Modeling expert risk assessments in utility tunnels with deep learning

In urban operations, public utility tunnels are exposed to numerous risks, necessitating effective risk assessment strategies. Existing semi-automated risk assessment models are constrained by the need for expert involvement, leading to limited response speeds, while fully automated models lack generalization capabilities due to scarce historical accident data. This paper focuses on merging the strengths of both models by developing a machine learning-based, fully automated risk assessment model capable of learning from historical data and simulating expert knowledge and assessment processes. To tackle the challenges of subjectivity, heterogeneity, and group decision-making in simulating expert decision processes, our proposed framework utilizes BNNs, self-attention mechanisms, and GCNs to effectively model the complexities of group decision-making. This study employs data from 13 real-world scenarios for experimental analysis. Quantitative comparisons with baseline models show that our framework significantly reduces RMSE and MAE by 25.9 % and 24.4 %, respectively, demonstrating considerable performance advantages. The findings reveal: 1) Bayesian Neural Networks effectively manage uncertainty and subjectivity, thereby enhancing the accuracy of simulating expert decisions; 2) Self-attention mechanisms successfully emulate the heterogeneity of experts, recognizing decision-making preferences across different backgrounds; 3) Graph Convolutional Networks proficiently facilitate the simulation of group decision dynamics, capturing intricate interactions among experts in detail.