A solution was proposed to address the issue of low generalization ability of diagnostic models for electric vehicle power battery faults caused by sparse data, which utilized a data augmentation method based on Generative Adversarial Networks (GAN). According to the augmented data, a fault diagnosis scheme was designed using the Random Forest (RF) model combined with the Bayesian Optimization (BO) method to form a GAN-RF-BO battery fault diagnosis framework. The proposed fault diagnosis approach was compared with the common Multilayer Perceptron (MLP), Support Vector Machine (SVM) and Gradient Boosting Decision Tree (GBDT) model on a real fault dataset. The results show that the accuracy of the proposed method is improved by 19.66%, 19.71% and 16.31% compared to the MLP, SVM, and GBDT models respectively. The GAN-RF-BO framework can better utilize sparse data to troubleshoot problems with power batteries.

Based on the multi-point preview model, a trajectory tracking model optimized by Grated Re-circulated Unit (GRU) neural network was designed to improve trajectory tracking accuracy of intelligent vehicle. Firstly, on the vehicle’s 2 degree of freedom model, 3 preview trajectory tracking models were established based on the preview theory. The simulation results show that the multi-point preview model has the best tracking effect. Then, the parameters such as multi-point preview lateral displacement deviation and steering wheel angle were used as the inputs of GRU neural network. After training, the optimized steering wheel angle was as output to control the driving direction of the vehicle. The verification results show that the trajectory tracking model optimized by GRU has better tracking effect under double shift path and S-curve path.

For the active obstacle avoidance of vehicle under different road conditions, this paper proposed an obstacle avoidance path planning method in driving risk field considering road adhesion coefficient. Firstly, the driving risk fields including road boundary risk field, target gravitational field and obstacle risk field were established. The road adhesion coefficient was estimated in real time based on the volumetric Kalman filter algorithm, and the driving risk field function considering the road adhesion coefficient was derived with negative gradient derivative, and the obstacle avoidance path with the lowest risk was obtained. Then, the obstacle avoidance reference path satisfying the vehicle constraints was obtained by 5-degree polynomial fitting optimization. Finally, the model predictive control algorithm was utilized to track the obstacle avoidance path. The simulation results show that at the same speed, the smaller the road adhesion coefficient, the smaller the lateral acceleration is, the smaller the standard deviation of the lateral acceleration is, and the more stable the obstacle avoidance effect will be.

To solve the problem that the path tracking controller has a large tracking error under different road adhesion coefficients and vehicle speeds based on traditional Model Predictive Control (MPC), this paper proposed a path tracking control strategy for autonomous vehicles based on Particle Swarm Optimization (PSO)-BP neural network optimization MPC. Firstly, a path tracking controller based on MPC was designed; Secondly, PSO-BP was used to optimize MPC, and the controller accuracy and vehicle stability were taken as evaluation functions to obtain the offline optimal time domain parameters of PSO. Finally four conditions were selected for comparison and simulation verification of double shift lane tracking. The results show that the lateral control accuracy of double shift lane tracking under the 4 conditions, including low adhesion at low speed, high adhesion at low speed, high adhesion at high speed and medium adhesion at medium speed, is improved by 50%, 55%, 9% and 20% respectively.

By summarizing the laws, regulations and standards formulated in China and foreign countries during the development of intelligent and connected vehicles, this paper sorted out the standardization construction system of intelligent and connected vehicles, which included 4 parts: admission, term definition, testing system and accident liability of intelligent and connected vehicles. Through the interpretation of standards, the understanding of the evolution of relevant laws and regulations in the process of the transformation of traditional vehicles to intelligent and connected vehicles was deepened, the technical requirements and social problems faced by the current intelligent and connected vehicle industry were clarified.

The existing passive safety related standard restricts the development of new intelligent vehicle technologies. To solve this problem, this research team developed China construction scheme of intelligent vehicles passive safety standards and the development roadmap according to the revision of the relevant passive safety standards of intelligent vehicles by the international standard organizations and the American standards institution, and then took one national mandatory standards as an example. This research can provide necessary support for the applicability developing and revising China’s passive safety standards.

In order to realize the accurate prediction of the trajectory of surrounding vehicles, a driving intention recognition and trajectory prediction model based on graph neural network and Gated Recurrent Unit (GRU) was designed by using deep learning method. The driving intention recognition model constructed the interaction relationship between vehicles into a space-time graph, used the graph neural network to learn its interaction rules and used Softmax function to calculate the probability of different driving intentions. The trajectory prediction model adopted an encoded-decoded GRU network and the encoder encoded the vehicle history trajectory information and fused the recognized driving intention information and then realized trajectory prediction through the decoder. Finally, the Next Generation Simulation (NGSIM) dataset was used to train and verify the model and the results show that the proposed model can better identify the driving intention of the vehicle, and the vehicle trajectory prediction model considering the driving intention can effectively improve the prediction accuracy.

In order to improve the accuracy and robustness of intelligent vehicle path tracking, a robust feedback path tracking controller was designed based on Lyapunov Stability Theory. By using Schur complement lemma and solving Linear Matrix Inequality (LMI), the feedback matrix of the control system was obtained. By building a CarSim/Simulink joint simulation platform, the robust feedback controller was compared with the Linear Quadratic Regulator (LQR). The performance of the designed controller was verified by simulation tests on the roads with different adhesion coefficients at different speeds. The simulation results show that, compared with LQR controller, the robust feedback controller based on Lyapunov stability theory not only has higher control accuracy, but also has stronger robustness when the vehicle is running on different roads at different speeds.

This paper elaborated the overall framework of commercial vehicle fatigue early warning system, and analyzed in details the research status of monitoring system, human machine interface and fatigue detection method. The paper pointed out that the future monitoring system should have the ability of high stability, short delay and massive data processing, and divided the warning into two levels, and defined the effective human-computer interaction respectively. The paper then analyzed 4 categories of fatigue detection methods, and indicated that the fatigue detection method based on multi-feature information fusion will be the main research direction in the future. The paper finally revealed the difficulties of the current research, and prospected research of the commercial vehicle fatigue warning system from 3 aspects, i.e. obtaining more driver information, extracting more fatigue features, and reducing the dependence on specific fatigue features.

For the situation that the hybrid A* algorithm is difficult to accelerate the search using collision-free Reeds-Shepp (RS) curves due to the complex environment near the destination in short distance parking scenarios, this paper proposed an improved hybrid A* algorithm that reversely searched the path from the target posture, combined with the cost map look up calculated by A* algorithm to obtain heuristic values. Collision detection was performed by judging whether the car body contour line intersecting with the simplified obstacle to save search time, and by setting suitable vehicle steering angle resolution to increase number of the node expansion direction to ensure the smoothness of the path. Finally, MATLAB was used to simulate and compare the improved algorithm with the original algorithm. The results showed the improved hybrid A * algorithm effectively shortened the path search time both in parallel and vertical parking scenario, resulting in shorter and smoother paths.