Domain adaptation-based multi-task model incorporating mixture-of-experts and feature decoupling for unseen compound fault diagnosis of gearbox

Abstract The high coupling and strong concealment characteristics of compound faults (especially multi-component compound faults) make them a 'hidden threat' to the safe service of equipment. In the absence of well-labeled compound fault data, how to attain accurate recognition of rotating machinery compound faults has consistently been a challenge. Consequently, a domain adaptation-based multi-task model incorporating mixture-of-experts and feature decoupling is investigated to diagnose unseen multi-component compound faults in gearboxes. Firstly, an attention-guided mixture-of-experts network, composed of shared experts and private experts, is designed for low-level feature learning, so that the model can better capture both commonalities and differentiated features according to different sub-tasks. Subsequently, a private-task feature extraction module based on convolutional residual blocks is constructed, and by introducing feature decoupling operations, the distinctiveness of features across different tasks is increased, ensuring each sub-task focuses more intently on its relevant key features. Meanwhile, a subtask-oriented domain adaptation learning strategy is designed by using training data with single faults and test data with unseen compound faults, which can equip the model to capture task-specific domain-invariant knowledge, thereby enhancing its ability for decoupled identification of unknown compound faults. Finally, a gearbox fault dataset is adopted to analyze the studied method in detail. Experimental validation against some advanced methods confirms the model's superior capability in identifying unseen compound faults, thus establishing a viable methodology for fault diagnosis of complicated mechanical system.