Abstract: To address the difficulty of accurately identifying driving behavior preferences in highway scenarios, vehicle trajectory data from NGSIM US-101are used in this paper.The data features including speed, acceleration, jerk, time headway, spacing, and lane-changing behavior are extracted. A three-class labeling method for driving behavior preferences, namely cautious, normal, and aggressive, is constructed by combining the density-based spatial clustering of applications with noise (DBSCAN) algorithm with an aggressiveness score. On this basis, a bidirectional long short-term memory (BiLSTM) identification model integrated with an attention mechanism is established to classify driving behavior sequences. The results show that, among 2 490 valid vehicles, the proportions of the three classes are 26.1%, 44.7%, and29.1%, respectively. Significant differences are observed among the classes in terms of speed level, longitudinal control intensity, and car-following safety margin. The proposed model achieves an accuracy of 90.25% in driving behavior preference identification. This method provides a reference for behavior prediction, decision-making and planning, and safety control of intelligent connected vehicles.

Key words: transportation engineering, driving behavior preference, DBSCAN algorithm, attention mechanism, BiLSTM model, supervised learning