A Study on Domain Generalization Methods for Rolling Bearing Fault Diagnosis Based on Mean Differences and Domain Adversarial Learning
TL;DR
A multiscale one-dimensional convolutional network is employed to extract fault impact features and periodic features across different time scales, and mean difference constraints between source domains are introduced in the feature space to reduce the offset in the centers of feature distributions across different source operating conditions.
Abstract
To address the issue that intelligent fault diagnosis models for rolling bearings are susceptible to the effects of load and speed variations, as well as shifts in data distribution, when applied across different operating conditions, this paper proposes a fault diagnosis domain generalization method that combines mean difference constraints with domain adversarial learning. Taking raw one-dimensional vibration signals as input, this method first employs a multiscale one-dimensional convolutional network to extract fault impact features and periodic features across different time scales; Subsequently, mean difference constraints between source domains are introduced in the feature space to reduce the offset in the centers of feature distributions across different source operating conditions; simultaneously, a domain-adversarial discriminator based on gradient inversion layers is constructed to weaken the operating condition discriminative information in the feature representations, thereby prompting the model to learn domain-invariant features that possess both fault class discriminability and operating condition insensitivity. Experiments were conducted using two rolling bearing datasets from Case Western Reserve University (CWRU) and Paderborn University (PU) to establish a leave-one-out cross-condition diagnosis task. The results show that the proposed method achieves an average accuracy of 99.08% ± 0.19% on the CWRU dataset, and an average accuracy of 91.86% ± 0.34% on the PU dataset, both outperforming comparison methods such as SVM, 1D-CNN, ResNet1D, TCN, CORAL, MMD, and DANN. Furthermore, a Welch’s t-test based on summary statistics indicates that the improvement over the best baseline method is statistically significant. Ablation experiments and parameter sensitivity analysis further validate the effectiveness of multiscale feature extraction, mean difference constraints, and domain adversarial learning in enhancing the model’s generalization capability across different operating conditions.