GIS gas-component detection based on a WGM gas-sensing microcavity and a one-dimensional convolutional neural network

Quantitative detection of SF₆ decomposition gases in gas-insulated switchgear is affected by nonlinear multicomponent spectral responses and environmental variations. A fiber-coupled whispering-gallery-mode (WGM) gas-sensing system integrated with a multi-scale residual one-dimensional convolutional neural network (MSR-1D-CNN) was developed for simultaneous concentration inversion of SO₂, H₂S, and CF₄. Parallel convolution branches with kernel sizes of 3, 5, and 7 provide complementary spectral receptive fields, while depthwise-separable convolution controls computational complexity and residual connections preserve lower-level spectral information. Temperature, relative humidity, and pressure are further fused with the spectral representation to compensate for environmental variations. The concentration-grid dataset was partitioned according to independent gas-preparation batches, and additional environmental conditions were independently prepared for calibration and held-out generalization evaluation. The proposed model achieved an RMSE of 1.82 ± 0.05 ppm, an MAE of 1.52 ± 0.04 ppm, and an R 2 of 0.995 ± 0.001 on the internal test set. The independent-date test yielded an average RMSE of 2.14 ppm. Under the held-out conditions of 55 °C, 70% relative humidity, and 0.13 MPa, RMSE values of 2.26, 2.14, and 2.31 ppm were obtained, respectively. The deployed FP32 model contained 0.186 M parameters and required 37.8 ms per model inference on a Cortex-M7 processor. These results demonstrate the feasibility of WGM-based multicomponent gas concentration inversion and embedded inference under the investigated laboratory conditions.