Gaseous pollutant Fourier transform infrared spectroscopic analysis method based on deep learning

Abstract

The invention provides a gaseous pollutant Fourier transform infrared spectroscopic analysis method based on deep learning. The gaseous pollutant Fourier transform infrared spectroscopic analysis method comprises the following steps: acquiring infrared absorbance spectrums of different gases to be detected by utilizing a Fourier transform infrared gas analyzer; the method comprises the following steps of: preprocessing the collected spectral data, namely carrying out segmented denoising on the original spectral data by adopting a self-adaptive Savitzky-Golay filter, and carrying out normalization processing to balance a data range; carrying out feature extraction on the preprocessed spectral data by utilizing a one-dimensional convolution deep learning model so as to adapt to one-dimensional FTIR spectral features and realize feature extraction of multiple components of gas; mapping the extracted features to a prediction result of the to-be-detected component by applying a full connection layer of the model; and training and optimizing the model by adopting an MSE regression loss function. According to the invention, the processing capability of the gaseous pollutant spectral signal can be effectively improved, and high-precision multi-component quantitative analysis is realized.

Description

Technical Field

[0001] The present invention relates to the fields of environmental monitoring, artificial intelligence and spectral analysis technology, and in particular to a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning. Background Art

[0002] Fourier transform infrared spectroscopy (FTIR), a technique for qualitative and quantitative analysis of the chemical composition of samples based on absorption spectral characteristics, has been widely used in various fields. In environmental pollutant analysis, it can accurately detect various pollutants; in industrial production process monitoring, it can provide real-time insights into changes in gas composition during production; and in the field of chemical testing, it also plays an important role. In particular, in atmospheric environmental quality monitoring, gaseous pollutants such as sulfur dioxide, nitrogen oxides, methane, and carbon monoxide are important targets for detection. FTIR technology, with its advantages of non-destructive, rapid, and real-time detection, has become an effective means of detecting the composition and concentration of these pollutants.

[0003] In multi-component, complex gas analysis scenarios, the high overlap of spectral absorption peaks of various pollutants is a common and challenging problem. For example, common industrial waste gas or atmospheric mixed gases may contain multiple pollutants, including sulfur dioxide, nitrogen oxides, methane, and carbon monoxide. The infrared absorption spectra of these pollutants overlap in specific wavelength bands. For example, the absorption peaks of CH4 and C3H8 at certain wavelengths may be very close or even partially overlap. Traditional spectral analysis methods, such as manual rule matching, linear regression, or least squares and partial least squares methods, primarily process spectral data based on simple mathematical models and preset rules. These methods struggle to accurately identify and separate highly overlapping spectral absorption peaks, resulting in low accuracy in detecting pollutant components and concentrations. In practical applications, this can lead to misjudgments of the concentration of certain pollutants, impacting the accuracy of environmental monitoring data and the reliability of industrial process control.

[0004] The actual detection environment is often complex and changeable, and the spectral intensity in different scenarios will be affected by many factors, such as the stability of the light source, changes in temperature and humidity of the detection environment, and the flow state of the gas. In addition, external interference, such as the presence of other non-target gases and electromagnetic interference, will also affect the spectral data. Traditional analysis methods are unable to adapt well to these changes due to the relatively fixed model structure and lack of adaptability to different scenarios and interference factors. For example, in different industrial production workshops, due to different production processes and equipment, the composition and concentration range of the gas may vary greatly, and the temperature, humidity and other environmental conditions of the workshop are also different. The traditional analysis model may work normally in one workshop, but in another workshop, there will be large detection errors, resulting in insufficient robustness of the model, which limits the widespread application of FTIR technology in the detection of multi-component complex gases.

[0005] Existing technologies for multi-component complex gas detection suffer from low accuracy and insufficient model robustness, resulting in numerous adverse impacts in areas such as environmental monitoring and industrial production process control. In environmental monitoring, inaccurate detection data can lead to misjudgments of environmental quality, impacting the formulation of environmental policies and the implementation of environmental protection measures. In industrial production process control, the inability to accurately and accurately monitor gas composition and concentration changes in real time can lead to production instability, increased production costs, and even safety accidents. Therefore, developing new technologies that can improve the accuracy and model robustness of multi-component complex gas detection is of great practical significance.

