An online SO2 and CS2 detection system combining weighted elliptical dual-spectrum reconstruction (WEDSR) with convolutional neural network (CNN).

Sulfur dioxide (SO2) and carbon disulfide (CS2) are indicator gases that are typically employed to assess the insulation condition of gas-insulated switchgear (GIS). However, there is significant spectral overlap between SO2 and CS2 in the ultraviolet (UV) band, which makes it difficult to quantify them simultaneously using UV differential optical absorption spectroscopy (UV-DOAS). In this paper, we propose a weighted elliptical dual-spectrum reconstruction (WEDSR) method, which enhances the feature signals by performing shunt spectral reconstruction on SO2 and CS2 mixture spectra. Our approach enables decoupling of the two spectral signals. We also introduce a weighted fitting and iterative optimization synergistic algorithm in the spectral reconstruction process to solve the problem of ellipse fitting accuracy defects triggered by the density of data points. It suppresses the noise interference under ultra-low concentration conditions effectively. On this basis, a convolutional neural network (CNN) model is established to realize the quantitative detection of SO2 and CS2. The experimental results show that the system can effectively invert SO2 and CS2 concentrations with mean absolute percentage errors (MAPE) of 0.818 % and 0.963 % for the test samples and detection limits of 7.71 ppb · m and 0.35 ppb · m, respectively. With high accuracy and stability, this system has potential application value in early warning and monitoring of faults during GIS operation.