Prediction and Analysis of Expressway Tunnels Crash Based on M-CNN and SHAP Techniques

The prediction and analysis of traffic crashes in expressway tunnels plays a pivotal role in enhancing tunnel safety. This study introduces the modified Convolutional Neural Network (M-CNN) for tunnel traffic crash prediction. The Synthetic Minority Oversampling Technique (SMOTE) is utilized to address the issue of imbalanced crash data. Based on the prediction results, this study identifies sections of high risk in tunnels and utilizes SHapley Additive Explanations (SHAP) to enhance the interpretability of M-CNN. The results demonstrate that the prediction accuracy of M-CNN is significantly higher at 74.62%, surpassing the baseline models including Convolutional Neural Network (CNN), back propagation neural network (BPNN), random forest (RF), long short-term memory (LSTM) and support vector machine (SVM). Moreover, zone 1, zone 2, zone 6, and zone 7 are identified as tunnel risk zones. In addition, driver’s operation, tunnel grade and vehicle speed have the greatest impact on rear-end crash, sideswipe crash, hit-guardrail crash respectively. This study also reveals intricate interaction effects between the variables and (Skidding resistance index) SRI exhibits a negative correlation with crash risk. The research findings have significant implications for the future implementation of machine learning models in crash studies, with practical applications for reducing crash rates.