[0006] In summary, the existing multi-component complex gas analysis methods based on FTIR technology have obvious limitations, and an innovative solution is urgently needed to overcome these problems to meet the needs of high-precision and high-robust gas detection in practical applications. Summary of the Invention

[0007] In response to the shortcomings of the existing technology, the present invention provides a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning. This method can effectively improve the processing capability of gaseous pollutant spectral signals and achieve high-precision multi-component quantitative analysis.

[0008] The present invention achieves the above-mentioned purpose through the following technical solutions:

[0009] A Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning, comprising the following steps:

[0010] The infrared absorbance spectra of different gases to be tested are collected using a Fourier infrared gas analyzer;

[0011] The collected spectral data were preprocessed, including using an adaptive Savitzky-Golay filter to denoise the raw spectral data in sections, automatically adjusting the filter window size and fitting polynomial order according to the local noise level to reduce noise, and performing normalization to balance the data range;

[0012] A one-dimensional convolutional deep learning model is used to extract features from the preprocessed spectral data. The model consists of multiple one-dimensional convolutional layers, ReLU activation functions, and pooling layers to adapt the one-dimensional FTIR spectral features and extract features from multiple gas components.

[0013] Applying the fully connected layer of the one-dimensional convolutional deep learning model to map the extracted features to the predicted results of the components to be tested;

[0014] The MSE regression loss function is used to train and optimize the one-dimensional convolutional deep learning model.

[0015] According to a deep learning-based Fourier transform infrared spectroscopy analysis method for gaseous pollutants provided by the present invention, the adaptive Savitzky-Golay filter is used to segmentally denoise the original spectral data, including:

[0016] The original spectral data is segmented and the local noise level of each segment is calculated. The local noise level is measured by calculating the standard deviation or coefficient of variation of the data in the segment.

[0017] Automatically adjust the Savitzky-Golay filter window size and fitting polynomial order based on the detected local noise level;

[0018] For each segment of the original spectral data, an adaptive Savitzky-Golay filter with corresponding parameters is applied to remove noise and retain the integrity of the spectral absorption peak.

[0019] According to a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning provided by the present invention, the normalization processing in the preprocessing of the collected spectral data is specifically as follows:

[0020] Each spectral sample data output by the adaptive Savitzky-Golay filter is normalized using the following formula:

[0021]

[0022] Where x represents a value in the original spectral data, xmin and xmax represent the minimum and maximum values in the spectral sample data, respectively, and xnorm represents the normalized value.

[0023] According to the present invention, a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning is provided. When using a one-dimensional convolution deep learning model to extract and predict features from preprocessed spectral data, the method includes:

[0024] The one-dimensional convolutional layer has a convolution kernel size of 9 and a channel number of 32, which is used to extract spectral features. Each convolutional layer is connected to a ReLU activation function to increase the nonlinear expression ability of the model. After the convolutional layer, a pooling layer is set to reduce the dimensionality of the features, reduce the amount of computation and avoid overfitting. Finally, a fully connected layer is used as the classification layer to map the features extracted by the convolutional and pooling layers to the confidence level of each component to be measured, and output the prediction results for each gas to be measured.

[0025] According to a deep learning-based Fourier transform infrared spectroscopy analysis method for gaseous pollutants provided by the present invention, an adaptive weight adjustment mechanism is introduced in the fully connected layer. This mechanism dynamically adjusts the weights of the fully connected layer through the following formula to optimize the model's prediction accuracy for different gas components:

[0026]

[0027] Among them, W is the original fully connected layer weight, y pred is the component vector predicted by the model, y true is the actual component vector, and σ is an adjustable hyperparameter used to control the sensitivity of weight adjustment.

