Developing a deep reinforcement learning model for safety risk prediction at subway construction sites

Abstract

Underground construction work is heavily affected by surrounding hydrogeology, adjacent pipelines, and existing subway lines, which can lead to a high degree of uncertainty and generate safety risk on site. In order to overcome rigid thinking of causal factors within a structured framework and incorporate features of different accidents, this study adopted grounded theory for the investigation on factors contributing to workplace accidents in subway construction. The deep reinforcement learning model of double deep Q-network (DDQN) was developed for predicting subway construction safety risk, which integrated the advantage of reinforcement learning in decision making with the advantage of deep learning in objection perception. The findings denoted that DDQN performed better than other machine learning models inclusive of random forest, extreme gradient boosting, k-nearest neighbor, and support vector machine. Contributing factors relevant to subway construction accidents were quantitatively analyzed using permutation importance of attributes. It was beneficial for determining how the 37 contributing factors had negative effects on subway construction safety risk. Safety measures for risk reduction and controlling could be optimized according to permutation importance of individual contributing factor, which paved a new way for the promotion of safety management performance at subway construction sites.