Domain-Specific Dual Network With Unsupervised Domain Adaptation for Transfer Fault Prognosis Across Machines Using Multiple Source Domains

Deep learning (DL)-based fault prognostic methods often require extensive training data, which can be challenging to obtain due to time, cost, or safety constraints. However, labeled data from related but distinct machines, such as from accelerated degradation experiments, can offer valuable insights. Most transfer learning (TL) methods focus on single-source domain adaptation (DA) within the same machine, overlooking variability within the source domain and missing the opportunity to leverage multisource historical data effectively for application on different machines lacking run-to-failure data. To address these limitations, a domain-specific dual network (DSDN) is proposed for transfer fault prognosis across heterogeneous machines using multisource domain data. The DSDN consists of two feature extraction modules: domain-specific degradation feature extraction (DSDFE) and domain-specific high-level feature extraction (DSHLFE). These modules incorporate a deep convolutional neural network (DCNN) enhanced with multiple convolutional block attention modules (CBAMs) and a multilayer perceptron (MLP). The outputs from modules are concatenated to generate domain-invariant representations, effectively capturing both time-series and time-frequency spectrum features. By incorporating the DA technique, this approach mitigates discrepancies between the source and target domains, enabling accurate remaining useful life (RUL). The proposed method enhances the extraction of robust domain-invariant features for each source-target domain pair separately, improving prognostic accuracy by mitigating negative transfer. The effectiveness of the DSDN is validated through extensive case studies on single-source and multisource bidirectional transfer fault prognosis tasks across machines, utilizing bearing datasets from the IEEE prognostic and health management (PHM) Challenge 2012 and XJTU-SY. A comparative analysis further highlights its superiority over existing methods.