[0028] According to a gaseous pollutant Fourier transform infrared spectroscopy analysis method based on deep learning provided by the present invention, the model effect evaluation adopts the MSE regression loss function, which is expressed as the following formula:

[0029]

[0030] y i is the true value, is the predicted value, n is the number of samples, The square error of each sample represents the square of the deviation between the predicted value and the true value.

[0031] According to a method for Fourier transform infrared spectroscopy analysis of gaseous pollutants based on deep learning provided by the present invention, a batch training strategy is adopted during the model training process, and the entire data set is traversed from beginning to end multiple times. Each traversal divides the data into multiple batches for training. The specific steps include:

[0032] Divide the entire training dataset into multiple small batches, each batch contains a certain number of samples;

[0033] In each batch, forward propagation is first performed to pass the input spectral data through the feature extraction layer and classification layer to obtain the prediction result of each sample; then, the loss value is calculated based on the prediction result and the true label;

[0034] Then backpropagation is performed to calculate the gradient based on the loss value and update the model parameters to reduce the loss value; this process is repeated in each batch until all samples in the batch are processed;

[0035] The entire dataset is traversed from beginning to end multiple times, and each traversal is called an epoch. Through training over multiple epochs, the model parameters gradually converge to reach the optimal or suboptimal state.

[0036] According to the present invention, a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning is provided. In the preprocessing stage, the spectrum is dynamically divided into sub-intervals based on the physical characteristics of the spectrum, and the parameters of the Savitzky-Golay filter are adaptively adjusted for different intervals. Specifically, the method includes:

[0037] The absorption peak boundary is detected using the first-order derivative or second-order derivative of the spectrum, and the spectrum is divided into multiple sub-intervals;

[0038] The local signal-to-noise ratio (SNR) is calculated for each subinterval, and the filter window size and polynomial order are dynamically adjusted according to the SNR, including using a low order in the baseline region and a high order in the absorption peak region.

[0039] Wavelet transform or empirical mode decomposition (EMD) is introduced to perform secondary denoising on the residual noise.

[0040] It can be seen that compared with the prior art, the present invention has the following beneficial effects:

[0041] 1. The present invention effectively overcomes the problem of overlapping component features in traditional methods by combining data preprocessing, adaptive filtering and deep learning algorithms. In multi-component complex gas detection scenarios, the spectral absorption peaks of different gaseous pollutants are often highly overlapping, and traditional methods are difficult to accurately distinguish. The deep learning model of the present invention has been trained with a large amount of multi-component gaseous pollutant spectral data and is able to learn the subtle differences in the spectral characteristics of each component. For example, for pollutants such as methane (CH4) and propane (C3H8) whose spectral absorption peaks are prone to overlap, the model can accurately identify their respective unique characteristic patterns, thereby achieving effective separation of overlapping component features and greatly improving the accuracy of detection.

[0042] 2. Unlike traditional methods that rely on manually designed feature extraction rules, the model of the present invention can adaptively learn deep-level features in spectral data. These features can better reflect the essential information of pollutants. By accurately extracting and analyzing these key features, the components and concentrations of each pollutant can be determined more accurately, significantly improving the detection accuracy of multi-component gaseous pollutants and providing more reliable data support for environmental monitoring and industrial production process control. In practical applications, the detection method of the present invention has obvious accuracy advantages over traditional methods, greatly reducing the detection error of major pollutants, and can more accurately reflect the actual content of each component in the exhaust gas, providing a more accurate basis for the environmental protection management of enterprises and the law enforcement of regulatory authorities.

[0043] 3. The deep learning model of the present invention has strong adaptability and can adapt to different experimental conditions. During the actual detection process, the experimental conditions may change due to various factors, such as the stability of the light source, the temperature and humidity of the detection environment, etc. Traditional methods have poor adaptability to these changes due to the fixed model structure, which easily leads to fluctuations in the detection results. The model of the present invention can automatically adapt to different experimental conditions through continuous learning and adjustment. For example, in different seasons or climatic conditions, the temperature and humidity of the detection environment may change significantly, but the model can still maintain stable detection performance, ensuring the accuracy and reliability of the detection results.

