Few-labeled bearing fault diagnosis based on attention mechanism time-frequency prediction self-supervised learning framework

Bearings, as critical components of rotating machinery, have operational states that directly impact the safety of the entire mechanical system. However, due to the low frequency of faults in industrial equipment during actual operations, obtaining high-quality labeled fault data is time-consuming and costly. This study proposes a Self-Supervised Learning Framework for Time-Frequency Prediction based on an Attention Mechanism (SE-TFPred) to address the challenge of bearing fault diagnosis under few-labeled conditions. By extracting latent fault representations from unlabeled fault data, this method aims to achieve bearing fault classification. Time-domain and frequency-domain features are extracted, and frequencydomain features are further processed for model pre-training using the SE-TFPred framework, which is composed of a prediction network, a temporal encoder, and a frequency encoder. A little quantity of labeled data is then used to refine the previously trained model. The framework was validated on the Paderborn University Datasets (PU), and experimental results demonstrate that SE-TFPred excels in feature extraction and fault classification, achieving significant performance using only 1% of labeled data, surpassing current mainstream framework models.