Prior Knowledge-Augmented Meta-Learning for Fine-Grained Fault Diagnosis

In existing fault diagnosis methods, fault categories are generally coarse-grained, which may result in failure to precisely identify fault details. Therefore, fine-grained fault diagnosis is explored in this study. In this scenario, the following three challenges arise, first, few-shot learning, second, less distinguishable, third, new fault emerging. To cope with these challenges, a prior knowledge-augmented meta-learning method for fault diagnosis is proposed in this study. The process of prior knowledge augmentation occurs in both the training and test phases. In the training phase, there are two processes: fault mechanism pretraining and true label fine-tuning. The fault mechanism pretraining process is employed to integrated prior knowledge of fault mechanism, where the pseudolabels are constructed by the values of 14 time-domain indicators. Subsequently, in the true label fine-tuning process, the pretrained model is fine-tuned by the training set with class labels. In the test phase, the heatmaps generated by gradient-weighted class activation mapping serve as prior knowledge. By integrating prior knowledge in the training and test phases, the proposed method presents a novel approach to addressing the challenge of fine-grained fault diagnosis in industrial applications. In addition, the deployment of prior knowledge to both the training and testing sets enhances the interpretability of the method. The proposed method is evaluated on rolling bearing and gearbox datasets and exhibits a remarkable capability to identify new fine-grained fault categories from only a few samples.