Spectral decomposition method and device based on interpretable neural network based on transfer learning

Abstract

The present invention discloses a spectral decomposition method and device of an interpretable neural network based on transfer learning, comprising: using a first spectral data set to perform supervised learning on a neural network based on a first label set and on a task of spectral decomposition of absolute concentration of pollutant gas; using a second spectral data set to perform transfer learning on the neural network after supervised learning based on a second label set and on a task of spectral decomposition of absolute concentration of pollutant gas, to obtain a spectral decomposition model; inputting spectral data to be predicted into the spectral decomposition model to perform spectral decomposition to obtain the absolute concentration of pollutant gas, and using an interpretable analyzer to analyze the interpretability relationship between different pollutant gas spectra and absolute concentrations in the spectral decomposition model. The method and device can realize the prediction from the original spectrum to the absolute concentration of pollutant gas, and provide a scientific basis for improving the accuracy of remote sensing passive ultraviolet imaging and increasing the credibility of monitoring data.

Description

Technical Field

[0001] The present invention belongs to the technical field of optical measurement, and in particular relates to a spectral decomposition method and device of an interpretable neural network based on transfer learning. Background Art

[0002] In the field of traditional remote sensing passive ultraviolet imaging, due to the uncertainty of the atmospheric photon path, only SCD (slant column concentration) can be measured, and the required absolute concentration cannot be determined.

[0003] A patent application with publication number CN110031412A discloses a method for obtaining the emission flux of atmospheric pollutants based on mobile AHSA observations. It uses mobile AHSA observations to obtain the differential oblique column concentration results of spatially distributed atmospheric pollutants (nitrogen dioxide, sulfur dioxide, formaldehyde, etc.), uses meteorological and chemical model simulations to obtain atmospheric information, inputs it into the atmospheric radiation transmission to calculate the atmospheric photon path information, and further converts the differential oblique column concentration results into tropospheric vertical column concentration results. Combined with the wind field information simulated by the meteorological and chemical model, the emission flux results of the pollution source are calculated.

[0004] The patent application with publication number CN115824979A discloses a method for quantitatively calculating the emission intensity of active nitrogen from agricultural sources based on stereo remote sensing, including: multi-dimensional remote sensing measurement, collecting solar scattering spectra, obtaining the gas oblique column concentration after inversion, and obtaining the vertical column concentration after conversion using the atmospheric photon scattering path; combining the model wind field data to quantify the total amount of agricultural main nitrogen-containing components emitted at each grid point in the region, and then subtracting the exchange amount of active nitrogen between all agricultural areas and outside the agricultural areas to obtain the total active nitrogen emissions from agricultural sources.

[0005] In the above two technical solutions, only SCD (slant column concentration) can be measured in gas pollutant emission detection, but the required absolute concentration cannot be determined. With the rise of deep learning in recent years, it has become possible to directly invert the spectrum to the absolute concentration. Summary of the invention

[0006] In view of the above, the purpose of the present invention is to provide a spectral decomposition method and system of an interpretable neural network based on transfer learning, combining spectral remote sensing monitoring technology, transfer learning, and interpretable deep learning to achieve prediction from raw spectra to absolute concentrations of polluted gases, providing a scientific basis for improving the accuracy of remote sensing passive ultraviolet imaging and increasing the credibility of monitoring data.

[0007] To achieve the above-mentioned object of the invention, the embodiment provides a spectral decomposition method of an interpretable neural network based on transfer learning, comprising the following steps:

[0008] The first spectral data of standard gases with different concentrations at different angles and the corresponding first label data are measured using the constructed experimental platform. At the same time, the second spectral data of standard gases with different concentrations at different angles and the corresponding second label data are obtained based on the radiation transmission model simulation. The first spectral data and the corresponding first label data, the second spectral data and the corresponding second label data are randomly mixed to obtain the first spectral data set and the corresponding first label set.

[0009] Using the first spectral data set, supervised learning is performed on a neural network based on a first label set and on a task of spectrally decomposing absolute concentrations of pollutant gases;

[0010] The spectral data collected on site are constructed into a second spectral data set and a corresponding second label set, and the second spectral data set is used to perform transfer learning on the neural network after supervised learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, so as to obtain a spectral decomposition model;

[0011] The spectral data to be predicted is input into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of the pollutant gas. At the same time, the interpretable relationship between the spectra and absolute concentrations of different pollutant gases in the spectral decomposition model is analyzed using an interpretable analyzer.