[0044] 4. The model of the present invention has a good analytical effect on complex multi-component gaseous pollutant spectral data. The spectral data of multi-component gases usually has the characteristics of nonlinearity, high dimensionality and large noise interference. Traditional methods are often unable to handle this type of data. The deep learning model can handle complex nonlinear relationships and, by learning from large amounts of data, uncover the underlying patterns in the data. Even if there is noise interference or outliers in the spectral data, the model can effectively suppress and correct them through its own robustness mechanism, thereby accurately analyzing the information of each component and improving the model's ability to process complex data.

[0045] 5. The model of this invention has strong generalization capabilities and can maintain good detection performance in different scenarios. For example, in different scenarios such as urban atmospheric environment monitoring, industrial pollution source emission monitoring, and rural ambient air quality monitoring, the model can quickly adapt and accurately detect the content of multi-component gaseous pollutants, greatly improving the applicability and practicality of the detection method and meeting the gas detection needs of different fields.

[0046] 6. The end-to-end analysis and detection mode of the present invention is easy to integrate with other systems, such as environmental monitoring systems, industrial production control systems, etc. Through seamless connection with these systems, it can realize real-time, online monitoring and control of multi-component gaseous pollutants, providing strong technical support for the intelligent management of environmental protection and industrial production.

[0047] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of an embodiment of a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning of the present invention.

[0049] Figure 2 This is an absorbance spectrum of 10 analytes in an embodiment of a Fourier transform infrared spectroscopy method for gaseous pollutants based on deep learning of the present invention.

[0050] Figure 3 This is an absorbance spectrum of a mixture in an embodiment of a Fourier transform infrared spectroscopy method for gaseous pollutants based on deep learning of the present invention.

[0051] Figure 4 It is a schematic diagram of a one-dimensional convolution deep learning model in an embodiment of a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning of the present invention.

[0052] Figure 5 It is a loss curve diagram of a model of an embodiment of a Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning of the present invention.

[0053] Figure 6 This is a schematic diagram of the output of prediction results in an embodiment of a Fourier transform infrared spectroscopy method for gaseous pollutants based on deep learning of the present invention. DETAILED DESCRIPTION

[0054] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the embodiments described are only some of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0055] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0056] See also Figures 1 to 6 This embodiment provides a method for analyzing gaseous pollutants using Fourier transform infrared spectroscopy based on deep learning, the method comprising the following steps:

[0057] Step S1, using a Fourier transform infrared gas analyzer to collect infrared absorbance spectra of different gases to be tested;

[0058] Step S2, preprocessing the collected spectral data, including using an adaptive Savitzky-Golay filter to segmentally denoise the raw spectral data, automatically adjusting the filter window size and the fitting polynomial order according to the local noise level to reduce noise, and performing normalization to balance the data range;

[0059] Step S3, using a one-dimensional convolutional deep learning model to extract features from the preprocessed spectral data. The model includes multiple one-dimensional convolutional layers, ReLU activation functions, and pooling layers to adapt the one-dimensional FTIR spectral features and realize feature extraction of multiple gas components;

[0060] Step S4, applying the fully connected layer of the one-dimensional convolutional deep learning model to map the extracted features to the predicted results of the components to be tested;

[0061] Step S5: Use the MSE regression loss function to train and optimize the one-dimensional convolutional deep learning model.

