Research On Infrared Spectroscopy For AI-Enhanced Gas Sensing Based On Deep Learning

Gas detection technology has a wide range of applications in various fields such as power grids,industrial and chemical production,ecological environment monitoring,and medical auxiliary diagnosis,and has broad prospects for development.With the development of science and technology and social progress in China,there is a high demand and significance for real-time online gas detection in many aspects of these fields.As a new type of gas sensing technology,laser infrared absorption spectroscopy gas sensing technology can make up for the shortcomings of traditional gas sensing technology based on chemical methods in various aspects of performance.On the one hand,most common gases in production and daily life have spectral absorption characteristics in the infrared band,which makes laser infrared absorption spectroscopy capable of specific(selective)detection based on the absorption characteristics of sample molecules.On the other hand,the development of fiber-optic communication and semiconductor laser technology enables infrared spectroscopic gas detection technology to achieve rapid,non-contact,and high-sensitivity detection.The instrument performance based on laser infrared spectroscopy has a crucial connection with the properties of the laser light source itself.For the detection of target gases with single components,selective absorption of gas molecules to the spectral lines is usually utilized to achieve targeted detection of specific molecules.Narrowband light sources,such as distributed feedback lasers,are used to finely scan the specific absorption peaks of gas molecules to achieve concentration detection.For multi-component gas sensing,there are two methods,namely specific detection reuse and broadband spectroscopic detection.Specific detection reuse uses multiple specific light sources to detect multiple target gases,which is essentially the same as single-component specific detection.In comparison,broadband spectroscopic technology encompasses the absorption characteristics of multi-component gases through a wider spectral window,achieving simultaneous sensing of gas mixtures.However,whether it is accurate detection of specific target gases or simultaneous detection of multi-component gas mixtures,how to further improve their detection performance is still a challenge.Firstly,for achieving precise single-component gas composition detection technology,although this type of technology has been developed early and is mature,the various detection indicators of gas sensors still need to be rapidly improved to meet the rapid progress of various industries.Therefore,the development of new sensors with higher resolution,stronger stability,and higher cost-effectiveness has a more important significance.In this article,we take tunable diode laser absorption spectroscopy(TDLAS)technology as the research object.For direct absorption spectroscopy(DAS)technology without calibration,accurate measurement results highly depend on the data processing of the original transmission signal.So far,this process still involves too much manual operation,and this uncertainty leads to the introduction of human error.Moreover,another problem with near-infrared DAS is its extremely low signal-to-noise ratio,which greatly limits its detection limit for trace gas detection.How to ensure the improvement of stability,sensitivity,detection limit,and other indicators in the process of developing near-infrared semiconductor laser absorption spectroscopy gas sensors is the key to our research on near-infrared DAS gas sensors.Next,for the precise detection of multi-component gas mixtures,our focus will be on the dual frequency comb technology.The dual frequency comb system provides a unique set of characteristics that are particularly suitable for sensing multi-component gas mixtures,combining wide spectral span,high spectral power and high spectral resolution with high spatial coherence,stability and ease of use.However,the ill-posed problem of blind deconvolution of mixture components and their concentrations from given overlapping absorption spectra cannot be solved purely optically and requires the use of algorithms.In addition,the interference of unknown gases other than the target gas presents another challenge to the detection accuracy.Furthermore,existing direct detection techniques for multi-component gas mixtures based on spectral quantification analysis algorithms suffer from large prediction errors due to the use of spectral data with erroneous prior knowledge during algorithm training.Moreover,they cannot identify the types of mixture components and are not suitable for the complex situation where multiple components coexist in gas mixtures.This article addresses key issues that hinder the further development of laser infrared spectroscopic gas sensing technology and presents a series of studies aimed at promoting intelligent and precise spectral gas sensing technology,providing new research directions and technical support.The specific research contents and innovative points of this thesis are as follows:(1)In response to the problem of introducing errors caused by manual baseline fitting in direct absorption spectroscopy technology,we first proposed a direct absorption spectroscopy gas sensing system based on an end-to-end deep neural network.This system directly maps the experimentally collected transmission signal to the corresponding concentration information,avoiding a large number of intermediate calculation processes and thus avoiding the introduction of human error.We solved the negative impact of insufficient experimental data on neural network training results through