SS2Rep: Driving Style Recognition for Vehicles Based on a Dual Encoder Architecture

With the continued growth in global vehicle ownership, road safety has become an increasingly pressing issue. Driving behavior recognition, as a key technology for enhancing traffic safety, optimizing vehicle performance, and improving the driving experience, has garnered more and more attention. Different drivers face varying levels of traffic accident risk and fuel consumption patterns due to their unique driving habits. Therefore, accurately identifying driving styles not only helps in developing personalized driver assistance systems but also provides insurance companies with more precise risk assessment data, thus promoting the further development of intelligent transportation systems. This study proposes a sensor signal-to-feature representation model (SS2Rep) based on a dual-encoder architecture, which effectively integrates multi-sensor data using deep learning methods to improve the accuracy and robustness of driving style recognition. First, a three-layer convolutional network is used to convert the time segments of each sensor into token representations. These are then encoded into target representations by the target encoder and multi-modal perceptual information is constructed using the context-aware encoder. Finally, a dynamic cross-attention mechanism is employed for feature integration. Experimental results show that, compared to other advanced models such as ResNet, Transformer, ViT, RNN, LSTM, and SVM, SS2Rep demonstrates higher accuracy and stronger generalization capabilities in handling driving style classification tasks under complex road conditions, proving its potential value in a variety of application scenarios.