Hearth combustion stability quantitative evaluation method based on flame image and stack convolution self-coding

Abstract

The invention provides a hearth combustion stability quantitative evaluation method based on flame images and stack convolution self-coding, and relates to the field of industrial hearth operation monitoring and intelligent control. According to the method, a hearth flame image is collected through camera equipment, after image preprocessing, a deep learning model fusing a convolution block attention module and a stacked convolution auto-encoder is adopted to carry out extraction and dimensionality reduction on flame image features, and a combustion state index and a combustion performance index are generated. And calculating the variance of the combustion state index by using a sliding time window, defining the variance as a combustion fluctuation degree, and reflecting the dynamic change characteristic of the combustion state. Combustion stability comprehensive evaluation is constructed by combining the combustion state index, the combustion performance index and the combustion fluctuation degree, the combustion state is quantitatively evaluated in real time, and early warning is provided when the combustion state is abnormal. Compared with a traditional flame detection system, the system can remarkably improve the accuracy and real-time performance of combustion state monitoring, and has the capability of early warning of abnormal combustion in advance.

Description

Technical Field

[0001] The present invention relates to the field of industrial furnace operation monitoring and intelligent control, and in particular to a method for quantitatively evaluating furnace combustion stability based on flame images and stacked convolution autoencoders. Background Art

[0002] In industrial boilers, thermal power plants and other combustion equipment, combustion stability is a key factor in ensuring efficient operation, safety and reducing pollutant emissions. Instability in the combustion process can lead to increased emissions, equipment damage, and even serious safety accidents. Therefore, accurate and real-time evaluation of the combustion state is of great significance for improving equipment operation efficiency and reducing environmental pollution. At present, the monitoring and evaluation technologies for combustion stability mainly include the following methods: (1) Monitoring based on flame detection signals. The flame detection system usually provides a simple binary signal (such as "flame exists" or "flame is extinguished") by monitoring whether the flame exists. Although this method can quickly determine the presence or absence of combustion, it cannot capture the dynamic changes of the combustion process, lacks sensitivity to the dynamic fluctuations of the combustion state, and has a delayed response, making it difficult to achieve early warning of combustion anomalies; (2) Analysis based on operating parameters. This method analyzes and predicts the combustion state by collecting combustion operation data (such as coal powder concentration, primary air flow, load, etc.). This type of method relies on physical models or empirical formulas, and usually requires simplifying complex combustion mechanisms, resulting in insufficient description of nonlinear combustion processes; (3) Combustion monitoring based on image processing. With the rapid development of machine vision technology, significant breakthroughs have been made in combustion monitoring methods based on flame images. This type of method evaluates the combustion state by extracting features such as flame brightness, geometric shape, and color distribution. However, traditional image processing technology relies on manual feature extraction and cannot fully reflect the complex dynamic characteristics underlying the combustion process. It also has poor adaptability to different combustion scenarios. (4) Combustion state assessment based on deep learning. Deep learning technology has been gradually applied to the monitoring of industrial combustion processes due to its powerful feature extraction capabilities. However, most existing methods use supervised learning and require a large amount of manually labeled flame image data. This not only increases the cost of data preparation, but also easily introduces human subjective bias, limiting the generalization ability of the model.

[0003] In summary, existing combustion state monitoring technologies suffer from issues such as delayed response, insufficient adaptability to dynamic changes, and a high reliance on manual experience. These issues make it difficult to meet the practical needs of real-time, accurate, and quantitative assessment of industrial furnace combustion states. Therefore, an unsupervised assessment method combining advanced image processing and deep learning techniques is urgently needed to achieve efficient feature extraction of combustion states, quantitative assessment of combustion stability, and anomaly warnings. This approach would fill the gaps in existing technologies and provide reliable support for the intelligent operation of combustion equipment. Summary of the Invention

[0004] Purpose of the invention: To address the problems existing in the prior art, the present invention combines flame image feature extraction with quantitative analysis to propose a quantitative evaluation method for furnace combustion stability based on flame images and stacked convolutional autoencoders. The specific technical solution is as follows:

