Research On Multi-features Fusion Fault Diagnosis Method For Rotating Machinery Based On Time-frequency Image Recognition

Time-frequency analysis can reveal the frequency components and time-varying feature of rotating machinery nonstationary signal. The time-frequency images(TFIs) constructed by mechanical vibration signal contain rich feature information of working states. Based on investigation on sparse essence of the TFIs about rotating machinery, three stages in the process of fault diagnosis, namely construction of TFIs, feature extraction and multi-features decision fusion diagnosis, were studied by using compressed sensing, sparse representation dimension reduction, locality constrainted low rank coding, intelligence pattern recognition, evidence theory and multi-objective optimization. And we proposed the multi-features decision fusion diagosis method for rotating machinery based on TFIs recognition. This sudy has important scientific significance in the theory development of time-frequency analysis. It also has broad application prospect for feature extraction and classification of nonstationary signal in other engineering fields.Because of being produced from Wigner-Ville distribution, most TFIs suffer inherent cross-term interference, reducing the TFIs resolution. In addition, the TFIs contain a large number of Fourier samples that are not good for real-time processing and long-distance transmission. For these reasons, we proposed the TFIs reconstruction method based on compressed sensing theory for rotating machinery vibration signal. With this method, the observation vector is obtained from the signal fuzzy domain close to the origin, and the adjustable radial Gaussian kernel function adaptively acts on the fuzzy function in order to suppress the cross-term interference, and the observed estimation value is obtained from the two-dimensional Fourier transform of sparse time-frequency representation. The improved gradient projection method is used to solve the reconstruction model, and the projection rule is used to get sparse TFIs. The performance of the proposed method has been verified through simulation and measured data of rot or. The results show that the proposed method not only increases the resolution and compression ratio, but also enhances the reconstruction accuracy and noise immunity performance.For feature extraction from TFIs problem of big sample, the dimension reduction methods based on sparse representation can effectively extract the effective features from these images, but there exist low discriminant and high computational complexity about current methods. For this reason, the method of feature extraction from TFIs of rotating machinery vibration signal based on improved sparse preserving projection is proposed. Firstly, the locality linear coding is used to obtain the sparse representation structure of TFIs, for the sake of enhancing discrimination of sparse representation structure and computation efficiency. At the same time, within-class scatter matrix and between-class scatter matrix are integrated into sparse preserving projection model in order to increase discrimination ability of the model. The proposed method is applied in the feature extraction from TFIs of rotor vibration signals under the circumstances of large sample, and the results show that low dimension feature of test samples has good separability.As we can’t acquire enough samples in job site, especially the labeled sample, the small sample recognition became a hot issue of research, but there exist some problems such as high computational complexity, inaccurate model with the current methods. For these reasons, the method of feature extraction from TFIs of rotating machinery vibration signal based on trace ratio sparse regularization discriminant analysis is proposed. Firstly, the locality linear coding is used to obtain the sparse representation structure. Secondly, the model of sparse regularization discriminant analysis is converted to the trace ratio problem. Thirdly, the projection matrix can be calculated through Dinkelbach’s algorithm. Finally, high dimensional samples are mapped into low dimensional space by the projection matrix. The proposed method is applied in the feature extraction from TFIs of rotor vibration signals under the circumstances of small sample, and the results show that low dimension feature of test samples has good separability.Due to the different degree of difficulty in obtaining different fault s, so it appears the imbalance sample condition in fault diagnosis. And most of dimension reduction methods, which are designed based on the assumption of the balanced sample, are not appropriate methods for the sake of feature extraction. For this reason, we propose the feature extraction method from TFIs of rotating machinery vibration signal based on locality-constrained low rank coding. Firstly, the low rank coding dictionary learning model containing locality-constrained regularization item is built. Secondly, the augmented Lagrange multiplier method and alternative optimization method are adopted to learn dictionary. Thirdly, the learned dictionary is used to obtain low rank coding. Finally, the spatial pyramid matching is employed to extract the feature from TFIs. The proposed method is applied in the feature extraction from TFIs of rotor vibration signals under the circumstances of imbalance sample, and the results show that low dimension feature of test samples has good separability.The accuracy rate using the single feature recognition method is low, and stability also is poor. For these reasons, a multi-features decision fusion fault diagnosis method about TFIs of rotating machinery, which is based on optimized support vector machine(SVM) and weighted evidence theory, is proposed. Firstly, the parameters of SVM model in each local diagnosis are optimized by multi-objective particle swarm optimization(MOPSO) in order to obtain the most optimal SVM prediction model. Secondly, the testing samples with single feature are input into the optimal SVM prediction model to obtain the weighted coefficients and basic probability assignment of this evidence on each fault, and the weighted probability assignments are acquired by weighted evidence theory. Finally, the combination rule and the decision rule are used to obtain the final diagnosis results. The feasibility and effectiveness of the proposed method are validated by the fault diagnosis about rotor.