A novel incremental method for bearing fault diagnosis that continuously incorporates unknown fault types

When they encounter a continuous flow of unlabeled bearing faults, deep learning-based diagnostic methods fail to recognize unknown features, significantly compromising the accuracy of fault diagnosis. To address this problem, we developed an Incremental Novelty Discovery (IND) method. Continuous convolutional neural networks serve as the model backbone, considering the constraints imposed by limited numbers of past samples. We then introduce a novel form of clustering loss based on robust rank ordering to initially classify unlabeled data. The generated pseudo-labels enable the weakly supervised learning of the model to discover the novelty of unlabeled faults. To mitigate catastrophic forgetting, we implemented memory replay and a novel distillation loss approach, thus facilitating knowledge transfer. IND of unlabeled faults was enabled via staged training. Experimentally, our proposed method overcame the challenge posed by catastrophic forgetting and outperformed other methods of bearing fault diagnosis when a continuous flow of unlabeled faults must be managed over time.