[0005] A method for quantitatively evaluating furnace combustion stability based on flame images and stacked convolutional autoencoders comprises the following steps:

[0006] S1: creating different furnace combustion operating conditions, collecting flame video data, smoothing the flame video data and collecting flame images at predetermined frame rate intervals;

[0007] S2: using a bilinear method to reduce the dimension of the flame image and normalize the pixel values ​​to the range of [0, 1];

[0008] S3: Establish a stacked convolutional autoencoder to extract deep features of the flame image. The stacked convolutional autoencoder includes two convolutional autoencoders and one autoencoder. A convolutional block attention module based on graph convolution is introduced into the first convolutional autoencoder to improve feature learning capability.

[0009] S4: using the unlabeled flame image to train the stacked convolutional autoencoder, optimizing hyperparameters during the training process, and further reducing the dimensionality of the deep features extracted by the stacked convolutional autoencoder using a principal component analysis method, thereby obtaining core features of the flame image;

[0010] S5: Exponential smoothing is used to reduce the noise of the core features and normalize them to the [0, 1] range. Based on this, the combustion state index (CSI) and combustion performance index (CPI) are established to evaluate the combustion state. High CSI and CPI values ​​indicate that the combustion state is stable.

[0011] S6: Use a sliding window to calculate the variance of the combustion state index (CSI) to form the combustion fluctuation degree (CFD). A high CFD value indicates that there is a combustion anomaly.

[0012] S7: Based on the combustion state index CSI, combustion performance index CPI, and combustion fluctuation degree CFD, a combustion stability comprehensive evaluation index S is established. A high S value means stable combustion, and a combustion abnormality warning is provided when the S value continuously decreases or approaches a threshold.

[0013] In a further embodiment, the pixel averaging method in step S1 is based on The frame rate is averaged over the video data, so each flame image reflects the burning state per second.

[0014] In a further embodiment, the graph convolution-based convolution block attention module in step S3 is composed of a channel graph convolution module and a spatial graph convolution module. After performing channel weighting and spatial weighting on the input features, the results are fused to effectively extract information that is highly relevant to the combustion state.

[0015] In a further embodiment, the channel graph convolution module is implemented as follows:

[0016] The dimension of the input feature F is in is the height, is the width, is the channel dimension;

[0017] The channel dimension of the input feature F Treated as nodes of the graph, the Euclidean distance between channels is calculated based on the feature distribution of each channel, and the adjacency matrix A is constructed based on the similarity c , where each element A c (i, j) represents the degree of correlation between channel i and channel j;

[0018] The input feature F is operated through the graph convolutional network to obtain the new channel feature F c =θ(A c FW c ), where A c represents the adjacency matrix, W c represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function;

[0019] F c Perform global average pooling and global maximum pooling to obtain the feature vector F avg 、F max :

[0020] F avg =GAP(F c )

[0021] F max =GMP(F c )

[0022] The eigenvector F avg 、F max Input shared multilayer perceptron, output channel weight M c :

[0023] M c =θ(MLP(F avg )+MLP(F max ))

[0024] The channel weight M cMultiply it with the original input feature F to get the weighted feature F channel =M c ·F.

[0025] In a further embodiment, the spatial graph convolution module implements the following steps:

[0026] The dimension of the input feature F is in is the height, is the width, is the channel dimension;

[0027] Each pixel of the input feature F is regarded as a node in the graph, and only the height and width of the spatial dimension are concerned. Therefore, the number of spatial nodes is

[0028] Calculate the adjacency matrix A based on the cosine similarity of pixels in space s , this matrix represents the relationship between nodes in the graph;

[0029] The input feature F is operated by the graph convolutional network to obtain the new feature F s =θ(A s FW s ), where A s represents the adjacency matrix, W s represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function;

[0030] For the weighted feature F s Perform global average pooling and global maximum pooling on the channel dimension to generate two two-dimensional feature maps, namely

[0031] These two 2D feature maps After splicing, a 1×1 convolution operation and a Sigmoid function are performed to obtain the spatial weights.

