Fault Feature Extraction And Diagnosis Approach Of Mechanical Based On Sparse Representation

To extract the weak fault features under complex working conditions,reduce the influence of redundant information on feature extraction,and recognise the mechanical fault status under sparse model.Inspired by sparse representation alogorithm,a sparsity-promoted strategy based on Majori-zation-Minimization for weak fault feature enhancement and a termination criterion improved K-SVD dictionary learning strategy for sparse feature extraction were investigated in this paper,moreover a weighted sparse representation fault classification method based on dictionary learning was proposed.The details are presented as follows:(1)The sparsity-promoted strategy based on Majorization-Minimization for weak fault feature enhancement was studied.As is known,it is difficult to extract the fault features accurately under strong background noise or in the early stage of fault occurrence.In order to enhance the fault features,a sparse feature enhancement model based on wavelet basis and MM was established.Firstly,the optimal wavelet basis atoms were selected by correlation filtering,and then the atoms were expanded into sparse dictionaries.Secondly,an optimization function was designed to transform the l1 norm problem into a series of convex optimization iteration problems.Finally,the sparse noise reduction and feature enhancement were achieved through continuous optimization iteration.Besides that,aiming at the low efficiency of dictionary construction and the unsatisfactory effect of optimization iteration,an improved MM algorithm was explored.The unit matrix was used to replace the wavelet dictionary to simplify the iteration function,which improved the effect of feature enhancement and reduced the computational complexity.(2)The sparse feature extraction method based on termination criterion improved K-SVD dictionary learning strategy was investigated.Aiming at the problem that sparse representation based on analysis dictionaries has poor adaptability,a dictionary learning algorithm was studied.Since the sparsity of target signal is unknown,the results of dictionary learning convergence error and sparse representation are poor using the traditional K-SVD.Therefore,a novel sparse extraction model was proposed based on improved K-SVD dictionary learning algorithm.The sparse objective function and constraints were optimized,for which the dictionary matched the fault impulse components could be constructed without setting the sparsity.At the same time,an improved orthogonal matching pursuit algorithm with residual threshold was constructed to sparse coding.In addition,for the sake of reducing the noise infuences to dictionary atoms,a sparse feature extraction strategy based on variational mode decomposition(VMD)and improved K-SVD was proposed.The results of simulating and experimental signals were employed to validate the effectiveness of traditional K-SVD algorithm,improved K-SVD algorithm and combined with VMD and improved K-SVD algorithm.(3)The weighted sparse representation fault classification method based on dictionary learning was proposed.Sparse classification is inspired by the theory of sparse representation,which can classify the status by the linear combination results of different kinds of signals in sparse dictionary.Since the traditional SRC neglects the local feature information of samples,which leads to the reduction of fault classification accuracy,and the training samples contains the large amount of redundant information which leads to the reduction of classification efficiency,therefore a weighted sparse representation fault classification model based on dictionary learning was proposed in this paper.An improved K-SVD dictionary learning algorithm was proposed to construct training samples,which can effectively reduce redundant components of samples and improve classification efficiency.The K-means clustering algorithm was used to select the optimal time-domain feature parameters as weighting coefficients.The sparse coding of test samples was acheived by solving the weighted l0 norm problem,which effectively enhances the local features of samples.In addition,in order to overcome the influence of time-shift deviation of vibration signals,correlation analysis was used as the criterion of fault classification.