Hierarchical assessment method for vehicle accident risk classification based on temporal misalignment causal variables and sub-objective training

In response to the problem of low assessment accuracy brought by the existing model using the relationship of "historical factors-historical accidents" to replace the relationship of "historical factors-future accidents", a hierarchical evaluation method was proposed for vehicle future accident risk oriented toward traffic management practice. Based on massive static traffic management data, characterization indicators were designed and determined around vehicle attributes, historical violations, and historical accidents. Historical static indicators were mismatched with future accident risk to construct a temporal misalignment causal dataset. Coupled with the multi-layer and multi-dimensional needs of accident risk prevention and control from traffic management department, a sub-objective training method was applied. Weighted samples were used to train high-, medium-, and low-accuracy sub-models. Logical rules were then integrated to output four risk levels: high, medium, low, and extremely low. Analysis results show that the temporal-misaligned modeling approach significantly improves the recognition performance compared to its temporal-aligned counterpart. The designed fusion model exhibits excellent overall performance across five vehicle types. The accuracy rate and recall rate for each vehicle type are 78.78% - 93.80% and 72.01% - 93.98%, respectively, while the precision of very-low-risk identification exceeds 97%. A graded change in prediction precision across risk levels meets the requirement for graded identification of vehicles with different accident risks. The model demonstrates good robustness. When migrated and applied to different years and specific vehicles, the model has an overall accuracy rate of 82.71% and an overall recall rate of 93.67%. Although recognition performance is somewhat affected, the final results still adequately meet practical requirements. In particular, the model's identification precision for extremely low-risk vehicles is maintained at a high level of 99.60%. Therefore, the construction of the temporal misalignment dataset and the sub-objective training method effectively improve the recognition performance of vehicle future accident risk, and provide a feasible technical path for traffic management department to carry out hierarchical and graded risk prevention and control.