[0012] Preferably, the experimental platform comprises, from bottom to top: a bracket, a pan-tilt connected to the bracket through a connector, an optical plate connected to the pan-tilt through an adapter, a light-proof air cavity fixed to the optical plate, a light-collecting device disposed at the rear end of the air cavity, and a spectrometer connected to the output end of the light-collecting device;

[0013] The constructed experimental platform is used to measure the first spectrum data and the corresponding first label data of standard gases of different concentrations at different angles, including:

[0014] Air is pumped into the air cavity, and a reference spectrum is collected by a spectrometer facing the sky; a mixed gas of standard gases of different concentrations is introduced into the air cavity, and the air cavity is adjusted to different elevation angles and azimuth angles, and an absorption spectrum of the mixed gas is collected by a spectrometer; the absorption spectrum is used to remove the background of the absorption spectrum using the reference spectrum to form the first spectrum data, and at the same time, the absolute concentration true value of the standard gas is obtained by the standard gas controller to construct the first label data corresponding to the first spectrum data.

[0015] Preferably, the second spectrum data of the standard gas with different concentrations at different angles and the corresponding second label data are obtained based on the radiation transmission model simulation, including:

[0016] The constructed experimental platform is used to measure the first spectral data of standard gases with different concentrations at different angles, and at the same time, the vertical profile data of aerosols at the same time and angle are recorded. The absorption spectrum and reference spectrum of the standard gas prepared during the experiment are simulated based on the vertical profile data through the radiation transfer model. The absorption spectrum obtained after background removal from the reference spectrum constitutes the second spectral data. At the same time, the absolute concentration true value of the standard gas corresponding to the spectral data is recorded as the second label data corresponding to the second spectral data.

[0017] Preferably, the neural network adopts a CNN-transformer-based neural network, including: a convolution module, a transformer encoder, a transformer decoder, and a fully connected layer;

[0018] The input spectral data is processed through a convolution module to extract data features. The extracted data features are converted through the transformer encoder and transformer decoder in turn and then input into the fully connected layer. The fully connected layer performs a mapping prediction from the features to the absolute concentrations of multiple pollutant gases to obtain the absolute concentration prediction results of different pollutant gases.

[0019] Preferably, when supervised learning is performed on the neural network, the loss function used is the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the true value of the absolute concentration of the pollutant gas in the first label set.

[0020] Preferably, constructing the spectral data collected on site into a second spectral data set and a corresponding second label set includes:

[0021] In the actual industrial scene, field data collection is conducted on polluted gas emission outlets, including: sampling the polluted gas emission outlets, and quantitatively testing the sampled samples to obtain the absolute concentration true value of the pollutant gas to construct a second label set, and remote sensing spectral collection of the emission spectra of the same polluted gas emission outlet at the same location and time to obtain spectral data to construct a second spectral data set.

[0022] Preferably, when transfer learning is performed on the neural network after supervised learning, the loss function used is the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the true value of the absolute concentration of the pollutant gas in the second label set.

[0023] Preferably, the interpretable relationship between the spectra and absolute concentrations of different pollutant gases in the spectral decomposition model is analyzed using an interpretable analyzer, including:

[0024] The SHAP interpreter is used to calculate the SHAP value corresponding to each feature in the spectral data to be predicted, and based on the SHAP value, the influence relationship between the input features and the absolute concentration output by the spectral decomposition model is displayed. This influence relationship can analyze the interpretability between the spectra and absolute concentrations of different pollutant gases.

