Initial fault feature extraction via sparse representation over learned dictionary

In the initial fault of rolling bearing, the useful weak impulses reflecting fault feature in measured vibration signal are usually corrupted by strong background noise. Sparse representation over learned dictionary is taken to extract the initial fault feature. Firstly, K-SVD learning algorithm is employed to obtain an adaptive dictionary matching the impulses. Then Batch Orthogonal Matching Pursuit (Batch-OMP) is utilized in sparse-coding stage, and kurtosis is introduced to determine the iteration stop condition in sparse approximation. The simulate data and real bearing data tests validate the proposed method.