Self-Supervised Learning Denoising Network For Intelligent Fault Diagnosis With Limited Labeled Data

While supervised learning has received tremendous attentions in many real applications due to its successful use of huge labeled data, in fault diagnosis it is impossible to obtain enough labeled fault data samples under each operating condition for training deep learning models. To address the issue of limited labeled data, we propose a few-shot learning approach for efficient fault diagnosis, which is referred to as Denoising Simple Siamese (DNSS). The proposed DNSS involves a denoising module that effectively integrates soft thresholding and deep learning techniques, leading to exceptional performance in noise reduction, feature extraction, and fault identification. We conducted extensive experiments to evaluate the effectiveness of DNSS under various noise conditions and with varying amounts of labeled and unlabeled data. The experimental results demonstrate that DNSS can maintain high diagnostic accuracy with limited labeled data under a wide range of operating conditions.