[0025] To achieve the above invention, an embodiment of the present invention further provides a spectral decomposition device of an interpretable neural network based on transfer learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the following steps are implemented:

[0026] Using the first spectral data set, supervised learning is performed on a neural network based on a first label set and on a task of spectrally decomposing absolute concentrations of pollutant gases;

[0027] Using the second spectral data set, the supervised learning neural network is subjected to transfer learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, so as to obtain a spectral decomposition model;

[0028] The spectral data to be predicted is input into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of pollutant gas, and the interpretable relationship between the spectra of different pollutant gases and the absolute concentration in the spectral decomposition model is analyzed by using an interpretable analyzer;

[0029] The first spectral data set and the corresponding first label set are constructed in the following manner, including: using the constructed experimental platform to measure the first spectral data of the standard gas with different concentrations at different angles and the corresponding first label data, and at the same time simulating the radiation transmission model to obtain the second spectral data of the standard gas with different concentrations at different angles and the corresponding second label data, and randomly mixing the first spectral data and the corresponding first label data, the second spectral data and the corresponding second label data to obtain the first spectral data set and the corresponding first label set;

[0030] The second spectral data set and the corresponding second label set are constructed based on spectral data collected in the field.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] For the first time, the use of interpretable neural networks in spectral decomposition was proposed. For gas inversion, it no longer relies solely on the DOAS method or the traditional neural network method, which enhances the transparency and credibility of spectral decomposition, making the monitoring results not only accurate but also interpretable. At the same time, the transfer learning method was adopted. For the actual industrial scene samples of polluted gases that are difficult to collect, a small range of actual samples were used to optimize the entire neural network, making the deployment easier and the results more reasonable, thereby realizing the prediction from the original spectrum to the absolute concentration of polluted gases. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0034] Figure 1 Flow chart of the spectral decomposition method based on transfer learning interpretable neural network;

[0035] Figure 2 Schematic diagram of the experimental platform for measuring different concentrations at different angles;

[0036] Figure 3 This is a structural diagram of a neural network based on CNN-transformer;

[0037] Figure 4 Flowchart for transfer learning of neural network M1 for a second spectral dataset D2 and a second label set L2 collected in a field industrial scene;

[0038] Figure 5 Absolute concentration results for plume observations in actual industrial scenarios. DETAILED DESCRIPTION

[0039] To make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the scope of protection of the present invention.

[0040] As an effective machine learning strategy, transfer learning can accelerate and improve the learning efficiency of new tasks by leveraging models pre-trained on related tasks. It does not require a large amount of data and can solve the problem of difficulty in collecting actual data. Interpretable neural networks further enhance the transparency and credibility of the model, making the monitoring results not only accurate but also interpretable. This feature is particularly important for understanding complex atmospheric chemical reactions and pollutant distribution. By explaining the decision-making process of the model, we can better understand the source and propagation mechanism of pollutants. It not only improves the accuracy and efficiency of monitoring technology, but also increases the credibility and practicality of monitoring data.

[0041] Based on the above analysis, an embodiment of the present invention provides a spectral decomposition method of an interpretable neural network based on transfer learning, which combines spectral remote sensing monitoring technology with interpretable deep learning, and performs transfer learning based on a self-designed experimental platform, standard gas experimental spectral observation data, radiation transfer model simulation, and field sampling data to achieve prediction from the original spectrum to the absolute concentration of the polluted gas, overcoming the problems of insufficient accuracy of polluted gas inversion, uncertain photon path, difficulty in actual sample collection, and neural network black box in the field of environmental optical monitoring.

[0042] like Figure 1 As shown, the spectral decomposition method of an interpretable neural network based on transfer learning provided by the embodiment includes the following steps:

[0043] Step 1: Use the constructed experimental platform to measure the first spectral data of standard gases with different concentrations at different angles and the corresponding first label data. Simultaneously, based on the radiation transmission model, simulate the second spectral data of standard gases with different concentrations at different angles and the corresponding second label data. Randomly mix the first spectral data and its corresponding first label data, the second spectral data and its corresponding second label data to obtain the first spectral data set and the corresponding first label set.

[0044] In the embodiment, Figure 2 As shown, the constructed experimental platform includes the following parts from bottom to top: the bottom is supported by a tripod, the tripod and the pan-tilt are connected by a connector, the pan-tilt is connected to the optical plate through an adapter, two light-proof air cavities (such as external light-proof glass air cavities) are fixed on the optical plate, and there are three holes between the air cavity and the outside, namely, the air inlet, the temperature and pressure gauge interface, and the air outlet. A light collecting device is connected to the rear end of the air cavity, and the received scattered light is then transmitted to the spectrometer through the optical fiber.

