible Event Stream) (van der Aalst et al., 2012) format by Disco 2.0.0 before applying ProM6.4.1, for the reason that XLSX format cannot be identified by ProM. To show how the log is transformed into XES format, a snippet of the transformation is presented as follows: (1) after starting the Disco 2.0.0 software, click the button of “open file” to read the event log with XLSX format; (2) set properties (case, activity, timestamp, resource, other property) for each column of data, add or delete one column; (3) click the button of “start import”; (4) click the button of “export” on the lower right corner; (5) select the format of the log which will be exported; (6) click the button of “export XES file”; (7) save the exported event log.
2.5. Process improvement The ultimate purpose of this study is to characterize and improve the process of GEA emergency rescue. Accordingly, positioning the deviations and bottlenecks of the GEA emergency rescue process from the mining results is the core task, which can provide some serviceable advices for future GEA emergency rescue in underground coal mines. The perspectives of control-flow, case, helicopter view, organization and performance are covered. When relating to case perspective and performance perspective, data granularity will be considered, which is related to the quality of mining results and improvement of business processes (van der Aalst, 2016). Dimensions of accident grade, accident site, the region of coal mine and coal mine ownership are often involved in coal mine accident statistical analysis (SACMS, 2015), and they are considered as data granularity factors in the analysis of case and performance perspectives. When considering the effect of accident grade, the original event log will be divided into two event logs based on considerable and major accidents; when studying the effect of accident site, the original event log will be divided into three event logs based on coal face, heading face and other accident sites; when studying the factor of coal mine region, the original event log will be divided into two event logs based on the location of coal mine in north or south of China; when studying the factor of coal mine ownership, the original event log will be divided into two event logs based on state and private ownership. All the event logs mentioned here will be transformed into XES format by Disco before applying ProM.
2.4. Data reprocessing and modeling process If the PM result is in form of spaghetti, it will be difficult to understand. Furthermore, the algorithms of PM merely provide syntactical analysis with activities, where activities are just alphabetic string (Montani et al., 2017). In reality, the elements of the process model which are higher than 50 will be difficult to comprehend (Mendling et al., 2010). In consideration of different user requirements, one just needs to display the most relevant elements in the model, and the less relevant elements should be hidden (Smirnov et al., 2010). Business process model abstraction is needed by the light of nature. Abstracting activities continually in data reprocessing to simplify the process model until the PM result is practicable (Montani et al., 2017). Further, the abstraction of activities here is the merger of activities with similar meaning. On account of three fundamental principles to the utmost, the work of abstracting activities is carried out. The three principles are as follows: (1) we take consideration of mine accident contingency plans (SAWS, 2006) maximally; (2) we comply with the original intention of the accident investigation reports to the maximum; (3) we follow the seven process modeling guidelines (7PMG) (Mendling et al., 2010) to the utmost. The detailed description of 7PMG consists of seven aspects: (1) the model contains as few elements as possible; (2) each element should possess the minimum routing paths; (3) there are one start and one end event in each model; (4) we should structure the model as far as possible; (5) the OR routing elements should be avoided; (6) verbobject constitutes the activity labels; (7) if the model have more than 50 elements, it should be decomposed. When implementing abstracting based on these principles, the activities with similar meaning in the process of GEA emergency rescue will be merged, such as “find part of the dead miners”, “find all of the dead miners”, “transport part of the dead miners”, “transport all of the dead miners”, any of the four activities has a bearing on searching and rescuing dead miners. Eventually, 117 kinds of activities are abstracted into 17 types of activities. Again, a new event log with 50 cases, 1186 events, 17 kinds of activities, 122 types of resources is generated. Of course, the data format remains the same as XLSX. Fig. 5 is the flowchart of data reprocessing. Both the work of data preprocessing and data reprocessing are completed together with related business personnel. The process map generator of Disco helps us with the confirmation of model availability, which facilitates the visualization of typical workflows to a great extent. In other words, the optimal model will be found out through adjusting
3. Mining control-flow perspective 3.1. Discovery of process model by inductive miner As well known, control-flow view is the most important dimension of PM (van der Aalst, 2016). To have an overview understanding the process model structure, we elicit the emergency rescue process model with the inductive miner (Leemans et al., 2013, 2014a, 2014b, 2015; Leemans, 2017) embedded in ProM 6.4.1. The process modeling language here is Petri net (Murata, 1989). The imported format of event log with 17 types of activities in ProM 6.4.1 is XES. The mining result is shown in Fig. 6, which is the process model based on Petri net, and follows the rule of Petri net as well. Accordingly, the round in the net represents the place, which sends out the branches with mutual exclusion (XOR-split); the square (both black and white) in the net represents the transition, which sends out the branches with concurrency relation (AND-split). The abbreviations and full names of activities are displayed in Table 2. As we can see from the diagram, there are some concurrency structures, selection structures and invisible tasks (black square, which does not have real semantics and they play a routing role) in the model. FA is always the first activity in each of the 50 cases, but the activity of FA is not located in the beginning of the net, which indicates that a new accident might be found in the process of the GEA rescue. After the activity of FA, we can find that the activity of RCMSDAL, RSLCM, 5
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
Fig. 5. Flowchart of data reprocessing.
