Image type combustion NOx emission detection method, medium and equipment

Abstract

The invention provides an image type combustion NOx emission detection method, medium and equipment, and the method comprises the steps: collecting a flame image of a hearth combustor, and carrying out the preprocessing through a bilateral filtering noise reduction algorithm; establishing an empirical formula, and extracting statistical characteristics of the flame image; building convolution self-coding, and extracting deep features of the flame image; carrying out dimensionality reduction on the flame image by adopting a principal component analysis method to obtain principal component features; splicing the statistical features, the deep features and the principal component features into comprehensive features; establishing full-connection self-coding, and extracting advanced features of the comprehensive features; inputting the extracted advanced features into a long short-term memory network, and extracting time features; and a self-attention mechanism is introduced to process time characteristics, and finally the output combustion NOx concentration is predicted. According to the method, the combustion state of the flame can be described more accurately, the capturing capacity of the data change trend is improved, the generalization capacity and the adaptive capacity of the model are improved, and accurate monitoring and prediction of combustion NOx emission are achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of combustion emission monitoring, and in particular relates to an image-based combustion NOx emission detection method, medium and equipment. Background Art

[0002] Nitrogen oxide (NOx) is a common pollutant in the combustion process, which seriously threatens the ecological environment and human health. Timely and accurate information on combustion NOx emissions provides an important data basis for the formulation of combustion optimization strategies and the implementation of flue gas treatment plans. Traditional NOx emission monitoring mainly relies on gas analyzers, but due to factors such as high maintenance costs and frequent offline calibration, analytical instruments are often difficult to achieve economical and continuous measurements. In contrast, soft measurement methods are more popular because of their convenient maintenance and low operating costs. Practice has proved that data-driven models can predict NOx emission concentrations using only combustion process data without the need to know the combustion mechanism knowledge, and have the advantages of high computational efficiency and strong generalization performance. However, the generation of combustion NOx is affected by many factors, so how to accurately establish a nonlinear mapping between flame images and combustion NOx emission concentrations faces great challenges. Therefore, there is an urgent need for an efficient NOx emission detection method to improve the existing detection scheme in order to achieve high-precision monitoring of combustion NOx emissions. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention provides an image-based combustion NOx emission detection method, medium and equipment.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] An image-based combustion NOx emission detection method, characterized in that it comprises the following steps:

[0006] Collect the furnace burner flame image and pre-process it using bilateral filtering noise reduction algorithm;

[0007] Establish empirical formulas to extract statistical features of flame images;

[0008] Convolutional autoencoder CAE is established to extract deep features of flame images;

[0009] The principal component analysis method is used to reduce the dimension of the flame image and obtain the principal component features;

[0010] Combine statistical features, deep features and principal component features to form comprehensive features;

[0011] Establish a fully connected autoencoder FAE to extract high-level features of comprehensive features;

[0012] The extracted high-level features are input into the long short-term memory network LSTM to extract the time features;

[0013] A self-attention mechanism is introduced to process temporal features and finally predict the output combustion NOx concentration.

[0014] To optimize the above technical solutions, the specific measures taken also include:

[0015] Furthermore, the bilateral filtering denoising algorithm is expressed as:

[0016]

[0017] Where I′(α) represents the pixel value of the filtered flame image at position α, I(β) represents the pixel value of the original flame image at position , S represents a local neighborhood centered on α, W(α) represents the normalization coefficient, and σ s represents the spatial Gaussian kernel parameter, the initial value is a constant, and is passed through Dynamic adjustment, where width×height represents the width and height of the image, 1024×768 represents the base resolution, and σ r (β)=σ r ×(1+δ×var(β)), where σ r represents the intensity Gaussian kernel parameter, var(β) represents the local variance of position β, and δ represents the adjustment coefficient.

[0018] Furthermore, the statistical characteristics include grayscale mean, flame brightness, flame effective area, grayscale standard deviation, flame skewness and flame kurtosis.

