Mixed gas concentration detection method based on overlapping time multiplexing direct absorption spectrum

Abstract

The invention discloses a mixed gas concentration detection method based on an overlapped time multiplexing direct absorption spectrum. A multi-component gas concentration detection scheme of OM-DAS is designed to solve the problem that multi-component gas detection lacks synchronism and universality. A spectrum separation module and a concentration regression module based on a depth residual multilayer perceptron are designed, so that the algorithm is suitable for separating spectrum data with overlapped spectrum peak height and regressing gas concentration from a transmission spectrum; establishing an end-to-end mixed gas analysis network, and integrating the network into two modules, namely a spectrum separation module and a concentration regression module, so that the network is suitable for directly regressing multi-component gas concentration from overlapped spectrum data; oM-DAS signals in the experimental environment are simulated and generated to serve as training data of a deep learning algorithm, so that concentration prediction of real data in the experimental environment is achieved; the thought of transfer learning is introduced in the training process, and the concentration regression generalization performance of the mixed gas analysis network is improved.

Description

Technical Field

[0001] This invention proposes a method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectra, belonging to the field of spectroscopic analysis and gas detection technology. Background Technology

[0002] In recent years, the demand for gas concentration detection has increased significantly in industries such as industry, medicine, and environmental protection chemistry. In scenarios such as petrochemical production process control, syngas component analysis, volatile organic compound (VOC) emission monitoring, and greenhouse gas emission verification, high-precision, high-sensitivity, and rapid-response quantitative detection of target gases is often required. Especially in certain specific situations, higher demands are placed on the ability to simultaneously measure multiple components in gas detection. Traditional gas detection methods, such as electrochemical and chromatographic methods, are insufficient to meet the needs of real-time detection, while optical sensors based on absorption spectroscopy are widely used in gas detection due to their advantages such as high sensitivity, non-contact operation, and fast response speed.

[0003] Direct absorption spectroscopy (DAS) is a gas detection technique based on tunable diode laser absorption spectroscopy (TDLAS). Due to its simple system setup, high resolution, real-time performance, and calibration-free operation, it is widely used in environmental monitoring and industrial production. However, because DAS systems have a limited wavelength scanning range, the absorption lines of different gases often approach or even overlap in multi-component gas detection, severely limiting the direct analysis of component concentrations. To address this issue, existing methods often try to avoid spectral line overlap as much as possible, or employ passive spectral line de-overlap schemes when overlap is unavoidable. For example, combining DAS with time-division multiplexing (TDM) or using inherently overlapping absorption lines for de-overlapping can achieve the measurement of absorption lines in multi-component gases. While these methods can effectively detect the concentration of multi-component gases, they also limit the overall detection performance of the DAS system. For example, TDM-based schemes cannot meet the application requirements for simultaneous detection of multiple components; while schemes based on de-overlapping of inherently overlapping absorption lines are limited by the types and combinations of naturally occurring overlapping gases, thus limiting the number of gas components that can be detected simultaneously and making them unsuitable for application scenarios requiring flexible switching of detection targets. Actively overlapping different gas absorption lines is a novel approach for the simultaneous detection of multi-component gas concentrations. Through flexible selection and active overlap of gas absorption lines, it holds promise as a universal method for simultaneous detection of multi-component gas concentrations.

[0004] Furthermore, a common method for predicting DAS concentration is to convert the absorption spectrum to an absorbance spectrum and then fit it using a nonlinear Vogit model to obtain gas parameters. However, the iterative process of nonlinear fitting is very time-consuming, and the selection of initial values ​​can also affect the fitting results. Currently, deep learning methods based on neural networks are increasingly being introduced into spectroscopy, allowing this data-driven approach to overcome the aforementioned problems.

[0005] Therefore, the introduction of deep learning and active overlap of absorption spectra is of great research significance for the synchronous measurement of multi-component gases.

[0006] A search revealed Chinese invention patent CN108007901B, which discloses a method and apparatus for applying laser absorption spectroscopy to the detection of trace gas concentrations. A tunable laser is used as the light source. The measurement wavelength is selected based on the characteristic absorption lines of the target gas molecules. The measured data is compared with the HITRAN standard database, and the concentration is inverted using the measured absorption spectral signal. The apparatus has a simple structure, is easy to operate, and has a short response time. By placing a total reflection mirror behind the absorption cell, the laser light projected through the resonant cavity is reflected and re-enters the cavity to interact with the gas inside, thus improving the system sensitivity. By changing the emission wavelength of the laser, precise detection of various trace gases in multi-component gases can be achieved.

