Learning from undesired events in the iron and steel industry using machine learning techniques

Safety in the iron and steel industry requires utmost attention: it is one of the most hazardous industries in the world, whose activities may expose workers to a wide range of hazards (e.g., the presence of molten metal). In such a context, it is particularly important to learn from what happened in past incidents and accidents to avoid the same dynamics happening again in the future. Machine Learning (ML) techniques could assist in this task because they permit discovering patterns and correlations from existing data and speeding up the identification of the relevant characteristics of undesired events. The scientific literature offers some contributions dealing with the analysis of incidents and accidents in the iron and steel industry through ML techniques, but they are based only on data collected from one or a few specific steel plants. Therefore, the generalisability of their results is limited. For this reason, this paper aims to learn from safety-related undesired events happened in various iron and steel plants by analysing a broader set of incidents and accidents, collected from relevant international data sources (i.e., Analysis, Research and Information on Accident (ARIA), Chemical Safety Board (CSB), and Occupational Health and Safety Administration (OSHA)). A dataset of incidents and accidents was then built, and a set of features and labels was defined to structure this dataset. We used Orange software to train and test well-known ML classification algorithms (e.g., Logistic Regression (LR), Random Forest (RF), k-Nearest Neighbors (kNN), and Artificial Neural Network (ANN)) to predict the type of occurred incident or accident and its degree. From the analysis of the dataset, we obtained that 45% of incidents were categorised as “caught in”, 20% as “struck by”, and 9% as “fall”. The main triggers associated to these types of incidents were “moving object” in 35% of the cases, and “operating machine” in another 35% of the cases.