Fault feature enhancement of rotating machinery via shift-invariant time-frequency manifold self-learning

In order to overcome the difficulties of traditional sparse methods in dictionary construction and the limitations of sparse representation results, a new adaptive fault diagnosis method, named as shift-invariant time-frequency manifold self-learning, is proposed by introducing time-frequency manifold learning in the framework of shift-invariant sparse learning. This method accomplishes the enhancement and extraction of shift-invariant intrinsic modes by establishing TFM learning on a local signal. Envelope entropy is used to adaptively output the optimal envelope basis component from the local TFM. In this manner, the global envelope of the raw signal can be reconstructed and strengthened via this local manifold envelope mode. Finally, signal enhancement and diagnosis of rotating machinery fault based on the reconstruction expression and reinforcement learning is achieved by combining phase preserve and a series of inverse transforms. Experimental results show that this method achieves effective suppression of strong background noise and mining of nonlinear transient features, which is conducive for efficient and accurate fault diagnosis researches. © 2020, Nanjing Univ. of Aeronautics an Astronautics. All right reserved.