An unsupervised multi-level fusion domain adaptation method for transfer diagnosis under time-varying working conditions

Unsupervised multi-source domain adaptation can overcome the limitations associated with insufficient information diversity in single-source domain adaptation for intelligent transfer diagnosis. However, the challenges of time-varying working conditions in practical industrial applications, limitation in single-level information fusion along with lack of multi-level information fusion restrict effective applications of unsupervised multi-source domain adaptation in transfer diagnosis. To address these challenges, this research presents a novel unsupervised multi-level fusion domain adaptation methodology for transfer diagnostics under time-varying working conditions, which employs a multi-level fusion domain adaptation network (MLFDAN). Firstly, a multi-sensor data enhancement and fusion module is proposed by combining continuous wavelet transform with an RGB information fusion, which integrates time–frequency and spatial information from multi-sensors. Then, a squeeze and excitation feature fusion module is designed for feature fusion across both time–frequency and spatial domains, which effectively emphasizes domain-invariant features while suppressing less relevant ones. Subsequently, an adaptive collaborative decision module is developed, which employs a weighted fusion strategy to address strong conflicts among multi-subnet predictions and utilizes consensus-based fusion strategy when multi-subnet predictions align, thus ensuring reliable and robust diagnostics decisions. Finally, a promising MLFDAN framework for transfer diagnosis is proposed by incorporating a dual-component domain adaptation approach that integrates a domain discriminator and multi-kernel maximum mean discrepancy. Numerous experiment results show that the presented MLFDAN methodology effectively adapts to transfer diagnosis scenarios from steady to time-varying working conditions, achieving impressive performances and outperforming several prominent unsupervised transfer diagnosis methodologies.