Risk assessment of utility tunnels through risk interaction-based deep learning

Due to the nonlinear, dynamic and multirisk coupling characteristics of utility tunnels, it is important to construct a continuous risk assessment method that can consider risk interactions. This paper proposes a model that incorporates uncertain parameters into graph convolutional neural networks. The model can use the convolutional calculation of topological graphs to learn the relationship among risk indicators, interaction strength and risk level and can obtain the ability to infer the risks not covered by the training set. The introduction of uncertain parameters to the complex interaction computation process can mitigate the negative effects of overfitting and domain bias. The method is evaluated using data from 13 different real utility tunnels in Beijing, China, and the experimental results show the superiority of the model proposed in this paper. This study provides not only a feasible and effective method to construct a risk assessment system for the risk control departments of utility tunnels but also a reference for constructing continuous dynamic risk assessment models for similar engineering scenarios.