Abstract: In this study some data mining techniques for accident investigation and risk analysis is proposed. Function of most of accident investigation and risk analysis methodologies have been based on establishment of different scenarios of accident occurrence and simulation of accidents situation and so far no fundamental action for the analysis of remained data from accident has taken place. This study with the approach of data analysis and using different techniques of data mining can eliminate deficiencies of other techniques therewith covers theirs advantages. In this study factor analysis utilized to identify effective factors on occurrence of accidents. Cluster analysis utilized to classify accidents. A case study in a petrochemical company has been done in order to execute and investigate proposed methodology. The results show four different factors effecting on accident`s occurrence and ten different clusters of accidents recognized. Also association rules exposed to discover all patterns and rules that cause occurrence of accidents.
INTRODUCTION
Iran geographic conditions and availability of huge crude oil reservoirs, Iran also has faced with creation and increasing development of oil, gas and petrochemical industries. As a direct consequence of ever-increasing variety of products and processes in these companies, the distances between such industries and human settlements have become lesser and lesser. Obviously growing these industries, the probability of occurring frightening and terrible accidents, treating life and financial capitals will increase. The damage potential of process industries is a cause of concern in many countries.
Strupczewski (2003) has made a comparison between nuclear power plant and oil, gas and other process industries. It was found that the accidents risk due to oil, gas and petroleum accidents are much larger than the risk due to nuclear-power accidents.
Reviewing published literature during years 1998 till 2007 clears that extended attempts for controlling effective factors on incidents of accidents in different countries of world’s have been done. Some methodologies introduced and developed such as the worst case methodology that have been provided by Carter and Hirst (2000), PROTEUS provided by Stam et al. (2000), SCAP provided by Khan et al. (2002), PRISMA provided by Dye and Schaaf (2002), Haswim provided by Reniers et al. (2005) and finally MRRA methodology provided by Hu et al. (2007). More over some attempts for standardization of reporting and follow-up systems after accident taking places have been executed that can mention Berentsen and Holmboe (2004) method.
Function of most of accident investigation and risk analysis methodologies have been based on establishment of different scenarios of accident occurrence and simulation of accidents situation and so far no fundamental action for the analysis of remained data from accident has taken place, even though huge volume of data have been gathered and stored in data centers of HSE databases in petrochemical company. In methodologies used from 1998 till 2007, one case of using factor analysis technique for recognition of effective factors in incidents in a construction site by Sawacha (1998), one case of using cluster analysis for classification of cultural factors effective on accidents by Oltedal and Rundmo (2007), one case of using bayes theory by Trucco et al. (2008) and one case by Meel et al. (2007) have been seen. Although in recent years specially 2007, tendency for using some techniques of data mining such as bayes theory is obvious, but non utilization of data mining techniques and it’s execution for risk analysis and accident investigation both for accident’s pertinent to other sectors such as nuclear industries, chemical industries, food industries, public transportation and air industry is intensely felt.
The goal of this study is a desired reduction of this gap. This research proposed some data mining techniques for accident investigation and risk analysis. Factor analysis utilized to identify effective factors on occurrence of accidents. Cluster analysis utilized to classify accidents. Also, association rules exposed to discover all patterns and rules that cause occurrence of accidents. A case study in a petrochemical company has been done in order to execute and investigate proposed techniques. Results show four different factors effecting on accident’s occurrence and ten different clusters of accidents recognized and discovered patterns and rules have been considered.
MATERIALS AND METHODS
The structure of this study is as shown in Fig. 1. Four steps are involved in the methodology. The first step includes identifying effective criteria on accident occurrence. The second step involves reducing effective criteria by using factor analysis. At the third step cluster analysis was utilized for accident clustering. Finally fourth step presents discovering patterns and rules by using association rules algorithm.
A short description of using techniques
Data mining: Also a non-sensitive and non-infectious skin problem that does
not affect other body organs. An individual with psoriasis usually has patches
of red areas with clear boundaries and layers of data mining is an important
part of information management technology. Simply put, it is a method to extract
and analyze meaningful patterns and correlations in a large relational database
(Frawley et al., 1991).
Factor analysis: Factor analysis is a general term that is given to a group of statistical multivariate methods which their primary goal is to define covert structure in data. In general terms by defining a collection of joint covert dimensions that are referred to as factors, it analysis relations structure (correlation) amongst great volume of variables (Lattin et al., 2003).
Cluster analysis: Cluster analysis is the name of a group of multivariate techniques that their primary goal is to classify things based on their specifications. Clustering is concerned with grouping together objects that are similar to each other and dissimilar to the objects belonging to other clusters. Similarity can be expressed as functions specified by the users or experts. Good clusters show high similarity within a group and low similarity between any two different groups (Lattin et al., 2003).
Association rules: The association algorithm is a basic data mining algorithm and nothing more than a correlation counting engine. This method discovers all possible and interesting patterns in a database. Association rules can be used to discover the relationships and potential associations, of attributes among huge amounts of data.
| Fig. 1: | Structure of methodology |
These rules can be effective in uncovering unknown relationships and provide results that can be the basis for forecasting and decision making (Tsay and Chiang, 2005; Chae et al., 2001).
Case study: According to the literature reviewed between 1998 till 2007, although using of data mining techniques for risk analysis increased nowadays, but most of these methodologies focus in limited aspects or study some cases. The methodology represent in this study try to propose a combined techniques for risk analysis. By using this methodology, researchers able to present a comprehensive view of accidents. For executing and investigating the proposed methodology, a case study in an Iranian petroleum company has been applied. The database contains 275 records of accidents that were registered in HSE department of a petroleum company for the period 1993-2007.
The structure of database is as shown in Table 1.
