FFNN Application In Diesel Fuel System Faults Diagnosis

Diesel is the complex structure system with many moving components and disturbing incentive sources. The relations between the symptoms and faults in diesel engine are not linear and assured, for example, the relations between the fuel pressure waveform and diesel fuel system faults, at the same time the signals are easy to be distorted because of the poor environment, therefore, it is difficult to diagnose the faults in diesel engine. However, the artifical neural networks technology is a new solved method to diagnose the diesel faults, and they can be the important tools to diagnose the diesel faults.There are the input-output nonlinear mapping, parallel processing and other features in artifical neural networks, especially the self-adaption and self-learning capability, with learning the nonlinear mapping problems can be solved. Thus, the artifical neural networks technology is the effective method to diagnose the faults. Especially, the most commoned neural networks, feed forward neural network (FFNN), for instance, BP (Back Propagation),RBF (Radial Basis Function) neural networks. But about BP neural network, there are something also should be improved, while the training number is so large that the learning efficiency and the convergence speed are low, so BP algorithm has been improved by adding momentum and adjusting the learning rate in this paper. However, the drawbacks in BP neural network can not be overcome completely. Howerver, RBF neural network doesn't have the drawbacks in BP neural network, because the hidden layer active functions is different to BP neural network.Through the experiments, the neural network performance has been improved effectively with the improved BP algorithm. Moreover, the neural network with improved BP algorithm also has been adopted to diagnose the diesel fuel system fault, the faults diagnosis results are satisfactory. If the centers can be elected correctly, RBF neural network can achieve the performance in BP neural network, the learning speed is still greater, the intense jounces cannot occur in the convergence process, and the faults in engine can be diagnosed effectively as well.