[0007] The technical comparison between this application and the aforementioned documents is as follows:

[0008] 1. The patent "A method and apparatus for applying laser absorption spectroscopy to detect trace gas concentrations" focuses on the high-sensitivity detection of mixed trace gases. This approach uses a single continuous laser to scan and measure the mixed gas. The measurement band needs to be carefully selected to avoid cross-interference of absorption lines from different gases. It relies on the nearest-neighbor spectral characteristics of the mixed gas itself, requires a wide scanning range and a long scanning time, and is applicable to a limited range of gas types. In contrast, this patent uses multiple continuous lasers, with no restrictions on the measurement band. It can collect strong absorption characteristic signals of different types of gases and then achieve synchronous concentration detection of the mixed gas through active overlap and algorithmic deoverlap, demonstrating higher detection efficiency and greater versatility. The two patents differ significantly in their core working mechanisms.

[0009] 2. The patent "A method and apparatus for applying laser absorption spectroscopy to detect trace gas concentrations" uses a HITRAN database parameter fitting method to invert the concentration of experimentally measured spectra in its data processing algorithm. Its accurate concentration detection results are highly dependent on the acquisition of interference-free absorption lines and a meticulous data preprocessing process. This patent, on the other hand, uses a data-driven deep learning method to train an end-to-end neural network model on simulated data, directly inverting the accurate concentration results from unprocessed experimental data. The data processing algorithms of the two are significantly different.

[0010] A search revealed Chinese invention patent CN114235700B, which discloses a device and method for detecting the concentration of multi-component gases. The trace gas concentration detection device provided by this invention comprises a broadband tuned laser source, an optical fiber collimator, a beam-splitting sampling plate, a laser wavelength monitoring module, a modulation signal generator, a laser control circuit module, a lens group, a ring-down cavity with precise temperature control, a ring-down cavity temperature control module, a focusing lens, a detector, and a data collection and analysis processing circuit module. Broadband tuning of the semiconductor laser is achieved using current or temperature control, covering the characteristic absorption spectra of the multi-component gases to be measured. Simultaneously, the refractive index of the ring-down cavity is adjusted by controlling its temperature, achieving mode matching between the laser wavelength and the ring-down cavity, thus solving the problem that existing laser absorption spectroscopy gas detection technologies cannot detect multi-component gases.

[0011] The technical comparison between this application and the aforementioned documents is as follows:

[0012] 1. The patent "A Multi-Component Gas Concentration Detection Device and Method" uses time-controlled scanning of multiple continuous-wave lasers of different wavelengths to construct a broadband tuned DFB laser source, achieving broadband coverage of laser wavelengths to match the absorption spectrum of the mixed gas. However, its continuous broadband scanning characteristics can easily introduce redundant absorption information in specific wavelength bands, thereby reducing the resolution of high-sensitivity spectral lines. This patent, on the other hand, uses highly selective laser tuning control, precisely tuning only for the required spectral line wavelength range. The two patents differ significantly in their methods of acquiring gas characteristic absorption spectra.

[0013] 2. The detection scheme of the patent "A multi-component gas concentration detection device and method" is essentially a time-division multiplexing technology, which sequentially starts and stops each laser within a complete broadband scanning cycle, resulting in a time delay in the absorption characteristic spectral lines of each component. This patent, on the other hand, synchronously modulates multiple continuous wave lasers to overlap them on the time scale, thereby achieving the synchronous acquisition of all gas absorption information. The two detection schemes are completely different.

[0014] 3. The technical principle of the patent "A Multi-component Gas Concentration Detection Device and Method" is based on cavity ring-down absorption spectroscopy (CRDS). It uses preset laser scanning current and cavity endoscope temperature parameters, combined with cavity calibration data, to measure the ring-down curve, and then inverts the gas concentration by comparing the ring-down time difference between the gas-filled cavity and the empty cavity. The technical principle of this patent is based on direct absorption spectroscopy (DAS), which directly inverts the concentration based on Beer-Lambert's law from the measured gas absorption spectrum. The detection principles of the two are completely different. Summary of the Invention

[0015] The purpose of this invention is to propose a method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy;

[0016] This paper proposes a multi-component gas concentration detection scheme by designing an overlapped and multiplexed direct absorption spectroscopy (OM-DAS) method to address the lack of synchronicity and versatility in multi-component gas detection.

