Cyclic Auto-encoder: A Novel Semi-supervised Method for Rotating Component Fault Diagnosis with Low Label Rate

Deep learning-based intelligent fault diagnosis methods have gained significant attention in fault diagnosis for rotating components. However, the limited labeled data in practical engineering possess a challenge in achieving satisfactory diagnosis results. To overcome this challenge, this study proposes a novel semi-supervised fault diagnosis method using a cyclic auto-encoder (CAE) network. Firstly, the CAE utilizes the constraint on the feature of the reconstruction samples to learn the feature distribution of limited labeled samples through unsupervised training. Additionally, a classifier is constructed based on the encoder of the well-trained CAE. Finally, fine-tuning is performed based on the limited labeled data, which allows for effective fault recognition even under low label rate conditions. Experimental data from bearings is used to validate the effectiveness of the proposed method. The results demonstrate that the proposed method can significantly improve the performance of the auto-encoder. Furthermore, it outperforms other comparable semi-supervised learning methods, particularly in scenarios with low label rates.