Abstract Rapid advancements in autonomous driving technology are reshaping the automotive industry, making vehicle intelligence and road safety critical benchmarks of sector-wide progress. However, due to the challenges in obtaining crash data, existing research on the severity and causal mechanisms of autonomous vehicle crashes remains insufficient. Moreover, the interactive effects among factors influencing crash severity are not fully understood. To address this, this study proposes a crash severity analysis method that accounts for interaction effects. This approach holds significant importance for analyzing influencing factors and improving overall traffic safety. This study explores the use of machine learning to model crash severity in autonomous vehicles (AVs) with a hybrid framework combining interpretable machine learning techniques and the Tree-Augmented Naive Bayes (TAN) network. Using real-world crash data from SAE Level 4 ADS vehicles in the AVOID dataset, classification models are employed to predict crash severity and identify significant factor interactions through TAN. The SHapley Additive exPlanations (SHAP) method is applied to improve model interpretability by quantifying feature contributions. The results show that Random Forest (RF) outperforms other algorithms, including XGBoost, K-Nearest Neighbors (KNN), and Support Vector Machines (SVM), achieving high accuracy (0.9759), precision (0.9765), recall (0.9759), and F1-score (0.9730) after applying the SMOTE technique for data balancing. Significant interactions are identified between factor pairs, including speed and pre-collision motion, rear-side collision and pre-collision stopping, and safety belt usage and sunlight conditions. By uncovering the interaction mechanisms behind crash severity, this research offers a key to understanding AV failure modes, paving the way for the development of safer, more resilient next-generation autonomous systems.
Wang et al. (Thu,) studied this question.