A test-time adaptation method using evidential deep learning for online machinery fault diagnosis

The real-world deployment of deep learning models for online machinery fault diagnosis faces significant challenges due to dynamic domain shifts and the emergence of previously unseen fault types during continual operation. Conventional domain adaptation diagnosis methods often fail in such scenarios due to their reliance on extensive target data for offline training. To address these limitations, a test-time adaptation method based on an evidential deep learning framework is proposed for online fault diagnosis under dynamic operating conditions. Specifically, a spectral-entropy-based label calibration strategy is proposed to mitigate the adverse influence of overconfident pseudo-labels. Furthermore, a Fisher information-based evidential deep network is employed to model prediction uncertainty, thereby calibrating diagnosis confidence and facilitating the detection of novel faults. In addition, an information maximization objective enhances discriminative confidence and preserves output diversity. Lastly, an evidence consistency anti-forgetting mechanism is incorporated to preserve previously learned knowledge to alleviate catastrophic forgetting during continual adaptation. Comprehensive evaluations across three benchmark rotating machinery datasets verify the effectiveness of the proposed method in handling dynamic industrial data streams compared with existing fault methods.