[0032] The spatial weight M s Multiply with the original input feature F to generate the final weighted feature F spatial =M s ·F.

[0033] In a further embodiment, after performing channel weighting and spatial weighting on the input features, the results are fused, specifically including:

[0034] The feature map output by the channel graph convolution module is F channel , the feature map output by the spatial graph convolution module is F spatial ;

[0035] Fchannel and F spatial Add the corresponding position values ​​to get the fused feature, i.e. F fused =F channel +F spatial , its feature dimension is still

[0036] In a further embodiment, in step S5, an exponential smoothing method is used to perform noise reduction on the core features, and the expression is as follows:

[0037] z′ i =αz i +(1-α)z′ i-1

[0038] Where z′ i represents the core features after smoothing, z i represents the core feature, and α represents the smoothing factor.

[0039] In a further embodiment, the combustion state index λ in step S5 i Calculated according to the following formula:

[0040]

[0041] Where z max and z mmn Indicates the maximum and minimum values ​​of the core features after smoothing;

[0042] The combustion performance index δ i Calculated according to the following formula:

[0043]

[0044] Where v(i) and μ(i) represent the i-th flame image z′ i The standard deviation and mean of Represents the weight parameter, satisfying

[0045] In a further embodiment, the sliding time window in step S6 is set to T, and the combustion state index within the time window is expressed as λ1, λ2, ..., λ T , calculate the variance σ of the combustion state index in each time window i , and the degree of combustion fluctuation is obtained:

[0046]

[0047] Where, represents the average value of CSI within the window, that is,

[0048] In a further embodiment, the combustion stability comprehensive evaluation index S in step S7 is defined as follows:

[0049] S=ω1·λ i +ω2·δ i -ω2·σ i

[0050] Where λ i represents CSI, σ i represents CFD, ω1, ω2, and ω3 are weight parameters and satisfy ω1+ω2+ω3=1.

[0051] Compared with the prior art, the present invention has the following significant beneficial effects:

[0052] (1) Efficient feature extraction and unsupervised learning. This paper adopts a deep learning model that integrates the convolutional block attention module (CBAM) and the stacked convolutional autoencoder (SCAE). It enhances the key feature extraction capability through the channel attention and spatial attention mechanisms, and introduces a hybrid sparrow search algorithm to optimize the SCAE hyperparameters to improve robustness and adaptability. It achieves efficient feature extraction and dynamic change capture of flame images, greatly improving the accuracy and automation of combustion state assessment.

[0053] (2) Real-time quantification and early warning. The present invention establishes a comprehensive evaluation model for combustion stability through the joint analysis of the combustion state index (CSI), combustion performance index (CPI) and combustion fluctuation degree (CFD). This model realizes the real-time quantitative evaluation of the combustion state, timely identifies potential problems in the combustion state, and has the ability to provide early warning of combustion anomalies.

[0054] (3) Multi-dimensional dynamic evaluation and strong applicability. Compared with traditional flame detection systems, the present invention can provide richer combustion state information and more accurately identify combustion instability. In addition, the present invention is applicable to a variety of combustion equipment (such as power boilers and industrial kilns) and can maintain high efficiency and stable performance under different combustion equipment scales, fuel types, and operating conditions, thus having wide application value.

[0055] In summary, the present invention effectively improves the monitoring accuracy and dynamic analysis capability of the combustion state, provides strong support for the intelligent operation, safety management and energy efficiency optimization of combustion equipment, and has significant economic and social benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic flow chart of the overall technical solution of the present invention.

[0057] Figure 2 It is a schematic diagram of the stacked convolutional autoencoder structure of the present invention.

[0058] Figure 3 It is a structural diagram of the convolutional block attention module of the present invention.

[0059] Figure 4 It is a schematic diagram of flame images under different furnace operating conditions of the present invention. DETAILED DESCRIPTION

[0060] In the following description, numerous specific details are provided to provide a more thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced without one or more of these details. In other instances, certain technical features well known in the art have not been described to avoid confusion with the present invention.

