Coal-fired flue gas nitric oxide concentration prediction method based on bimodal feature fusion
Through the method of dual-modal feature fusion, one-dimensional photoacoustic spectral data is combined with two-dimensional photoacoustic spectral images to establish a deep learning model, solving the accuracy and robustness of nitric oxide concentration prediction in coal-fired flue gas, and achieving a more efficient prediction effect.
Patent Information
- Application Number
- CN202510518730.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-23
- Publication Date
- 2025-07-22
AI Technical Summary
The existing photoacoustic spectroscopy technology has problems with low prediction accuracy and robustness in the prediction of nitric oxide concentration in coal-fired flue gas. It is mainly because the photoacoustic signal processing method focuses on one-dimensional signal processing, making it difficult to fully tap rich information in complex environments.
The two-modal feature fusion method is used to combine the original one-dimensional photoacoustic spectral data with the two-dimensional photoacoustic spectral image processed by the Gram angle field, and feature extraction and fusion are performed through deep learning models to establish a prediction model for bimodal feature fusion.
It significantly improves the prediction accuracy and stability of nitric oxide concentration in coal-fired flue gas, optimizes the calculation efficiency, and improves the accuracy and reliability of prediction.
Smart Images
Figure CN120354364A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of predicting the concentration of nitrogen monoxide in coal-fired flue gas, and particularly to a method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on dual-modal feature fusion. Background Art
[0002] With the increasingly strict global environmental protection policies, harmful gases emitted by industrial facilities such as coal-fired power plants, especially nitrogen monoxide, have become serious environmental pollution sources. The concentration of nitrogen monoxide in coal-fired flue gas affects air quality and public health. Therefore, accurately predicting the concentration of nitrogen monoxide in coal-fired flue gas is crucial for effectively controlling emissions and optimizing treatment measures. Currently, common gas concentration measurement techniques include chemical absorption method, electrochemistry method, and infrared spectroscopy method, etc. However, the application of these traditional methods in complex industrial environments is restricted in many aspects. For example, the chemical absorption method requires the consumption of reagents and has a slow response speed; although the electrochemistry method is convenient, its stability and long-term reliability are poor; although the infrared spectroscopy method has high precision, the equipment cost is large, and it is severely interfered by environmental factors.
[0003] In recent years, photoacoustic spectroscopy technology has shown great potential in the field of gas monitoring due to its advantages such as non-contact, high sensitivity, and fast response speed. Photoacoustic spectroscopy technology monitors the gas concentration in real time through the photoacoustic effect caused by laser irradiation of target gas molecules. This technology is not only applicable to environmental monitoring but also can be used in complex industrial emission control. Especially in high-temperature and high-pressure coal-fired flue gas, accurate monitoring of the concentration of nitrogen monoxide is crucial. Although photoacoustic spectroscopy technology has significant advantages in gas detection, in practical applications, the composition of coal-fired flue gas is complex, and changes in environmental factors (such as temperature, humidity, air flow, etc.) will interfere with the photoacoustic spectroscopy signal. Therefore, how to accurately extract the effective information of the nitrogen monoxide concentration from these photoacoustic signals has become a key challenge in the application of this technology.
[0004] Currently, the existing photoacoustic spectroscopy data processing methods mainly focus on signal denoising and feature extraction, but mostly use one-dimensional signal processing methods, such as analysis based on spectral line features or simple processing of time-domain signals. These methods often have difficulty in fully excavating the rich information contained in photoacoustic signals, resulting in low prediction accuracy and robustness.
[0005] To overcome this problem, the present invention proposes a method for predicting the concentration of nitric oxide in coal-fired flue gas based on dual-modal feature fusion. This method combines the original one-dimensional photoacoustic spectroscopy data with the two-dimensional photoacoustic spectroscopy images obtained after Gramian angular field transformation, and uses these two different modal data for deep fusion. The Gramian angular field can effectively extract the time-frequency features in the photoacoustic signal, thereby generating two-dimensional images that contain more signal details and potential laws. By fusing the original one-dimensional data with the two-dimensional images, the present invention can make full use of the multi-faceted features of the photoacoustic spectroscopy signal and improve the accuracy and stability of predicting the concentration of nitric oxide in coal-fired flue gas. Summary of the Invention
[0006] The present invention provides a method for predicting the concentration of nitric oxide in coal-fired flue gas based on dual-modal feature fusion, which can improve the prediction accuracy of the nitric oxide concentration.
