Design And Realization Of Diesel Fuel System Fault Diagnosis Based On Wavelet Neural Networks

The research on fault diagnosis theories and methods for fuel system becomes not only popular but also challenging at present. As a representative of reciprocating machine, the faults of fuel system are bound to be complex and diverse, and traditional fault diagnosis theories and methods cannot meet actual requirements. With the maturity of Neural networks, new ideas and methods are provided for the fault diagnosis of diesel fuel system. Feature extraction is an important part of fault diagnosis, and determines whether or not the diagnosis would success. Fortunately, wavelet transform becomes an important method of feature extraction due to its space localization property. Therefore, the combination of wavelet transform and neural networks, so called the wavelet neural networks, becomes a focus in diesel fuel fault diagnosis. Based on modern measure technology, wavelet transform and neural networks, the system of diesel fuel fault diagnosis is designed in this thesis. The system is applied in practice and performs well. The main achievements can be summarized up as follows:1. A new solid state signal acquisition method without disintegration is adopted. This method uses clamp-on sensors to get pressure waveforms of the high-pressure oil chamber indirectly, and realizes on-line acquisition of oil pressure waveforms without disintegration.2. Feature extraction method based on wavelet transform is studied. The theories and methods based on wavelet transform are analyzed in detail. According to the characteristics of oil pressure waveforms, two methods of feature extraction are proposed.(1) Modulus maximum of wavelet coefficients. Diesel oil pressure signal often contains important fault features at the spray point and the maximum injection pressure point. In view of this characteristic, multi-scale signal edge detection is applied to extract fault features of oil pressure signal by modulus maximum of wavelet coefficients. The results show that the method can obtain the fault feature of oil pressure signals accurately.(2) Wavelet packet frequency-band power analysis. According to the power band analysis, the "power - fault" method is proposed, and features of oil pressure signal are extracted. The results show that frequency-band power analysis is better than modulus maximum of wavelet coefficients, and can get more suitable features which will be treated as the inputs of neural networks.3. Diesel fuel system fault diagnosis based on neural networks is studied. To overcome the disadvantages of BP-NN, such as low convergence rate and easy convergence to local minimum, some fault diagnosis methods for diesel fuel system are proposed.(1) The fuel system fault diagnosis based on RBF networks is studied in this thesis. The RBF networks have unique best approximation and have no local minimum. Using these important features, fault diagnosis of diesel fuel system is carried out. Diagnosis results show that RBF network can fundamentally overcome the problems of converge to the local minimum point and low convergence rate. Meanwhile, it can diagnose the fuel system faults more quickly and accurately.(2) Fault diagnosis of fuel system based on SOFM network is explored. According to SOFM theory, the network model is built and applied in diesel fuel fault diagnosis. The results show that SOFM network diagnostic model is very strict with the input sample vectors, but can achieve more accurate diagnosis results.4. A new method which adopts wavelet parket neural networks to fault diagnosis of diesel fuel is proposed. Methods of characteristics extraction using wavelet transform and the basic structure and theories of neural networks are expounded. Wavelet packet power and RBF networks are combined together to diagnose the system. The loose combination of wavelet analysis and neural networks largely improves the accuracy of diagnosis results.5. Taking 165-type diesel engine as the research object, the system of diesel fuel fault diagnosis is designed, and verified by experiments. The master system is developed with Visual C++ and the slave system is designed based on 89C52. The two systems communicate via the RS-232 serial communication. Experimental results show that the system is feasible, and can diagnose the fuel system online.At the last part of the thesis, the main contents are summarized and suggestions of the future work in this filed are given.