A Bearing Fault Feature Extraction Strategy Based on Finite Frequency Shift-Invariant Sparse Coding

Suffering from strong background noise, sparse representation cannot effectively reveal signal implicit features, which brings challenges in identifying mechanical fault. Therefore, a finite frequency shift-invariant sparse feature extraction strategy (FF-SISFES) for signal feature enhancement is proposed in this paper. The strategy consists of two parts. Firstly, a differential envelope energy (DEE) extraction method based on collaborative filtering (DEE-CF) is proposed. It is employed for estimating sensitive frequency band and extracting feature wavelet corresponding to the sub-band signal with the highest energy. Secondly, to effectively obtain objected feature and greatly reduce the interference of background noise, a feature wavelet enhancement generalized minimax-concave convolution dictionary learning method (FWE-GMCCDL) is developed. Wherein, the DEE is utilized for obtaining sparse parameter. Finally, feature frequencies contained in sparse coefficients are extracted by envelope spectrum. The experimental results verify the FF-SISFES can effectively learn and enhance shift-invariant feature from observed signal. The proposed method has better ability of fault extraction than available dictionary learning method like Laplace correlation filtering, group-sparsity learning (GSL), periodic GSL (P-GSL), and traditional convolutional dictionary learning (CDL).