The notion of fitness can be used to measure “the proportion of behavior in the event log possible in the light of the model”. According to the result, 0 of the 50 traces is “Non Fit Traces”. Therefore, the value of fitness is 1. That is to say, our mined model and our event log share perfect conformance.
RGCLCM, RIP, RCR, UERSAD, RDP and ERCM are mutually exclusive. Meanwhile, each of the eight activities has formed a cycle, which means that each activity of the eight might be executed multiple times during an emergency rescue of GEA. The activities of MAIP, EERH, RGAL and SERP hold a relation of mutual exclusion as well. There is a concurrency relationship between the activity of EPERR and any the four activities (MAIP, EERH, RGAL, SERP). In particular, MAIP, EERH, RGAL, SERP, EPERR, ARR and SAA have formed a cycle respectively as well. Obviously, the activity of ARR is after the activities of RCMSDAL, RSLCM, RGCLCM, RIP, RCR, UERSAD, RDP, ERCM, MAIP, EERH, RGAL and SERP, which indicates that the activity of assessing the risk of emergency rescue is based on a lot of various emergency rescue efforts. The intent of SAA is to prevent the expansion of the accident, which is the decision of having to be made. The activity of AER is a sign of the terminal point during the emergency rescue.
3.3. Discussion of external validity In terms of external validity, we take the measure of comparing the mined model with a real-life model. In this subpart, the mined model is compared with the model of GEA emergency rescue model which is being used in a particularly large coal mine in Shanxi province of China (real-life model). The real-life model is included in the special contingency plan of methane, especially for GEAs, which consists of 17 pages of documents. The simplified version of real-life emergency rescue model is shown in Fig. 7. Intuitively, the real-life model which contains three response levels differs from the hierarchical response of mine accident contingency plans (SAWS, 2006), which contains four response levels corresponding to four different accident levels. The possible reason for this situation is that the decision makers of coal mine always could not identify the accident severity at the initial stage of GEAs, which leads to result that coal mine can only start the emergency rescue relying on their own rescue power (rescue teams of the coal mine and group company) firstly. With the expanding of the accident, the coal mine will raise emergency response level. The two are not contradictory. When external rescue forces are needed to intervene, the mine will start the contingency plan of class one. This also reflects the empirical basis of contingency plans to a certain extent. Similarities and differences of the two models will be considered in the following part of this subsection. First of all, both the mined model and the real-life model cover the activities of FA, ERCM, RCR, RGCLCM, RSLCM, AER. Compared with the mined model, the real-life model covers a judging activity of CSA (confirm severity of the accident) and the activities of ERM (emergency rescue measures). In the real-life model, the activity of ERM refers to collecting information about coal mine, making emergency rescue plans and mobilizing emergency rescue supplies. The reason why the mined
3.2. Discussion of construct validity The technique of conformance checking can be used to measure the performance of process discovery algorithms (van der Aalst, 2016). In essence, conformance checking is the comparison of the event log and the model, which contains the comparison possibility of the old event log and the new model, the old event log and the old model, the new event log and the new model, the new event log and the old model. In this regard, the conformance checking algorithm of “Check Conformance using ETConformance” (Munoz-Gama, 2014; Munoz-Gama and Carmona, 2010, 2011) is used. The result of conformance checking is as follows: ● ● ● ● ●
Model: c82b8b1d-4b3a-449f-b134-9cb2386ad586 (by inductive miner) Log: Anonymous log imported from 50cases.xes (event log in this study) |Traces|: 50 |Non Fit Traces|: 0 of 50 Average Trace Size: 24.0
6
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
Fig. 6. The mining emergency rescue process model of GEAs by inductive miner.
