Dynamic weighted adversarial domain adaptation network with sparse representation denoising module for rotating machinery fault diagnosis

Domain adaptation effectively addresses the issue of varying vibration data conditions in fault diagnosis. However, current domain adaptation methods rarely consider the impact of imbalanced distributions and background noise, limiting their applicability. To alleviate these problems, a dynamic weighted adversarial domain adaptation network with the sparse representation denoising module (DWADAN-SRDM) is designed for rotating machinery fault diagnosis. First, a sparse representation denoising module is designed to effectively suppress noise, enabling more accurate extraction of vibration signal features. Then, an adaptive exponential loss is introduced address class imbalance. Finally, a weighted adversarial domain adaptation strategy, integrated with a dynamic sample weighting mechanism that can assign greater weight to source domain samples with higher relevance to the target domain, is used to extract domain-invariant features while reducing negative transfer. Extensive experiments on three unbalanced rotating machinery datasets demonstrate that the fault identification accuracy of DWADAN-SRDM improves by over 8.5% compared to existing methods.