A GNN-Based Enhanced Learning Framework for Node Classification in Radar Fault Knowledge Graphs

To address the dual challenges of insufficient fault node classification accuracy and low semi-automation efficiency in radar fault knowledge graph construction, this paper proposes a collaborative learning framework termed UL-GSE integrating unsupervised pre-training with graph structure enhancement. The framework enhances node representation through a dualmodule architecture: (1) The preprocessing module employs a two-layer fully connected network for feature dimensionality reduction, combined with unsupervised clustering for pretraining; (2) The feature enhancement module integrates multihead graph attention mechanisms with residual networks to prevent overfitting while preserving topological information. Experimental results on a self-constructed radar fault knowledge graph (containing <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$X$</tex> nodes and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$Y$</tex> edge types) demonstrate that the proposed method achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$97.19 \% (\pm 0.8 \%)$</tex> node classification accuracy, representing 6.29 -point and 34 -point improvements over MLP (90.9%) and GAT (63.19%) baselines, respectively. This framework provides an interpretable and resource-efficient solution for industrial knowledge graph construction under small-sample scenarios.