A self-supervised domain-adversarial graph convolutional network with temporal KNN graph construction for generalizable bearing condition recognition across multiple scenarios

Bearings are critical components in rotating machinery, and their condition recognition is essential to ensure operational safety and reliability. However, vibration signals collected under different operating scenarios often exhibit non-stationarity and distribution discrepancies, which limits the generalization ability of conventional diagnosis models. To address this problem, this paper proposes a self-supervised domain-adversarial graph convolutional network (DAGCN-SSL) with temporal KNN graph construction, for multi-scenario bearing condition recognition. First, vibration signals are segmented by a sliding window, and time-domain, frequency-domain, and wavelet packet energy features are extracted to construct multi-domain node representations. Then, a temporal KNN graph is built to model the sequential degradation relationships among samples. Based on the constructed graph, a weighted GCN is developed to learn discriminative degradation features. In addition, a domain-adversarial mechanism based on a gradient reversal layer is introduced to reduce distribution discrepancies across scenarios, while a masked feature reconstruction task is designed as a self-supervised branch to enhance feature robustness. Experiments are conducted on the self-collected low-speed heavy-load bearing dataset, the laboratory-made small-size high-speed bearing dataset, and the real-world wind turbine bearing monitoring dataset. These three datasets cover low-speed heavy-load operation, small-size high-speed accelerated degradation, and real-world wind turbine monitoring, respectively, providing complementary scenarios for evaluating the applicability of the proposed method. The proposed method achieves accuracies of 98.65%, 97.53%, and 98.76% on the three datasets, respectively. The results show that DAGCN-SSL can effectively recognize different degradation stages and is strongly applicable to bearing condition recognition under multiple scenarios.