Attention Encoder-Decoder Network Based Autonomous Risk Driving Identification for Connected Heavy-Duty Vehicles

Driving style identification plays a crucial role in enhancing traffic safety and reducing fuel consumption. While existing literature primarily focuses on analyzing driving styles in light-duty vehicles using simulation or questionnaire data, research on heavy-duty vehicles, despite their significant contribution to fatal accidents, remains scarce. To bridge this gap, this paper presents an attention encoder-decoder deep framework for the autonomous identification of risk driving in heavy-duty vehicles. The proposed model capitalizes on multi-source heterogeneous data collected under connected environment. High-level representations of driving style are learned through a BiLSTM-based deep architecture. Through the attention mechanism, crucial information for prediction is adaptively focused, thereby generating personalized latent semantic vectors at different time steps. The framework’s evaluation utilizes multi-source heterogeneous driving data from nearly 13,000 vehicles, conclusively demonstrating the model’s superiority over competing approaches in accurately identifying risk driving behaviors in connected heavy-duty vehicles.