[0007] The present invention provides a method for predicting the concentration of nitric oxide in coal-fired flue gas based on dual-modal feature fusion, comprising the following steps:
[0008] S1. Acquisition of one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas;
[0009] S2. Denoising the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas and converting the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas into two-dimensional images;
[0010] S3. Dataset division of the photoacoustic spectroscopy data samples of nitric oxide in coal-fired flue gas and the photoacoustic spectroscopy data samples that have been converted into two-dimensional images;
[0011] S4. Establishing a prediction model for the concentration of nitric oxide in coal-fired flue gas with dual-modal feature fusion;
[0012] S5. Testing the model effect.
[0013] Further, in S1, the photoacoustic spectroscopy technology is used to collect the spectroscopy data of multiple groups of experimental samples, obtain the photoacoustic spectroscopy characteristic information of each sample, and obtain the photoacoustic spectroscopy data of different concentrations of nitric oxide.
[0014] Further, in S2, the principal component analysis method is used to denoise the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas, specifically as follows:
[0015] S211. A sample matrix x of size n*p can be formed:
[0016]
[0017] S212. First, perform standardization processing on it:
[0018] Calculate the mean value column by column and the standard deviation The standardized data is calculated The original sample matrix becomes after standardization:
[0019]
[0020] S213. Calculate the covariance matrix of the standardized samples:
[0021]
[0022] Where
[0023] S214. Calculate the eigenvalues and eigenvectors of R:
[0024] Eigenvalues:
[0025] λ1≥λ2≥…≥λ p ≥0
[0026] Eigenvectors:
[0027]
[0028] S215. Calculate the contribution degree and cumulative contribution degree of the principal components:
[0029] Contribution degree:
[0030]
[0031] Cumulative contribution degree:
[0032]
[0033] S216. Write out the principal components
[0034] Generally, the first, second, …, m (m ≤ p) principal components corresponding to the eigenvalues with a cumulative contribution rate exceeding 80% are taken. The i-th principal component:
[0035] F i =a 1i X1 + a 2i X2 + … + a pi X p (i = 1, 2, …, m)
[0036] Furthermore, in S2, the one-dimensional photoacoustic spectroscopy data of nitrogen oxides in coal-fired flue gas is converted into a two-dimensional image using the Gramian angular field, specifically as follows:
[0037] S221. Normalize the denoised photoacoustic spectroscopy data:
[0038]
[0039] S222. Polar coordinate transformation:
[0040]
[0041] S223. Calculate the Gram matrix. For each pair of data points (x n , x p ), calculate GASF and GADF:
[0042] GASF = cos(φ i + φ j )
[0043] GADF = sin(φ i - φ j )
[0044] Furthermore, in S3, use a random seed to divide the dataset into a training set, a validation set, and a test set, ensuring that the batches of the two dataset divisions are the same.
[0045] Furthermore, in S4, define a prediction model, a dual-modal feature fusion prediction model for the nitric oxide concentration in coal-fired flue gas. This model is a dual-branch model. Its first branch mainly includes an MSA module and a Flatten layer. The second branch mainly includes a one-dimensional convolutional unit. At the end of the two branches, there is a feature fusion module, which consists of an AFF module and a fully connected layer, and a prediction model for the nitric oxide concentration in coal-fired flue gas based on dual-modal feature fusion is established. The schematic diagram of the model structure is as Figure 2 shown and is specifically described as follows:
[0046] S41. The one-dimensional convolutional unit is composed of three convolutional layers, three pooling layers, and a concatenation layer in Pytorch. Its structure is as Figure 3 shown and is specifically described as follows:
[0047] S411. The three convolutional layers are one-dimensional convolutional layers that use 16 convolutional kernels of sizes 1*3, 16 convolutional kernels of sizes 1*7, and
[0048] 16 convolutional kernels of sizes 1*13 respectively. The activation function uses the Relu activation function;
[0049] S412. The three pooling layers are one-dimensional pooling layers, all using max pooling with a size of 3;
[0050] S413. The concatenation layer concatenates the outputs of the three pooling layers into a one-dimensional output;
[0051] S42. The MSA module extracts the features of the two-dimensional image. First, the MSA module performs a linear transformation on the input to generate three components: query, key, and value. It calculates the dot product of the query and the key to obtain the attention scores, which determine how much attention should be given to each part of the input sequence when generating the output. Using these scores, a weighted sum of the values is performed to obtain the final output representation. The Flatten layer flattens the extracted two-dimensional features;
[0052] S43. The AFF module fuses the features from different sources through the attention mechanism and performs weighted fusion on the features extracted from the two branches.