[0017] By designing a spectral separation module and a concentration regression module based on a deep residual multilayer perceptron (DR-MLP), the algorithm is made suitable for separating spectral data with highly overlapping peaks and regressing gas concentrations from transmission spectra.

[0018] By constructing an end-to-end mixed gas analysis network (E2E-MAN), the network integrates two modules: a spectral separation module and a concentration regression module. This makes it suitable for directly regressing the concentrations of multi-component gases from overlapping spectral data.

[0019] By simulating and generating OM-DAS signals under experimental conditions as training data for deep learning algorithms, the concentration of real data under experimental conditions can be predicted.

[0020] By introducing the concept of transfer learning into the training process, the concentration regression generalization performance of the E2E-MAN model is improved.

[0021] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0022] A method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy is characterized by comprising the following steps:

[0023] S1. Construction of an overlapped time-multiplexed direct absorption spectroscopy detection system;

[0024] S2. Construct a mixed gas overlapping spectrum simulation dataset based on HITRAN;

[0025] S3. Construct a mixed gas analysis network based on a deep residual multilayer sensor;

[0026] S4. Training optimization of the mixed gas analysis network;

[0027] S5. Performance evaluation of the model.

[0028] As a preferred technical solution of the present invention, step S1 is specifically as follows:

[0029] S101. Set up the gas configuration module:

[0030] A vacuum gas pool environment is provided by a vacuum pump; at room temperature and normal pressure, a variety of standard gases are mixed and configured into mixed gases of different concentrations using a multi-channel flow controller, and then transported to the gas pool through pipelines.

[0031] S102. Construct the optical inspection module:

[0032] Multiple distributed feedback lasers of different wavelengths are coupled to a single optical path using an optical coupler to scan a mixed gas sample in a gas cell; the scanned optical signal is converted into an analog electrical signal by a photodetector after power control by an optical attenuator.

[0033] S103. Set up the signal acquisition and processing module:

[0034] Using a data acquisition card, multiple electrical signals are synchronously modulated, and the absorption spectral lines scanned by each laser are actively overlapped by a current driver; at the same time, the data acquisition card acquires the analog electrical signals output from the photodetector and converts them into digital signals for computer processing.

[0035] As a preferred technical solution of the present invention: in step S101, the room temperature is set to 25 degrees Celsius and the pressure is set to 1 atm.

[0036] As a preferred technical solution of the present invention, step S2 is specifically as follows:

[0037] S21. First, the composition mechanism and corresponding spectral parameters of the superimposed and multiplexed direct absorption spectrum under experimental conditions are analyzed. Then, based on the principle of superimposed and multiplexed direct absorption spectrum, the superimposed spectrum is deconstructed by analyzing and verifying the signals obtained by each laser under time-division scanning and synchronous scanning. Composition:

[0038] (1);

[0039] (2);

[0040] in, It represents a specific gaseous component. Indicates the quantity of different components. It is the first The baseline of each measurement channel, It refers to the modulation wavelength range of the laser. These are the transmission spectral lines of the components. It is random noise. It is the spectral overlap coefficient, defined as the peak spacing. and total wavelength range The ratio;

[0041] S22, Spectral overlap coefficient obtained under experimental conditions based on Equation 2 The transmission spectra of each component were simulated using the HITRAN database;

[0042] S23. Based on Equation 1, the simulated transmission spectra are combined into overlapping spectra, and Gaussian noise with a mean of 0 and a variance of 0.0028 is added to them to simulate the overlapping and multiplexing direct absorption spectrum signal collected under the experimental environment. At the same time, the transmission spectra of each component constitute a transmission spectrum set, and the corresponding concentration constitutes a concentration set, which is used as training labels.

[0043] As a preferred technical solution of the present invention, step S3 is as follows:

[0044] Step S31: Design the spectral separation module, as follows:

[0045] It includes spectral attention blocks and residual blocks. The front end of the spectral attention block uses a multilayer perceptron combined with skip connections to adjust the feature response of the entire spectral dimension, and the back end achieves feature fusion through a fully connected layer, thereby outputting spectral features with embedded attention mechanism.

