Heterophily-Adaptive Graph Neural Networks with Structural-Aware and Dynamic Residual Propagation for Intelligent Diagnosis

Accurate and efficient fault diagnosis for high end equipment remains a significant challenge, primarily due to-heterogeneous data and complex temporal dependencies in industrial monitoring environments. To address these challenges, a novel structural-aware graph neural network framework is proposed. It transforms multi-source time-series signals into graph-structured data using a hybrid construction strategy that integrates temporal sequencing with similarity-based relationships. This approach enables the model to capture both short-term temporal dependencies and long range global patterns within co- mplex mechanical systems. The framework incorporates two key mechanisms to enhance diagnostic performance: Structural-Aware Similarity Modeling (SASM) and Locally Adaptive Residual Feature Propagation (LARFP). SASM dynamically adjusts feature aggregation based on node heterogeneity, enhancing feature discriminability in heterogeneous graphs. LARFP mitigates the over-smoothing problem commonly found in deep graph neural network, ensuring robust feature preservation and effective multi-layer information fusion. Experimental results on three benchmark datasets demonstrate that the proposed framework consistently outperforms existing methods in classification accuracy, robustness, and computational efficiency. The results confirm the framework’s potential for deployment in intelligent health monitoring and predictive maintenance applications for high-end equipment.