Unsupervised Identification of Driving Styles from Naturalistic Driving Data Through a Context-Normalized Framework

Identifying driving styles is essential for personalizing driving assistance systems and enhancing intelligent transportation services. However, existing approaches predominantly rely on experience-driven feature engineering and annotated data, limiting objectivity and hindering the exploitation of unlabeled naturalistic driving datasets. To address these limitations, this paper proposes an unsupervised framework for driving style identification from naturalistic driving data through context normalization. A Constrained Convolutional Autoencoder (CCAE) integrated with a global self-attention mechanism is developed to map unlabeled driving sequences onto a standardized dynamic reference defined by the Worldwide Harmonized Light Vehicles Test Cycle (WLTC). This process extracts Driving Adaptability Characteristics (DACs) as WLTC-anchored latent representations that characterize normalized driver-specific response patterns across heterogeneous naturalistic contexts. To ensure feature robustness, frequency-domain refinement is applied to eliminate high-frequency noise. The extracted DAC sequences are subsequently partitioned into distinct driving styles using a kernel-mapped clustering algorithm. To evaluate the external relevance and physical interpretability of the identified styles, actual vehicle accident records and raw CAN-bus feature backtracking are introduced as validation evidence. The results show that the identified driving styles exhibit different historical accident probabilities. The proposed CCAE model achieves clearer cluster-level differentiation than traditional feature engineering and unconstrained deep learning models, and the ablation analysis confirms the contribution of the WLTC-based constraint. These findings indicate that the context-normalization framework can extract interpretable and externally relevant driving style representations from unlabeled naturalistic data.