government response), different forms of hierarchical response activities in the original mined model without abstracting are found as well. Nevertheless, similar activities have to be abstracted in order to simplify the mined model, such as RGAL (report governments at all levels) and RCMSDAL (report coal mine supervision departments at all levels) both of which cover several similar activities.
model doesn’t contain CSA and ERM is related to the brief description of considerable and major GEA investigation reports, in other words, the data quality is not ideal enough. In terms of the hierarchical response in the real-life model, such as CTRCMR (class three response-coal mine response), CTRGCR (class two response-group company response) and CORGR (class one response7
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
Table 2 Abbreviations and full names of the activities. Abbreviation
Full name
Abbreviation
Full name
Abbreviation
Full name
FA RCR RSLCM
Find accident Report control room Report senior leader of coal mine
RGAL SERP RGCLCM
SAA AER RIP
Seal accident area Announce the end of rescue Rescue of injured personnel
ERCM
Emergency response of coal mine
EERH
MAIP
Medical assistance of injured personnel
RCMSDAL
Report coal mine supervision departments at all levels External power emergency rescue response
RDP
Report governments at all levels Start emergency response plan Report group company leader of coal mine Establish emergency rescue headquarters Rescue of dead personnel
UERSAD
ARR
Assess rescue risk
Underground engineering repair and secondary accident defense /
EPERR
/
Fig. 7. Simplified GEA emergency rescue workflow of a large coal mine in China.
8
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
4.2. A helicopter view
In terms of differences between the two models, the mined model consists of cycles, which may contain multiple activities or a single activity. For example, the activity of FA allows emergency rescue personnel to find that accidents (GEAs or other types of accidents) could happen again during the emergency rescue processes intuitively. Besides, the mined model can help emergency rescue personnel find the logical relationship among the concrete emergency rescue activities. The cycle can make the emergency rescue personnel be aware of the time bottlenecks in the GEA emergency rescue processes. The optimization of the process model which is not directly related to the rescue of miners will help to improve the efficiency of emergency rescue. Meanwhile, the guideline of such optimization depends on the business demand of emergency rescue, and the PM technique is just responsible for the discovery of these bottlenecks. In other word, the decisionmaking power is in the hands of emergency rescue leaders. By comparison, the cycles in the real-life model is more abstract and play the role of increasing the accident response level. Quite evidently, the mined model is more suitable for the training of emergency personnel, which can facilitate the understanding of detailed emergency rescue process for the personnel.
The PM technique of dotted chart provides a general view of the GEA emergency rescue processes (Song and van der Aalst, 2007). In this subsection, the event log with the format of XES is imported into ProM 6.4.1 for the application of dotted chart technique. As shown in Fig. 10, each dot depicts an event. Owing to a classifier of Task ID (activity), each dot is related to an emergency rescue activity, and the dots in same line describe multiple events of one class. The gap between columns of dots reflects the time slots that are not involved in this study. When turning actual time into relative time (Fig. 11), the longest GEA emergency rescue time with more than 54 days is found, and the shortest time is less than half a day. A significant thinking of PM is to find out the deviation of the process. In Fig. 11, the emergency rescue time of case 48 is more than 54 days, which attracts our interest. The reason why the case takes the longest time to rescue is plenty of the missing underground miners, which reflects the personnel-oriented thinking in the emergency rescue process. Furthermore, Fig. 12 is the number of cases with different emergency rescue duration, 14% of the cases take more than 10 days, 16% of the cases take less than half a day, 38% of the cases take less than one day, 56% of the cases take less than two days.
