A Study On Engine Fault Diagnosis Based On Probabilistic Neural Networks

As the power source of the automobile, the running condition of engine can directly affect the whole working condition of the automobile. Engine faults constitute most proportion of automobile faults, and it takes high cost and long time to maintain it. While the automobile's structure becoming more and more complex, working condition becoming more and more severe, fault diagnosis becoming more and more difficult. The reliability and security of automobile will be enhanced if we can exactly judge the engine current condition in time, distinguish the positions of the faults and point out the reason and solving method with the engine being not disassembly. This paper is use of the intelligent method of wavelet packet analysis and neural network for fault diagnosis of diesel engines.Wavelet packet analysis is a way of handling signals, capable of distinguishing each wave frequency to the desired level of detail. It can not only resolve low-frequency signals further, but also do the same to high-frequency signals, and so has broad application in the diagnosis of machine faults. In this paper, wavelet db2 is chosen, and all signals are de-noised by an improved wavelet packet threshold has been designed. Then wavelet packet decomposes and constructs the energy eigenvectors, and extract the normalization eigenvector of the machine fault diagnosis.Owing to their strong sensitivity of non-linear reflection, neural networks are particularly suited to non-stationary signal analysis and complex pattern recognition, and have therefore become a powerful tool for recognition of the status of power machinery. In the paper, used the Probabilistic Neural Network as the tool for fault classification, according to the experimental requirements and the needing solve problems, ultimately determine the network structure is: 9 inputs layer neuron, 28 output layer neuron, that is also to say 28 pattern recognition.This paper is use of the professional analysis software DASP-V10 collect the vibration signal of Yunnei 4100QB diesel engine. DASP-V10 is a real time multi-channel signal acquisition and analysis system run on Window98/Me/NT/2000/XP/Vista platform, And through use the different hardware of China orient institute of noise & vibration, you can form a variety tests laboratory both of static and dynamic. DASP-V10 also has easy operating features.A self-adaptive machine fault diagnosis system was designed on the basis of experiments on the 4100QB diesel engine produced by the Kunming Yunnei Power Co. Ltd., based on wavelet analysis and neural networks. This system picks up the vibration signal of the diesel engine cylinder head, and filters the vibration noise through wavelet analysis. It extracts the eigenvector of the vibration signal that indicates the diesel engine fault, and processing the speed signal, also as a parameter of the eigenvector. Then use the eigenvector as a neural network experimental sample, eventually building a self-adaptive machine fault diagnosis system. By inputting the test sample into the neural networks self-adaptive machine fault diagnosis system to conduct a verification of the system, it showed that the system can effectively identify and classify the machine fault, and ultimately achieve a fault diagnosis. Also analysis the rate factors of the probability neural network fault diagnosis, the results showed that, the quantity and quality of training samples is the main factors affect the rate of fault diagnosis , with the number of training samples increase, the correct diagnosis rate will also be increase. While change the spread value which is the expansion constant were not affected the fault diagnosis rate in this paper.In the experiment, DASP-V10 professional software was used to acquire the vibration signal of the engine's cylinder head and cylinder wall under both normal and abnormal engine operating conditions. The wavelet analysis source program of MATLAB was applied to filter the vibration signal to extract the relevant eigenvector to serve as the experimental and verification sample for the neural network.The experiment results and analysis indicated that the method of fault diagnosis based on probabilistic neural network proposed in the paper is simple, and has wide application prospect.