Adaptively Optimized Gas Analysis Model with Deep Learning for Near-Infrared Methane Sensors

Abstract

Noise significantly limits the accuracy and stability of retrieving gas concentration with the traditional direct absorption spectroscopy (DAS). Here, we developed an adaptively optimized gas analysis model (AOGAM) composed of a neural sequence filter (NSF) and a neural concentration retriever (NCR) based on deep learning algorithms for extraction of methane absorption information from the noisy transmission spectra and obtaining the corresponding concentrations from the denoised spectra. The model was trained on two data sets, including a computationally generated one and the experimental one. We have applied this model for retrieving methane concentration from its transmission spectra in the near-infrared (NIR) region. The NSF was implemented through an encoder–decoder structure enhanced by the attention mechanism, improving robustness under noisy conditions. Further, the NCR was employed based on a combination of a principal component analysis (PCA) layer, which focuses the algorithm on the most significant spectral components, and a fully connected layer for solving the nonlinear inversion problem of the determination of methane concentration from the denoised spectra without manual computation. Evaluation results show that the proposed NSF outperforms widely used digital filters as well as the state-of-the-art filtering algorithms, improving the signal-to-noise ratio by 7.3 dB, and the concentrations determined with the NCR are more accurate than those determined with the traditional DAS method. With the AOGAM enhancement, the optimized methane sensor features precision and stability in real-time measurements and achieves the minimum detectable column density of 1.40 ppm·m (1σ). The promising results of the present study demonstrate that the combination of deep learning and absorption spectroscopy provides a more effective, accurate, and stable solution for a gas monitoring system.

Feature Domain Transform Filter for the Removal of Inherent Noise Bound to the Absorption Signal

“…Compared with the sensor only relying on the WMS technique, our sensor simply increases the computational cost without increasing the system cost and has made great progress. Our sensor also has a cost advantage compared with those sensors that combine multiple techniques. , For the sensor related to DAS technology, although low cost is a major advantage, it is difficult to achieve a very low detection limit.…”

“…Our sensor also has a cost advantage compared with those sensors that combine multiple techniques. 21,22 For the sensor related to DAS technology, 16 although low cost is a major advantage, it is difficult to achieve a very low detection limit.…”

“…Therefore, we believe that only by opening up a new domain will the information on noise and absorption signal be clearly distinguished. We propose a new domain named feature domain, which is composed of the principal components obtained by the principal component analysis (PCA) 16 algorithm based on singular value decomposition (SVD). The PCA is performed on the original noisy second harmonic signal and the corresponding no-noise second harmonic signal, and different feature domain signals are obtained.…”

Neural network multi-component gas mixture analysis with broadband dual-frequency comb absorption spectroscopy

“…In the present study, we established the dataset of blended absorption spectra of multi-component gas mixtures by proper simulation. It has been con rmed that the model which is pre-trained on the simulated dataset can also perform well on the real data, which is especially important in cases where it is di cult and time-consuming to collect experimental data 30,31 . This is due to the good reproduction of the distribution of experimental data by a well-designed simulation.…”

“…Consequently, the presence of the spectral contribution of an unidenti ed gas will inevitably lead to the prediction deviation in the determined concentrations of the selected set of gas components. With the adoption of machine learning and deep learning technology, intelligent algorithms have been widely used in various elds of optical applications, such as gas concentration retrieval 18 , signal ltering 30,31 , ultrashort pulse reconstruction 32 , hyperspectral image and material spectral classi cation 33 . For the purpose of gas sensing, traditional machine learning algorithms such as K-nearest neighbors, decision tree, random forest and support vector machine have been applied and compared 34,35 , but the results still require considerable improvements.…”