Meta-Learning Pseudo-Sample Task Resolves Zero-Shot Compound Fault Decoupling With Inconsistent Single–Fault Contributions

Mechanical system health monitoring faces significant challenges due to fault coupling effects and the combinatorial explosion of fault-type permutations. Given the relative accessibility of single-fault samples, recent research has focused on zero-shot learning for compound faults (ZSLCFs) through decoupling methodologies. However, current methodologies predominantly presume equal contribution coefficients across constituent faults, inducing feature extraction bias toward prominent signatures while overlooking weakly manifested fault patterns. Furthermore, existing evaluation protocols primarily target unseen compound faults, neglecting potential co-occurrence scenarios involving both trained and untrained single faults. To address these challenges, this study proposes a novel meta-learning-based paradigm for ZSLCF. First, we employ a feature pyramid network (FPN) to synthesize pseudo-ZSLCF tasks from single-fault samples, establishing a generalized feature space for untrained real-world compound faults through meta-learning. Second, to bridge the semantic gap between pseudo-incremental tasks and initial decoupling models, we introduce a graph neural network (GNN)-based meta-calibration module within the multistage meta-learning pseudo-sample task (MLPST) framework. This module facilitates information embedding between pretrained decoupling models and pseudo-sample collections. Finally, we incorporate a Shapley additive explanations (SHAPs)-driven feature contribution metric module within MLPST to dynamically update the weights in the compound fault decoupling network. The proposed MLPST achieves adaptive feature balancing among constituent faults while preventing catastrophic forgetting of known single faults. Extensive experimental validation on publicly available bearing compound fault datasets demonstrates that MLPST achieves state-of-the-art compound fault decoupling accuracy compared to existing advanced methods.