Bearing Fault Diagnosis Based on Variational Autoencoder Under Unbalanced Data

This paper proposes a high-dimensional variational autoencoder based on feature reconstruction to address the limited stability and accuracy of fault diagnosis models caused by the unbalanced samples. This model enhances the diversity and complexity of feature expression by constructing high-dimensional distribution functions of latent space variables, thereby improving the ability to mine deep discriminative information from data; At the same time, a feature reconstruction model is introduced to alleviate the problem of data imbalance by generating high-quality samples, further enhancing the generalization performance and robustness of the model. Taking rolling bearings as the research object, conducting experiments based on publicly available datasets, and comparing the results of fault diagnosis with other mainstream fault diagnosis methods. The results show that HDVAE-FR achieves a classification accuracy of up to 97.5% under imbalanced samples, outperforming other mainstream models. HDVAE-FR can effectively improve the accuracy of bearing fault diagnosis under sample imbalance conditions.