[0045] The constructed experimental platform is used to measure the first spectrum data and the corresponding first label data of standard gases of different concentrations at different angles, including:

[0046] The air cavity is vacuumed, and air is pumped into the glass air cavity, so that one end of the cavity faces the sky at 90 degrees, and the other end is connected to a spectrometer through a light collecting device to collect a reference spectrum S0; with NO2 and N2 standard gases as standards, a mixed gas of standard gases of different concentrations as shown in Table 1 is introduced into the cavity, and the cavity is adjusted to different elevation angles and azimuth angles, and the other end is connected to a spectrometer through a light collecting device to collect an absorption spectrum S of the mixed gas; the reference spectrum S0 and each band in the absorption spectrum S are subjected to corresponding background removal operations to obtain an absorption spectrum S' after removing the background concentration in the atmosphere; the background removal formula is shown in formula (1), and the absorption spectrum S' is composed of the first spectrum data, and the absolute concentration true value of the standard gas is obtained by the standard gas controller to construct the first label data corresponding to the first spectrum data.

[0047] Table 1

[0048]

[0049]

[0050] Where I0(λ) is the light intensity of the reference spectrum S0 at wavelength λ, I(λ) is the light intensity of the absorption spectrum S at wavelength λ, and I′0(λ) is the light intensity of the absorption spectrum at wavelength λ after removing the background.

[0051] In the embodiment, while measuring the first spectral data of the standard gas with different concentrations at different angles using the constructed experimental platform, the second spectral data of the standard gas with different concentrations at different angles and the corresponding second label data are also obtained based on the radiation transfer model simulation, including: recording the vertical profile data of the aerosol at the same time and the same angle, and simulating the spectral data of the NO2 standard gas prepared during the experiment based on the vertical profile data through the radiation transfer model (such as SCIATRAN), which are the reference spectra S 02 The absorption spectrum S2 is used to remove the background from each band. The principle is the same as that of formula (1). The absorption spectrum S2' is obtained after the background concentration in the atmosphere is removed. The absorption spectrum S2' is used to form the second spectrum data. At the same time, the absolute concentration true value of the standard gas corresponding to the spectrum data is recorded as the second label data corresponding to the second spectrum data.

[0052] In the embodiment, the first spectral data and its corresponding first label data, the second spectral data and its corresponding second label data are randomly mixed to obtain a first spectral data set D1 and a corresponding first label set L1. The formula used for random mixing is:

[0053] I′ i (λ) = x·I′ x (λ)+y·I′ y (λ) (2)

[0054] In the formula, I′i (λ) is the light intensity of the absorption spectrum S′ at wavelength λ in the first spectral data set D1, where x and y are the proportions, and x+y=1, I′ x (λ) represents the light intensity at wavelength λ of the absorption spectrum with a proportion of x from the first spectrum data, I′ y (λ) represents the light intensity at wavelength λ of the absorption spectrum with a proportion y from the second spectrum data.

[0055] Step 2: Using the first spectral data set, supervised learning is performed on a neural network based on the first label set and on the task of spectrally decomposing the absolute concentration of pollutant gases.

[0056] In the embodiment, the neural network adopts a CNN-transformer-based neural network, including: a convolution module composed of a convolution layer and a pooling layer, a transformer encoder, a transformer decoder, and a fully connected layer. The input spectral data is subjected to the convolution module to extract data features, and the extracted data features are successively subjected to feature conversion by the transformer encoder and the transformer decoder and then input to the fully connected layer, and the fully connected layer performs a mapping prediction from the features to the absolute concentrations of multiple pollutant gases to obtain the absolute concentration prediction results of different pollutant gases.

[0057] When the neural network is subjected to supervised learning, the first spectral data set D1 is input into the neural network, and the neural network calculates the output absolute concentration prediction result L'1 of the pollutant gas through forward reasoning. Then, the loss function is constructed by the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the true value of the absolute concentration of the pollutant gas in the first label set L1, and the network parameters are updated using the Adam optimizer based on the loss function.

