Fault Diagnosis of Few Sample Rolling Bearings Based on SE Attention and Similarity Contrastive Learning

In recent years, deep learning has demonstrated excellent performance in the diagnosis of rolling bearing faults due to its outstanding feature extraction capabilities. However, when faced with limited small sample datasets, this technique still faces challenges in feature learning and fault type discrimination. This study proposes a new learning framework, namely the SE Attention based similarity comparison method. In this framework, we extract features from scarce samples through an encoder, and then use the similarity and difference between labeled samples to construct positive and negative sample pairs. The encoder further extracts features with the assistance of contrastive learning, and then completes fault classification through a classifier. In addition, we introduced the SE Attention mechanism to enhance the efficiency of feature extraction. This framework significantly reduces the manpower and time required for annotating samples. Experimental results have shown that the bearing fault diagnosis model proposed in this study has high accuracy and effectiveness.