Predicting the Elements of Incident Path from Process Safety Reports Using BERT Embedded Feedforward Neural Network

Despite the established efficacy of machine learning (ML) in domains such as predictive maintenance, healthcare, and autonomous systems, its application within the safety domain particularly for incident analysis remains relatively underexplored. One of the persistent challenges faced by industries is the timely and accurate identification of the chain of events and their associated accident pathways, which is critical for enabling proactive accident prevention. One approach is to build a system that extracts information from the historical incident reports and build the chain of events for proactive prevention measures. However, there exist various computational challenges. These safety reports contain rich textual information, but processing and extracting the elements from the reports is time-consuming, error-prone, and not scalable. This study addresses this gap by proposing a transformer-based feedforward neural network architecture for the automated classification of key accident path elements from unstructured textual descriptions in process safety reports. The model focuses on extracting and predicting structured components such as the hazardous element, accident-causing mechanism, category of hazard, initiating mechanism, and risk control system directly from the incident narrative. To evaluate the effectiveness of the proposed model, its performance was benchmarked against several state-of-the-art machine learning algorithms, including Kernel Support Vector Machines (SVM), Multinomial Naive Bayes (MNB), XGBoost, and traditional feedforward neural networks (FFNN). Comparative results demonstrate that the proposed BERT-based model significantly outperforms these baseline models across multiple performance metrics, thereby validating its suitability for this complex classification task.