[0062] In the above step S1, the infrared absorbance spectra of different gases to be tested are collected using a Fourier infrared gas analyzer to obtain the following Figure 2 、 Figure 3 The absorbance spectra of different test gases and their mixtures are shown. The present invention uses propane (C3H8), methane (CH4), carbon monoxide (CO), carbon dioxide (CO2), water (H2O), nitrous oxide (N2O), ammonia (NH3), nitric oxide (NO), nitrogen dioxide (NO2), and sulfur dioxide (SO2) as examples, with a total of 10 test gases. The horizontal axis is the pixel point and the vertical axis is the absorbance value. The specific steps are as follows:

[0063] Construct a gas collection environment, introduce the gas to be measured into the gas flow cell of the Fourier infrared gas analyzer, and control the temperature and pressure of the gas flow cell within a preset stable range to reduce the impact of environmental factors on the infrared absorbance spectrum collection;

[0064] Set the scanning parameters of the Fourier transform infrared gas analyzer, including the scanning wavenumber range, resolution, and scan frequency. The scanning wavenumber range should cover the wavenumber interval where the main absorption peak of the gas to be measured is located. The resolution should be adjusted according to the detection accuracy requirements. The scan frequency should be no less than 10 times / second to improve the signal-to-noise ratio of the spectral data.

[0065] Start the Fourier infrared gas analyzer to perform spectral scanning, collect infrared absorbance spectrum data of different gases to be tested, and store the collected spectral data in the designated data storage module.

[0066] When constructing a gas collection environment, it is also necessary to set up a gas filter device at the air inlet of the gas circulation pool to remove particulate impurities in the gas to be tested and prevent impurities from interfering with the infrared absorbance spectrum collection. At the same time, an exhaust gas treatment device is set up at the air outlet of the gas circulation pool to harmlessly treat the collected gas to avoid pollution to the environment. When setting the scanning parameters of the Fourier infrared gas analyzer, considering the situation where there are multiple gases to be tested, the scanning wavenumber range covers the wavenumber interval where the absorption peaks of all gases to be tested are located. The resolution is selected based on the trade-off between the detection accuracy requirements and the performance of the instrument. When the detection accuracy requirements are high, a higher resolution is selected, but the scanning time will increase accordingly. The number of scans is set according to the signal-to-noise ratio requirements of the spectral data, and the signal-to-noise ratio is improved by taking the average value of multiple scans.

[0067] In the above step S2, the adaptive Savitzky-Golay filter is used to denoise the original spectral data in sections, including:

[0068] The original spectral data is segmented and the local noise level of each segment is calculated. The local noise level can be measured by calculating the standard deviation or coefficient of variation of the data in the segment.

[0069] Automatically adjust the Savitzky-Golay filter window size and fitting polynomial order based on the detected local noise level. Generally, a larger window and a lower order are selected when the noise level is high, and a smaller window and a higher order are selected when the noise level is low, so as to reduce noise while preserving the spectral information to the greatest extent possible.

[0070] For each segment of the original spectral data, an adaptive Savitzky-Golay filter with corresponding parameters is applied to remove noise and retain the integrity of the spectral absorption peak, providing a high-quality data basis for subsequent feature extraction and prediction.

[0071] In the above step S2, the normalization process in the pre-processing of the collected spectral data is specifically as follows:

[0072] Each spectral sample data output by the adaptive Savitzky-Golay filter is normalized using the following formula:

[0073]

[0074] Here, x represents a value in the original spectral data, xmin and xmax represent the minimum and maximum values in the spectral sample data, respectively, and xnorm represents the normalized value. This process scales all spectral sample data to between 0 and 1, helping to improve the deep learning model's recognition performance and prediction accuracy for different gas components.

[0075] In the above step S3, a multi-layer convolutional deep learning structure is used to extract spectral features. The model structure is as follows: Figure 4 When using a one-dimensional convolutional deep learning model to extract and predict features from preprocessed spectral data, it includes:

[0076] The one-dimensional convolutional layer has a convolution kernel size of 9 and a channel number of 32, which is used to extract features from the spectral data. Each convolutional layer is followed by a ReLU activation function to increase the nonlinear expression capability of the model. After the convolutional layer, a pooling layer is set to reduce the dimensionality of the features, such as using maximum pooling (Max Pooling) or average pooling (Average Pooling), to reduce computational complexity and avoid overfitting. Finally, a fully connected layer is used as the classification layer to map the features extracted by the convolutional and pooling layers to the confidence level of each component to be measured, and output the prediction results for each gas to be measured.

