Learning From Major Accidents: A Meta-Learning Perspective

Learning from the past is essential to improve safety and reliability in the chemical industry. In the context of Industry 4.0 and Industry 5.0, where Artificial Intelligence and IoT are expanding throughout every industrial sector, it is essential to determine if an artificial learner may exploit historical accident data to support a more efficient and sustainable learning framework. One important limitation of Machine Learning algorithms is their difficulty in generalizing over multiple tasks. In this context, the present study aims to investigate the issue of meta-learning and transfer learning, evaluating whether the knowledge extracted from a generic accident database could be used to predict the consequence of new, technology-specific accidents. To this end, a classi-fication algorithm is trained on a large and generic accident database to learn the relationship between accident features and consequence severity from a diverse pool of examples. Later, the acquired knowledge is transferred to another domain to predict the number of fatalities and injuries in new accidents. The methodology is eval-uated on a test case, where two classification algorithms are trained on a generic accident database (i.e., the Major Hazard Incident Data Service) and evaluated on a technology-specific, lower-quality database. The results suggest that automated algorithms can learn from historical data and transfer knowledge to predict the severity of different types of accidents. The findings indicate that the knowledge gained from previous tasks might be used to address new tasks. Therefore, the proposed approach reduces the need for new data and the cost of the analyses.