A novel meta fusion detective twin network approach for dynamic degradation process in rotating Machinery

Li Zou, Huai Chen, Jun Hu, Kejia Zhuang

Abstract: The generation of faults in rotating machinery is a continuously dynamic process that evolves over time. Under complex environmental factors, faults can develop in unpredictable directions, making the occurrence of rare fault instances (RFI). Moreover, due to the shared physical foundation, there may be some correlations between the characteristics of different fault modes. Therefore, the scarcity of RFI data, along with the feature overlap between data from different fault modes, presents significant challenges for fault diagnosis techniques. To address this challenge, this study proposes a meta fusion detective twin network (MFDTN). First, a twin network model is constructed, incorporating an adversarial training mechanism to learn invariant features from different but correlated fault data. Leveraging a meta-learning framework and a pre-trained invariant feature extractor, a classification model is then developed, allowing for rapid adaptation to new fault tasks based on existing fault data, thus improving diagnostic performance under RFI. Finally, tested on the rotor rolling bearing and computer numerical control machining center bearing datasets, the results of MFDTN outperformed related methods, demonstrating superior training efficiency (133 s and 120 s), fault recognition accuracy (99 % and 98 %), and Micro-F1 scores (0.99 and 0.98).