4. Mining other perspectives 4.3. Social network analysis 4.1. Case perspective Organizational mining is concerned with organizational dimensions (Song and van der Aalst, 2008), which reflects the property of resources in Table 1. One thing to notice is that each resource in the event log may contain several professional groups or departments, for instance, multiple mine rescue teams. There are 122 types of resources in the event log, and the four most frequent resources (relative frequency of the resources is more than 5%) are emergency rescue headquarters, municipal government, county government and barmaster. It seems that not all the last resources of cases are emergency rescue headquarters. Ordinarily, emergency rescue headquarters is the command center of emergency rescue, which issues the command of end. Handover-of-work social network (Song and van der Aalst, 2008) is mined by ProM6.4.1 in this subsection. Centrality can represent the position of resources in the social network, and there are various metrics to measure centrality, such as Bavelas-Leavitt index, closeness, betweenness, sociometric status and so on (van der Aalst et al., 2005). When applying the algorithm in ProM 6.4.1 to discover the social network, some parameters need to be set. To be specific, we select Degree to measure centrality; in the meantime, we select “Ranking View” as Layout, and “size by ranking”, “show edge” are selected as “View options”. Ultimately, the social network of GEA emergency rescue process is shown in Fig. 13. The nodes in the figure represent the resources (such as roles, groups, departments) and the arcs between nodes represent the relationship of different nodes (van der Aalst, 2016). In Fig. 13, there are 122 nodes (resources) in the social network, and the key resources will be tagged in the figure. The resources of emergency rescue headquarters share the highest centrality in social network, followed by barmaster, medical staff, and they are the three kinds of business staff who are located in the center of the network.
Case perspective is one of the dimensions PM concerns (van der Aalst, 2016). The final event log of this study consists of 50 cases, and each case corresponds to one emergency rescue process of a GEA. Case number is the number of cases in this study and case length is the number of events in each case. Fig. 8 shows the statistical characteristics of case length. The case of maximum length contains 45 events and 15 events are covered in the case of minimum length. The average length of cases is 23.72 events, 60% of the cases are below the average length and 40% of the cases exceed the average length. When considering the factors of accident grade, accident site, the region of coal mine and coal mine ownership, statistical distribution of case number and case length are shown in Fig. 9. The left four graphs are statistical characteristics of case number and the right four graphs are statistical characteristics of average case length, both of which are under the condition of considering these four factors. Conclusion can be drawn from the graph, no matter considerable or major GEAs, GEAs occurring in the heading face or coal face, the accident areas locating in north or south part of China, state or private ownership of coal mine, there’s little difference in the average length of cases.
4.4. Performance analysis In the case of GEA emergency rescue, the most important performance indicator is the response speed of emergency rescue, which can be measured by case duration. The factors of accident grade, accident site, the region of coal mine and coal mine ownership are considered to compare the case duration in this subsection as well. In this study, case duration is the time interval of FA and the end of the emergency rescue (the activity of AER is generally the terminal of the rescue, but some special circumstances are not excluded) in each case. Fig. 14 is the mean case duration considering the four factors compared with mean
Fig. 8. Statistical distribution of case length. 9
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
Fig. 9. Statistical distribution of case number and average case length considering four factors.
Fig. 10. Dotted chart of actual time.
Fig. 11. Dotted chart of relative time.
