Bayesian Deep Reliability Network for dynamic human reliability analysis in high-risk industrial environments

Abstract Industrial and mining accidents continue to pose significant threats to worker safety and production operations worldwide, with human error accounting for a substantial proportion of high-risk incidents. This paper addresses three critical challenges in contemporary Human Reliability Analysis (HRA): the inability of traditional methods to capture dynamic human-machine-environment interactions, the difficulty in integrating heterogeneous multi-source data, and the lack of interpretable uncertainty quantification in deep learning predictions. To tackle these issues, we propose the Bayesian Deep Reliability Network (BDRN), a novel framework that integrates a Graph Attention Temporal Encoder (GATE) for spatiotemporal dependency modeling, a Multimodal Feature Fusion Network (MFFN) for heterogeneous data integration, and an Evidential Deep Reinforcement Learning (EDRL) module for uncertainty-aware decision support. The proposed model is validated on the Mine Safety and Health Administration (MSHA) accident dataset and the New York City Motor Vehicle Collisions dataset. Experimental results demonstrate that BDRN achieves superior performance with an accuracy of 0.892, F1-score of 0.867, and AUC-ROC of 0.921 on the MSHA dataset, numerically higher than traditional Bayesian network HRA models, LSTM-based sequence models, and Transformer-based classifiers. Furthermore, the uncertainty calibration error of 0.068 confirms the model’s reliability in quantifying decision confidence. This research bridges the gap between probabilistic reasoning and deep learning, providing a practical tool for safety decision-making in high-risk industrial environments.