[0061] Example 1:

[0062] This embodiment discloses a method for quantitatively evaluating furnace combustion stability based on flame images and stacked convolutional autoencoders. The technical route is as follows: Figure 1 As shown, the specific steps include:

[0063] S1: Create different furnace combustion operating conditions, use the monitoring device to collect flame video data, and use the pixel averaging method to smooth the video data to obtain high-quality flame images;

[0064] S2: Use bilinear method to reduce the dimension of the original flame image and normalize the pixel values ​​to the range of [0,1];

[0065] S3: Build a stacked convolutional autoencoder (SCAE) to extract deep features of flame images. It consists of two convolutional autoencoders (CAEs) and one autoencoder (AE). The first CAE introduces a graph convolutional network-convolutional block attention module (GCN-CBAM) to improve feature learning capabilities.

[0066] S4: SCAE is trained using unlabeled flame images. A multi-strategy improved sparrow search algorithm (MISSA) is introduced during the training process to optimize SCAE hyperparameters. Principal component analysis (PCA) is then used to further reduce the dimensionality of the deep features extracted by SCAE, thereby obtaining the core features of the flame images.

[0067] S5: Exponential smoothing is used to reduce the noise of the core features and normalize them to the [0, 1] range. Based on this, the combustion state index (CSI) and combustion performance index (CPI) are then established to evaluate the combustion state. High CSI and CPI values ​​indicate that the combustion state is stable.

[0068] S6: Use a sliding window to calculate the CSI variance and form the combustion fluctuation degree (CFD). A high CFD value indicates that there is a combustion anomaly.

[0069] S7: A comprehensive combustion stability evaluation index S is established based on CSI, CPI, and CFD. A high S value indicates stable combustion, and a combustion abnormality warning is provided when the S value continuously decreases or approaches the threshold.

[0070] As an example, the pixel averaging method in step S1 is based on the pixel averaging method per second. The frame rate averages the video data, so each flame image reflects the burning state per second;

[0071] Preferably, the SCAE structure described in step S3 is as follows Figure 2 As shown, where x represents the preprocessed input image, e represents the convolutional encoder, h represents the deep features, d represents the convolutional decoder, and x rec Represents the reconstructed image. Table 2 summarizes the hyperparameters of SCAE, where e a represents the ath encoder, d bRepresents the b-th decoder, ' / ' represents no operation, Conv2d represents two-dimensional convolution, ReLU represents activation function, MaxPool2d represents two-dimensional maximum pooling, F represents flattening operation, f represents fully connected operation, Y represents expansion operation, Upsampling represents upsampling operation, ConvTranspose2d represents two-dimensional convolution transpose, C(Q@c×c+q) represents convolution operation with Q filters of size c×c and stride q; P(o×o+l) represents pooling operation with window size o×o and stride l; U(g×g) represents upsampling operation, which can transform the dimension of input data into g×g;

[0072] Preferably, the GCN-CBAM structure in step S3 is as follows Figure 3 As shown in the figure, it consists of a channel graph convolution module and a spatial graph convolution module. After performing channel weighting and spatial weighting on the input features, the results are fused to effectively extract information that is highly relevant to the combustion state.

[0073] Preferably, the implementation steps of the channel graph convolution module are as follows:

[0074] S3.1a, the dimension of the input feature F is in is the height, is the width, is the channel dimension;

[0075] S3.2a. Change the channel dimension of the input feature F Treated as nodes of the graph, the Euclidean distance between channels is calculated based on the feature distribution of each channel, and the adjacency matrix A is constructed based on the similarity c , where each element A c (i, j) represents the degree of correlation between channel i and channel j;

[0076] S3.3a. Operate the input feature F through the graph convolutional network to obtain the new channel feature F c =θ(A c FW c ), where A c represents the adjacency matrix, W c represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function;

[0077] S3.4a, for F c Perform global average pooling (GAP) and global max pooling (GMP) to obtain the feature vector F avg =GAP(F c ), Fmax =GMP(F c );

[0078] S3.5a. Input the two feature vectors into a shared multilayer perceptron (MLP) and output the channel weights M. c =θ(MLP(F avg )+MLP(F max ));

[0079] S3.6a, channel weight M c Multiply it with the original input feature F to get the weighted feature F channel =M c ·F.

