Frequency-Assisted Contrastive Learning With Cyclic Fine-Tuning for Rotating Machinery Fault Diagnosis Under Limited Labeled Data

Abstract

The scarcity of labeled data severely limits the performance of traditional supervised learning-based intelligent methods for rotating machinery fault diagnosis. Recently, contrastive learning (CL) has been successfully applied to rotating machinery fault diagnosis due to its capability to learn data representations without the need for labeling information. Nevertheless, existing CL-based diagnostic methods primarily suffer from two drawbacks. First, most methods merely depend on time-domain views of signals while neglecting the significance of the frequency domain, leading to weakly discriminative representations. Second, the conventional fine-tuning strategy only relies on labeled data while discarding the unlabeled one, impairing model generalization. To tackle these drawbacks, this article proposes a novel diagnostic method based on frequency-assisted CL (FACL) and cyclic fine-tuning (CFT). In the pretraining stage, FACL utilizes the time-frequency relational contrastive loss to correlate the time view with the frequency view, allowing the frequency-view data to effectively assist the time-view contrastive learning, thus yielding representations with enhanced discriminability. In the fine-tuning stage, CFT enables both labeled and unlabeled data to contribute to model fine-tuning through pseudo labeling, enhancing model generalization. Additionally, a weighted pseudo-labeling (WPL) strategy is devised to alleviate the defect of noisy labels. Comparative and ablation experiments are conducted on two public and one self-organized datasets. The superiority of the proposed method in diagnosing faults for rotating machinery with limited labeled data is evidenced by a 39.77% accuracy improvement over supervised learning and a 28.61% improvement over other CL-based methods.