transfer learning.To verify the universality of the algorithm,we selected methane and acetylene gases as target gases for detection,and the measurement results of the two gases showed that the direct absorption spectroscopy technology based on deep neural networks can be reliably applied to different gas molecules.In addition to comparing the performance of the proposed neural network algorithm with commonly used machine learning algorithms such as decision trees,AdaBoost decision trees,and KNN,the end-to-end direct absorption spectroscopy sensor system was also compared with a sensor system based on wavelength modulation spectroscopy.The developed gas sensing system exhibited a more accurate and faster concentration inversion calculation speed,as well as good robustness to noise,laser aging,and changes in circuit parameters.The combination of deep neural networks and direct absorption spectroscopy provides new ideas for further research on gas absorption spectroscopy.(2)A novel methane sensor based on neural network filtering(NNF)assisted direct absorption spectroscopy(DAS)was proposed and experimentally verified to address the problem of the stability and robustness of direct absorption spectroscopy technology,which is prone to noise interference.The developed detection device adds the advantages of a neural network-based digital filter,thus overcoming the shortcomings of traditional DAS.We overcame the problem of data scarcity in building and training NNF under practical experimental conditions by using simulated absorption spectra.Compared with several widely used filtering algorithms,the proposed NNF algorithm showed the best performance.We conducted a detailed evaluation of the detection system improved by NNF.The sensor has higher concentration inversion accuracy and better stability in real-time measurements.The minimum detectable limit of the method is 2.93 ppm·m(1σ),significantly higher than the near-infrared methane detection technology reported in the past using DAS.Finally,we systematically discussed the frequency principle of NNF to clearly explain the mechanism of generalized filtering.The improved methane sensor demonstrated the feasibility of using neural network algorithms to improve DAS technology performance,which has.broad applicability in high-sensitivity measurements of trace gases such as methane.(3)Cross interference of gas species in absorption spectra is one of the most challenging obstacles in analyzing multicomponent gas mixtures with overlapping absorption features(mixed spectra).To address this issue,we propose a multi-component gas mixture sensor that combines broadband absorption spectroscopy with spectroscopic analysis algorithms.The sensor uses a mid-infrared dual-comb laser source that enables sensitive measurements over a wide spectral range,and combines deep learning algorithms for spectroscopic analysis to accurately identify gas components and detect their concentrations in gas mixtures.The sensor has been tested on gas mixtures of three common gas species:methane,acetone,and water vapor.Architecture adjustment and model training were performed using a physically informative enhanced dataset.First,the proposed spectroscopic analysis model was compared and evaluated against two other advanced neural network algorithms(2L-ARNN and 1D-CNN).Then,the sensor’s performance was evaluated in real-time measurements in actual detection scenarios.In addition,we systematically analyzed and presented clear visualization results that explain the internal workings of the considered algorithms.The high performance of the proposed sensor demonstrates the feasibility of integrating broadband optical sensing and deep learning models to achieve a more general gas mixture analysis solution.(4)Although mid-infrared broadband spectroscopy technologies,such as the dual-frequency comb that we used,provide high sensitivity for the parallel detection of trace gases,they also face many challenges.For example,improper experimental equipment and operation introduce uncertainties in the data.This uncertainty manifests as a discrepancy between the preset concentration and the actual gas concentration obtained,as well as a discrepancy between the preset environmental temperature and pressure and the actual conditions in the gas chamber.In addition,there will be many unknown gases present under actual detection conditions,leading to problems such as baseline fluctuations and model errors.To address these issues,we have made improvements based on the research in Section 4,and proposed a new platform that combines the advantages of a mid-infrared dual-comb spectrometer based on two difference frequency generation combs pumped by a femtosecond Er-doped fiber comb oscillator and an unsupervised deep learning neural network consisting of information extraction and information mapping blocks.By combining unsupervised learning methods with a model-agnostic,physics-assisted data augmentation strategy and using simulated data from spectral databases,we overcome the inherent data sparsity problems and uncertainties of devices and manual operations in the analysis of multi-component gas mixtures.Our system provides reliable identification of gas species,concentration detection,and environmental pressure prediction,and eliminates negative effects on measurements,such as model errors,baseline fluctuations,and unknown absorbing substances.We demonstrated parallel optical detection of 31 different five-component gas mixtures in the 2900-3100 cm-1 spectral range,with sub-billionth level sensitivity,demonstrating its potential in various application fields,such as atmospheric monitoring,respiratory biomarker diagnosis,and rapid capture of chemical reaction kinetics.