Lessons from past hazardous events: data analytics for severity prediction

An increasing amount of information is collected by the monitoring systems within the process industry, especially concerning safety management. For instance, the Seveso III regulation on the control of major-accident hazards involving dangerous substances is the first version that refers to the collection of safety indicators for monitoring the performance of safety management systems. This leads to a call for improvement in learning past lessons and definition of techniques to process relevant data, in order to deal with unexpected events and provide the right support to safety management. Through this work, we suggest a data analytics approach for severity prediction of future hazardous events. The approach is twofold and is based on the use and comparison of multiple linear regression (MLR) and deep neural network (DNN) models. These models are developed and tested on the Major Hazardous Incident Data Service (MHIDAS) database. A set of simulations has been carried out not only to evaluate the models, but also to identify their limitations. The results show the capability of these models to manage heterogeneous data from past accident records and extract important information to support safety-related decision making. It must also be mentioned that intrinsic model limitations should be considered, and appropriate model selection and customization should be carefully carried out to deliver the needed support.