Short-term Driving Style Recognition Based on Improved Deep Clustering in High-speed Scenarios

The recognition result of the short-term driving style of the target vehicle is the key to formulate the warning information and decision-making task. Aiming at the problem that the style features extracted by current research are difficult to accurately express driving style, this paper designs the frequency and intensity based feature indicators of dangerous driving behavior; aiming at the problem that the traditional feature dimensionality reduction method compresses the driving data indiscriminately, this paper designs the feature attention mechanism and the temporal attention mechanism to improve the weight of the key features and the momentary data and combines them with the BiLSTM autoencoder to form the ATBiLSTM autoencoder to extract the temporal feature encoding; in order to improve the recognition effect, the ATBiLSTM autoencoder is combined with the improved deep clustering algorithm to form IDEC-ATBiLSTM, which is able to optimize the feature encoding and clustering results at the same time during the training process. The results of the study show that (1) the algorithm categorizes 4764 drivers with a total of 15,701 driving segment samples into aggressive, cautious, and normal types. After the characterization, the model's short-term driving style recognition results for each driving segment of the drivers are in line with the empirical cognition, which is reasonable and reliable; (2) The reconstruction loss of IDEC-ATBiLSTM designed in this paper decreases by 61.62% and 35.94% compared with IDEC-LSTM and IDEC-BiLSTM, which indicates that the ATBiLSTM autoencoder can reconstruct the original data better and improve the quality of feature encoding; (3) The IDEC-ATBiLSTM algorithm can effectively reduce the clustering loss, and the CH value is improved by 27.75% and the silhouette coefficient is improved by 16.42% compared with ATBiLSTM+ K-Means++ without deep clustering, and the clustering effect is improved obviously. This study has important reference value for the recognition of drivers' short-term driving style.