[0058] Specifically, the structure of the neural network is as follows Figure 3 As shown in Figure 2, the supervised learning process based on the neural network structure is:

[0059] (1) The input data extracted from the first spectral dataset D1 is denoted as X. After the first module, two one-dimensional convolutional layers are applied, each with 16 channels, using a convolution kernel of size 3 and setting an edge padding of 1 to keep the dimension of the data unchanged. Immediately after the convolutional layer, the data is passed through a maximum pooling layer, downsampled using a window size of 3, and the resulting data is denoted as F1(X).

[0060] (2) F1(X) is passed through the second module: two one-dimensional convolutional layers, each with 64 channels, using a convolution kernel of size 3 and setting an edge padding of 1 to keep the data dimension unchanged. Immediately after the convolutional layer, the data is passed through a maximum pooling layer and downsampled using a window size of 3. The resulting data is recorded as F2(X);

[0061] (3) F2(X) is passed through a convolutional layer with 64 channels, a convolution kernel of size 3, and an edge padding of 1. The resulting data is recorded as F3(X).

[0062] (4) F3(X) is passed through a Transformer Encoder layer with a model size of 64 and 8 heads, a Transformer Decoder layer with a model size of 64 and 8 heads, and a fully connected layer with a 64-dimensional input and a 2-dimensional output. The resulting data is recorded as L'1.

[0063] The loss function MSE used to train the CNN-transformer-based neural network M1 is:

[0064]

[0065] Where L' 1i represents the absolute concentration prediction result of the i-th absorption spectrum obtained by the neural network M1, L 1i represents the absolute concentration true value of the i-th absorption spectrum;

[0066] (5) According to the training error, Adam is used to automatically adjust the learning rate of each parameter at different training stages to perform gradient descent on the error of the entire neural network M1. The neural network after supervised learning is recorded as M2.

[0067] Step 3, constructing the spectral data collected on site into a second spectral data set and a corresponding second label set, and using the second spectral data set to perform transfer learning on the neural network after supervised learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, to obtain a spectral decomposition model;

[0068] In the embodiment, constructing the spectral data collected on site into a second spectral data set and a corresponding second label set includes:

[0069] In the field industrial scene, the polluted gas emission outlet is collected on the spot. The collection is divided into two parts, including: the first part is to sample the polluted gas emission outlet, and quantitatively detect the sampled samples to obtain the absolute concentration true value of the pollutant gas to construct the second label set L2, and the second part is to collect the emission spectrum of the same polluted gas emission outlet at the same location and time by remote sensing spectrum, and obtain the spectrum data to construct the second spectrum data set D2. Among them, the second label set L2 has the same data dimension size as the first label set L1, and the second spectrum data set D2 has the same data dimension size as the first spectrum data set D1, and the input and output structure of the neural network M2 does not need to be modified.

[0070] The second spectral data set D2 and the corresponding second label set L2 are used to perform transfer learning on the neural network M2 after supervised learning. The specific process is as follows: Figure 4 As shown, the neural network is fine-tuned to adapt to the new environmental variables, where the formula for adjusting the learning rate is:

[0071]

[0072] Where η old represents the learning rate before adjustment, η new represents the adjusted learning rate, θ represents the model parameters, and E represents the error.

[0073] In transfer learning, Adam is still used to automatically adjust the learning rate of each parameter at different training stages, so that the error of the entire neural network M2 is gradient decreased, and the spectral decomposition model M3 is obtained by transfer learning. The loss function used in transfer learning is still the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the absolute concentration true value of the pollutant gas in the second label set. The formula is:

[0074]

[0075] Where L' 2i represents the absolute concentration prediction result of the i-th absorption spectrum obtained by the neural network M2, L 2i Represents the absolute concentration true value of the i-th spectrum obtained by chemical quantitative sampling.

[0076] Step 4: Input the spectral data to be predicted into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of the pollutant gas, and use the interpretable analyzer to analyze the interpretability relationship between the spectra of different pollutants and the absolute concentrations in the spectral decomposition model.

[0077] In the embodiment, the spectral data to be predicted collected from the actual industrial scene is input into the spectral decomposition model M3, and the absolute concentration of the polluted gas is obtained through spectral decomposition by forward reasoning. The absolute concentration results of the NO2 plume observed in the actual industrial scene in the experimental example are shown in Figure 5 .

