Prior knowledge-informed multi-task collaborative learning for few-shot fault diagnosis of aero-engines

As the core power system of aircraft, the fault prediction and health management of aircraft engines are of great significance in ensuring flight safety and optimizing maintenance strategies. Existing research faces the dual challenges of scarcity of real flight fault data and cross-domain feature differences. At present, most few-shot cross-domain fault diagnosis methods focus on efficient fault feature extraction and model structure optimization, while insufficiently leveraging diagnostic knowledge accumulated in the field over time. A key issue that remains unresolved in few-shot cross-domain fault diagnosis for aero-engine bearings is how to integrate valuable prior knowledge with effective cross-domain feature alignment methods into the diagnostic model. This study introduces a prior knowledge-informed multi-task collaborative learning (PKMTCL) approach. First, a cosine contrastive loss is introduced to implicitly embed prior diagnostic knowledge into the model, thereby reducing its dependence on large training datasets. Then, a novel information entropy-based prototype construction and cross-domain feature alignment strategy for the target domain is designed, effectively alleviating feature shift under varying working conditions. Finally, a multi-task collaborative learning framework is developed, where the inductive bias provided by auxiliary tasks guides the main task to learn more generalizable feature representations, thereby effectively improving the generalization performance of the main task. Experiments on two aero-engine bearing datasets demonstrate that, compared with state-of-the-art methods, the proposed method achieves higher fault identification accuracy and lower volatility in diagnostic results. The related code can be downloaded from https://github.com/LZJHIT/PKMTCL .