Abstract: When abnormal conditions such as overload, harmonics, and non-periodic currents occur in electrical circuits, they can easily cause abnormal temperature rises, which may lead to electrical fires. Rapid and accurate identification of these over-temperature causes is crucial for improving the accuracy of fire warning systems and ensuring fire safety. This paper proposes an over-temperature cause identification method for electrical circuits based on Recurrence Graph-Graph Attention Network and Residual Network (ReGAT-ResNet). By leveraging the temporal dependencies of current signals under over-temperature conditions, the method maps them into high-dimensional trajectories through phase space reconstruction, and extracts temporal dynamic structural features using recurrence plot techniques. A graph neural network is then employed to model the graph-structured data, and a three-layer Graph Attention Network (GAT) is constructed. The introduction of a residual connection mechanism enhances the stable propagation and fusion of deep features, while global average pooling and a fully connected layer are used for classification prediction. The method is validated and analyzed using experimental datasets under different operating conditions involving overload, harmonics, and non-periodic currents. Experimental results show that the proposed model achieves an identification accuracy of 99.57%, can provide an effective technical foundation for early warning and fire risk prevention in electrical systems.

Key words: overtemperature triggers, trigger identification, recursive graphs, figure attention, residual networks, electrical circuits