An Adaptive Randomized Orthogonal Matching Pursuit Algorithm With Sliding Window for Rolling Bearing Fault Diagnosis

Sparse representation, as a powerful signal processing method, is widely used in rolling bearing fault diagnosis. As an algorithm for solving the sparse model, randomized orthogonal matching pursuit (RandOMP) algorithm shows excellent performance in impulse features extraction. However, the sparsity of impulse features is difficult to be determined accurately which influenced the effect of feature extraction. In order to more accurately obtain the fault information of rolling bearing, a method named adaptive RandOMP algorithm with a sliding window (AdRaOMP-SW) is proposed in this paper. The method consists of the adaptive RandOMP (AdRaOMP) algorithm and the sliding window. The AdRaOMP algorithm is an improvement of RandOMP, and it solves the problem that the sparsity of impulse features is difficult to determine and improves the efficiency of feature extraction by selecting a plurality of atoms in each iteration. The sliding window is a data segmentation strategy for the original signal which can be used to weaken the influence of noise. First, the vibration signal of the rolling bearing is segmented into multiple sub-signals by the constructed sliding window. Then, each sub-signal is processed by the AdRaOMP algorithm to obtain the corresponding sparse representation result. Finally, the sparse representation results of all sub-signals are spliced together to obtain the final fault features. The experiments and engineering application demonstrate that AdRaOMP-SW has better feature extraction performance compared with the RandOMP and spectral kurtosis.