A novel cross domain deep network for unsupervised fault diagnosis of rotating machinery

Deep learning-based industrial fault diagnosis usually necessitates the training of predictive models on a large volume of labeled data to achieve high performance. However, the signals obtained when a machine runs under various working conditions have different distributions, and labels cannot be provided for data under all working conditions in actual production. Therefore, we designed an unsupervised deep-learning fault-detection model. First, building on a one-dimensional residual convolutional network, the model's performance is enhanced by focusing on the global feature receptive field, adaptive noise reduction for signals, and mitigating the loss of long-range feature information, all of which contribute to improved feature engineering; second, the model optimization integrates the classification loss from the source domain with the domain transfer loss from the target domain, facilitating continuous enhancement of accuracy in the unsupervised transfer learning process. The ablation and comparison experiments were conducted to evaluate the performance of the proposed model. The proposed model achieved a maximum accuracy improvement of 37.20 % over the original ResNet structure and the highest improvement effect of 43.02 % compared to other comparative models. Across various experimental scenarios, it exhibited remarkable advantages and superior overall performance, thereby furnishing a practical unsupervised deep-learning approach for industrial applications.