[0046] The spectral features of the output embedded attention mechanism are fed into nine stacked residual blocks, each consisting of two linear layers using LayerNorm and GELU activation functions, respectively, and skip connections are introduced to facilitate gradient flow.

[0047] After removing the mixed baseline by the residual block, the independent transmission spectra corresponding to each component are separated and sent as input to the concentration regression module.

[0048] Step S32: Design the concentration regression module, as follows:

[0049] It includes a concentration feature extraction unit, a connection unit, and a fusion unit. The independent transmission spectra of the gas are sequentially processed by the linear transformation, BatchNorm, and ReLU activation layer in the concentration feature extraction unit to extract concentration features. The connection unit preserves global information. Finally, the fusion unit performs feature fusion to output the final concentration regression result.

[0050] As a preferred technical solution of the present invention, step S4 is specifically as follows:

[0051] The mixed gas analysis network was trained and optimized using a training set, as detailed below:

[0052] S41. Data Generation:

[0053] During the data generation phase, each input overlapping spectrum is normalized, and the corresponding ideal transmittance label and gas concentration label are normalized to the range of 0 to 1 according to each gas component.

[0054] S42. Model Training:

[0055] During the model training phase, the spectral separation module was first pre-trained using a combined loss function consisting of the coarse peak region weighted mean square error (MSE), the first-order difference MSE, and the correlation coefficient between components. Subsequently, the pre-trained module was loaded into the complete mixed gas analysis network for end-to-end training, using a loss function consisting of the concentration prediction MSE and the MSE between the reconstructed spectrum and the input spectrum. The Adam optimizer and learning rate scheduling strategy were used during training, and the training curves and visualization results were recorded periodically.

[0056] As a preferred technical solution of the present invention, step S5 is as follows:

[0057] A comprehensive performance evaluation was conducted on the mixed gas analysis network that met the preset requirements, as detailed below:

[0058] S51. Analyze the spectral separation performance of the experimental measurement data: Under strong overlap, calculate the root mean square error (RMSE) between the reconstructed spectrum based on the separated transmission spectrum and the measured overlapping spectrum, and analyze the stability of RMSE under continuous measurement.

[0059] S52. Perform concentration regression performance analysis on the network within a preset range: calculate the determination coefficient, absolute error, and relative error of the regression concentration, and analyze the overall concentration regression performance of the model.

[0060] S53. Analyze the network's resistance to cross-interference between components: With the concentration of one gas fixed, change the concentrations of other gases and continuously measure this dynamic process, and calculate the error of the model's regression to the fixed gas concentration under the dynamic process.

[0061] S54. Perform long-term stability analysis of the model for concentration detection: continuously detect a fixed concentration of mixed gas for a long period of time, calculate the standard deviation of the concentration regression error, and perform Allan bias analysis to calculate the minimum detection limit of each component gas.

[0062] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0063] This invention introduces a deep learning approach to achieve real-time transmission spectrum separation, baseline elimination, and concentration prediction of overlapping spectra based on the proposed OM-DAS detection method.

[0064] This invention utilizes transfer learning to enable the proposed mixed gas analysis network to directly process OM-DAS signals to obtain concentration data. This not only simplifies data processing and improves the efficiency of training and deployment, but also enhances the model's ability to generalize to new data.

[0065] Comparative analysis shows that the method proposed in this invention has a relatively simple system structure, higher detection efficiency, and excellent detection accuracy for multi-component gas concentration detection, providing a novel and universal method for simultaneous detection of multi-component gases. Attached Figure Description

[0066] Figure 1 This is a flowchart of the mixed gas concentration detection method based on overlap time multiplexing direct absorption spectroscopy in this invention.

[0067] Figure 2 This is a schematic diagram of the OM-DAS principle in this invention.

[0068] Figure 3 This is a structural diagram of the OM-DAS detection system in this invention.

[0069] Figure 4 This is a diagram of the mixed gas analysis network structure based on a deep residual multilayer sensor in this invention. Detailed Implementation

[0070] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are for illustrative purposes only and are not intended to limit the scope of the invention.