10
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
significant differences, considerable GEAs need more time for emergency rescue obviously. Similarly, compared with GEAs in south part of China, emergency rescue of GEA in north of China consumes more time. In contrast, the emergency rescue time durations of state and private coal mine have less difference. Moreover, during GEA emergency rescue processes, the accident site of heading face is slightly higher than coal face in emergency rescue time. From the perspective of the mean emergency rescue time, the emergency rescue performance is less affected by accident site and coal mine ownership, especially ownership. 5. Discussion Considering the control-flow perspective, the final emergency rescue process model is discovered by inductive miner from event log. Furthermore, both the construct validity (Section 3.2) and external validity (Section 3.3) are discussed. Compared with the high abstraction of real-life model, the mined process model can better reflect the detailed emergency rescue process. The display of details (for example, the self-circulation with one activity in the net, the location of FA) can facilitate the understanding of emergency rescue for the personnel participating in emergency rescue training. Deficiencies are also found in the mined model, for instance, secondary accidents can occur at any stage of the emergency rescue before the termination of the rescue, which hasn’t been presented for the reasons of data quality, readability and technical limitations. Similarly, hierarchical response is faced with the same problem. One of the critical work during data processing stage is activity abstraction, the efforts of this study is mainly based on qualitative research. The quantitative research of process model selection on the basis of activity abstraction is an interesting and novel topic, which will be involved in our following study. The empirical research method will be used, for example, the approach of questionnaire. Besides, the detailed guideline to improve the emergency rescue process needs to be fixed to the quantity as well. Considering case perspective, 60% of the cases are below the average case length, and the average case length is not significantly affected by the factors of accident grade, heading face or coal face, the accident region of China, state or private ownership. From performance perspective, compared with major GEAs, considerable GEAs need more time for emergency rescue, and this may be related to the greater damage to underground coal mines of considerable GEAs, which have raised the difficulty of emergency rescue; emergency rescue of GEAs in north part of China consumes more time compared with GEAs in south part of China, and this may be related to the geology differences of coal mines in different regions of China; meanwhile, the emergency rescue performance is less affected by accident site and coal mine ownership, especially ownership. When making comparative analysis on the dimensions of case length and case duration, the factors of accident grade, coal mine region and accident site have greater influence on case duration. As a whole, the perspectives of case and performance are mainly based on the statistical dimension. If possible, a larger sample size will certainly benefit the reliability of the statistical results. Considering the helicopter view, time bottleneck of GEA emergency rescue is visualized with the dotted chart technique as well. 56% of the cases take less than two days, 38% of the cases take less than one day, 16% of the cases take less than half a day, 14% of the cases take more than 10 days. Even so, overwhelming majority of the emergency rescue measures will be taken once the accidents occur. This can be regarded as the bottleneck of emergency rescue process. Considering organizational perspective, according to the description of the source data, if the government receives the accident report, the government will set up the emergency rescue headquarters in general. The case of which command of end is issued by other resources (for example, barmaster) possibly indicates that the accident is concealed illegally until the termination of the emergency rescue. Besides, the resource of emergency rescue headquarters shares the highest centrality and most intensive arrows in the social network, followed by barmaster and medical staff. The
Fig. 12. Number of cases with different emergency rescue duration.
Fig. 13. Social network of GEA emergency rescue processes.
Fig. 14. The mean case duration considering four factors.
case duration as a whole, and case duration refers to the mean emergency rescue time of per GEA. As can be seen from the graph, the mean case duration of considerable and major GEA emergency rescue has
11
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
61472207, 61325008). Moreover, we would like to thank the coal mine emergency rescue experts, coal mine safety experts, frontline technical staff of coal mine, safety supervisors of the government who provided us with business guidance. The comments from the editor and reviewers are all rich in sparkling insights and we believe that the quality of this manuscript has benefited a lot from them. Many thanks for the editor and all reviewers.
structure of the social network is closely related to the activities undertaken by the resources. To be specific, emergency rescue headquarters is the command center in the whole process of emergency rescue. Barmaster is the most in-depth understanding of the accident coal mine normally, and one of the basic principles of emergency rescue is that accident coal mine is the main body during emergency rescue. While the medical staff are related to personnel rescue, which is most important task throughout the rescue process. In what follows, the arrows in the social network reveal the handover of task among the resources. As can be seen, emergency rescue headquarters shares most intensive arrows, which indicates the most frequent handover of work. The organizational mining result evidently suggests that the coordination efficiency in the emergency rescue process is directly related to the efficiency of emergency rescue process.
