Open-set fault diagnosis of rolling bearings via class similarity-guided graph convolutional adversarial network with adaptive channel feature relation fusion

To address the challenges in open-set fault diagnosis of rolling bearings, such as the vague demarcation of features between known and unknown class faults, as well as the difficulty in capturing the latent correlations among samples, this paper proposes a class similarity-guided graph convolutional adversarial network with adaptive channel feature relation fusion. First, an adaptive channel–feature relation fusion graph generation strategy is designed. This strategy employs a dual-path approach combining topological encoding networks and multi-head attention networks to capture inter-sample feature correlations, along with learnable parameters for weighted fusion, thereby enhancing the modeling capability for complex data structures. Meanwhile, a multi-receptive field graph convolutional network is adopted to effectively aggregate multi-scale feature information and improve feature discriminability. Second, a class similarity-guided discrimination mechanism is introduced, which integrates Kullback–Leibler divergence and an auxiliary discriminator to quantify sample uncertainty and inter-domain similarity. A dynamic weight encoding strategy is constructed to optimize adversarial training and sharpen the decision boundary between known and unknown faults. Finally, multiple open-set transfer tasks are designed on two bearing fault datasets, and comparative experiments are conducted with several mainstream methods. Through experiments, it can be shown that the diagnostic accuracy of this method significantly outperforms that of the comparison methods, verifying its potential in practical industrial applications.