In order to realize the torque control accuracy of the electric drive system under the influence of the deviation of permanent magnet supply, the temperature deviation of permanent magnet, the current sensor accuracy deviation, the voltage sensor accuracy deviation, the calibration temperature management deviation and the initial angle detection influence deviation, simulation and test were conducted to obtain the influence factors under the independent action of the above factors, and statistical principle was used to obtain the torque control accuracy range of the comprehensive influence factors, which was tested and verified on the prototype. Test results show that the accuracy of the torque value calculated by this method is within the range of -1.1%~2.8%, which meets the product requirements.

This paper presentes a deep reinforcement learning based energy management strategy for Plug-in Hybrid Electric Vehicle (PHEV) of the THS-III platform. Firstly, a forward simulation model of the vehicle was built using MATLAB/Simulink. Secondly, a Markov process for vehicle energy management and a deep reinforcement learning algorithm were built. Finally, simulation and verification were carried out using WLTC-Class3 and ACC-60. The simulation results indicate that compared with the rule-based energy management strategy, the deep reinforcement learning-based energy management strategy saves 16.51% in cost and 15.56% in fuel consumption under WLTC-Class3, and saves 31.95% in cost and 29.96% in fuel consumption under ACC-60.

In order to improve the recognition accuracy and pre-judgment ability of autonomous vehicles in high-speed dynamic complex traffic scenarios, The lane-changing intention recognition model based on convolutional residual Bidirectional Long Short-Term Memory (BiLSTM) with fusion attention mechanism is proposed. It uses the one-dimensional Convolutional Neural Network (CNN) to extract the vehicle’s motion state features. The constructed feature vector is used as the input information of BiLSTM network. The residual connection is used to solve the problems of optimization bottlenecks and gradient disappearance in multi-layer BiLSTM network. It’s achieved to a adjust the weight of the output of the residual BiLSTM network at different moments with the attention mechanism. And the driving intent probability can be calculated by the Softmax function. The validity of the model is verified by using the expressway data set in NGSIM, the performance and effect of the other 4 models are compared with the model. The results show that the recognition accuracy of the lane-changing intention is the highest, which reaches 97.44%, and prediction accuracy of the vehicle’s lane-changing intention is 90% and higher within 2.5 s before the changing lanes, it shows that the model has better intent recognition accuracy and prediction ability.

For the transient current overloading when the permanent magnet fan for heat pump is powered down and resumed, this paper proposes a sensorless coasting startup strategy based on current resonant tracking control. Utilizing the current suppression control method and the quasi-resonant tracking method, a two-stage starting process was designed including the current-loop suppression tracking and the speed-loop synchronous recovery for motor free coasting condition, to quickly resume the current-speed closed-loop vector control and make the start smooth. The simulation and the testing results showed that in the medium and high-speed coasting, the proposed starting strategy can effectively reduce the steady-state error of the current and improve its dynamic response, and improve the estimation accuracy of the sensorless algorithm for the current tracking process, so as to enhance the smoothness and reliability of the coasting starting process.

To solve the problem that the traditional shift schedule of two-speed transmissions for battery electric vehicles cannot obtain the optimal shift performance under different conditions, this paper proposes a knowledge-based method for shift decision-making. Firstly, the dynamic model of battery electric vehicles was established, the optimal shift data was obtained by dynamic programming. The two-parameter shift schedule was extracted based on support vector machine. Secondly, the data of manual shift was collected to build an exclusive database. The intelligent shift decision model was built based on the long short-term memory network, and the shift decision model was updated online by over-the-air technology. Finally the proposed shift decision method was verified by simulation. The results show that long short-term memory network model has high shift decision accuracy, the proposed knowledge-based shift decision method has better shifting performance than traditional two-parameter shift schedule.

In order to detect the driver’s fatigue driving situation in time, this paper proposed a yawn detection method based on Dlib and variant Transformer. First, the yawn feature matrix of the driver’s eyes and mouth was constructed based on the face key point model of Dlib. Then a variant Transformer model was proposed in the field of video detection, to extract the yawn feature matrix and classify the results. Finally, it was verified based on the YawDD dataset. The results show that the yawn detection accuracy of the proposed algorithm is 96.8%, which is higher than the existing algorithms, and is suitable for the detection of yawning behavior when the driver is fatigued.

For the defects of traditional A* algorithm in unmanned vehicle path planning in structured road scene, such as multiple twists and turns of search path, close to obstacle boundary, unsmooth and exponential growth trend of search time with the increase of grid scale, this paper proposed an improved A* algorithm. Firstly, the map preview module was used to extract the key nodes in the grid map, then the collision field model based on the safe distance was introduced to adjust the cost function. The algorithm conducted incremental extended search based on the information of key nodes until the target node was identified. Finally, the generated path was smoothed by quasi uniform cubic B-spline curve to obtain the final planned path. The simulation results show that compared with the traditional A* and weighted-A* algorithm, the improved A* algorithm proposed in this paper improves the search efficiency, path security and feasibility.

In order to effectively solve the problem of limited amount of downloaded data due to the short travel time of vehicles in the coverage of Road Side Unit (RSU) during high-speed movement, this paper proposed a message transmission strategy of vehicle road cooperation mode based on ant colony algorithm. According to the characteristics that information such as vehicle data can be shared between RSUs, the corresponding heuristic function and the corresponding path pheromone update principle were designed to form multiple vehicle road cooperation communication groups, which increased the amount and types of data transmission in the network and avoid falling into the local optimal solution. SUMO simulation platform was utilized for experimental verification. The results show that, compared with the non-cooperation, Coalition Formation Games (CGS) and Multilevel Hyper-graph Partitioning Based on Heavy Edge Matching Scheme (MHEMs), the proposed strategy is better than the above strategies in terms of information transmission volume, road network revenue and operating time, which proves the effectiveness of this strategy.

For the problems of adaptive cruise control technology, including insufficient environmental adaptability of control algorithm for Deep Reinforcement Learning (DRL), poor model mitigation and generalization ability, this paper proposed the Soft Actor-Critic (SAC) control algorithm based on the principle of maximum entropy and stochastic off-line policy. SAC network was built to fit action value function and action policy function, and auto-adjusting temperature coefficient was used to improve the environmental exploration ability of intelligent agent. For the problem of sparse reward, the reward function was designed by using the idea of reward shaping. In addition, a new experience replay mechanism was proposed to improve the utilization rate of samples. The proposed control algorithm was simulated and tested in different scenes, and compared with Deep Deterministic Policy Gradient (DDPG). The results show that the algorithm has better model generalization ability and migration effect on real vehicles.

Based on the actual operating data of electric vehicles, this paper proposed an analysis method of in-service power battery usage behaviors, so as to quantitatively evaluate the charging behaviors and driving behaviors of vehicles, and provide effective support for battery fault diagnosis. Firstly, characteristic parameters of power battery use behaviors based on membership function were extracted, and then the accumulative risk score of using behaviors was defined and calculated. Finally, the battery use behavior differences between vehicles and vehicles in different time dimensions were quantitatively analyzed by using the idea of horizontal and vertical comparison. Experimental results show that there is a strong positive correlation between the battery using behavior score quantified in this paper and battery pack consistency, which can fully evaluate the using behavior of power battery, and provide data support for battery fault diagnosis.