پیدا کردن الگوهای حوادث شغلی در صنعت معدن با استفاده از روش داده کاوی سیستماتیک

This paper deals with occupational accident patterns of in the Portuguese Extractive Industry. It constitutes a significant advance with relation to a previous study made in 2008, both in terms of methodology and extended knowledge on the patterns’ details. This work uses more recent data (2005–2007) and this time the identification of the “typical accident” shifts from a bivariate, to a multivariate pattern, for characterising more accurately the accident mechanisms. Instead of crossing only two variables (Deviation x Contact), the new methodology developed here uses data mining techniques to associate nine variables, through their categories, and to quantify the statistical cohesion of each pattern. The results confirmed the “typical accident” of the 2008 study, but went much further: it reveals three statistically significant patterns (the top-3 categories in frequency); moreover, each pattern includes now more variables (4–5 categories) and indicates their statistical cohesion. This approach allowed a more accurate vision of the reality, which is fundamental for risk management. The methodology is best suited for large groups, such as national Authorities, Insurers or Corporate Groups, to assist them planning target-oriented safety strategies. Not least importantly, researchers can apply the same algorithm to other study areas, as it is not restricted to accidents, neither to safety.

Since the establishment of the International Labour Organisation (ILO), in the early decades of the 1900 s, the collection of accident data and production of statistical analysis has always been a privileged source of accident information, from where to derive prevention and (international) resolutions concerning safety at work. The aim of this work is to typify patterns of non-fatal accidents in the Portuguese mineral extractive industry, hereafter called simply Extractive Industry. This study continues and extends substantially a previous one by Jacinto & Guedes Soares in 2008 [1]. The novelty in 2008 was the ability to identify accident patterns, especially in the case of the so-called “typical accident”; such patterns were quite accurate and, above all, their relevance in terms of statistical association were clearly quantified at the level of each modality (or category) of the main variable, rather than simply associating the main variables themselves. This ability and the overall methodology applied at that time offered some novelty. In the present work, the authors intend to go further and establish accident patterns, which are even more accurate and also more complete, as they now encompass multiple variables; or best said: include specific categories of multiple variables (i.e., modalities from the main categorical variables). The first study covered the triennium 2001–2003 of Economic Activity – Sector C (Mining & Quarrying; also referred to as Extractive Industry), whilst this one covers 2005–2007. Again, all the data was supplied fist-hand (raw data) to the authors directly by the competent authority, i.e., the Office of Strategy & Planning (GEP), which is the national agency responsible for collecting and coding all data on accidents at work. The motivation for this second study was driven by the fact that availability of accident data is continuously increasing, partly because more countries are implementing the ESAW system (European Statistics of Accidents at Work), defined by the Eurostat in 2001 [2] and the 1998 ILO Resolution [3]. As stressed by Jørgensen in 2008 [4], who was for many years a leading member of the ESAW task-force, the statistical analysis of accidents at work constitutes an essential source of information to support the development of new prevention strategies. Hence, the higher is the availability of data, the higher is the need to explore new techniques and statistical tools for mining “hidden” details, which might help a better understanding of the phenomena; the main assumption is that understanding a phenomena is an essential condition to be able to control it. The novelty in this work lies on two aspects: the multivariate facet of the findings (i.e., much more informative accident patterns) and, equally important, on the data mining approach developed to find such patterns. The referred technique not only permitted more information about patterns, but also enabled measuring (quantitatively) their cohesion.

This work typifies occupational accident patterns. Such patterns have the ability of incorporating multiple variables and, simultaneously, quantifying the statistical strength of each association (each pattern). To attain this purpose, the paper describes the development of a data mining approach that allows multivariate analysis in the field of occupational accidents. In contrast with traditional univariate studies (analysing variables one at a time), or even with bivariate approaches (crossing two variables at a time), the proposed methodology provides a much more detailed and complete characterisation of accident patterns, since each pattern is now composed of multiple categories of factors that are strongly associated from a statistical point of view. An important contribution of this work, thus, is the ability to mine accident patterns not limited to 2 categories/modalities of factors. More informative patterns having 3, 4 or more categories are detected and their cohesion can be assessed. The advantage is the increased completeness and robustness of the resulting accident information, which in turn constitutes a central pillar for decision-making towards prevention and safety learning. The methodology was applied to the Portuguese Extractive Industry, over the triennium 2005–2007, to disclose the most relevant accident patterns in this sector. The current study corroborates the main findings of a previous one in the same sector (for the period 2001–2003), but it now adds relevant information, as the newly applied multivariate analysis allowed a step forward in the characterisation of accidents’ risk in this particular working activity. Each pattern identified here is “unique” and holds two main attributes: (1) it combines information from several main variables and identifies their specific modalities (categories), which typify the pattern, and (2) it allows to measure (quantify) the strength or “positive cohesion” of the entire association that characterises each specific pattern. The gain produced by this approach becomes immediately apparent from the contrast between the generic “typical accident” portrayed in Section 4.1 and the main findings discussed in Section 4.2, which unveil not one, but three distinctive patterns that are much more clear and concise. Not surprisingly, all main “ingredients” (or specific modalities) of the generic portrait are still present within these top-3 more frequent accidents, but the three patterns are significantly different from each other and show their own idiosyncrasies. Knowing the differences and “what” makes them different are vital information for planning prevention. This study used national accident data (ESAW data, compatible within the EU) and this means that only ESAW variables were included in the analysis; thus, the preventive suggestions emanating from the patterns identified are also broadband by nature (macro strategies). To produce more purposeful and tangible recommendations, one would need to incorporate more variables accounting for, say, human factors, technological specificities, and/or management organisational factors. This could be feasible at company level (or at least within Associations or large Corporate Groups) but it is not viable nation-wide, based solely on the ESAW variables, unless the ESAW system incorporates more harmonised variables, as suggested by Jacinto et al. [43]. The down side of this coin, however, is that data collection systems and maintenance of databases are costly and not all the countries can afford such an extra effort. Also noteworthy is that the methodology demonstrated here can be replicated in many other scientific fields, especially those that depend on a multitude of nominal (categorical) data, such as the social sciences (e.g. in “context analysis” of texts) and/or management systems (e.g.: quality attributes based on nominal variables, environmental, etc.). In future work, the authors will explore the possibility of optimising the last phase of the computational algorithm, i.e., adding a complete “automatic search” of the most relevant patterns without any human interaction. However, this apparent benefit must be balanced against the advantage of giving the annalist the flexibility of putting his/her human expertise/perception to make the final selection among the stronger pattern associations.