Step 1: Identifying effective criteria: This research commenced by reviewing the relevant literature on accident investigation and analysis published. This was followed by exploratory interviews which took place with operatives, managers and safety officer in a petroleum company.
Contributing factors to accident occurrence are shown in Table 2.
Step 2: Reducing effective criteria: After the exploratory interviews, as a pilot study questionnaire was designed and discussed with 100 personal. The response to each attitudinal question was measured on a five- point, strongly agree, agree, neither agree nor disagree, disagree, strongly disagree. The data was then analyzed utilizing the statistical computing package SPSS. As a result, the factor analysis technique was applied to reduce the large amount of criteria to a small number of factors.
The factor analysis technique was utilized to help identifying the underlying cluster of factors that dominate safety performance.
| Table 1: | Database structure |
| |
|
| Table 2: | Different criteria |
| |
|
| Table 3: | Results of factor analysis before rotation |
| |
|
The factor analysis on the 20 accident factors and Table 3 shows the factor matrix.
In some criteria, elements of the factor loading matrix for several factors would be close to one another and allocation of those criteria to a special factor would cause loss of some part of data. In this regard, in order to deducing better and better specifying the quantity of factors, factors were rotated based on varimax method.
| Fig. 2: | Screen plot |
| Table 4: | Results of factor after rotation |
| |
|
After factor rotation, allocation of criteria to factor is executed with higher accuracy and less error. We can also understand the number of factor by screen plot (Table 4). This diagram shows number of factors in comparison with eigenvalues. As it is seen in the Fig. 2, elbow point is placed at factor 4.
It is necessary to assign an identifiable name to the group of factors of high correlation coefficients. The researchers named the groups of factors as historical factors, environmental factors, operator behavior, organization and equipment factors. Allocated criteria to factors are shown in Table 5.
Step 3: Accident clustering by using cluster analysis: For attaining the best answer, 15 different structures including 2 step models (cluster-association rules) with consideration of input variables and different prediction variable were devised. Prediction variable with letter P and input variable with letter I has been shown as seen in Table 6.
| Table 5: | Allocated criteria to factors |
| |
|
| Table 6: | Data mining structure with various predict and input variables |
| |
|
From amongst below considered structures, as it follows we will focus of analyzing the best made models in cluster and association algorithms. SQL SERVER 2005 is used to execute proposed methodology.
In this step, the accidents were clustered by using cluster analysis. Two hundred records of accident had been used to build training model. Performing cluster analysis in first structure generates 10 clusters. In a cluster analysis, each object is assigned to precisely one of a set of clusters. The cluster diagram view of SQL SERVER 2005 presents each cluster as a single node. These nodes are scattered across a field and allowed to group based on similarities. It is attention worthy that in cluster analysis; the suitable answer is an answer that samples in a cluster are similar furthest possible and samples belonging to different clusters are separate from each other and not similar. There more thickness of cluster connection lines, the more strengths between clusters.
| Fig. 3: | Similarities between different clusters |
As it is observed in Fig. 3, only between cluster No. 1 and 5, there is weak connection.
| Table 7: | Cluster profiles |
| |
|
In Table 7, specifications of accidents assign to each clusters and its population is shown.
Seventy records of accident have been utilized to investigate the accuracy of clustering model and the model accuracy chart is shown in Fig. 4.
The blue line shows ideal model. An ideal model is a theoretical model that predicts the result correctly 100% of the time and red line is actual model can compare against the results of the ideal model to estimate accuracy of model. Maximum error of actual model as compared with ideal model is 6%.
Step 4: Discovering patterns by using association rules: Association rules show existing patterns in data without considering any special goal. Due to this reason, this algorithm is examples of undirected data mining. Considered structure for analyzing association rules is the model related to structure 12 that from precision point of view, has more accuracy in comparison with other models. Structure of association rules shows probability of incidence of rules and importance of its occurrence. In this research, probability of occurrence 1 and minimum importance has been considered as 0. Below refers to most important rules discovered in above model:
| • | Incase substance directly involved in accident are not toxic and flammable, at the occurrence of accident, national heritage doesn’t occurrence |
| • | Incase occurrence of accident is in transfer or storage, occurrence of accident doesn’t culminate in human death, material loss and community disruption |
| Fig. 4: | Accuracy chart of cluster analysis |
| • | Incidence of accident in the transfer stage without fire and explosion does not culminate in cutting off of community disruption |
| • | In case contamination by toxic takes place in storage, doesn’t culminate in human death |
| • | Incase involve materials in accident are toxic but explosion does not take place, accident does not culminate in human death |
Dependency network diagram pertinent to structure 12 association rules has been shown in Fig. 5.
Seventy five records of accident have been utilized to investigate the accuracy of association rules model and the model accuracy chart is shown in Fig. 6.
Maximum error of actual model as compared with ideal model is 14%.
| Fig. 5: | Dependency network diagram of association rules model |
| Fig. 6: | Accuracy chart of association rules |
CONCLUSION
The aim of this study was to show a new application of data mining techniques in the field of accident investigation and risk analysis where, they have never been used and to show results and benefits. This research commenced by reviewing the relevant literature on accident investigation and analysis published. This was followed by exploratory interviews which took place with operatives, managers and safety officer in a petroleum company. In this case, 20 different effective factors have been found. Then factor analysis utilized to reducing effective factors on occurrence of accidents. Therefore as a pilot study questionnaire was designed and discussed with 100 personal. Four different factors recognized and labeled. A case study in a petrochemical company has been done in order to execute and investigate proposed techniques. Cluster analysis utilized to classify accidents. Results show 10 different clusters of accidents recognized. Also association rules exposed to discover all patterns and rules that cause occurrence of accidents.