Research On Diesel Engine Vibration Signal Feature Extraction And Fault Diagnosis Methods

As power machinery, the diesel engine is widely used in industrial and agriculturalproduction. Its operation status directly affects the performance of the entire unit. Therefore, itis of important significance to monitor the operation status and carry out fault diagnosis sothat the diesel engine operates in the normal state.Diesel engine fault diagnosis technology is an integrated technology based onmultidisciplinary which identifies the status of diesel engine and predicts the abnormal faultcondition by analyzing and processing the real-time status information of the diesel engine.Diesel engine cylinder head vibration signals contain a lot of information of operating status,so it is an effective method to diagnose diesel engine fault using cylinder head vibrationsignals.Extracting features from the cylinder head vibration signals and fault type recognitionaccording to the extracted fault features are the two key topics in the field of fault diagnosis ofdiesel engine. From the engineering application perspective, this thesis systematicallyinvestigated the fault feature extraction and fault type recognition methods of the dieselengine applying testing technology, wavelet analysis, empirical mode decomposition, chaoticnumerical characteristics, BP neural network, and support vector machine theory. The mainresearch work is as follows:(1) A diesel engine cylinder head vibration signal acquisition experimental platform wascomposed. The diesel engine cylinder head vibration signals in normal and different faultconditions of S195were collected, which can be used for feature extraction and faultdiagnosis of diesel engine fault vibration signal.(2) This thesis has studied application of the wavelet analysis in the cylinder headvibration signal feature extraction. The wavelet packet energy distribution and the continuouswavelet scale energy distribution of diesel engine cylinder head vibration signal underdifferent states were analyzed, and the results showed that wavelet packet energy distributionand continuous wavelet scale energy distribution could be characterized as fault features ofthe diesel engine failure. Taking into account the time-domain characteristic parameters of thecylinder head vibration signals, this thesis put forward four methods to compose diesel engine fault feature vector, including the wavelet packet energy distribution, the wavelet packetenergy entropy distribution, the wavelet packet energy distribution combined with signal timedomain characteristics, and the continuous wavelet scale energy distribution.(3) Feature extraction of diesel engine cylinder head vibration signals using EMD wasdiscussed. For the pulse noise interference appearing in the cylinder head vibration signal, thisthesis put forward an improved signal noise reduction method based on EMD. The thesis usedmarginal spectrum, the power spectrum and the AR model spectrum estimation methodcombined with EMD to illustrate the characterization of diesel engine fault featureinformation. Taking into account the signal time-domain characteristic and frequency domaincharacteristics, three composing diesel engine fault feature vector methods were presented.They are the time-domain features combined with IMF energy, the time-domain featurescombined with Hilbert marginal spectrum and the time-domain features combined with theAR model spectrum estimation.(4) The chaotic numerical characteristics correlation dimensions and the analysis andcalculation methods of the maximum Lyaponov exponent were studied, and the Chaoticnumerical characteristics of the diesel engine fuel system fault vibration signal were analyzed.The results showed correlation dimension values computed with the original signals can notdistinguish whether the diesel engine fuel system operates properly while the maximumLyapunov exponents can. So this thesis put forward a method of using the wavelet packet todeal with the cylinder head vibration signal firstly, and then computing correlation dimensionvalues of the signals to diagnose fault. The results showed that the correlation dimensionvalue obtained by this method can judge diesel engine fuel system to operate properly or not.(5) BP neural network and support vector machine method of diesel engine faultidentification were studied. Diesel engine fault feature vector were constructed using the fourmethods (wavelet packet energy distribution; wavelet packet energy entropy; wavelet packetenergy distribution combined with time-domain characteristics; continuous wavelet energydistribution) based on wavelet analysis for five types of abnormalities of diesel engine whichare the normal operating conditions, the abnormal exhaust valve clearance, the abnormalintake valve clearance, the abnormal fuel supply advance angle and the abnormal pressurefuel injection conditions. Diagnostic classification recognition accuracy rate by BP networkwere79.39%,85.67%,87.03%,91.15%respectively; and diagnostic classification recognitionaccuracy rate using support vector machine were82.67%,90.33%,87%and92.50%respectively. The fault diagnosis results showed that four types of feature vector constructingmethods based on wavelet analysis can be used in feature vector constructing for a dieselengine cylinder head vibration signal. Feature vector constructing using continuous wavelet scale energy distribution and classification diagnosis using support vector machine canachieve the best diagnosis effect. For the same method to feature vector constructing, theeffect of diagnosis classification using support vector machine is better than using BP neuralnetwork.(6) Diesel engine fault feature vector were constructed using the three methods (thetime-domain features combined with IMF energy; the time-domain features combined withHilbert marginal spectrum; the time-domain features combined with the AR model spectrumestimation.) based on EMD for five types of abnormalities of diesel engine which are thenormal operating conditions, the abnormal exhaust valve clearance, the abnormal intake valveclearance, the abnormal fuel supply advance angle and the abnormal pressure fuel injectionconditions. Diagnostic classification recognition accuracy rate by BP neural network were80.2%,85.5%, and87.8%respectively and diagnostic classification recognition accuracy rateusing support vector machine were80.83%,83.67%and91.0%respectively. The faultdiagnosis results showed that three types of feature vector constructing methods based onEMD can be used in feature vector constructing for a diesel engine cylinder head vibrationsignal. Feature Extraction using the time-domain features combined with the AR modelspectrum estimation and classification diagnosis using support vector machine can get thebest diagnosis effect.