Machinery cross domain degradation prognostics considering compound domain shifts

Nowadays, data-driven based decision-making mode significantly promotes machinery prognostics and health management (PHM), but are also profoundly affected by domain shift problems. As a promising transfer learning methodology, domain adaptation has matured in machinery fault classifications under variable operating conditions and achieved tremendous success in aligning data distribution discrepancies among multiple domains, called inter-domain shifts in this study. However the monitoring signal presents nonlinear and non-stationary characteristics when the machine is degraded, its period-level and chronological discrepancy within the given set of monitoring time series is often ignored. This intra discrepancy is not conducive to maintaining the time series consistency and may fail to build an accurate prediction model. Therefore, this study, for the first time, formulas this kind of shift and comprehensively resolves the compound shifts originating from inter and intra domains. Firstly, the domain-invariant degradation indicator is constructed through the designed double adversarial learning based multi-source domain adaptation module. Then the obtained indicators are segmented into distinct degradation periods according to outlier detections for subsequent intra-domain level alignments, quantifying as the regularization term for matching the discrepancies between degradation periods. Finally, realmeasured run-to-failed experiments are applied to test the effectiveness of our proposed method.