Research On Sparse Feature Extraction And Diagnosis Methods For Mechanical Fault

Mechanical equipment is a vital tool for modern production. It is important to carry out operation condition monitoring and fault diagnosis for the safety production.Identifying incipient signs of failure and extracting them are the key to condition monitoring and fault diagnosis for mechanical equipment. In recent years, the rapid development of signal sparse representation theory, provides a powerful tool for incipient fault feature extraction and diagnosis methods based on vibration analysis. Based on project " Research of Incipient Fault Diagnosis of Low-speed and Heavy-load Machinery based on Sparse Feature Identification" supported by National Natural Science Foundation of China, with signal sparse representation as the main theoretical tool, the paper deeply studies on extracting incipient fault features and diagnosing faults methods for the easily faulty components, bearings and gears. The main research contents are as follows.Aiming at the problem that the incipient fault feature is easy to be covered by the noise and difficult to extract accurately, an incipient fault feature extraction method based on Tunable Q-factor Wavelet Transform(TQWT) is proposed. Firstly, the method utilises TQWT to decompose equipment vibration signal under different Q- factors and scales, and the best Q-factor and scales are determined by the kurtosis maximum principle. Then, neighboring coefficients denoising method is used to handle the selected scales of transform coefficients. At last, fault feature is gotten by implementing inverse TQWT. Experimental results show that the method can effectively extract the early fault of equipment in medium-speed operation. Compared with the traditional wavelet method,the extraction results are less noisy and the fault feature frequency is more prominent.Aiming at the problem that the dictionary is difficult to be constructed in extracting incipient fault feature via sparse representation, an initial fault feature extraction method based on dictionary learning is proposed. In the method, the differences between faultsignals and normal ones are used as training samples, and the dictionary learning algorithm, improved K-means singular value decomposition(IK-SVD), is applied to construct the dictionary to match diagnostic component. In sparse decomposition process,the kurtosis of each iteration approximation signal is calculated, and the sparse decomposition cut-off point for fault feature and noise is adaptive determined by finding the maximum kurtosis of the approximation signals. Comparative results indicate that the method can effectively extract early fault feature of equipment in low-speed operation.Compared with the parametric dictionary, the fault feature extracted by the proposed method is of higher accuracy.A single fault diagnosis method based on group sparse classification is proposed to solve the problem of single fault diagnosis with unknown parameters of fault components. In the method, the known fault type of training samples and unknown fault type of tested samples are transformed to the frequency domain, and the frequency coefficients of training samples are combined as a dictionary. Then, the frequency coefficients of tested samples are group sparse decomposed on the dictionary. And their fault types are determined by finding the minimum value of reconstruction errors. Fault diagnosis experiments proved that the method can accurately diagnose the single fault type of rolling bearing and gear with unknown feature parameters.A complex fault diagnosis method based on sparse classification of wavelet packet coefficients is proposed to solve the problem of complex fault diagnosis with unknown parameters of fault components. In the method, Wavelet Packet Transform(WPT) is implemented to translate the known single fault type of training samples into Wavelet Packet Coefficients(WPCs), and the frequency bands with class diffidence are screened out by distance evaluation parameters. Then, the WPCs of selected frequency bands are used to construct a group of dictionaries for sparse decomposition, and the WPCs of selected frequency bands of a tested sample are sparse decomposed on the corresponding dictionaries. At last, the complex fault types of the tested sample are determined one by one through the category of each minimum reconstruction error. The effectiveness of theproposed method is verified by examples of complex fault diagnosis of bearings and gears.