Could vehicles analyze driving risks using human fuzzy semantic logic? A data-knowledge-driven new perspective

Accurate risk identification is crucial for ensuring the safe operation of Host vehicles (HoVs) in environments shared with Neighboring vehicles (NeVs). Traditional risk identification mechanisms typically rely on large amounts of precise numerical data, making it difficult to comprehensively and accurately identify traffic crash risks under conditions of imperfect data associated with fuzzy information. However, human drivers rely on knowledge-driven, subjective assessments using fuzzy descriptors like distance and speed semantics to evaluate driving risk. These insights provide significant value for addressing the limitations of precise data-driven methods. This study proposes a novel traffic crash risk analysis framework called Token Tree Generation and Parsing (TTGP). It integrates knowledge-driven insights from human drivers with data-driven methods. TTGP includes the Token Tree Generation Module (Module 1) and the Token Tree Parsing Module (Module 2). In Module 1, we apply the token-tree-of-thoughts method to transform natural language traffic regulations and vehicles’ traffic behaviors and attribute parameters into token tree based on semantic rules. This module simulates the generation of human fuzzy semantics in traffic scenarios. In Module 2, we integrate three encoders and decoders to extract traffic crash risk semantic features and identify traffic crash risk level from the digitized token tree. Experiments in the highway and urban expressway interweaving areas demonstrate that TTGP can accurately analyze risk using imprecise data. The TTGP performs better than traditional methods such as Tree, Naïve Bayes, RUSBoost and Efficient Logistic Regression models. This study significantly enhances the flexibility, generalization, and reliability of risk assessment. It bridges the gap in how HoVs handle fuzzy information in risk analysis.