Peak-Hold-Aided Evaluation Approach for the Insulation State of Vehicle Cable Terminals via Adaptive Spatial–Temporal Graph-Gated Attention Networks

Existing assessment methods for vehicle cable terminals suffer from accuracy degradation due to insufficient ability to extract spatial and temporal features, and inability to be installed on resource-constrained high-speed electric multiple units (high-speed EMUs) due to huge parameters of the neural network model. A peak-hold-aided evaluation approach to acquire the insulation state of vehicle cable terminals via an adaptive spatial–temporal graph-gated attention network (AdaSTGGAT) is proposed in this article. Initially, an approach for the construction of temporal graph structures based on peak hold and Jensen-Shannon divergence is proposed to address the issues of large amounts of data and the difficulty of analyzing graph structures. The graph-gated attention block is subsequently introduced for feature extraction to extract the spatial-temporal features of the signal accurately. Furthermore, AdaSTGGAT is designed as the backbone for feature extraction and aggregation. Specifically, the supervision mechanism is introduced to stabilize the training process, which enables the model to be pruned. Ultimately, the proposed approach is experimentally validated via HFCT experiments on vehicle cable terminals. The proposed methodology demonstrates robust spatial-temporal feature representation capabilities for insulation condition assessment for vehicle cable terminals, achieving average performance metrics of 97.40% accuracy, 97.50% precision, 97.40% recall, 97.40% F1-score, and 96.90% Kappa coefficient. Notably, the architecture incorporates a data-driven dynamic pruning mechanism that optimizes model complexity while maintaining diagnostic reliability, thereby advancing the practical implementation of online insulation monitoring systems in industrial applications.