Switching thermal cycle modulation and cycle feature fusion for mixed gas concentration prediction

Accurate prediction of methane and acetylene concentrations supports insulation condition assessment and fault diagnosis in power equipment. Metal oxide semiconductor (MOS) sensors offer compact, low-cost sensing, but cross-sensitivity limits their accuracy in mixed-gas concentration prediction. To improve prediction accuracy while simplifying temperature modulation hardware, a Switching Thermal Cycle Modulation (STCM) device and a Cycle Feature Fusion Network (CFFNet) are developed. STCM uses a timer-based circuit to periodically switch the sensor heaters on and off, producing temperature variations that induce periodic changes in gas response waveforms. CFFNet uses cycle means and fixed cosine projections to represent response intensity and waveform shape, respectively, and a Transformer encoder to model their temporal relationships across cycles during gas exposure and recovery. Performance was evaluated using nested leave-one-batch-out cross-validation on STCM and constant-temperature (CT) datasets, each comprising five acquisition batches. On the STCM dataset, CFFNet achieved mean coefficients of determination (R 2 ) of 0.9973 for methane and 0.9974 for acetylene, outperforming all four baseline models on this metric for both gases. Compared with CT, STCM improved concentration prediction accuracy across all evaluated models, demonstrating the value of modulation-induced response information for distinguishing gas concentrations. Ablation experiments supported the complementary roles of cycle mean and shape features, with cross-cycle temporal modeling providing a greater improvement in methane prediction. These results support combining simplified temperature modulation with cycle feature modeling to improve gas concentration prediction under the evaluated conditions.