|
Engine, as the core component of a car, its performance directly affects on the safetyand reliability of vehicles. Thus, the techniques for engine fault diagnosis have become animportant issue in the field of vehicle engineering. Based on the previous research findings,this thesis paper studies and presents a noise-based technology for engine fault diagnosis,using the Hilbert-Huang Transform (HHT) and the support vector machine (SVM).In this thesis, firstly, the common engine faults are investigated. A data acquisitionscheme is determined by analyzing the generation mechanism and propagation path ofengine noises. Subsequently, some noise measurements are performed by setting a sampleengine into different working conditions, including the normal and fault states, thereby, adatabase establishment of the measured engine noise signals. Secondly, according to thetheories in nonstationary signal processing, a study on time-frequency analysis techniquesfor signal denoising and feature extraction of the engine noise is conducted. Based on a setof assumed signals and the measured noise signals, the short-time Fourier transform(STFT), Wigner-Ville distribution (WVD), wavelet transform (WT) and the HHT, areinvestigated and compared. The wavelet packet analysis and HHT are selected for signaldenoising and calculation of fault feature vectors of the engine. Based on the above results,finally, the artificial intelligence theories, such as machine learning, statistics learning, andSVM, and their implementation procedure are discussed in detail. The nonlinear SVM andmulti-class SVM are used in pattern recognition of the calculated fault feature vectors, thus,the completion of a HHT-SVM modeling got intelligent engine noise diagnosis.The above studies suggest that, for the engine noise signals, the wavelet packetdenoising method has an obvious advantage; the HHT is suitable for analyzing thenonstationary signals, due to its high time-frequency resolution and without cross terms;the SVM with a good trait in training and classification is effective for signal patternrecognitions. The experimental verifications imply the nine-dimensional vectors computed by the HHT can be used to describe engine fault features; and the newly developedHHT-SVM model is accurate and feasible for the engine noise diagnosis. The noise-basedHHT-SVM algorithm, which can be extended to other sound-related fields for failuredetections and recognitions, may be a promising approach in both the theoretical researchand the practical application in engineering.
|