Temperature prediction of cable overheating: Semiconductor gas sensor with attention-based fusion model

Accurate early warning of cable overheating is essential for reducing electrical fire risks. Semiconductor gas sensors provide a promising non-contact monitoring route because cable insulation releases characteristic decomposition gases during overheating. However, reliable localization and temperature prediction from long-sequence multi-sensor olfactory signals remain challenging. In this work, we first evaluate 23 commercial semiconductor gas sensors through multi-cycle response tests and select four sensors, namely MP503, TGS2600, TGS2620, and TGS2610, to construct a gas sensor array for cable overheating monitoring. Based on the acquired multi-sensor signals, we propose an Electrical Signal Fusion Transformer (ESFT) model. ESFT adopts a cascaded two-stage architecture that first predicts the overheating distance and then uses it as a static spatial prior to calibrate temperature prediction. A Masked Fusion Multi-Head Attention module is introduced to fuse concurrent multi-sensor features, while a gating network adaptively selects informative temporal segments and sensor channels. Experiments show that ESFT achieves an MAE of 3.793 °C and an RMSE of 5.371 °C for temperature prediction, together with an MAE of 0.021 m and an RMSE of 0.052 m for distance prediction. Compared with M-Mamba, ESFT reduces the temperature MAE from 6.570 °C to 3.793 °C, while maintaining a lightweight deployment cost of 0.358 MB memory, 10.40 M FLOPs, and 291.1 ms inference latency. Cross-cavity validation and H₂S/CO interference tests further demonstrate its robustness and practical potential for continuous cable overheating monitoring.