A Hybrid Supervised Approach for Fault Diagnosis Based on Temporal and Frequency Domains

Recent intelligent deep-learning-based fault diagnosis methods can achieve great progress. However, in real-world industrial applications, the challenge to acquire a substantial amount of labeled data and the presence of non-stationary data environments make the extracted features easy to overfitting, resulting in low accuracy and efficiency of fault diagnosis. Drawing inspiration from the efficacy of artificially derived features across diverse domains, we introduce an innovative approach designed to execute hybrid supervised learning simultaneously in both the temporal and frequency domains. In the hybrid supervised learning, we leverages the contrastive learning approach from both the temporal and frequency domains to learn the disentangled features of unlabeled fault data. Meanwhile, the learned encoder is employed to process few labeled data in the feature space to achieve the final fault diagnosis. Experiments on two challenging bearing datasets highlight the superior performance of our proposed framework compared to other self-supervised methods.