Intelligent Fault Diagnosis Using Graph Neural Networks

Intelligent fault diagnosis harnesses advanced machine-learning techniques to detect and classify anomalies in industrial and mechanical systems, thereby safeguarding performance and preventing unplanned downtime. Traditional methods often treat sensor data as independent time series, overlooking the complex interconnections among components. Graph neural networks (GNNs) address this limitation by representing measurement points or signal features as nodes and their relationships as edges, enabling end-to-end learning on non-Euclidean data structures. By propagating and aggregating information across a graph, GNN-based approaches can capture both local and global patterns of system behaviour, improving sensitivity to subtle fault signatures. Recent developments integrate attention mechanisms, adaptive signal decomposition and domain-adaptation strategies to enhance robustness under noise, variable operating conditions and scarce labelling. These innovations have broadened the applicability of intelligent fault diagnosis to rotating machinery, bearings, gearboxes and other critical systems, delivering higher accuracy and faster response times in real-world industrial settings.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Intelligent Fault Diagnosis Using Graph Neural Networks publication trend

The graph below shows the total number of articles in intelligent fault diagnosis using graph neural networks across all publications each year (not limited to Nature Index journals).

Technical terms

Graph Neural Network (GNN): A neural-network framework designed to operate on graph-structured data, learning representations by iteratively propagating information between connected nodes.

Graph Convolutional Network (GCN): A subclass of GNNs that generalises the convolution operation to graphs, aggregating neighbourhood node features to update each node’s representation.

Graph Attention Network (GAT): A GNN variant that employs attention scores to weight the influence of neighbouring nodes dynamically during the aggregation process.

Variational Mode Decomposition (VMD): An adaptive signal-decomposition technique that separates a complex signal into variationally determined intrinsic mode functions.

Unsupervised Domain Adaptation (UDA): A strategy to transfer knowledge from a labelled source domain to an unlabelled target domain, minimising performance loss due to domain shift.

Adjacency Matrix: A matrix representation of a graph where each element indicates the presence or weight of an edge between a pair of nodes.

References

1. A two-stage importance-aware subgraph convolutional network based on multi-source sensors for cross-domain fault diagnosis. Neural Networks (2024).
2. A Rolling Bearing Fault Diagnosis Method Based on the WOA-VMD and the GAT. Entropy (2023).
3. The combination model of CNN and GCN for machine fault diagnosis. PLOS ONE (2023).