[0019] Furthermore, the convolutional autoencoder CAE is composed of an encoder and a decoder, and the specific implementation process is as follows:

[0020] The flame image x is processed by the encoder e1, specifically including: firstly, feature extraction is performed by Q1 convolution filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the convolution feature h1 with the output dimension of a1×b1×c1 is output;

[0021] The convolution feature h1 is processed by encoders e2, e3, and e4 in turn. The data processing process of encoders e2, e3, and e4 is the same as that of encoder e1, and the final output is a deep feature h with a dimension of a4×b4×c4. deep ;

[0022] Deep feature h deepAfter being processed by the decoder d1, specifically including: firstly, the data dimension is increased by the upsampling layer U1 with a window size of g×g, then the feature is extracted by Q5 convolution filters C5 with a window size of c×c and a step size of q, and finally the feature is activated by the ReLU activation function, and the convolution feature h4 with the output dimension of a3×b3×c3 is output;

[0023] The convolution feature h4 is processed by encoders d2, d3, and d4 in sequence. The data processing process of encoders d2, d3, and d4 is the same as that of decoder d1, and the final output is a reconstructed image z with a dimension of O×P×3.

[0024] Furthermore, the convolutional autoencoder CAE adopts the following loss function during training:

[0025] L CAE =L MSE +L PSNR +L edge ;

[0026]

[0027]

[0028]

[0029] Where, L CAE represents the loss function, L MSE represents the mean square error, L PSNR represents the peak signal-to-noise ratio, L edge represents the edge preservation loss, M represents the number of training images, and x i and z i Respectively represent the i-th input flame image and the reconstructed image, ξ represents the maximum grayscale level of the image, represents the gradient operator, Indicates the image index position, Indicates at location The pixel value of ||·|| 2 Represents the square of the Euclidean norm.

[0030] Furthermore, the fully connected autoencoder FAE consists of an input layer, a hidden layer and an output layer, and the specific implementation process is as follows:

[0031] m-dimensional comprehensive feature h com Mapped to n-dimensional high-level features h by the fully connected encoder adv , the mapping process is represented by h adc =S(Wh com +b), where m>n, W represents the weight matrix, b represents the bias parameter, and S represents the Sigmoid function;

[0032] n-dimensional high-level features h adv Reconstructed into m-dimensional reconstructed feature r by the fully connected decoder com , the reconstruction process is represented by r com =S(W′h adc +b′), where W′ represents the weight matrix connecting the hidden layer and the output layer, and b′ represents the bias parameter.

[0033] Furthermore, the long short-term memory network LSTM is composed of an input gate i, a forget gate f and an output gate o, and the specific calculation process is as follows:

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040] In the formula, Indicates that the high-level features of the moment, Indicates the time characteristics of the moment, represents the time characteristics at time t-1, i t 、f t and t Represent the activation vectors of the input gate, forget gate, and output gate respectively, tanh represents the hyperbolic sine function, represents the candidate storage state, c t Indicates the internal state; U i , U f , U o , U C and W i , W f , W o , W c Both represent weight matrices, b i , b f , b o , b c represents the bias, and ⊙ represents the dot product of the matrix.

[0041] Furthermore, the self-attention mechanism is introduced to process the time feature, and finally the combustion NOx concentration is predicted and outputted. The specific processing process is as follows:

[0042]

[0043]

[0044]

[0045] In the formula, the attention score function e t represents the correlation between the output vector of the long short-term memory network LSTM and the combustion NOx concentration at time t, u and ω both represent weight coefficients, and ε represents the bias term; α t represents the probability distribution of the attention mechanism output to the hidden layer of the long short-term memory network LSTM; s t Represents the output of the attention layer at time t, which is the final predicted combustion NOx concentration.

[0046] Accordingly, the present invention provides a computer-readable storage medium storing a computer program, characterized in that the computer program enables a computer to execute the above-mentioned image-based combustion NOx emission detection method.

[0047] Accordingly, the present invention provides an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the image-based combustion NOx emission detection method as described above is implemented.

[0048] The beneficial effects of the present invention are as follows: the convolutional autoencoder CAE and the fully connected autoencoder FAE established by the present invention can extract deep features and high-level features of the flame image, thereby more accurately describing the combustion state of the flame; the long short-term memory network LSTM and the self-attention mechanism introduced by the present invention significantly enhance the processing capability of time series data and greatly improve the ability to capture data change trends; the present invention realizes accurate monitoring and prediction of combustion NOx emissions by integrating a variety of improved image analysis and machine learning technologies; at the same time, the present invention also effectively improves the generalization ability of the model and the ability to adapt to complex changing environments by combining static features and dynamic features. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the structure of the convolutional autoencoder CAE of the present invention.