Appendix A. Supplementary material Supplementary data associated with this article can be found, in the online version, at https://doi.org/10.1016/j.ssci.2018.07.006. References
6. Conclusions and future study
Barga, R.S., Goldstein, J., Ali, M.H., Hong, M., 2007. Consistent streaming through time: a vision for event stream processing. CIDR 363–374. Bogarín, A., Cerezo, R., Romero, C., 2018. A survey on educational process mining. Wiley Interdiscipl. Rev. Data Min. Knowl. Discov. 8 (1), 1–17. Bohannon, P., Flaster, M., Fan, W., Rastogi, R., 2005. A cost-based model and effective heuristic for repairing constraints by value modification. SIGMOD 143–154. Chen, H., Qi, H., Feng, Q., 2013. Characteristics of direct causes and human factors in major gas explosion accidents in Chinese coal mines: case study spanning the years 1980–2010. J. Loss Prev. Process Ind. 26, 38–44. Cui, Y.H., 2015. Application and Research on Shared Ontology Model of Coal Mine Emergency Cases. Atlantis Press, pp. 405–409. der Weerdt, J., Schupp, A., Vanderloock, A., Baesens, B., 2013. Process mining for the multi-faceted analysis of business processes-a case study in a financial services organization. Comput. Ind. 64 (1), 57–67. Friedrich, F., 2010. Automated Generation of Business Process Models from Natural Language Input. Master thesis. Humboldt-Universität zu Berlin. Gao, Y., Fu, G., Nieto, A., 2016. A comparative study of gas explosion occurrences and causes in China and the United States. Int. J. Min. Reclam. Environ. 30, 269–278. Günther, C.W., van der Aalst, W.M.P., 2007. Fuzzy Mining – Adaptive Process Simplification Based on Multi-perspective Metrics. In: International Conference on Business Process Management, vol. 4714. pp. 328–343. Guo, D.Y., Du, B., Wang, H.W., 2013. Analysis of Coal Mine Emergency Rescue in China. China Coal Industry Publishing House, Beijing. Gupta, S., 2007. Workflow and Process Mining in Healthcare. Eindhoven University of Technology, Eindhoven. Heinrich, H.W., Petersen, D., Roos, N., 1980. Industrial Accident Prevention: A Safety Management Approach, fifth ed. McGraw-Hill Book Company, New York. Intayoad, W., Becker, T., 2018. Applying process mining in manufacturing and logistic for large transaction data. Lecture Notes in Logistics In: International Conference on Dynamics in Logistics. Springer International Publishing AG, pp. 378–388. Kelly, D.L., 2011. Incorporating Process Mining into Human Reliability Analysis. Eindhoven University of Technology, Netherlands. Kowalski-Trakofler, K.M., Vaught, C., Brnich, M.J., Jansky, J.H., 2010. A study of first moments in underground mine emergency response. J. Homel. Secur. Emerg. Manage. 7, 1271–1283. Leemans, S.J.J., 2017. Robust process mining with guarantees. Printed by Ipskamp Printing, Enschede, the Netherlands. Technische Universiteit, Eindhoven. Leemans, S., Fahland, D., van der Aalst, W.M.P., 2013. Discovering Block-Structured Process Models from Event Logs - A Constructive Approach Applications and Theory of Petri Nets 2013, vol. 7927, 311–329. Leemans, S.J.J., Fahland, D., van der Aalst, W.M.P., 2014b. Discovering block-structured process models from incomplete event logs. In: International Conference on Applications & Theory of Petri Nets & Concurrency, vol. 8489. Springer, Berlin, pp. 91–110. Leemans, S., Fahland, D., van der Aalst, W.M.P., 2014b. Discovering Block-Structured Process Models from Event Logs Containing Infrequent Behaviour. Lecture Notes in Business Information Processing, vol. 171, 66–78. Leemans, S.J.J., Fahland, D., van der Aalst, W.M.P., 2015. Scalable Process Discovery with Guarantees. Lecture Notes in Business Information Processing, vol. 214, 85–101. Mans, R.S., van der Aalst, W.M.P., Vanwersch, R.J.B., 2015. Process Mining in Healthcare Evaluating and Exploiting Operational Healthcare Process. Springer-Verlag, Berlin. Mendling, J., Reijers, H.A., van der Aalst, W.M.P., 2010. Seven process modeling guidelines (7pmg). Inf. Softw. Technol. 