[0071] like Figure 1 As shown, the mixed gas concentration detection method based on overlapping time-multiplexed direct absorption spectroscopy proposed in this invention includes the following steps:

[0072] S1. Construction of the Overlapping Time-Multiplexed Direct Absorption Spectroscopy (OM-DAS) Detection System. Step S1 involves constructing the detection system from three components: a gas configuration module, an optical detection module, and a signal acquisition and processing module. The hardware structure diagram is shown below. Figure 3 As shown, the details are as follows:

[0073] S101. Set up the gas configuration module:

[0074] A vacuum gas pool environment is provided by a vacuum pump; at room temperature (25 degrees Celsius) and atmospheric pressure (1 atm), a multi-channel flow controller is used to mix various standard gases (gas 1, 2, 3, ...) into a mixed gas of a certain concentration combination, which is then transported to the gas pool through pipelines;

[0075] S102. Construct the optical inspection module:

[0076] Multiple distributed feedback lasers of different wavelengths (each laser is responsible for acquiring the spectral absorption lines of the corresponding components in the gas mixture) are coupled to an optical path using an optical coupler to scan the gas mixture sample in the gas cell; the scanned optical signal is converted into an analog electrical signal by a photodetector through power control of an optical attenuator.

[0077] S103. Set up the signal acquisition and processing module:

[0078] Using a data acquisition card, multiple electrical signals are synchronously modulated, and the absorption spectral lines scanned by each laser are actively overlapped by a current driver; at the same time, the data acquisition card acquires the analog electrical signals output from the photodetector and converts them into digital signals for computer processing.

[0079] S2. Construct a mixed gas overlapping spectrum simulation dataset based on HITRAN, as detailed below:

[0080] S21. First, analyze the formation mechanism and corresponding spectral parameters of overlapping and multiplexed direct absorption spectra under experimental conditions. Then, based on the principle of overlapping and multiplexed direct absorption spectra (such as...), Figure 2 As shown, multiple absorption spectral lines are actively overlapped through time-series control. The overlapping spectra are deconstructed by analyzing and verifying the signals obtained from time-division and synchronous scanning of each laser. Composition:

[0081] (1);

[0082] (2);

[0083] in, It represents a specific gaseous component. Indicates the quantity of different components. It is the first The baseline of each measurement channel, It refers to the modulation wavelength range of the laser. These are the transmission spectral lines of the components. It is random noise. It is the spectral overlap coefficient, defined as the peak spacing. and total wavelength range The ratio;

[0084] S22, Spectral overlap coefficient obtained under experimental conditions based on Equation 2 The transmission spectra of each component were simulated using the HITRAN database. Taking a mixture of methane and carbon dioxide as an example, 20,000 sets of simulated transmission spectra of methane and carbon dioxide at different concentrations were generated. The methane concentration ranged from 1 to 100 ppm, and the carbon dioxide concentration ranged from 1% to 5%. Each data point consisted of 1,600 sampling points.

[0085] S23. Based on Equation 1, the simulated transmission spectra are combined into an overlapping spectrum, and Gaussian noise with a mean of 0 and a variance of 0.0028 is added to it to simulate the OM-DAS signal collected under the experimental environment. At the same time, the transmission spectra of each component constitute a transmission spectrum set, and its corresponding concentration constitutes a concentration set, which is used as training labels.

[0086] S3. Construct a mixed gas analysis network based on a deep residual multilayer perceptron. The network structure diagram is shown below. Figure 4 As shown, the details are as follows:

[0087] Step S31: Design the spectral separation module, as follows:

[0088] It includes spectral attention blocks and residual blocks. The front end of the spectral attention block uses a multilayer perceptron (MLP) combined with skip connections to adjust the feature response of the entire spectral dimension, and the back end achieves feature fusion through a fully connected layer, thereby outputting spectral features with embedded attention mechanism.

[0089] To further enhance feature representation capabilities and preserve long-range dependencies in the spectrum, the output spectral features embedded with the attention mechanism are fed into nine stacked residual blocks. Each residual block consists of two linear layers, employing LayerNorm and GELU activation functions respectively. Skip connections are introduced to promote gradient flow and alleviate the gradient vanishing problem during training, thereby improving the accuracy of spectral separation. The introduced LayerNorm plays a crucial role in suppressing baseline drift-induced overlap of transmission spectral frame differences, accelerating model training, and improving model performance and generalization ability.

[0090] After removing the mixed baseline by the residual block, the independent transmission spectra corresponding to each component are separated and sent as input to the concentration regression module.