[0050] Figure 2 It is a schematic diagram of the structure of the fully connected autoencoder FAE of the present invention.

[0051] Figure 3 It is a schematic diagram of the structure of the long short-term memory network LSTM of the present invention.

[0052] Figure 4 It is a structural schematic diagram of the self-attention mechanism of the present invention.

[0053] Figure 5 These are the furnace flame images under eight combustion conditions of the present invention.

[0054] Figure 6 This is the combustion NOx concentration prediction result of the present invention. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0056] In one embodiment, the present invention proposes an image-based combustion NOx emission detection method, which specifically includes the following steps.

[0057] S1: Collect the furnace burner flame image and pre-process it using the bilateral filtering noise reduction algorithm.

[0058] In this embodiment, the bilateral filtering algorithm is expressed as:

[0059]

[0060] Where I′(α) represents the pixel value at position α (such as (o, p), o represents the row where the pixel is located, and p represents the column where the pixel is located) after filtering, I(β) represents the pixel value at position β of the original image, S represents a local neighborhood centered on α, and W(α) represents the normalization coefficient. s represents the spatial Gaussian kernel parameter, the initial value is a constant, and is passed through Dynamic adjustment, where width×height represents the width and height of the image, and 1024×768 represents the base resolution; σ r (β) represents the intensity Gaussian kernel parameter based on local variance adjustment, defined as σ r (β)=σ r ×(1+δ×var(β)), where σ r represents the original intensity Gaussian kernel parameter (preset to a fixed constant), var(β) represents the local variance of position β, and δ represents the adjustment coefficient (usually set to 0.1).

[0061] In this embodiment, 8000 furnace flame images are collected, which are from eight combustion conditions, such as Figure 5 As shown, the bilateral filtering denoising algorithm is used for preprocessing.

[0062] S2: Establish six empirical formulas to extract the statistical features of flame images h static .

[0063] In this embodiment, the six statistical features are grayscale mean, flame brightness, flame effective area, grayscale standard deviation, flame skewness and flame kurtosis, and the calculation formula is shown in Table 1.

[0064] Table 1 Empirical formulas for six statistical characteristics

[0065]

[0066] S3: Establish convolutional autoencoder CAE to extract deep features h of flame image deep .

[0067] In this embodiment, the structure of the convolutional autoencoder CAE is as follows: Figure 1 As shown, it consists of an encoder and a decoder, and the specific implementation process is:

[0068] The flame image x is processed by the encoder e1, specifically including: firstly, feature extraction is performed by Q1 convolution filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the convolution feature h1 with the output dimension of a1×b1×c1 is output;

[0069] The convolution feature h1 is processed by encoders e2, e3, and e4 in sequence. The data processing process is similar to that of encoder e1, and the final output is a deep feature h with a dimension of a4×b4×c4. deep ;

[0070] Deep feature h deep After being processed by the decoder d1, specifically including: firstly, the data dimension is increased by the upsampling layer U1 with a window size of g×g, then the feature is extracted by Q5 convolution filters C5 with a window size of c×c and a step size of q, and finally the feature is activated by the ReLU activation function, and the convolution feature h4 with the output dimension of a3×b3×c3 is output;

[0071] The convolution feature h4 is processed by encoders d2, d3, and d4 in sequence. The data processing process is similar to that of decoder d1, and the final output is a reconstructed image z with a dimension of O×P×3.

[0072] In this embodiment, the loss function is used in the convolutional autoencoder CAE training process:

[0073] L CAE =L MSE +L PSNR +L edge ;

[0074] The mean square error Peak signal-to-noise ratio Edge retention loss Where M represents the number of training images, x i and z i denote the i-th input image and the reconstructed image respectively, ξ denotes the maximum grayscale level of the image (usually set to 255), represents the gradient operator, Indicates the image index position, Indicates at location The pixel value of ||·|| 2 Represents the square of the Euclidean norm.

[0075] S4: Use the principal component analysis method to reduce the dimension of the flame image and obtain the principal component feature h pca .

[0076] S5: The statistical feature h static , deep features h deep , principal component feature h pca Splice to form comprehensive features h com .

