Non-stationary Signal Feature Extraction Method And Its Application In Internal Combustion Engine Fault Diagnosis

Internal combustion engine health assessment and fault diagnosis without disassembly are maintenance and important safeguards of safe operation as the power source of mechanical systems with the minimum cost of fault diagnosis and repairs. Internal combustion engine surface vibration signal contains abundant state information. The surface vibration signals show the complex non-stationary and nonlinear time-varying features. It is difficult to be directly serve as the basis of internal combustion engine health status evaluation and fault diagnosis. Therefore, internal combustion engine surface vibration signal with deep processing and features extraction which can represent the key features of its running states have become the key problem in the study of the internal combustion engine fault diagnosis method without disassembly.This dissertation researchs the non-stationary and non-linear vibration signal feature extraction method and pattern recognition theory. Through the analysis and deep processing of the internal combustion engine vibration signals, key features which can represent internal combustion engine running status information are extracted. Through characteristic information extracted contrasting and patterns classifying with recognizers, the judgement of internal combustion engine work state and fault types can be realized.Based on the study of the theory of wavelet threshold de-noising, this paper proposes EEMD-wavelet energy and LMD-wavelet energy feature extraction methods.In view of the blind source separation cannot be achieved in the process of single channel signal input and deficiency of empirical mode decomposition(EMD) and independent component analysis(ICA) based on linear assumption in analysis of nonlinear signal, based on local mean decomposition(LMD) and kernel independent component analysis(KICA) collaborative signal feature extraction methods(LMD-KICA) is proposed, and be applied in the diesel engine fault diagnosis. Signal feature extraction method(KICA-LMD-CD) based on the KICA-LMD and fractal theory is proposed.Adaptive wavelet packet threshold method and mathematical morphological filtering method are used to de-noise diesel engine fault signals, and then after the KICA-LMD decomposition, fractal correlation dimension of components(PFs) are computed. Through the analysis of correlation dimension numerical size and tendency to judge states of diesel engine fuel injection advance angle. The noise and local wave decomposition for the calculation of correlation dimension of diesel engine vibration signal are discussed. The results show that the noise have strong impact on numerical calculation of correlation dimensions, and the noise reduction is the precondition of calculating correlation dimension of KICA-LMD decomposition components. The values of PFs correlation dimensions can be used as a judgment of the status of the diesel engine.Support vector machine(SVM) as the diagnosis identifier of the diesel engine valve clearance fault. An signal feature extraction method based on the KICA-LMD and correlation coefficient is put forward, which is used to extract a six-cylinder diesel engine vibration signal decomposition component correlation coefficients of PFs. And the correlation coefficients of PFs are used as characteristics of support vector machine(SVM) for diesel engine fault classification.A multiple signal feature extraction method of feature extraction and classification recognition is proposed. Extract the feature information of cylinder head vibration signal under seven states of a six-cylinder diesel engine. Fault pattern recognition results show that, the correlation coefficient, correlation dimension, energy feature can be very good characterizations of nonlinear state of the system, and recognition rate is 99.4351%; while data statistical characteristic is poor, and recognition rate is 87.1429%. Also, after dealing with the dimension reduction of classifier than without dimension reduction classifier recognition rate is higher.