Updatable Online Learning Successive Difference Mode Decomposition for Rotating Machine Fault Diagnosis

Signal processing methods are widely used in fault diagnosis and are known for their strong interpretability. Among them, signal adaptive decomposition algorithms are used to extract the features of fault signals. As an effective adaptive decomposition algorithm, difference mode decomposition (DMD) divides the signals into three components using spectrum weighting. However, it can only separate mixed fault components and is not suitable for multiclass fault diagnosis tasks. This article presents a successive DMD (SDMD) method. The reference component (RC) and concerned components (CCs) (fault features) are defined based on the differences in faults. Then, the filters corresponding to different components are obtained through iterative convex optimization at each layer. Finally, using these filters, signals are decomposed into multiple fault components corresponding to different fault sources. Furthermore, the white noise replacement module is proposed to solve the gradient vanishing problem introduced by successive decompositions. In addition, an updatable online learning framework is proposed for the incremental demand scenario, providing data efficiency and interpretability. The effectiveness of this method is validated on real datasets.