Diffusion-Enhanced Dual-Domain Adversarial Network: A Zero-Shot Fault Diagnosis Method for Electrohydrostatic Actuators

To address the zero-shot fault diagnosis challenge in electrohydrostatic actuator (EHA), a diffusion-enhanced dual-domain adversarial network (DEDDAN) is proposed. Acquired experimental signals were stacked and reorganized in the time domain into grayscale images. A denoising diffusion probabilistic model (DDPM) was then used to augment the simulation image dataset generated from an EHA simulation model following identical preprocessing. The training set for DEDDAN was constructed by combining experimental normal-condition grayscale images with the augmented simulation dataset. Two domain discriminators were designed to focus on global and local features, while weighting of loss functions balanced training objectives. Experimental results show DEDDAN achieves 95.71% accuracy, outperforming purely data-driven and single-discriminator adversarial methods.