[0078] In an embodiment, an interpretable is used to analyze the interpretability relationship between different pollutant gas spectra and absolute concentrations in a spectral decomposition model, including: calculating the SHAP value corresponding to each feature in the spectral data to be predicted through a SHAP interpreter, and displaying the influence relationship between the input features and the absolute concentration output by the spectral decomposition model based on the difference between the SHAP value and the background data set, and the influence relationship can analyze the interpretability between different pollutant gas spectra and absolute concentrations.

[0079] The specific process is:

[0080] (1) The second spectral data set D2 is used as the background data set, the spectral data to be predicted D3 is input into the spectral decomposition model M3, and the SHAP value of the spectral data D3 is calculated by the SHAP interpreter;

[0081] Define the model function: f(S) represents the model function, which includes a subset S of the feature set. S can be any combination of features in the model.

[0082] Marginal contribution: For feature j, its marginal contribution refers to the change in model output f(S) when j is added to a feature subset S that does not contain j;

[0083] Average marginal contribution: Calculate the average prediction change caused by adding feature j to all possible subsets S. The calculation formula is as follows:

[0084]

[0085] Where F is the set of all features, S is any subset that does not contain feature j, |S| is the size of set S, |F| is the total number of all features, and φ j Represents the SHAP value of feature j for the prediction result.

[0086] (2) Compare the SHAP values ​​of the spectral data D3 with those of the background dataset to show how each input feature affects the output of the spectral decomposition model M3.

[0087] The embodiment further provides a spectral decomposition device of an interpretable neural network based on transfer learning, comprising a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, the following steps are implemented:

[0088] Using the first spectral data set, supervised learning is performed on a neural network based on a first label set and on a task of spectrally decomposing absolute concentrations of pollutant gases;

[0089] Using the second spectral data set, the supervised learning neural network is subjected to transfer learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, so as to obtain a spectral decomposition model;

[0090] The spectral data to be predicted is input into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of the pollutant gas. At the same time, the interpretable relationship between the spectra and absolute concentrations of different pollutant gases in the spectral decomposition model is analyzed using an interpretable analyzer.

[0091] Each step performed in the device is the same as each step in the method and will not be repeated here.

[0092] The specific implementation methods described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A spectral decomposition method of interpretable neural network based on transfer learning, characterized in that: The following steps are involved: The first spectral data of standard gases with different concentrations at different angles and the corresponding first label data are measured using the constructed experimental platform. At the same time, the second spectral data of standard gases with different concentrations at different angles and the corresponding second label data are obtained based on the radiation transmission model simulation. The first spectral data and the corresponding first label data, the second spectral data and the corresponding second label data are randomly mixed to obtain the first spectral data set and the corresponding first label set. Using the first spectral data set, supervised learning is performed on a neural network based on a first label set and on a task of spectrally decomposing absolute concentrations of pollutant gases; Constructing the spectral data collected on site into a second spectral data set and a corresponding second label set, including: conducting field collection of polluted gas emission outlets in a field industrial scene, specifically including: sampling the polluted gas emission outlets, and conducting quantitative detection on the sampled samples to obtain the absolute concentration true value of the pollutant gas to construct the second label set, and conducting remote sensing spectral collection on the emission spectra of the same polluted gas emission outlet at the same position and time to obtain spectral data to construct the second spectral data set; Using the second spectral data set, the supervised learning neural network is subjected to transfer learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, so as to obtain a spectral decomposition model; The spectral data to be predicted is input into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of the pollutant gas. At the same time, the interpretable relationship between the spectra and absolute concentrations of different pollutant gases in the spectral decomposition model is analyzed using an interpretable analyzer.

2. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: The experimental platform includes, from bottom to top, a bracket, a pan-tilt connected to the bracket through a connector, an optical plate connected to the pan-tilt through an adapter, a light-proof air cavity fixed to the optical plate, a light-collecting device disposed at the rear end of the air cavity, and a spectrometer connected to the output end of the light-collecting device; The constructed experimental platform is used to measure the first spectrum data and the corresponding first label data of standard gases of different concentrations at different angles, including: Air is pumped into the air cavity, and a reference spectrum is collected by a spectrometer facing the sky; a mixed gas of standard gases of different concentrations is introduced into the air cavity, and the air cavity is adjusted to different elevation angles and azimuth angles, and an absorption spectrum of the mixed gas is collected by a spectrometer; the absorption spectrum is used to remove the background of the absorption spectrum using the reference spectrum to form the first spectrum data, and at the same time, the absolute concentration true value of the standard gas is obtained by the standard gas controller to construct the first label data corresponding to the first spectrum data.

3. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: Based on the radiation transmission model simulation, the second spectrum data of the standard gas with different concentrations at different angles and the corresponding second label data are obtained, including: The constructed experimental platform is used to measure the first spectral data of standard gases with different concentrations at different angles, and at the same time, the vertical profile data of aerosols at the same time and angle are recorded. The absorption spectrum and reference spectrum of the standard gas prepared during the experiment are simulated based on the vertical profile data through the radiation transfer model. The absorption spectrum obtained after background removal from the reference spectrum constitutes the second spectral data. At the same time, the absolute concentration true value of the standard gas corresponding to the spectral data is recorded as the second label data corresponding to the second spectral data.

4. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: The neural network adopts a CNN-transformer-based neural network, including: a convolution module, a transformer encoder, a transformer decoder, and a fully connected layer; The input spectral data is processed through a convolution module to extract data features. The extracted data features are converted through the transformer encoder and transformer decoder in turn and then input into the fully connected layer. The fully connected layer performs a mapping prediction from the features to the absolute concentrations of multiple pollutant gases to obtain the absolute concentration prediction results of different pollutant gases.

5. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: When the neural network is supervised for learning, the loss function used is the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the true value of the absolute concentration of the pollutant gas in the first label set.

6. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: When performing transfer learning on the neural network after supervised learning, the loss function used is the mean square error between the absolute concentration prediction result of the pollutant gas output by the neural network and the true value of the absolute concentration of the pollutant gas in the second label set.

7. The spectral decomposition method of interpretable neural network based on transfer learning according to claim 1, characterized in that: The interpretable relationship between the spectra and absolute concentrations of different pollutant gases in the spectral decomposition model is analyzed using the interpretable tool, including: The SHAP interpreter is used to calculate the SHAP value corresponding to each feature in the spectral data to be predicted, and based on the difference between the SHAP value and the background data set, the influence of the input features on the absolute concentration output by the spectral decomposition model is demonstrated. This influence relationship can analyze the interpretability between the spectra and absolute concentrations of different pollutant gases.

8. A spectral decomposition device of an interpretable neural network based on transfer learning, comprising a memory and one or more processors, wherein the memory stores executable code, characterized in that: When the one or more processors execute the executable code, the one or more processors are used to implement the following steps: Using the first spectral data set, supervised learning is performed on a neural network based on a first label set and on a task of spectrally decomposing absolute concentrations of pollutant gases; Using the second spectral data set, the supervised learning neural network is subjected to transfer learning based on the second label set and on the task of spectral decomposition of absolute concentration of pollutant gas, so as to obtain a spectral decomposition model; The spectral data to be predicted is input into the spectral decomposition model for spectral decomposition to obtain the absolute concentration of pollutant gas, and the interpretable relationship between the spectra of different pollutant gases and the absolute concentration in the spectral decomposition model is analyzed by using an interpretable analyzer; The first spectral data set and the corresponding first label set are constructed in the following manner, including: using the constructed experimental platform to measure the first spectral data of the standard gas with different concentrations at different angles and the corresponding first label data, and at the same time simulating the radiation transmission model to obtain the second spectral data of the standard gas with different concentrations at different angles and the corresponding second label data, and randomly mixing the first spectral data and the corresponding first label data, the second spectral data and the corresponding second label data to obtain the first spectral data set and the corresponding first label set; The second spectral data set and the corresponding second label set are constructed based on spectral data collected in the field, including: field collection of polluted gas emission outlets in actual industrial scenes, specifically including: sampling the polluted gas emission outlets, and quantitatively detecting the sampled samples to obtain the absolute concentration true value of the pollutant gas to construct the second label set, and remote sensing spectral collection of the emission spectrum at the same location and time of the same polluted gas emission outlet to obtain spectral data to construct the second spectral data set.

Images