Dual-domain fused vision transformer for mechanical fault diagnosis under fluctuating working conditions

Abstract Signal instability, noise interference and the inability of individual signals to accurately represent the operational state of the equipment present substantial challenges for traditional fault diagnosis methods under fluctuating conditions. Furthermore, the over-globalization of the self-attention mechanism in deep networks undermines feature diversity, while conventional inter-domain distance metrics lack the robustness required to address complex scenarios. To tackle these issues, we propose an innovative adversarial training network model based on a dual-domain fused vision transformer, which processes time-domain and frequency-domain signals in parallel. The model enhances feature diversity through a residual attention mechanism and dynamically fuses signals via self-attention. By co-optimizing a dual-domain distance metric that incorporates maximum mean discrepancy and Wasserstein distance to reduce domain distance, the ability of the model to transfer across domains is strengthened. Experimental results from three sets of tests demonstrate that the proposed model significantly improves fault diagnosis accuracy under fluctuating conditions, effectively mitigates instability and noise interference, and exhibits enhanced stability and generalization capabilities.