[0053] Furthermore, in S4, the training set and the validation set processed in step S3 are used for training and validation to obtain the test model.
[0054] Furthermore, in S5, the test set is used to test the prediction effect of the model.
[0055] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0056] By fusing one-dimensional photoacoustic spectroscopy data and two-dimensional Gram angular field images, the present invention makes full use of the multi-faceted features of photoacoustic signals, significantly improving the prediction accuracy and stability of the nitric oxide concentration in coal-fired flue gas. At the same time, deep learning is used to improve the prediction accuracy and optimize the calculation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0057] Figure 1 It is a flowchart of a method for predicting the nitric oxide concentration in coal-fired flue gas based on dual-modal feature fusion provided by the present invention.
[0058] Figure 2 It is a structural schematic diagram of a network for feature extraction and feature fusion provided by the present invention.
[0059] Figure 3 It is a structural schematic diagram of a convolutional unit provided by the present invention. DETAILED DESCRIPTION OF THE INVENTION
[0060] The following Figures 1-3 , in conjunction with the attached drawings, describes in detail a specific embodiment of the present invention, but it should be understood that the protection scope of the present invention is not limited by the specific embodiment.
[0061] The experimental operating environment is windows11, python2024, the development framework is Pytorch, and the Nvidia Gtx3050 GPU is used for accelerated training during the training process.
[0062] As Figure 1 shown, the present invention provides a method for predicting the nitric oxide concentration in coal-fired flue gas with dual-modal feature fusion, including the following steps:
[0063] S1. Acquisition of one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas;
[0064] S2. Denoise the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas and convert the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas into a two-dimensional image;
[0065] S3. Divide the datasets of the photoacoustic spectroscopy data samples of nitric oxide in coal-fired flue gas and the photoacoustic spectroscopy data samples that have been converted into two-dimensional images;
[0066] S4. Establish a prediction model for the concentration of nitric oxide in coal-fired flue gas with dual-modal feature fusion;
[0067] S5. Test the model effect.
[0068] Preferably, in step S1, the photoacoustic spectroscopy technology is used to collect spectral data of multiple experimental samples, obtain the photoacoustic spectroscopy characteristic information of each sample, and obtain the photoacoustic spectroscopy data of different concentrations of nitric oxide.
[0069] Preferably, in step S2, the principal component analysis method is used to denoise the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas, and the denoised one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas is converted into a two-dimensional image through the Gramian angular field.
[0070] Preferably, in step S3, a random seed is used for dataset division, which is divided into a training set, a validation set and a test set to ensure that the batches of the two dataset divisions are the same.
[0071] Preferably, in step S4, a prediction model, a prediction model for the concentration of nitric oxide in coal-fired flue gas with dual-modal feature fusion, mainly includes an MSA module, a Flatten layer, a one-dimensional convolutional unit and a feature fusion module, and a prediction model for the concentration of nitric oxide in coal-fired flue gas based on dual-modal feature fusion is established.
[0072] Preferably, in step S4, the model is trained to obtain a test model.
[0073] Preferably, in step S5, the trained model is tested through the test set.
