Adaptive multi-scale graph fusion for fault diagnosis in concurrent variable conditions and unlabeled data

Abstract

Listen

Translate article

Deep learning methods, enhanced by unsupervised domain adaptation (UDA) for fault diagnosis, effectively address individual challenges of variable conditions or unlabeled data, yet encounter difficulties when these conditions are concurrent. The separation of feature extraction from domain adaptation in these methods not only creates information silos but also diminishes diagnostic accuracy. This situation is further aggravated by the generation of unreliable pseudo-labels for unlabeled data. This paper introduces a new fault diagnosis method based on adaptive multi-scale graph fusion (AMSGF). Firstly, the method uses multi-scale wavelet-optimized sparse filtering for feature extraction and optimal scale determination through cross-validation. Secondly, it constructs data graphs, refined with maximum mean discrepancy, to strengthen feature associations. Lastly, a classifier is trained on these enhanced features for fault diagnosis. Experiments with electro-hydraulic actuators and bearing dataset show AMSGF’s superiority. It notably enhances diagnostic accuracy in variable conditions with unlabeled data, outperforming current UDA methods.