[0080] Preferably, the implementation steps of the spatial graph convolution module are as follows:

[0081] S3.1b, the dimension of the input feature F is in is the height, is the width, is the channel dimension;

[0082] S3.2b, each pixel of the input feature F is treated as a node in the graph. Here we only focus on the height and width of the spatial dimension, so the number of spatial nodes is

[0083] S3.3b, calculate the adjacency matrix A based on the cosine similarity of pixels in space s , this matrix represents the relationship between nodes (pixels) in the graph;

[0084] S3.4b, operate the input feature F through the graph convolutional network to obtain the new feature F s =θ(A s FW s ), where A s represents the adjacency matrix, W s represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function;

[0085] S3.5b, weighted feature F s Perform global average pooling and global maximum pooling on the channel dimension to generate two two-dimensional feature maps, namely

[0086] S3.6b, after concatenating these two feature maps, pass a 1×1 convolution operation and a Sigmoid function to obtain the spatial weight M s ,Right now

[0087] S3.7b, the spatial weight M s Multiply with the original input feature F to generate the final weighted feature F spatial =M s ·F.

[0088] As a preferred method, the implementation steps of feature fusion are as follows:

[0089] S3.1c, the feature map output by the channel graph convolution module is F channel , the feature map output by the spatial graph convolution module is F spatial ;

[0090] S3.2c, F channel and F spatial Add the corresponding position values ​​to get the fused feature, i.e. F fused =F channel +F spatial , its feature dimension is still

[0091] As a preferred method, the MISSA described in step S4 solves the problems of the traditional sparrow search algorithm being prone to falling into local optimality and limited convergence speed during the high-dimensional space optimization process through mechanisms such as elite individual mutation, hybrid attention search, and multimodal perturbation. The specific steps are as follows:

[0092] S4.1. Initialize parameters. Set the size of the sparrow population N and the maximum number of iterations. Set the search range of hyperparameters such as convolution filter size and number of filters, set the adaptive weight ψ and dynamic adjustment coefficient ξ, and use Logistic chaotic mapping to generate the initialized sparrow population. The Logistic chaotic mapping formula is x n+1 =ηx n (1-x n ), where η represents the chaos parameter;

[0093] S4.2. Establish fitness function. For SCAE, the fitness function is defined as Fitness = MSE (X rec ,X input ), where MSE represents the reconstructed image X rec With the input image X input The mean square error between

[0094] S4.3, Sparrow search process.

[0095] (1) Discoverer update. The discoverer is responsible for global search, and its position update formula is Where Xbestt represents the current optimal solution, ψ represents the adaptive weight, and rand is a random number;

[0096] (2) Follower update. The follower is responsible for local search, and its position update formula is

[0097] (3) Update of the alerter. The alerter is responsible for jumping out of the local optimum, and its position update formula is Where ξ represents the dynamic adjustment coefficient, randn represents the random number of the standard normal distribution;

[0098] S4.4, Adaptive weight update. The adaptive weight ψ is dynamically adjusted with the number of iterations and is defined as where ψ max and ψ min Represents the maximum and minimum values ​​of the adaptive weight, t represents the current number of iterations, Indicates the maximum number of iterations;

[0099] S4.5, Dynamic adjustment coefficient update. The dynamic adjustment coefficient ξ is dynamically adjusted with the number of iterations and is defined as where ξ min and ξ max Indicates the maximum and minimum values ​​of the dynamic adjustment coefficient ξ;

