Fast Iterative Dictionary Learning And Its Application In Transient Fault Feature Extraction

Sparse representation method has been widely used in various fields where data processing is needed.Learning from the actual data to obtain the adaptive dictionary can usually get better sparse representation performance.However,the amount of training data needs to be large enough to obtain the comprehensive features of the actual data,as a result,there exists issues such as large amount of calculation and long time consuming.Exploring fast and accurate dictionary learning algorithm has always been an important research topic in the field of sparse representation.In order to solve the problem of time-consuming in the current adaptive dictionary method for transient impulse signal extraction in rotating machinery fault diagnosis,this paper mainly studies the adaptive dictionary learning method in sparse representation based on nonconvex regular terms.By using the techniques of variable separation and reorganization,and combining the coordinate descent algorithm,fast dictionary learning and precise sparse reconstruction are realized.The main contents are as follows:Firstly,in the sparse coding stage,we update the coefficients in block-wise based on coordinate descent method and adopt the generalized mini-max concave function as the nonconvex sparse regularization,which avoids the systematic underestimation characteristic of L1 norm.Secondly,we improve the atom update step through simultaneously updating the corresponding nonzero entries,leading to more accuracy in sparse reconstruction,this is addressed by the least square solutions.Thirdly,sparsity-based adaptive penalty parameter control method is utilized,and the further version which plays a tradeoff between the sparsity and error is proposed,achieving better performance in the signal extraction with noise;Finally,Simulation and application experiments are carried out to validate the effectiveness of proposed algorithm.Results show that the proposed algorithm can effectively avoid the underestimation of the amplitude,leading to better performance than the L1 norm based BCD dictionary learning method in the extraction of impulsive signal,and its running time is much less than that of the K-SVD algorithm.