[0077] Through the above steps, the one-dimensional convolutional deep learning model can efficiently adapt to the one-dimensional FTIR spectral characteristics, realize the feature extraction and prediction of multi-component gas, and improve the detection accuracy of multi-component gaseous pollutants and the robustness of the model.

[0078] In the above step S4, the fully connected layer adopts a multi-layer perceptron MLP structure, the number of nodes in the fully connected layer is set according to the number of components to be measured, and the output layer adopts a linear activation function, wherein the fully connected layer includes at least one hidden layer, and a batch normalization layer is added after each hidden layer to normalize the hidden layer output. The number of hidden layer nodes is set according to the feature complexity and model capacity requirements, and the ReLU activation function is used to increase the nonlinear expression capability.

[0079] In this embodiment, an adaptive weight adjustment mechanism is introduced in the fully connected layer. This mechanism dynamically adjusts the weights of the fully connected layer using the following formula to optimize the model's prediction accuracy for different gas components:

[0080]

[0081] Among them, W is the original fully connected layer weight, y pred is the component vector predicted by the model, y true is the actual component vector, and σ is an adjustable hyperparameter that controls the sensitivity of weight adjustment. When the difference between the predicted and actual results is large, the corresponding weight adjustment will also increase, thereby accelerating model convergence and improving prediction accuracy.

[0082] During the forward propagation of the model, the input spectral data is processed by the following steps:

[0083] Feature extraction layer processing: The input spectral data first passes through the feature extraction layer, which includes multiple convolutional layers, ReLU activation functions, and pooling layers. Through layer-by-layer convolution and nonlinear transformation, the global and local features in the spectral data are extracted to finally obtain the feature vector;

[0084] Classification layer processing: The feature vector output by the feature extraction layer is input to the classification layer, which adopts a fully connected layer structure;

[0085] Output confidence: The classification layer calculates the confidence that each sample belongs to each category through linear transformation and activation function (such as softmax function, if normalization to probability distribution is required), and finally outputs the prediction result for each gas to be tested.

[0086] Through the above-mentioned forward propagation process, the model can efficiently extract features from spectral data and accurately predict the results of each component in each sample, thereby achieving efficient qualitative and quantitative analysis of multi-component gaseous pollutants.

[0087] In the above step S5, if Figure 5 As shown, the model effect evaluation uses the MSE regression loss function, which is expressed as the following formula:

[0088]

[0089] y i is the true value (actual value), is the predicted value (model output), n is the number of samples (number of data points, or batch size), The square error of each sample represents the square of the deviation between the predicted value and the true value.

[0090] The trained model is used to predict the concentrations of 10 analytes and the model convergence status is as follows: Figure 5 As shown, the model effect is evaluated as follows Figure 6 shown.

[0091] During the model training process, a batch training strategy is adopted to traverse the entire dataset from beginning to end multiple times. Each traversal divides the data into multiple batches for training. The specific steps include:

[0092] Divide the entire training dataset into multiple small batches, each batch contains a certain number of samples;

[0093] In each batch, forward propagation is first performed to pass the input spectral data through the feature extraction layer and classification layer to obtain the prediction result of each sample; then, the loss value is calculated based on the prediction result and the true label;

[0094] Then, backpropagation is performed to calculate the gradient based on the loss value and update the model parameters to reduce the loss value. This process is repeated in each batch until all samples in the batch are processed.

[0095] The entire dataset is traversed from beginning to end multiple times, each traversal is called an epoch. Through multiple epochs of training, the model parameters gradually converge to the optimal or suboptimal state. At the same time, after each batch, the model is evaluated, the training loss and test loss are recorded, and the learning rate is adjusted based on the test loss.

[0096] Furthermore, in the data preprocessing stage, the spectrum is dynamically divided into sub-intervals based on the physical characteristics of the spectrum, and the parameters of the Savitzky-Golay filter are adaptively adjusted for different intervals, including:

[0097] The absorption peak boundary is detected using the first-order derivative or second-order derivative of the spectrum, and the spectrum is divided into multiple sub-intervals;

[0098] The local signal-to-noise ratio (SNR) is calculated for each subinterval, and the filter window size and polynomial order are dynamically adjusted according to the SNR, including using a low order in the baseline region and a high order in the absorption peak region.