[0100] S4.6, Hybrid mutation search. In the hyperparameter search process, a hybrid mutation strategy consisting of Cauchy mutation and Gaussian mutation is introduced to effectively enhance the overall search accuracy, that is, in and θ represent the coefficient of variation, r represents the random number, L c Follows Cauchy distribution (L c ~Cauchy(0,1)), P c represents the probability of Cauchy mutation, N(0,1) represents the random number of standard normal distribution;

[0101] S4.7, elite retention mechanism. In each iteration, the current optimal solution is retained And update the global optimal solution X best ;

[0102] S4.8. Repeat steps S4.3 to S4.7 until the maximum number of iterations T is reached. max or the fitness function converges;

[0103] S4.9. Output the optimal hyperparameters. Output the global optimal solution X best ,The stacked convolutional autoencoder hyperparameters are shown in Table 1.

[0104] Table 1: Stacked Convolutional Autoencoder Hyperparameters

[0105]

[0106] As an example, the exponential smoothing calculation process in step S5 is: i =αz i +(1-α)z′ i-1 , where z′ i represents the core features after smoothing, z i represents the core feature, and α represents the smoothing factor.

[0107] Preferably, the CSI in step S5 is defined as where z max and z min Indicates the maximum and minimum values ​​of the core feature after smoothing.

[0108] Preferably, the CPI in step S5 is defined as Where ν(I) and μ(i) represent the i-th flame image z′ i The standard deviation and mean of Represents the weight parameter, satisfying

[0109] Preferably, the sliding time window in step S6 is set to T, and the combustion state index within the time window is expressed as λ1, λ2, ..., λ T , then calculate CFD in each time window, that is in represents the average value of CSI within the window, that is,

[0110] As a preference, the combustion stability comprehensive evaluation index S in step S7 is defined as S=ω1·λ i +ω2·δ i -ω2·σ i , where λ i represents CSI, σ i represents CFD, ω1, ω2, and ω3 are weight parameters and satisfy ω1+ω2+ω3=1.

[0111] Example 2:

[0112] S1. In the furnace of a 660MW coal-fired boiler, a CCD camera installed at the burner flame observation hole was used to collect flame video data under different operating conditions. The frame rate was 25 frames per second, the resolution was 960×576 pixels, and the image type was RGB color image.

[0113] S2. The resolution of the flame image is adjusted to 128×128 pixels using a bilinear method, and the pixel value of the flame image is linearly normalized from [0,255] to [0,1]. Finally, pixel averaging is performed on consecutive frames of the flame video to eliminate image noise interference.

[0114] S3. Establish a SCAE model consisting of two CAEs and one AE, and introduce GCN-CBAM into the first CAE to improve the extraction ability of key features.

[0115] S4. Use unlabeled flame images to train SCAE, and introduce MISSA to optimize SCAE hyperparameters during the training process. Then use PCA to reduce the dimensionality of the deep features extracted by SCAE to obtain the core feature z of the flame image. i , whose dimension is 10.

[0116] In this embodiment, the steps of MISSA optimizing SCAE hyperparameters are summarized as follows:

[0117] S4.1. Set MISSAE parameters, i.e., the size of the sparrow population N = 30, the maximum number of iterations Adaptive weight range ψ max =1,ψ max =0.2, dynamic adjustment coefficient ξ max =1,ξ min = 0.1, Gaussian coefficient of variation ζ = 0.1, Cauchy coefficient of variation θ = 0.1, the probability of Cauchy variation P c =0.3;

[0118] S4.2. Set the SCAE hyperparameter search range, i.e., convolution filter size c∈[3,7], number of filters Q∈[16,128], learning rate α∈[0.0001,0.01], batch size ∈[25,75], and use Logistic chaotic mapping to generate the initial sparrow population x. n+1 =ηx n (1-x n ), where η = 4, x n ∈(0,1), and establish the fitness function Fitness=MSE(X rec ,X input );

[0119] S4.3, Sparrow search process. (1) Discoverer update, (2) Follower updates, (3) Update of the vigilant,

[0120] S4.4, hybrid mutation search. In the later stage of the search (such as t>0.8T max), perform mixed mutation on some sparrows and generate random numbers r∈[0,1], if r <P c Cauchy variation If r≥P c Gaussian variation

[0121] S4.5, elite retention mechanism. In each iteration, the current optimal solution is retained And update the global optimal solution X best ;

[0122] S4.6. Repeat steps S4.3 to S4.5 until the maximum number of iterations T is reached. max or the fitness function converges;

[0123] S4.9. Output the optimal hyperparameters. Output the global optimal solution X best , that is, the optimal hyperparameter combination, convolution filter size c = 5, number of filters Q = 64, learning rate α = 0.002, batch size batch size = 50.

