Fault Diagnosis Method for Rotating Machinery Based on Attention Adversarial Transfer Networks Under Variable Operating Conditions

Recently, deep learning-based fault diagnosis methods for rotating machinery have achieved promising classification performance under the independent and identical distribution of training and test data. However, the complex and variable operating conditions, including bearing speed, working load, and load torque, of rotating machinery in practical applications lead to the problem of inconsistent distribution between training and testing data. To address the above issue, an attentional adversarial transfer network is developed to improve the accuracy of rotating machinery fault diagnosis under variable operating conditions by using adversarial transfer learning techniques. First, a feature extractor based on self-normalized convolutional neural networks (SCNN) is constructed to provide effective features for different domains. A lightweight attention layer is further designed to reinforce the structural information in the features while reducing the redundant disturbances. Then, a discriminator based on probability calculation is employed to perform adversarial learning with the feature extractor so as to achieve the alignment of feature distribution between different domains. Finally, an efficient loss function is designed to drive efficient adversarial transfer learning for the whole model. To verify the effectiveness and superiority of the proposal, the Paderborn University (PU) bearing dataset is selected for fault diagnosis experiments, where proposed method achieves an average cross-domain fault diagnosis accuracy of 88.36%. The experimental results indicate that this approach can implement accurate fault identification for rotating machinery under variable operating conditions.