In order to solve the problem of unstable cooperative control caused by unsatisfactory Vehicle-to-Vehicle (V2V) communication environment in the process of vehicle formation lane change, a cooperative control method combining Deep Deterministic Policy Gradient (DDPG) and Model Predictive Control (MPC) is proposed. Firstly, a non-ideal V2V communication model including delay, noise interference and random interruption is constructed based on the wireless channel fading theory. Then, the multi-agent DDPG algorithm is optimized, and a multi-dimensional reward function including communication quality index, lateral tracking deviation and other factors is designed to output the future reference trajectory under collaborative decision-making. Finally, smooth steering angle and acceleration control sequences are generated and executed based on MPC. The simulation results show that compared with the traditional method, the formation spacing error and lateral tracking error of the proposed method are reduced by 59.11% and 48.81% respectively in the communication delay environment. In the scenario of communication interference and interruption, the lane change success rate of this method is increased by 16.73% at most, which can maintain the stability of vehicle formation and realize safe lane change.

To address the limited generalization ability of rule-based methods in long-tail scenarios, as well as the overfitting issue and poor interpretability of data-driven models in autonomous driving, this paper proposes an autonomous driving framework, based on a Large Language Model (LLM) and Deep Deterministic Policy Gradient (DDPG). The framework integrates commonsense knowledge with environmental information to generate interpretable driving decisions, while the DDPG algorithm is employed to achieve low-level continuous control. In the memory module, a hybrid retrieval mechanism that combines numerical-feature similarity and semantic-feature similarity is designed to match similar historical cases in real time, thereby supporting dynamic decision-making. Experimental results show that compared with Rule-Driving and DQN-Driving, the proposed method achieves relative improvements of approximately 33% and 14% in success rate, respectively, in heterogeneous test Scenario A. In heterogeneous test Scenario B, the relative improvements further increase to approximately 70% and 42%, demonstrating stronger cross-scenario generalization capability and environmental adaptability.

To address insufficient feature representation and poor robustness caused by viewpoint changes, occlusions, and visual similarity, this paper proposes an adaptive local partitioning Vision Transformer for vehicle re-identification. Firstly, an Adaptive Local Partitioning (ALP) module is designed to perform dynamic regional division. Secondly, by integrating low-level visual details with high-level global semantics through a Hierarchical Attention Fusion (HAF) module, it provides feature guidance for adaptive region segmentation. Finally, a Multi-Feature Embedding (MFE) module is introduced to enhance discriminability using a dynamic weighting mechanism for camera and viewpoint information. The experimental results show that in the vehicle re-identification task, the proposed method achieved an mAP of 81.0% and a Rank-1 of 97.1% on the VeRi-776 dataset, and a Rank-1 of 80.2% on the VehicleID dataset, significantly improving the model's identification accuracy and robustness.

To enhance the accuracy and adaptability of electric heavy-duty truck speed prediction under various operating conditions, a speed prediction method based on information fusion from vehicles-road-traffic is proposed. A real-road traffic simulation model is established using SUMO to acquire vehicle and road traffic data. Principal components analysis is applied for data dimensionality reduction, with input features for the speed prediction model selected based on contribution rates. A speed prediction algorithm integrating Radial Basis Function-Linear Regression is designed, utilizing a sliding time window to update the prediction model. Joint simulation results using MATLAB and SUMO demonstrate that, compared to using vehicle speed information alone, the proposed method reduces the mean root-mean-square error and mean absolute error in speed prediction by 39.86% and 40.48%, respectively; compared to backpropagation neural networks, the average reduction is 40.30% and 41.84%; compared to generalized regression neural networks, the average reduction is 31.78% and 35.28%. Through Hardware-in-the-Loop simulation, the deployability and real-time capability of the vehicle speed prediction method are further validated.

To address the issue of attenuation and waveform distortion in laser echo signals, laser echo signal model is constructed. Lightweight CNN-Transformer hybrid network is built to extract both local details and global temporal features of the signal, and a dynamic adjustment mechanism is designed to adaptively fuse local and global features. Furthermore, cognitive enhancement module is introduced, which utilizes an external memory bank to store and retrieve prior information from historical echoes, and optimizes feature inference through knowledge distillation loss. The results indicate that the attenuation of the laser echo is positively correlated with the severity of adverse weather conditions. Through distillation learning, the student network fully retains the core feature extraction logic of the teacher network. The proposed Dynamic Cognitive Enhancement Network algorithm reduces Mean Squared Error (MSE) by 40.2% and improves Peak Signal to Noise Ratio (PSNR) by 2.21 dB compared to the suboptimal model.

Aiming at the problem of large torque ripple during the operation of high-speed motors, this paper starts from the perspective of motor structure optimization, takes key geometric parameters such as stator slot width, permanent magnet thickness, sheath thickness and pole arc coefficient as optimization variables, and aims to minimize torque ripple and maximize average torque as optimization goals, establishing a high-precision mapping model between structural parameters and motor performance. Firstly, the Hodrick-Prescott decomposition method is adopted to separate the trend and periodic components of the torque data in order to extract the key features; Secondly, a BiLSTM deep learning model integrating the cross-attention mechanism is constructed to characterize the nonlinear characteristics of the motor. On this basis, the improved butterfly optimization algorithm is introduced to optimize the model hyperparameters and obtain the optimal combination of structural parameters. Finally, the effectiveness of the proposed method is verified through simulation and experiments. The results show that this method can significantly reduce the torque pulsation of the motor and improve the smoothness and reliability of high-speed motor operation.

To address the stringent robustness requirements of electromagnetic compatibility (EMC) test equipment for automotive steering systems in complex environments, this paper proposes a novel active disturbance rejection control strategy that integrates fuzzy control with a super-twisting sliding mode algorithm. Building upon the linear active disturbance rejection control framework, the super-twisting sliding mode algorithm is incorporated into both the extended state observer and the error feedback control law, thereby enhancing the system's dynamic response and disturbance rejection capability. Furthermore, fuzzy control is employed to dynamically optimize key parameters in real time. Matlab simulations demonstrate that the proposed method achieves faster response and stronger anti-interference performance compared to traditional linear active disturbance rejection control. Additionally, to cope with complex EMC environments, an improved power filter is designed to suppress noise. Finally, EMC anechoic chamber tests validate that the developed equipment exhibits high testing accuracy and strong robustness, meeting the testing requirements for vehicle steering systems.