Human risk prediction and key factors identification for advanced operation systems based on Attention-STGRU model

Polymorphic and interactive information clusters increase the complexity of human cognition in complex operating scenarios, and human errors are more significant than before. The use of traditional HRA methods has limitations in integrating multiple and heterogeneous scenario data and predicting large-scale events. Therefore, this study proposes a method for predicting human errors in sparse time events based on deep learning algorithms. First, a human error probability prediction network based on the Attention-STGRU model is proposed, which consists of a variant STGRU model of a sequence model (GRU) and an attention mechanism. It has accurate network training efficiency and memory simulation ability for predicting events with irregular time characteristics. In addition, a Granger Causality inference method based on the Attention-STGRU model is proposed. This method can identify the key factors affecting risk tasks and effectively learn the nonlinear and dynamic features contained in the scenario information data, thereby forming a more reliable causal relationship structure. Our results show that the model proposed in this study has the performance of enhancing the influence of time and characterizing the attention characteristics of the brain.