Uncertainty-guided adversarial augmented domain networks for single domain generalization fault diagnosis

•Novel Uncertainty-Guided Adversarial Network: The paper proposes an Uncertainty-guided Adversarial Augmented Domain (UAAD) network designed for single-domain generalization in sensor fault diagnosis. The method introduces uncertainty estimation in the domain adversarial network, an uncertainty domain adversarial network (UDANet). UDANet will guides the construction of the augmented domain through uncertainty estimation, enhancing fault diagnosis performance under unknown conditions.•Special generation module and Out-of-domain auxiliary domain module: The UAAD network has a generation module constrained by the Wasserstein distance and an out-of-domain auxiliary module using multiscale random convolution. It expands adversarial samples by projecting source samples onto drifting domains, creating pseudo-samples and forming a "fictional domain" for learning.•Domain Adversarial Training with Uncertainty Estimation: The method incorporates domain adversarial training balanced by uncertainty estimation. This iterative optimization process ensures the augmented domain effectively simulates unseen domains, improving the model’s ability to generalize and maintain robust diagnostic performance.•Experimental validation under unknown working conditions that include known and unknown fault modes: Experiments using real industrial sensor data from a nickel flash furnace validate the proposed method’s superiority. The UAAD network demonstrates significantly higher accuracy in classifying known and unknown fault modes under various working conditions compared to other state-of-the-art methods, achieving an average H-scores of 90%.•Single domain generalization more closely with real-world industrial processes: To relaxing the requirement for multiple source domains and focusing on single domain generalization, the UAAD method offers practical value in industrial applications where data from multiple conditions are challenging to obtain. This approach allows for effective fault diagnosis with limited training data, enhancing its applicability in real-world industrial processes. Domain generalization methods can address cross-domain diagnostic issues in industrial sensors. However, in practical industrial environments, gathering sensor data across difference working conditions is particularly challenging, and new sensor fault modes may arise under unknown working conditions. This paper proposes an uncertainty-guided adversarial augmentation domain network for fault diagnosis