Dynamic risk assessment and mitigation effect evaluation of TBM jamming based on BPNN-BN coupling model

Tunnel boring machine (TBM) jamming is a severe hazard that impedes the efficient construction of long tunnels. Existing risk assessment methods are mostly static and fail to quantify the dynamic intervention effects of mitigation measures. To address this limitation, this study proposes a novel model that couples a Back Propagation Neural Network (BPNN) with a Bayesian Network (BN). The BPNN module uses 8 parameters to predict the initial jamming probability P₀. The model achieves a high prediction accuracy with R² = 0.90 and MSE = 0.023 on the test set. This output serves as the prior probability for a pre-constructed BN, which introduces mitigation measures as evidence nodes to dynamically infer risk attenuation under various intervention scenarios. The BN module demonstrates strong inference capability, with a Risk Attenuation Prediction Accuracy (RAPA) of 86.5% and a Measure Recommendation Accuracy (MRA) of 92.3%. In a case study of a plateau soft-rock tunnel, the model predicted a high initial risk, P₀ = 0.82, and effectively simulated the risk reduction process through three-phase mitigation measures, ultimately supporting the successful extrication of the TBM. Compared to static models, the proposed BPNN-BN coupling model provides superior dynamic reasoning, quantifiable intervention effects, and closed-loop optimization, offering a reliable decision-support tool for TBM jamming risk management in complex geological conditions.