[0077] S6: Establish a fully connected autoencoder FAE and extract comprehensive features h com The high-level features h adv .

[0078] In this embodiment, the structure of the fully connected autoencoder FAE is as follows: Figure 2 As shown, it consists of an input layer, a hidden layer, and an output layer. The specific implementation process is:

[0079] m-dimensional comprehensive feature h com Mapped to n-dimensional high-level features h by the fully connected encoder adv , the mapping process is represented by h adc =S(Wh com +b), where m>n, W represents the weight matrix, b represents the bias parameter, and S represents the Sigmoid function;

[0080] n-dimensional high-level features h adv Reconstructed into m-dimensional reconstructed feature r by the fully connected decoder com , the reconstruction process is represented by r com =S(W′h adc +b′), where W′ represents the weight matrix connecting the hidden layer and the output layer, and b′ represents the bias parameter.

[0081] S7: Extract the high-level features h adv Input to the long short-term memory network LSTM to extract the time feature h time .

[0082] In this embodiment, the structure of the long short-term memory network LSTM is as follows Figure 3 As shown, it consists of input gate i, forget gate f and output gate o. The specific calculation process is as follows:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089] In the formula, represents the high-level features at time t, represents the time characteristics at time t, represents the time characteristics at time t-1, i t 、f t and t Represent the activation vectors of the input gate, forget gate, and output gate respectively, tanh represents the hyperbolic sine function, represents the candidate storage state, c t represents the internal state; U and W both represent weight matrices, b represents the bias, and ⊙ represents the dot product of the matrices.

[0090] S8: Introducing self-attention mechanism to process temporal features h time , and finally predict the output combustion NOx concentration.

[0091] In step S8, the self-attention mechanism structure is as follows Figure 4 As shown, it is used to predict the combustion NOx concentration. The specific calculation process is as follows:

[0092]

[0093]

[0094]

[0095] In the formula, the attention score function e t represents the correlation between the output vector of the long short-term memory network LSTM and the combustion NOx concentration at time t, u and ω both represent weight coefficients, and ε represents the bias term; α t represents the probability distribution of the attention mechanism output to the hidden layer of the LSTM network; s t represents the output of the attention layer at time t, that is, the final predicted combustion NOx concentration.

[0096] S9: Use the test set to test the NOx concentration prediction accuracy, and use the mean absolute error MAE, root mean square error RMSE and determination coefficient R 2 Evaluate the effectiveness of your practice.

[0097] In this embodiment, in represents the number of test samples, Indicates the actual value, Represents the predicted value.

[0098] In this embodiment, the combustion NOx concentration prediction result is as follows: Figure 6 As shown in the figure, it can be seen that the prediction results of the image-based combustion NOx emission detection method proposed in this embodiment are completely consistent with the actual observation values, thereby achieving accurate monitoring and prediction of combustion NOx emissions.

[0099] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the image-based combustion NOx emission detection method as described in the first embodiment.

[0100] In another embodiment, the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the image-based combustion NOx emission detection method as described in the first embodiment is implemented.

[0101] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0102] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0103] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. An image-based combustion NOx emission detection method, characterized in that: The steps include: Collect the furnace burner flame image and pre-process it using bilateral filtering noise reduction algorithm; Establish empirical formulas to extract statistical features of flame images; Convolutional autoencoder CAE is established to extract deep features of flame images; The principal component analysis method is used to reduce the dimension of the flame image and obtain the principal component features; Combine statistical features, deep features and principal component features to form comprehensive features; Establish a fully connected autoencoder FAE to extract high-level features of comprehensive features; The extracted high-level features are input into the long short-term memory network LSTM to extract the temporal features; A self-attention mechanism is introduced to process temporal features and finally predict the output combustion NOx concentration.

2. The image-based combustion NOx emission detection method according to claim 1, characterized in that: The bilateral filtering denoising algorithm is expressed as: Where I′(α) represents the pixel value of the filtered flame image at position α, I(β) represents the pixel value of the original flame image at position β, S represents a local neighborhood centered on α, W(α) represents the normalization coefficient, and σ s represents the spatial Gaussian kernel parameter, the initial value is a constant, and is passed through Dynamic adjustment, where width×height represents the width and height of the image, 1024×768 represents the base resolution, and σ r (β)=σ r ×(1+δ×var(β)), where σ r represents the intensity Gaussian kernel parameter, var(β) represents the local variance of position β, and δ represents the adjustment coefficient.

