Bearing fault diagnosis under variable speed conditions based on time series mixup and unsupervised domain adaptation

Abstract The operation of rolling bearings under practical conditions often involves variable speed, which results in non-stationary vibrations. Extracting time-varying fault frequencies effectively from these non-stationary vibration signals due to speed fluctuations poses a significant challenge when assessing model performance and stability. In order to solve this problem, this paper proposes and constructs a multi-head attentive residual network MH-ResNet18 as a feature extraction, encoder and classifier fault diagnosis method MR-CoTMix by improving the backbone network consisting of encoders and classifiers on the basis of residual network. The proposed method is applied to the fault diagnosis under variable-speed conditioned task with an unsupervised domain adaptive contrast learning method to make the training process simpler and the diagnosis more effective. This method utilizes time series Mixup data augmentation and a multi-head self-attention residual network to eliminate the need for complex manual feature extraction. By introducing unsupervised domain adaptation transfer learning, it achieves unsupervised cross-domain fault diagnosis of unlabeled target domain data under variable speed conditions using labeled source domain data. The results of two sets of experiments demonstrate that the average accuracy and MF1 score remain above 92% and 96%, respectively. The MR-CoTMix method is more effective than current mainstream unsupervised domain-adaptive transfer learning methods. The proposed model exhibits excellent performance and stability in diagnosing bearing faults under non-stationary speed conditions.