<p>Conventional graph construction for GNNs hinders fault diagnosis, as rigid structures fail to adapt to variable conditions or effectively integrate local and global information. To address this issue, this paper proposes the trilateral adaptive graph neural network, a novel model that follows a three-stage architecture. First, a hierarchical wavelet CNN front-end extracts multi-scale deep features from raw signals. These features are then transformed into a comprehensive graph by a novel trilateral adaptive construction strategy that incorporates spatial, temporal, and data-driven non-local similarities. Finally, a hierarchical GNN learns high-order patterns from this graph to perform fault classification. Experiments on the KAIST dataset demonstrate that the proposed model achieves an average diagnostic accuracy of 99.954 % under various operating conditions, outperforming existing methods. These results highlight the potential of the proposed adaptive graph construction strategy for enhancing the robustness and accuracy of GNN-based fault diagnosis in complex industrial scenarios.</p>