[0074] Example 1:
[0075] In step S1, the present invention uses the quantum cascade laser developed by Chengdu Multi-Spectrum Optoelectronics Technology Co., Ltd. as the core light source of the photoacoustic spectroscopy experiment module, and collects spectral data of multiple coal-fired flue gas samples in the wavelength range of 600 - 875 nm, and successfully obtains the high-resolution photoacoustic spectroscopy data of each sample.
[0076] Example 2:
[0077] In step S2, n samples and photoacoustic spectroscopy data of p characteristic wavelengths of nitric oxide obtained in step S1 are used, and then the principal component analysis method is used to denoise the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas. The specific method is as follows:
[0078] S211. Construct a sample matrix x of size n*p from the photoacoustic spectroscopy data of n samples and p characteristic wavelengths of nitric oxide:
[0079]
[0080] S212. Perform standardization processing on the sample matrix:
[0081] Calculate the mean value column by column and the standard deviation Calculate the standardized data The original sample matrix becomes after standardization:
[0082]
[0083] S213. Calculate the covariance matrix of the standardized samples:
[0084]
[0085] where
[0086] S214. Calculate the eigenvalues and eigenvectors of R:
[0087] Eigenvalues:
[0088] λ1≥λ2≥…≥λ p ≥0
[0089] Eigenvectors:
[0090]
[0091] S215. Calculate the contribution degree of the principal components and the cumulative contribution degree:
[0092] Contribution degree:
[0093]
[0094] Cumulative contribution degree:
[0095]
[0096] S216. Write out the principal components
[0097] Generally, the first, second, …, the mth (m≤p) principal components corresponding to the eigenvalues with a cumulative contribution rate exceeding 80% are taken. The ith principal component:
[0098] F i = a 1i X1 + a 2i X2 + … + a pi X p (i = 1, 2, …, m)
[0099] Complete the denoising of the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas.
[0100] Example 3:
[0101] In step 2, use the Gram angular field to convert the denoised one-dimensional photoacoustic spectroscopy data of nitric oxide into a two-dimensional image. The specific method is as follows:
[0102] S221. Normalize the denoised one-dimensional photoacoustic spectroscopy data of nitric oxide:
[0103]
[0104] S222. Polar coordinate transformation:
[0105]
[0106] S223. Calculate the Gram matrix. For each pair of data points (x n , x p ), calculate GASF and GADF:
[0107] GASF = cos(φ i + φ j )
[0108] GADF = sin(φ i - φ j )
[0109] Complete the conversion of the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas into a two-dimensional image.
[0110] Example 4:
[0111] In step 3, use a random seed to divide the dataset into a training set, a validation set, and a test set. The random seed seed is set to 49, with 70% for the training set, 20% for the validation set, and 10% for the test set.
[0112] Example 5:
[0113] In step 4, a prediction model for the concentration of nitrogen monoxide in coal-fired flue gas with bimodal feature fusion is defined. This model is a two-branch model. Its first branch mainly includes an MSA module and a Flatten layer. The second branch mainly includes a one-dimensional convolutional unit. At the end of the two branches, there is a feature fusion module, which consists of an AFF module and a fully connected layer. A prediction model for the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion is established. The specific method is as follows:
[0114] S41. The one-dimensional convolutional unit is composed of three convolutional layers, three pooling layers and a concatenation layer in Pytorch. Its structure is as Figure 3 shown, and the specific description is as follows:
[0115] S411. The three convolutional layers are one-dimensional convolutional layers, which use 16 convolutional kernels of size 1*3, 16 convolutional kernels of size 1*7 and
[0116] 16 convolutional kernels of size 1*13 respectively. The activation function uses the Relu activation function;
[0117] S412. The three pooling layers are one-dimensional pooling layers, and all use max pooling with a size of 3;
[0118] S413. The concatenation layer concatenates the outputs of the three pooling layers into a one-dimensional output;
[0119] S42. The MSA module extracts the features of the two-dimensional image. The MSA module first performs a linear transformation on the input to generate three components: query, key and value. It calculates the dot product of the query and the key to obtain the attention scores, which determine how much attention should be given to each part of the input sequence when generating the output. These scores are used to weight and sum the values to obtain the final output representation. The Flatten layer flattens the extracted two-dimensional features;
[0120] S43. The AFF module dynamically fuses features from different sources through the attention mechanism and performs weighted fusion on the features extracted from the two branches.
