Clustering for Learning from Safety-Related Undesired Events: Application to the Iron and Steel Industry

Safety-related undesired events can cause different kinds of workers' injuries and fatalities. Learning from incidents is a key step in safety risk management, which guides the exploitation of information for implementing effective safety-related decision-making processes. To accelerate the overall process and mitigate the impact of potential human biases, Machine Learning (ML) techniques may be adopted. However, available sources of safety incident reports frequently collect brief unstructured narratives with significant missing data, which are also phrased in a no standardised structure and language. In such a context, relying only on outcomes provided by ML techniques is risky, highlighting the need for human intervention to ensure meaningful results. For such reason, this paper proposes a multi-step approach integrating a hierarchical clustering and subject matter expert evaluations for learning from incidents. The proposed approach has been applied to examine undesired events happened in the iron and steel industry, i.e., one of the most hazardous industries in the world, where a multitude of risks can potentially give rise to a wide range of accidental scenarios. A set of 24 clusters were identified, providing insights into relationships among consequences, number of events, and operating conditions.