Ultratrace CO/SF6 Detection System Based on Differential Fourier Transform Infrared Spectroscopy Combined with the Weighted Sine Spectral Reconstruction Convolutional Neural Network (WSSR-CNN) Model

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