Towards Machine Learning Application for Safety in Confined Spaces: Creating an Incident Database

In Occupational Safety and Health (OSH), confined spaces represent high-risk working areas, where fatal and nonfatal serious incidents frequently happen. Reports of such incidents are a key resource to improve safety because they permit learning the contributory factors and causes of their occurrence. Classical statistical analyses have been conducted over the years, but, to the best of our knowledge, only one study applies advanced data-driven approaches based on Machine Learning (ML) techniques for learning from confined space incidents. These techniques can extract relations in data and predict incident outcomes. In such a context, this paper addresses the first steps of a ML-based approach for learning from confined space incidents: (1) the collection and analysis of data recorded in national and international databases (e.g. US Occupational Safety and Health Administration and Italian National Institute for Insurance against Accidents at Work), and (2) the creation of a single incident database suitable for the ML application (dataset). Our dataset includes 5,054 incidents occurring from 1984 to 2021 in confined spaces and causing workers' fatalities and/or severe injuries. It is structured through 20 features related to relevant dimensions for the analysis, e.g. confined space type, performed task, number of workers fatally and non-fatally injured, direct cause, existing controls. Starting from the dataset, the application of possible supervised ML algorithms able to support risk management in confined spaces is proposed.