Self-Supervised Learning for Machinery Fault Diagnosis Under Strong Background Noise and Limited Epochs

Abstract Self-supervised learning (SSL) has been widely used for fault diagnosis of rotating machinery. However, real industrial environments often generate a large amount of noise, and existing SSL-based fault diagnosis methods are, firstly, difficult to apply to fault diagnosis under highly noisy working conditions due to the inherent scarcity of labeled samples. Secondly, most of the existing methods add prediction head during pre-training and classification head during fine-tuning, such a design makes the number of parameters and computations of the model excessive, which requires a lot of time for training and cannot meet the needs of low-end edge device deployment. To solve the above problems, this paper proposes a multi-scale channel-mixed depthwise separable convolutional neural network (MSCM-DSCNN)-based fault diagnosis model under SSL. Experimental results on bearing datasets show that the model has a significant advantage over other methods in terms of strong noise fault diagnosis accuracy under limited labeled samples, and also the number of parameters and computational cost of the model are minimized. Specifically, with only 10 pre-training epochs and 50 fine-tuning epochs under 5% labeled data, the proposed method can achieve up to 96.27% average accuracy on the PU dataset with strong noise interference.