52 (2), 127–136. Montani, S., Leonardi, G., Striani, M., Quaglini, S., Cavallini, A., 2017. Multi-level abstraction for trace comparison and process discovery. Expert Syst. Appl. 81, 398–409. Munoz-Gama, J., Carmona, J., 2010. A fresh look at precision in process conformance. In: Business Process Management (BPM 2010), volume 6336 of Lecture Notes in Computer Science. Springer-Verlag, Berlin, pp. 211–226. Munoz-Gama, J., Carmona, J., 2011. Enhancing precision in process conformance: stability, confidence and severity. IEEE Symposium on Computational Intelligence and Data Mining (CIDM 2011), Paris, France, April 2011. IEEE. Munoz-Gama, J., 2014. Conformance Checking and Diagnosis in Process Mining. (PhD Thesis). Murata, T., 1989. Petri nets: properties, analysis and applications. Proc. IEEE 77 (4), 541–580. National People’s Congress of the People’s Republic of China (NPC), 2014. Law of the
As well known, emergency rescue training and exercise is an important part at home and abroad (SAWS, 2006; Kowalski-Trakofler et al., 2010). This research can provide interesting materials for the emergency rescue training of coal mine GEAs and serve as a reference for emergency rescue decision makers from different perspectives. The following statements will make a summary of the contributions in this study: (1) From control-flow perspective, the feasibility and advantage of extracting emergency rescue process model for GEAs from historical events log is validated and the emergency rescue process is characterized in the meantime. (2) From the case perspective, the average length of cases is 23.72 events, 60% of the cases are below the average length, and 40% of the cases exceed the average length; (3) From performance perspective, considerable GEAs consume more time for emergency rescue compared with major GEAs, emergency rescue of GEAs in north part of China consumes more time compared with south part of China, the accident site of heading face is slightly higher than coal face in emergency rescue time, the emergency rescue duration is least affected by coal mine ownership; compared with case length, the factors of accident grade, coal mine region and accident site possess greater influence on case duration; (4) From the helicopter view, 56% of the cases take less than two days, 38% of the cases take less than one day, 16% of the cases take less than half a day, while 14% of the cases take more than 10 days; (5) From organizational perspective, the resources of emergency rescue headquarters, barmaster and medical staff share a better centrality in GEA emergency rescue processes. The main objective of PM in this study is to discover process model, bottleneck and characterize the GEA emergency rescue, and we will focus on the following directions in the future: (1) We will do some work on automatic workflow elicitation for process mining from accident report documents of GEAs as well; (2) GEA is only one type of coal mine accidents, and we also hold the ambition to apply the PM to other accident emergency rescue processes’ promotion, such as roof accidents, flood accidents, fire accidents etc.; (3) The longer term, we even hold the view that PM can make a contribution to the improvement of other industries’ emergency rescue processes, such as traffic accident emergency rescue, chemical accident emergency rescue etc. Acknowledgements The research is supported by four National Natural Science Foundation of China (NSFC) grant projects (71532008, 71661147003, 12
