Hybrid Semi-supervised Fault Diagnosis Based on Asymmetric and Multi-scale Pooling Module

To address the problem that fault diagnosis methods based on deep learning networks have difficulty in learning vibration image features that have large differences from visual images, this paper proposes a fault diagnosis network structure (FDNet) based on asymmetric pooling and multi-scale pooling module (MSPM). Meanwhile, a hybrid semi-supervised fault diagnosis method (HSFD) is constructed by introducing contrast learning. This method not only improves the model's feature learning capability but also improves the diagnostic performance of the model under limited labeling. In addition, to better balance the relationship between multiple losses in the semi-supervised task, this paper proposes an adaptive loss weight (ALW) calculation method, which uses the number of labeled samples and the total number of categories to calculate the weights of contrast losses. In this paper, experiments related to the proposed method are conducted using typical motor-bearing datasets CWRU and MFPT. The experimental results show that the proposed method in this paper has better performance under the same conditions.