Reference-periodic-similarity based reweighted singular value decomposition signal denoising scheme and its application for weak fault feature diagnosis in industrial rotating machinery

Reweighted singular value decomposition (RSVD) has demonstrated its effectiveness for signal denoising and fault feature diagnosis for rotating machinery in recent years. Different from traditional energy-based SVD denoising approaches which usually retain singular components (SCs) with higher singular values, RSVD leverages various indicators such as periodic modulation intensity (PMI) and Gini index (Gini) to selectively pick SCs that contain richer fault-related information. However, it is shown that PMI-based RSVD and Gini-based RSVD are not as effective in some weak fault feature scenarios. To overcome this issue, a reference-periodic-similarity (RPS) based RSVD scheme, which leverages fault period T as prior knowledge, is proposed for signal denoising and weak feature diagnosis. A new indicator, namely RPS, is introduced to select the most suitable SCs by evaluating the periodic similarity with a pre-defined reference signal. Using RPS, interferences from non-fault related signal components become less influential and SCs with very weak fault signature can still be selected for signal reconstruction. The advantages of RPS-based RSVD are verified by both simulated analysis and real experimental signals. The results show that the proposed method is powerful in signal denoising even when fault features are very weak.