3. The image-based combustion NOx emission detection method according to claim 1, characterized in that: The statistical characteristics include grayscale mean, flame brightness, flame effective area, grayscale standard deviation, flame skewness and flame kurtosis.

4. The image-based combustion NOx emission detection method according to claim 1, characterized in that: The convolutional autoencoder CAE consists of an encoder and a decoder, and the specific implementation process is as follows: The flame image x is processed by the encoder e1, specifically including: firstly, feature extraction is performed by Q1 convolution filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the convolution feature h1 with the output dimension of a1×b1×c1 is output; The convolution feature h1 is processed by encoders e2, e3, and e4 in turn. The data processing process of encoders e2, e3, and e4 is the same as that of encoder e1, and the final output is a deep feature h with a dimension of a4×b4×c4. deep ; Deep feature h deep After being processed by the decoder d1, specifically including: firstly, the data dimension is increased by the upsampling layer U1 with a window size of g×g, then the feature is extracted by the convolution filter C5 with Q5 c×c window size and q step size, and finally the feature is activated by the ReLU activation function, and the convolution feature h4 with the output dimension of a3×b3×c3 is output; The convolution feature h4 is processed by encoders d2, d3, and d4 in sequence. The data processing process of encoders d2, d3, and d4 is the same as that of decoder d1, and the final output is a reconstructed image z with a dimension of O×P×3.

5. The image-based combustion NOx emission detection method according to claim 4, characterized in that: The convolutional autoencoder CAE uses the following loss function during training: L CAE =L MSE +L PSNR +L edge ; Where, L CAE represents the loss function, L MSE represents the mean square error, L PSNR represents the peak signal-to-noise ratio, L edge represents the edge preservation loss, M represents the number of training images, and x i and z i Respectively represent the i-th input flame image and the reconstructed image, ξ represents the maximum grayscale level of the image, represents the gradient operator, Indicates the image index position, Indicates at location The pixel value of ||·|| 2 Represents the square of the Euclidean norm.

6. The image-based combustion NOx emission detection method according to claim 1, characterized in that: The fully connected autoencoder FAE consists of an input layer, a hidden layer and an output layer. The specific implementation process is as follows: m-dimensional comprehensive feature h com Mapped to n-dimensional high-level features h by the fully connected encoder adv , the mapping process is represented by h adc =S(Wh com +b), where m>n, W represents the weight matrix, b represents the bias parameter, and S represents the Sigmoid function; n-dimensional high-level features h adv Reconstructed into m-dimensional reconstructed feature r by the fully connected decoder com , the reconstruction process is represented by r com =S(W′h adc +b′), where W′ represents the weight matrix connecting the hidden layer and the output layer, and b′ represents the bias parameter.

7. The image-based combustion NOx emission detection method according to claim 1, characterized in that: The long short-term memory network LSTM is composed of an input gate i, a forget gate f and an output gate o. The specific calculation process is as follows: In the formula, Indicates that the high-level features of the moment, Indicates the time characteristics of the moment, represents the time characteristics at time t-1, i t 、f t and t Represent the activation vectors of the input gate, forget gate, and output gate respectively, tanh represents the hyperbolic sine function, represents the candidate storage state, c t Indicates the internal state; U i , U f , U o , U C and W i , W f , W o , W c Both represent weight matrices, b i , b f , b o , b c represents the bias, and ⊙ represents the dot product of the matrix.

8. The image-based combustion NOx emission detection method according to claim 7, characterized in that: The self-attention mechanism is introduced to process the time feature, and finally the combustion NOx concentration is predicted and output. The specific processing process is as follows: In the formula, the attention score function e t represents the correlation between the output vector of the long short-term memory network LSTM and the combustion NOx concentration at time t, u and ω both represent weight coefficients, and ε represents the bias term; α t represents the probability distribution of the attention mechanism output to the hidden layer of the long short-term memory network LSTM; s t Represents the output of the attention layer at time t, which is the final predicted combustion NOx concentration.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the image-based combustion NOx emission detection method as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the image-based combustion NOx emission detection method as described in any one of claims 1 to 8 is implemented.

Images