Rotating Machinery Fault Diagnosis with Weighted Variational Manifold Learning

The transient characteristics attributed to partial fault of rotating machinery are of technical evidences for health diagnosis, which are unavoidably overwhelmed by much noise disturbance in complex modulation bands. Usually, only the optimal component with a selected scale is filtrated and then extracted to be the evidence for fault diagnosis, which may ignore some important features distributed in the other scales. Focusing on the feature distribution in a whole-band, this paper proposed a novel weighted variational manifold learning (WVML) for transient features reinforcement and fault diagnosis in the view of multi-scale analysis. First, a series of intrinsic mode functions, containing different modulated frequency information, are automatically obtained via an improved variational mode decomposition (IVMD) approach, where a signal difference average criterion is introduced to determine the appropriate decomposition level. Time–frequency manifold (TFM) learning is referred to then sequentially mine their intrinsic structures in the corresponding time–frequency domain. Therewith, the reinforced signal for each variational mode can be achieved via phase preserve and a series of inverse transforms in a suppression of the in-band noise. Combining the reinforced mode signals with different weight coefficients, the reconstructed waveform can be finally synthesized. In these manner, the desired features of all scales can be soundly reinforced with mode manifolds mined and dynamically weighted in a self-learning way. The consequence of proposed WVML scheme is verified through experimental gear defective signal and result reveals this method yields an excellent accuracy in health diagnosis of rotating machinery.