[0099] Wavelet transform or empirical mode decomposition (EMD) is introduced to perform secondary denoising on the residual noise.

[0100] In summary, this embodiment effectively overcomes the problem of overlapping component features in traditional methods by combining data preprocessing, adaptive filtering and deep learning algorithms. In multi-component complex gas detection scenarios, the spectral absorption peaks of different gaseous pollutants are often highly overlapping, and traditional methods are difficult to accurately distinguish. The deep learning model of the present invention is trained with a large amount of multi-component gaseous pollutant spectral data and is able to learn the subtle differences in the spectral characteristics of each component. For example, for pollutants such as methane-CH4 and propane-C3H8 whose spectral absorption peaks are prone to overlap, the model can accurately identify their respective unique characteristic patterns, thereby achieving effective separation of overlapping component features and greatly improving the accuracy of detection.

[0101] Unlike traditional methods that rely on manually designed feature extraction rules, the model of this embodiment can adaptively learn deep-level features in spectral data. These features can better reflect the essential information of pollutants. By accurately extracting and analyzing these key features, the components and concentrations of each pollutant can be determined more accurately, significantly improving the detection accuracy of multi-component gaseous pollutants and providing more reliable data support for environmental monitoring and industrial production process control. In practical applications, the detection method of the present invention has obvious accuracy advantages over traditional methods, greatly reducing the detection error of major pollutants, and can more accurately reflect the actual content of each component in the exhaust gas, providing a more accurate basis for the environmental protection management of enterprises and the law enforcement of regulatory authorities.

[0102] During actual testing, experimental conditions may vary due to various factors, such as the stability of the light source, the temperature and humidity of the testing environment, and so on. Traditional methods, due to their fixed model structure, have poor adaptability to these changes, which can easily lead to fluctuations in test results. However, the model of this embodiment can automatically adapt to different experimental conditions through continuous learning and adjustment. For example, the temperature and humidity of the testing environment may vary significantly depending on the season or climate, but the model can still maintain stable detection performance, ensuring the accuracy and reliability of the test results.

[0103] The model of this embodiment has a good analytical effect on complex multi-component gaseous pollutant spectral data. The spectral data of multi-component gases usually has the characteristics of nonlinearity, high dimensionality and large noise interference. Traditional methods are often unable to cope with this type of data. The deep learning model can handle complex nonlinear relationships and mine the potential laws in the data by learning from large amounts of data. Even if there is noise interference or outliers in the spectral data, the model can effectively suppress and correct them through its own robustness mechanism, thereby accurately analyzing the information of each component and improving the model's ability to process complex data.

[0104] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0105] The above embodiments are only preferred embodiments of the present invention and cannot be used to limit the scope of protection of the present invention. Any non-substantial changes and replacements made by technicians in this field on the basis of the present invention fall within the scope of protection required by the present invention.

Claims

1. A Fourier transform infrared spectroscopy analysis method for gaseous pollutants based on deep learning, characterized in that: The following steps are involved: The infrared absorbance spectra of different gases to be tested are collected using a Fourier infrared gas analyzer; The collected spectral data were preprocessed, including using an adaptive Savitzky-Golay filter to denoise the raw spectral data in sections, automatically adjusting the filter window size and fitting polynomial order according to the local noise level to reduce noise, and performing normalization to balance the data range; A one-dimensional convolutional deep learning model is used to extract features from the preprocessed spectral data. The model consists of multiple one-dimensional convolutional layers, ReLU activation functions, and pooling layers to adapt the one-dimensional FTIR spectral features and extract features from multiple gas components. Applying the fully connected layer of the one-dimensional convolutional deep learning model to map the extracted features to the prediction results of the components to be tested; The MSE regression loss function is used to train and optimize the one-dimensional convolutional deep learning model.

