End-to-End Methane Concentration Measurement Model With Deep Residual Shrinkage Network for Direct Absorption Spectroscopy

Noise interference severely limits the accuracy and stability of gas concentration measurement in direct absorption spectroscopy (DAS). In this article, an end-to-end methane concentration measurement model (EoE-MCMM) consisting of a gas spectrum filter (GSF) and a gas concentration predictor (GCP) is established. The GSF is implemented by deep residual shrinkage network (DRSN) and multilayer perceptron. Based on the advantage of DRSN in effectively processing high-noise data, the GSF can mine effective absorption information from the noisy methane transmission spectrum. In addition, the GCP consists of three fully connected layers, which directly inverts the output spectrum from the GSF. It is worth noting that in the training stage of the GCP, the pretrained weights of the GSF are loaded by transfer learning to accelerate the convergence speed of the GCP module training process. The results show that the proposed GSF outperforms other commonly used filtering methods and improves the signal-to-noise ratio of the experimental data by 10 dB. In terms of concentration prediction, the gas concentration determination coefficient ( $R^{2}$ ) is 99.9%, which is higher than the traditional filtering methods. In addition, the measurement precision of the method is 0.26 ppm by continuous measurement of methane for 1 h, and the limit of detection (LOD) of methane is 0.052 ppm at an integration time of 30 s and 0.008 ppm at an integration time of 400 s according to Allan deviation analysis. These results show that the model provides a more effective, accurate, and convenient solution for DAS gas concentration measurement.