Dual-domain Signal Fusion Adversarial Network Based Vision Transformer for Cross-domain Fault Diagnosis

Many of the current fault diagnosis methods for intelligence rely on either a signal in the time domain or a signal in the frequency domain. However, frequency-domain signals often fail to fully capture the operational condition of bearings, despite their intuitive nature. On the other hand, time-domain signals contain valuable equipment status information, but their complexity poses challenges for learning by networks. To overcome the challenge, this paper proposes a dual-domain signal fusion adversarial network (DSFAN) based on Vision Transformer. The network is constructed with a parallel feature extractor based on the Vision Transformer. The time domain and frequency domain signal features are integrated considered after parallel extraction and fusion to achieve cross-domain fault diagnosis of bearings. Meanwhile, to enhance the network's ability to extract domain-invariant features, an adversarial training strategy is employed in conjunction with the Wasserstein distance. In order to evaluate the performance of DSFAN, six cross-domain fault diagnosis tasks were conducted for bearing analysis. Compared to other comparison methods, the experimental results demonstrate that the proposed method exhibits superior diagnostic accuracy and domain alignment performance.