2. The method according to claim 1, characterized in that The adaptive Savitzky-Golay filter is used to denoise the original spectral data in sections, including: The original spectral data is segmented and the local noise level of each segment is calculated. The local noise level is measured by calculating the standard deviation or coefficient of variation of the data in the segment. Automatically adjust the Savitzky-Golay filter window size and fitting polynomial order based on the detected local noise level; For each segment of the original spectral data, an adaptive Savitzky-Golay filter with corresponding parameters is applied to remove noise and retain the integrity of the spectral absorption peak.

3. The method according to claim 2, characterized in that The normalization process in the preprocessing of the collected spectral data is specifically as follows: Each spectral sample data output by the adaptive Savitzky-Golay filter is normalized using the following formula: Where x represents a value in the original spectral data, xmin and xmax represent the minimum and maximum values in the spectral sample data, respectively, and xnorm represents the normalized value.

4. The method according to claim 1, wherein When using a one-dimensional convolutional deep learning model to extract and predict features from preprocessed spectral data, it includes: The one-dimensional convolutional layer has a convolution kernel size of 9 and a channel number of 32, which is used to extract spectral features. Each convolutional layer is connected to a ReLU activation function to increase the nonlinear expression ability of the model. After the convolutional layer, a pooling layer is set to reduce the dimensionality of the features, reduce the amount of computation and avoid overfitting. Finally, a fully connected layer is used as the classification layer to map the features extracted by the convolutional and pooling layers to the confidence level of each component to be measured, and output the prediction results for each gas to be measured.

5. The method according to claim 4, characterized in that: The fully connected layer adopts a multi-layer perceptron MLP structure, and the output layer adopts a linear activation function to directly output the predicted value of each component to be tested.

6. The method according to claim 5, characterized in that: An adaptive weight adjustment mechanism is introduced in the fully connected layer. This mechanism dynamically adjusts the weights of the fully connected layer through the following formula to optimize the model's prediction accuracy for different gas components: Among them, W is the original fully connected layer weight, y pred is the component vector predicted by the model, y true is the actual component vector, and σ is an adjustable hyperparameter used to control the sensitivity of weight adjustment.

7. The method according to claim 1, wherein: The model effect evaluation uses the MSE regression loss function, which is expressed as the following formula: y i is the true value, is the predicted value, n is the number of samples, The square error of each sample represents the square of the deviation between the predicted value and the true value.

8. The method according to any one of claims 1 to 7, characterized in that During the model training process, a batch training strategy is adopted to traverse the entire dataset from beginning to end multiple times. Each traversal divides the data into multiple batches for training. The specific steps include: Divide the entire training dataset into multiple small batches, each batch contains a certain number of samples; In each batch, forward propagation is first performed to pass the input spectral data through the feature extraction layer and classification layer to obtain the prediction result of each sample; then, the loss value is calculated based on the prediction result and the true label; Then backpropagation is performed to calculate the gradient based on the loss value and update the model parameters to reduce the loss value; this process is repeated in each batch until all samples in the batch are processed; The entire dataset is traversed from beginning to end multiple times, and each traversal is called an epoch. Through training over multiple epochs, the model parameters gradually converge to reach the optimal or suboptimal state.

9. The method according to any one of claims 1 to 7, characterized in that: In the preprocessing stage, the spectrum is dynamically divided into sub-intervals based on the physical characteristics of the spectrum, and the parameters of the Savitzky-Golay filter are adaptively adjusted for different intervals. Specifically, The absorption peak boundary is detected using the first-order derivative or second-order derivative of the spectrum, and the spectrum is divided into multiple sub-intervals; The local signal-to-noise ratio (SNR) is calculated for each subinterval, and the filter window size and polynomial order are dynamically adjusted according to the SNR, including using low order in the baseline area and high order in the absorption peak area; wavelet transform or empirical mode decomposition (EMD) is introduced to perform secondary denoising on the residual noise.

Images