Adversarial domain adaptation convolutional neural network for intelligent recognition of bearing faults

Varying working condition leads to the data distributions offset between training (source domain) and testing (target domain), which results in insufficient capability of traditional intelligent recognition methods for rotating machinery. Aiming at this problem, a novel intelligent recognition method based on adversarial domain adaptation convolutional neural network (ADACNN) is proposed for bearing faults. First, a weight-sharing convolutional neural network (CNN) is constructed to map training data from both source and target domains to a feature and predicted label spaces, respectively. Then, adversarial learning and maximum mean discrepancy (MMD) are separately introduced in the feature and predicted label space for domain adaptation, and the ADACNN model is established. Finally, a method based on ADACNN, which can realize domain adaptation in two different spaces, is proposed for intelligent recognition of bearing faults. The proposed method is validated by various cross-domain bearing fault recognition tasks under variable speeds and loads. Compared with shallow models, CNN, and commonly used domain adaptation methods, the proposed method has more than a 4% fault recognition accuracy advantage under varying working conditions.

Automated summary

This study proposes an intelligent recognition method based on ADACNN to overcome the limitation of traditional methods in bearing fault detection caused by variations in working conditions.