SVNT-DKB: A Data-Knowledge-Bayesian Integrated Safety Evaluation Framework for Autonomous Driving Systems

The autonomous driving system (ADS) faces significant challenges in safety verification: due to the high cost and inherent risks of large-scale real-world environment testing, safety-critical data in the real world is still scarce. This makes traditional data-based methods ineffective. To address this gap, this paper proposes the novel safety evaluation framework SVNT-DKB, which integrates data-based learning, knowledge-guided modeling and Bayesian probability reasoning structurally, resolving the contradiction between small sample limitations and cross-layer risk modeling. The core innovations of this framework include: (1) a four-dimensional hierarchical SVNT model (system-vehicle-network-traffic) based on system safety principles, capable of capturing causal risk propagation between the technical and environmental layers; (2) the DKB fusion strategy using Dempster–Shafer (DS) evidence theory, integrating multiple expert knowledge and reducing small sample bias; (3) an improved NewBIC scoring function with an adaptive scaling factor, capable of dynamically balancing expert knowledge and data reliability. Comprehensive experimental verification shows that SVNT-DKB has significant advantages in three aspects: (1) small sample robustness: with only 30 training samples, its overall accuracy reaches 86.7%, which is 33.4% higher than BPNN and 13.4% higher than SVM and the accuracy in high-risk scenarios is 86.3%; (2) industrial feasibility: compared with ANSYS Medini Analyze, it reduces data preprocessing time by 66.7% (10 ± 2[Formula: see text]ms versus 30 ± 5[Formula: see text]ms), increases system-level risk path coverage by 27% (92% versus 65%) and keeps diagnostic delay at 95 ± 3[Formula: see text]ms, meeting the real-time requirements of ISO 26262; (3) scenario generalization: it maintains reliable performance in urban intersections, highways and rainy conditions and the risk distribution is consistent with the accident statistics in the real world. These findings confirm the excellent performance of SVNT-DKB in small sample robustness, interpretability and real-time performance, effectively bridging the gap between limited data and the comprehensive safety guarantee requirements of ADSs.