Fault Diagnosis For ESP Using BP Neural Networks

As the rapid development of automobile industry,the improvement of vehicle safety and maneuverability is urgent to be worked out. Automobile and components manufacturers all over the world pushed out some electronic stability control systems,in which ESP is a revolutionary production which is based on ABS and TCS. As we all known,the more complex a electronic system is,the more difficult to ensure its safety and stability it is. So when we equip the ESP system in a vehicle,the research of its fault diagnosis is of great importance in the field of vehicle active safety.This paper concentrates on estimating the important state parameters that evaluate the vehicle stability: the yaw rate,the lateral acceleration,the side slip angle of the mass center and the road friction coefficient,and then diagnosing the fault of the yaw rate sensor,the lateral acceleration sensor and the side slip angle of mass center sensor. As neural networks have a strong fitting ability to nonlinear mapping relation of MIMO(multiple input multiple output) system. So BP neural networks were used to estimate fault-free output of the sensor in ESP system,and then compared the fault-free estimation with the output of fault sensor to get the residual and diagnose the fault. Afterwards,analyzing and evaluating the results. The proposed fault diagnosis method in this paper is very simple and practical. It can be used to diagnose a single sensor’s fault and two sensors’ faults. It can not only figure out if there is a fault and the time when it happens,but also reflect the size of the fault. Before using BP neural networks to estimate the states,this paper also gives a proof to the stability of BP neural network using Lyapunov theory which provides a theoretical basis to the proposed fault diagnosis method.Owing to four kinds of roads that are existing in reality: dry asphalt road,wet asphalt road,snow road and ice road. When estimating the output of fault-free sensor,only one neural network cannot include the condition of four kinds of different roads. To work it out and increase the precision of the network estimator,the influence of road friction coefficient to the driving condition should be taken into consideration. So we have trained four neural networks respectively to be a network group. In addition,adding a selecting module to the system,so we can pick out the network corresponding to the certain road friction coefficient to get the exact state estimation. We have also finished training of 81 networks to propose a estimation method of road friction coefficient. The car model of 15 dofs was built by using AMESim software and the experiments were finished by co-simulation of AMESim and MATLAB. The fault diagnosis of yaw rate sensor,lateral acceleration sensor,side slip angle of the mass center sensor has been realized. The proposed estimation method of road friction coefficient has some certain reference significance. The simulation results verified the effectiveness of the proposed method.