Fault Diagnosis of Motor Bearing Under Variable Operating Conditions Based on CDAN-SM

The vibration signal data distribution varies greatly when the motor bearing fails under different working conditions and the signal acquisition is also susceptible to noise interference. These two factors often lead to low accuracy of the cross-domain fault diagnosis task. In order to handle these problems, a novel motor bearing fault diagnosis method based on conditional domain adversarial network with smoothing optimization and minimum class confusion is proposed in this paper. First, the conditional adversarial domain adaptation regulates the domain discriminator and reduces the feature distribution differences between the source and target domains. Second, smoothing optimization is introduced into the network to make the classifier task loss converge to the smoothed minimum value, which stabilizes the adversarial training. Finally, the minimum class confusion is utilized to further improve the classifier’s ability to discriminate different classes of samples in the target domain. Based on the minimum class confusion loss and the smoothed classification loss, the network parameters can be updated via backpropagation progress to effectively reduce the target risk of the classifier. Extensive experiments are conducted on the public Case Western Reserve University (CWRU) dataset, and experimental results show that the proposed method outperforms the compared methods in solving the problem of electric motor bearing fault diagnosis under variable operating conditions.