Self-supervised contrastive learning with time-frequency consistency for few-shot bearing fault diagnosis

Abstract Deep learning technology has made significant progress in fault diagnosis. However, in real-world industrial settings, most existing methods require substantial labeled data for training, while harsh operating conditions and data collection constraints often result in scarce fault samples. This limitation significantly impairs their diagnostic performance in practical applications. To address this challenge, we propose a few-shot fault diagnosis approach based on a time-frequency contrastive learning (TF-CL) framework. The TF-CL framework adopts a pre-training and downstream task pipeline, enabling the model to automatically learn and extract multi-perspective features from unlabeled data in self-supervised conditions. During the pre-training, dedicated encoders separately extract time-domain and frequency-domain feature representations from abundant unlabeled samples. The extracted features are then projected into a shared time-frequency space using a projector. To ensure that multi-perspective features can be extracted from unlabeled data, this paper introduces a time-frequency consistency loss function, constructed using novel positive and negative sample pairs. In the downstream task, the TF-CL model is combined with a multilayer perceptron classifier and optimized fine-tuned end-to-end using the limited labeled data. Gradient updates during downstream training further refine the learned feature representations, enhancing their adaptability to target classification tasks. The superiority of TF-CL was demonstrated through a variety of fault diagnosis experiments conducted on both public and self-collected datasets.