A Novel Hierarchical Graph Neural Network With Heterogeneous Multisensor Data for Fault Diagnosis of Complex Mechanical Equipment

Modern mechanical equipment increasingly relies on multi-sensor data for health monitoring, which generates complex, heterogeneous, non-Euclidean data that conventional methods struggle to process. While graph neural networks (GNNs) are promising for such data, existing GNN-based fault diagnosis methods often suffer from two key limitations: 1) neglecting morphological details when mapping time-series data to graph structures, resulting in the loss of critical information; and 2) The lack of domain knowledge regarding fault propagation to guide the graph construction process makes it difficult to capture complex correlations among heterogeneous sensor data. To overcome these challenges, this paper proposes a hierarchical graph neural network (HGNN) for the dynamic fusion of heterogeneous data. Specifically, our method first builds sensor-level graphs by extracting discriminative shapelets and applying a pattern-distance-basedk-nearest neighbor algorithm to capture fine-grained local features. Subsequently, these individual graphs are dynamically interconnected into a comprehensive system-level graph using a dual-metric key-node screening method and a novel Feature-Type Dynamic Weighted Similarity (FTDWS) metric. This hierarchical approach preserves critical local fault signatures while effectively modeling system-wide fault propagation and sensor correlations across heterogeneous sensors. Finally, a graph attention network performs classification on this system-level representation. Extensive experiments on two public datasets and one real-world industrial coal mill dataset demonstrate the effectiveness and superiority of our proposed method.