A Multi-View Graph Learning Framework for Bearing Fault Diagnosis with Adaptive Fusion

Bearing fault diagnosis methods based on single sensors often suffer from reduced accuracy due to limited information. Although multi-sensor systems provide richer vibration information, the high dimensionality and complexity of these signals still pose challenges for effective feature extraction and fusion. In addition, many existing deep learning-based fusion methods rely on a single analysis domain or simple feature concatenation, making it difficult to fully exploit the complementarity among raw temporal signals, time-domain statistical features, and frequency-domain characteristics. To address these issues, this paper proposes a multi-view graph-based fault diagnosis framework with adaptive fusion, termed MDEGCN, for bearing condition identification. Specifically, non-overlapping vibration windows are treated as graph nodes, and three graph views are constructed to capture temporal proximity, time-domain similarity, and frequency-domain correlation, respectively. Each graph view is processed by an enhanced graph neural network branch to learn view-specific representations, and an adaptive, differentiable fusion mechanism is introduced to integrate complementary information from different views for final fault classification. Experiments on the Northeast Forestry University and Politecnico di Torino bearing datasets were conducted under a purged blocked split protocol to reduce potential information leakage between adjacent windows. Additional hard settings with a low training ratio further evaluate the robustness of the proposed framework under limited labelled data. The experimental results demonstrate that MDEGCN achieves competitive diagnostic performance and provides an effective multi-view representation learning strategy for bearing fault diagnosis.