Ultratrace\nCO/SF<sub>6</sub> Detection System Based\non Differential Fourier Transform Infrared Spectroscopy Combined with\nthe Weighted Sine Spectral Reconstruction Convolutional Neural Network\n(WSSR-CNN) Model

Carbon monoxide (CO), as an indicator gas for fault diagnosis\nof\ngas-insulated switchgear, has the strongest absorption coefficient\nin the mid-infrared region, making Fourier transform infrared (FTIR)\nspectroscopy an ideal method for its concentration detection. However,\nthere is extremely abundant absorption of SF<sub>6</sub> in this region,\nresulting in a great challenge for CO/SF<sub>6</sub> detection. To\naddress this challenge, we report a high sensitivity online detection\nsystem for ultratrace levels of CO, which combines differential Fourier\ntransform infrared (DFTIR) spectroscopy with a weighted sine spectral\nreconstruction convolutional neural network (WSSR-CNN). The differential\ntechnique is first introduced into FTIR for baseline correction of\nCO absorption signals. The novel WSSR method achieves feature extraction\nand denoising of weak spectral signals in strong interference by discretizing\ninterfering signals and enhancing target signals. On this basis, the\ndetection of CO concentration is realized using the CNN model. Finally,\nWSSR-CNN is compared with four other methods. Results showed that\nWSSR-CNN outperforms all four models with evaluation index <i>R</i><sup>2</sup> reaching 0.99982 and 0.99999 in the range\nof low concentration (0.096–0.986 ppm) and high concentration\n(1.984–52.193 ppm) and mean absolute percentage error of 0.97%\nand 0.22%, respectively. The lowest detection limit of the system\nat 1σ is 13 ppb, which is the best result reported so far for\ndetecting CO/SF<sub>6</sub> concentration based on FTIR.