Safety Science xxx (xxxx) xxx–xxx
Z. He et al.
glance. In: Proceedings of the Seventeenth Annual Workshop on Information Technologies and Systems, pp. 139–145. Song, M., van der Aalst, W.M.P., 2008. Towards comprehensive support for organizational mining. Decis. Support Syst. 46 (1), 300–317. State Administration, of Coal Mine Safety (SACMS), 2015. Chinese Coal Mine Accident Investigation Reports, 2004–2014. State Administration of Coal Mine Safety. State Administration of Work Safety (SAWS), 2006. Mine Accident Contingency Plans. SAWS. State Administration of Work Safety (SAWS), 2014. Bulletin of Quality Standardization Examination Grade of Mine Rescue Teams in 2013. SAWS. State Council of the People’s Republic of China (SCPRC), 2007. Regulations on Reporting and Investigation of Production Safety Accidents. State Council of the People’s Republic of China. United States Department of Defense (DoD), 2012. Department of Defense Standard Practice System Safety, MIL-STD-882E. United States Department of Defense. van Beest, N.R.T.P., Maruster, L., 2007. A process mining approach to redesign business processes-a case study in gas industry. SYNASC 541–548. van der Aalst, W.M.P., 2016. Process mining: Data Science in Action. Springer-Verlag, Berlin. van der Aalst, W.M.P., Adriansyah, A., et al., 2012. Process mining manifesto. In: BPM 2011 Workshops Part I, LNBIP, vol. 99. pp. 169–194. van der Aalst, W.M.P., Reijers, H.A., Song, M., 2005. Discovering social networks from event logs. Comput. Support. Coop. Work 14, 549–593. ver der Aa, H., Leopold, H., Reijers, H.A., 2017. Comparing textual descriptions to process models–the automatic detection of inconsistencies. Inf. Syst. 64, 447–460. Wang, L., Cheng, Y.P., Liu, H.Y., 2014. An analysis of fatal gas accidents in Chinese coal mines. Saf. Sci. 62, 107–113. Yang, H.D., Wen, L.J., Wang, J.M., Wong, R.K., 2014. CPL+: an improved approach for evaluating the local completeness of event logs. Inf. Process. Lett. 114, 607–610. Yin, W.T., Fu, G., Yang, C., Jiang, Z.A., Zhu, K., Gao, Y., 2017. Fatal gas explosion accidents on Chinese coal mines and the characteristics of unsafe behaviors: 2000–2014. Saf. Sci. 92, 173–179.
People’s Republic of China on Safety in Production. Law Press, Beijing. Nieto, A., Gao, Y., Grayson, L., 2014. A comparative study of coal mine safety performance indicators in China and the USA. Int. J. Min. Miner. Eng. 5, 299–313. Niu, H.Y., Deng, J., Zhou, X.Q., Wang, H.Q., 2012. Association analysis of emergency rescue and accident prevention in coal mine. Procedia Eng. 43, 71–75. Park, J., Jung, J., Jung, W., 2016. The use of a process mining technique to characterize the work process of main control room crews: a feasibility study. Reliab. Eng. Syst. Saf. 154, 31–41. Park, M., Song, M., Baek, T.H., Son, S.Y., Ha, S.J., Cho, S., 2015. Workload and delay analysis in manufacturing process using process mining. In: Asia-Pacific Conference on Business Process Management, AP-BPM, vol. 219. pp. 138–151. Pei, J.S., Wen, L.J., Yang, H.D., Wang, J.M., Ye, X.J., 2018. Estimating global completeness of event Logs: a comparative study. IEEE Trans. Serv. Comput. 11, 1–17. Perimal-Lewis, L., Qin, S., Thompson, C., Hakendorf, P., 2012. Gaining insight from patient journey data using a process-oriented analysis approach. In: Australian Workshop on Health Informatics & Knowledge Management, vol. 129. pp. 59–66. Poelmans, J., Dedene, G., Verheyden, G., Herman, V.D.M., Viaene, S., Peters, E., 2010. Combining Business Process and Data Discovery Techniques for Analyzing and Improving Integrated Care Pathways, vol. 6171, 505–517. Rebuge, Á., Ferreira, D.R., 2012. Business process analysis in healthcare environments: a methodology based on process mining. Inf. Syst. 37, 99–116. Rojas, E., Munoz-Gama, J., Sepúlveda, M., Capurro, D., 2016. Process mining in healthcare: a literature review. J. Biomed. Inform. 61, 224–236. State Administration of Work Safety (SAWS), 2017. Online Accident Inquiry System of SAWS. http://media.chinasafety.gov.cn:8090/iSystem/shigumain.jsp. State Council Information Office of the People’s Republic of China (SCIOPRC), 2014. Coal Mine Safety has Always been the Most Important Part of China's Safety Work. http:// www.scio.gov.cn/video/gxsp/Document/1359138/1359138.htm. Smirnov, S., Reijers, H.A., Nugteren, T., Weske, M., 2010. Business Process Model Abstraction: Theory and Practice. Technical report. Song, S.X., Cao, Y., Wang, J.M., 2016. Cleaning Timestamps with Temporal Constraints. In: PVLDB, vol. 9. pp. 708–719. Song, M., van der Aalst, W.M.P., 2007. Supporting process mining by showing events at a
13