[0091] Step S32: Design concentration regression modules. Design multiple dedicated concentration regression modules to effectively maintain the independence of the absorption characteristics of each gas, as detailed below:

[0092] It includes a concentration feature extraction unit, a connection unit, and a fusion unit. The independent transmission spectra of the gas are sequentially processed by the linear transformation, BatchNorm, and ReLU activation layer in the concentration feature extraction unit to extract concentration features. The global information is preserved by a connection unit similar to residuals. Finally, after feature fusion and further processing by the fusion unit, the final concentration regression result is output.

[0093] S4. Training optimization of the mixed gas analysis network, as detailed below:

[0094] The mixed gas analysis network was trained and optimized using a training set, as detailed below:

[0095] S41. Data Generation:

[0096] During the data generation phase, each input overlapping spectrum is normalized, and the corresponding ideal transmittance label and gas concentration label are normalized to the range of 0 to 1 according to each gas component.

[0097] S42. Model Training:

[0098] During the model training phase, the spectral separation module was first pre-trained using a combined loss function consisting of the coarse peak region weighted mean square error (MSE), the first-order difference MSE, and the correlation coefficient between components. Subsequently, the pre-trained module was loaded into the complete mixed gas analysis network for end-to-end training, using a loss function consisting of the concentration prediction MSE and the MSE between the reconstructed spectrum and the input spectrum. The Adam optimizer and learning rate scheduling strategy were used during training, and the training curves and visualization results were recorded periodically.

[0099] S5. Model performance evaluation, as detailed below:

[0100] A comprehensive performance evaluation was conducted on the mixed gas analysis network that met the preset requirements, as detailed below:

[0101] S51. Analyze the spectral separation performance of the experimental measurement data: Under strong overlap, calculate the root mean square error (RMSE) between the reconstructed spectrum based on the separated transmission spectrum and the measured overlapping spectrum, and analyze the stability of RMSE under continuous measurement.

[0102] S52. Perform concentration regression performance analysis on the network within a preset range: calculate the determination coefficient, absolute error, and relative error of the regression concentration, and analyze the overall concentration regression performance of the model.

[0103] S53. Analyze the network's resistance to cross-interference between components: With the concentration of one gas fixed, change the concentrations of other gases and continuously measure this dynamic process, and calculate the error of the model's regression to the fixed gas concentration under the dynamic process.

[0104] S54. Perform long-term stability analysis of the model for concentration detection: continuously detect a fixed concentration of mixed gas for a long period of time, calculate the standard deviation of the concentration regression error, and perform Allan bias analysis to calculate the minimum detection limit of each component gas.

[0105] Based on the above technical solution, the effectiveness of this invention can be summarized by three key factors. First, the invention uses an OM-DAS detection scheme, enabling simultaneous measurement of the absorption spectra of multi-component gases. Second, by employing a DR-MLP as the backbone network, the spectral separation module can eliminate baselines and separate transmission spectra from spectral data with strong overlap characteristics. Finally, through a novel end-to-end structure integrating the spectral separation module and the concentration regression module, the accuracy of detecting the concentration of multi-component gases in overlapping states can be significantly improved.

[0106] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.

Claims

1. A method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy, characterized in that, Includes the following steps: S1. Construction of an overlapped time-multiplexed direct absorption spectroscopy detection system; S2. Construct a mixed gas overlapping spectrum simulation dataset based on HITRAN; S3. Construct a mixed gas analysis network based on a deep residual multilayer sensor; S4. Training optimization of the mixed gas analysis network; S5. Performance evaluation of the model.

2. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 1, characterized in that, Step S1 is as follows: S101. Set up the gas configuration module: A vacuum gas pool environment is provided by a vacuum pump; at room temperature and normal pressure, a variety of standard gases are mixed and configured into mixed gases of different concentrations using a multi-channel flow controller, and then transported to the gas pool through pipelines. S102. Construct the optical detection module: Multiple distributed feedback lasers of different wavelengths are coupled to a single optical path using an optical coupler to scan a mixed gas sample in a gas cell; the scanned optical signal is converted into an analog electrical signal by a photodetector after power control by an optical attenuator. S103. Set up the signal acquisition and processing module: Using a data acquisition card, multiple electrical signals are synchronously modulated, and the absorption spectral lines scanned by each laser are actively overlapped by a current driver; at the same time, the data acquisition card acquires the analog electrical signals output from the photodetector and converts them into digital signals for computer processing.

3. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 2, characterized in that, In step S101, the room temperature is set to 25 degrees Celsius and the pressure is set to 1 atm.

4. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 1, characterized in that, Step S2 is as follows: S21. First, the composition mechanism and corresponding spectral parameters of the superimposed and multiplexed direct absorption spectrum under experimental conditions are analyzed. Then, based on the principle of superimposed and multiplexed direct absorption spectrum, the superimposed spectrum is deconstructed by analyzing and verifying the signals obtained by each laser under time-division scanning and synchronous scanning. Composition: （1）； （2）； in, It represents a specific gaseous component. Indicates the quantity of different components. It is the first The baseline of each measurement channel, It refers to the modulation wavelength range of the laser. These are the transmission spectral lines of the components. It is random noise. It is the spectral overlap coefficient, defined as the peak spacing. and total wavelength range The ratio; S22, Spectral overlap coefficient obtained under experimental conditions based on Equation 2 The transmission spectra of each component were simulated using the HITRAN database; S23. Based on Equation 1, the simulated transmission spectra are combined into overlapping spectra, and Gaussian noise with a mean of 0 and a variance of 0.0028 is added to them to simulate the overlapping and multiplexing direct absorption spectrum signal collected under the experimental environment. At the same time, the transmission spectra of each component constitute a transmission spectrum set, and the corresponding concentration constitutes a concentration set, which is used as training labels.

5. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 1, characterized in that, Step S3 is as follows: Step S31: Design the spectral separation module, as follows: It includes spectral attention blocks and residual blocks. The front end of the spectral attention block uses a multilayer perceptron combined with skip connections to adjust the feature response of the entire spectral dimension, and the back end achieves feature fusion through a fully connected layer, thereby outputting spectral features with embedded attention mechanism. The spectral features of the output embedded attention mechanism are fed into nine stacked residual blocks, each consisting of two linear layers using LayerNorm and GELU activation functions, respectively, and skip connections are introduced to facilitate gradient flow. After removing the mixed baseline by the residual block, the independent transmission spectra corresponding to each component are separated and sent as input to the concentration regression module. Step S32: Design the concentration regression module, as follows: It includes a concentration feature extraction unit, a connection unit, and a fusion unit. The independent transmission spectra of the gas are sequentially processed by the linear transformation, BatchNorm, and ReLU activation layer in the concentration feature extraction unit to extract concentration features. The connection unit preserves global information. Finally, the fusion unit performs feature fusion to output the final concentration regression result.

6. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 1, characterized in that, Step S4 is as follows: The mixed gas analysis network was trained and optimized using a training set, as detailed below: S41. Data Generation: During the data generation phase, each input overlapping spectrum is normalized, and the corresponding ideal transmittance label and gas concentration label are normalized to the range of 0 to 1 according to each gas component. S42. Model Training: During the model training phase, the spectral separation module was first pre-trained using a combined loss function consisting of the coarse peak region weighted mean square error (MSE), the first-order difference MSE, and the correlation coefficient between components. Subsequently, the pre-trained module was loaded into the complete mixed gas analysis network for end-to-end training, using a loss function consisting of the concentration prediction MSE and the MSE between the reconstructed spectrum and the input spectrum. The Adam optimizer and learning rate scheduling strategy were used during training, and the training curves and visualization results were recorded periodically.

7. The method for detecting the concentration of mixed gases based on overlapping time-multiplexed direct absorption spectroscopy according to claim 1, characterized in that, Step S5 is as follows: A comprehensive performance evaluation was conducted on the mixed gas analysis network that met the preset requirements, as detailed below: S51. Analyze the spectral separation performance of the experimental measurement data: Under strong overlap, calculate the root mean square error (RMSE) between the reconstructed spectrum based on the separated transmission spectrum and the measured overlapping spectrum, and analyze the stability of RMSE under continuous measurement. S52. Perform concentration regression performance analysis on the network within a preset range: calculate the determination coefficient, absolute error, and relative error of the regression concentration, and analyze the overall concentration regression performance of the model. S53. Analyze the network's resistance to cross-interference between components: With the concentration of one gas fixed, change the concentrations of other gases and continuously measure this dynamic process, and calculate the error of the model's regression to the fixed gas concentration under the dynamic process. S54. Perform long-term stability analysis of the model for concentration detection: continuously detect a fixed concentration of mixed gas for a long period of time, calculate the standard deviation of the concentration regression error, and perform Allan bias analysis to calculate the minimum detection limit of each component gas.

Images