Self-Organizing Map and clustering algorithms for the analysis of occupational accident databases

Data mining techniques are a powerful method for extracting information from large databases. Among these techniques, clustering and projection of data from high-dimensional spaces hold a main role, since they allow to discover hidden structures in the data set. Following this approach, this paper presents a data analysis method designed to help the management and investigation of occupational accident databases. The purpose is to discover the most common sequences of events leading to accidents for devising preventive actions. To this aim, we developed a two-level approach based on the joint use of the Kohonen's Self-Organizing Map and the k-means clustering algorithm. This approach allows not only to group the accidents in different classes but also to visualize them in a way understandable for the analyst. The method has been applied with satisfactory results to a large database of occupational accidents occurred in the Italian wood processing industry. A comparison with the Hierarchical Clustering method confirmed the effectiveness of the proposed approach.