Intelligent Fault Diagnosis for Variable Working Conditions Based on SAAFN and BICP

The transferability and discriminability of unsupervised adversarial domain adaptation (UADA) significantly affect the performance of diagnosis models. Existing UADA models ignore the effects of transferability and discriminability on the model, which results in poor fault diagnosis accuracy. To address this issue, a UADA model by enhancing discriminability and transferability is proposed. The model is constructed based on two metrics in two steps to improve transferability and discriminability respectively. First, batch intraclass penalization (BICP) is used to reduce the intraclass distance and increase the interclass distance for feature vectors with large singular values, thereby increasing the discriminability of the model. Second, the stepwise adaptive align feature norm (SAAFN) is used to increase the transferability of the model by reducing the feature norm difference between the source and target domains. Four bearing datasets and one real wind turbine dataset are used and five advanced models are applied to these datasets for comparison. The results all indicate that transferability and discriminability of the proposed model both achieve the highest performance among all the used models, and its fault diagnosis accuracy is 1%–5% higher than other comparative models. The model proposed in this article can achieve better accuracy in the production environment with variable working conditions.