[0121] The prediction model for the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion in the present invention solves the deficiencies existing in the existing prediction methods for the concentration of nitrogen monoxide in coal-fired flue gas; a new bimodal model is proposed, which can make feature extraction more effective and comprehensive, improve the prediction accuracy and optimize the calculation efficiency, and is very suitable for the prediction of the concentration of nitrogen monoxide in coal-fired flue gas.
[0122] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above-described exemplary embodiments, and the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, in any aspect, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be embraced within the present invention. Any reference signs in the claims should not be construed as limiting the claims involved.
[0123] In addition, it should be understood that although this specification is described according to embodiments, not every embodiment only contains an independent technical solution. This narrative manner of the specification is only for clarity. Those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.
Claims
1. A method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion, characterized in that, It includes the following steps: S1. Acquisition of one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas; S2. Denoising the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas and converting the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas into a two-dimensional image; S3. Dataset division of the photoacoustic spectroscopy data samples of nitric oxide in coal-fired flue gas and the photoacoustic spectroscopy data samples that have been converted into two-dimensional images; S4. Establish a prediction model for the concentration of nitric oxide in coal-fired flue gas with dual-modal feature fusion; S5. Test the model effect.
2. The method for predicting the nitric oxide concentration in coal-fired flue gas based on bimodal feature fusion according to claim 1, wherein In the step S1, the photoacoustic spectroscopy technology is used to collect spectral data of multiple groups of experimental samples, obtain the photoacoustic spectroscopy characteristic information of each sample, and obtain the photoacoustic spectroscopy data of different concentrations of nitric oxide.
3. A method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion according to claim 1, characterized in that, In the step S2, the principal component analysis method is used to denoise the one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas, and the denoised one-dimensional photoacoustic spectroscopy data of nitric oxide in coal-fired flue gas is converted into a two-dimensional image through the Gram angular field.
4. The method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion according to claim 1, wherein In the step S3, the random seed is used for dataset division, which is divided into a training set, a validation set and a test set. Among them, the random seed seed is set to 49, the training set is 70%, the validation set is 20%, and the test set is 10%.
5. The method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion according to claim 1, wherein In the step S4, a prediction model for the concentration of nitric oxide in coal-fired flue gas with dual-modal feature fusion is defined. This model is a dual-branch model. Its first branch mainly includes an MSA module and a Flatten layer. The second branch mainly includes a one-dimensional convolutional unit. There is a feature fusion module at the end of the two branches, which is composed of an AFF module and a fully connected layer. A prediction model for the concentration of nitric oxide in coal-fired flue gas based on dual-modal feature fusion is established. The specific method is as follows: S41. The one-dimensional convolutional unit is composed of three convolutional layers, three pooling layers and a concatenation layer in Pytorch. Its structure is specifically described as follows: S411. The three convolutional layers are one-dimensional convolutional layers, which respectively use 16 convolutional kernels of size 1*3, 16 convolutional kernels of size 1*7 and 16 convolutional kernels of size 1*13, and the activation function uses the Relu activation function; S412. The three pooling layers are one-dimensional pooling layers, and all use max pooling with a size of 3; S413. The concatenation layer concatenates the outputs of the three pooling layers into a one-dimensional output; S42. The MSA module extracts the features of the two-dimensional image. The MSA module first performs a linear transformation on the input to generate three components: query, key and value. It calculates the dot product of the query and the key to obtain the attention scores. These scores determine how much attention should be given to each part of the input sequence when generating the output. These scores are used to weighted sum the values to obtain the final output representation. The Flatten layer flattens the extracted two-dimensional features; S43. The AFF module fuses the features from different sources through the attention mechanism and performs weighted fusion on the features extracted by the two branches.
6. The method for predicting the concentration of nitrogen monoxide in coal-fired flue gas based on bimodal feature fusion according to claim 1, wherein In the step S5, the test set is used to test the prediction effect of the model.