Real-time equipment health prediction on industrial edge devices requires predictive models that remain accurate under tight memory, energy, and latency constraints. Graph neural networks are attractive for this task because degradation signals often depend on relations among sensor channels, extracted features, or equipment units, yet high-capacity graph models are frequently too expensive for embedded deployment. This paper presents LiteGKD , a relation-preserving graph knowledge distillation framework for real-time health prediction. The framework constructs a hybrid health graph by combining weak prior topology with window-level data-driven correlations, trains a topology-biased graph Transformer teacher to capture relational degradation patterns, and distills its feature, relation, and response knowledge into a compact GraphSAGE-based student with lightweight temporal convolution. Unlike output-only distillation, LiteGKD explicitly aligns support-restricted relation distributions so that the student preserves the teacher’s dependency structure on the sparse graph it can actually use at inference time. Experiments on NASA C-MAPSS and IEEE PHM 2012 bearing datasets show that LiteGKD achieves an average C-MAPSS RMSE of 15.45, close to the teacher’s 15.05, while reducing model size to 1.62 MB and Raspberry Pi 4B inference latency to 7.8 ms. Multi-seed evaluation, additional prognostic metrics, end-to-end edge measurements, graph-corruption analysis, and teacher-quality sensitivity further show that multi-level graph distillation provides a robust accuracy–efficiency trade-off for edge-intelligent condition monitoring.
