Graph-Based Model Compression for HSR Bogies Fault Diagnosis at IoT Edge via Adversarial Knowledge Distillation

Deep graph neural networks (GNNs) have demonstrated their exceptional expressive capability in detecting fault features from multisensor signals for fault diagnosis. However, the excessive depth of these models often hinders their deployment on Internet of Things (IoT) systems. In order to facilitate the implementation of graph-based models on IoT devices for fault diagnosis of HSR bogie, this paper proposes a novel compression technique called GraKD. GraKD distills the latent knowledge from the teacher model to a lightweight student model in an adversarial manner. The discriminator of GraKD comprises a representation identifier and a logit identifier. The former effectively distinguishes between the node representations of the student and the teacher by discerning the local affinity of connected node patch-patch pairs and the global affinity represented by the patch-global pairs. The latter employs a residual multi-layer perceptron block to differentiate between the logits of the student and the teacher. The effectiveness of GraKD is validated using data collected from test rigs of bogies. A series of experiments substantiate the broad applicability of GraKD in compressing various GNN-based models.