[0124] S5, core feature z i Perform exponential smoothing to obtain z′ i =αz i +(1-α)z′ i-1 , where the smoothing factor α is set to 0.7;

[0125] For the smoothed core feature z′ i Normalize and get CSI, that is where z max and z min Represent the maximum and minimum values ​​of the feature respectively;

[0126] Calculate z′ i The standard deviation v(i) and mean μ(i) are used to calculate CPI, that is, in Represents the weight parameter, satisfying

[0127] S6. Set the sliding time window T to 1 minute and select the CSI within the time window, i.e., λ1, λ2,…, λ 60 , calculate CFD in each time window, that is in represents the average value of CSI within the window, that is,

[0128] S7. Establish a comprehensive evaluation index of combustion stability based on CSI, CPI and CFD, that is, S = ω1·λ i +ω2·δ i-ω2·σ i , where the weight parameters ω1 = 0.4, ω2 = 0.6, ω3 = 0.5;

[0129] When S>0, it means that the combustion state is stable, and when S≤0, it means that there is a risk of instability in the combustion state. The quantitative characterization of combustion stability helps to guide the adjustment of the combustion state and avoid abnormal combustion accidents.

[0130] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes may be made to it in form and detail without departing from the spirit and scope of the present invention as defined in the appended claims.

Claims

1. A method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder, characterized in that: The steps include: S1: creating different furnace combustion operating conditions, collecting flame video data, smoothing the flame video data and collecting flame images at predetermined frame rate intervals; S2: using a bilinear method to reduce the dimension of the flame image and normalize the pixel values ​​to the range of [0, 1]; S3: Establish a stacked convolutional autoencoder to extract deep features of the flame image, wherein the stacked convolutional autoencoder includes two convolutional autoencoders and one autoencoder, wherein the first convolutional autoencoder introduces a convolutional block attention module based on graph convolution; S4: using the unlabeled flame image to train the stacked convolutional autoencoder, optimizing hyperparameters during the training process, and further reducing the dimensionality of the deep features extracted by the stacked convolutional autoencoder using a principal component analysis method to obtain the core features of the flame image; S5: Use exponential smoothing to reduce the noise of the core features and normalize them to the [0,1] numerical range. On this basis, the combustion state index and combustion performance index are established to evaluate the combustion state. S6: Calculate the variance of the combustion state index using a sliding window to form the degree of combustion fluctuation; S7: Based on the combustion state index, combustion performance index, and combustion fluctuation degree, a combustion stability comprehensive evaluation index is established, and a combustion abnormality warning is provided when the combustion stability comprehensive evaluation index continuously decreases or approaches a threshold value.

2. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: In step S1, the video data is smoothed using pixel averaging method, with a resolution of The frame rate is averaged over the video data, and each flame image reflects the burning state every second.

3. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: The graph convolution-based convolution block attention module described in step S3 consists of a channel graph convolution module and a spatial graph convolution module. After performing channel weighting and spatial weighting on the input features, the results are fused to extract information related to the combustion state.

4. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 3, characterized in that: The implementation steps of the channel graph convolution module are as follows: The dimension of the input feature F is in is the height, is the width, is the channel dimension; The channel dimension of the input feature F Treated as nodes of the graph, the Euclidean distance between channels is calculated based on the feature distribution of each channel, and the adjacency matrix A is constructed based on the similarity c , where each element A c (i, j) represents the degree of correlation between channel i and channel j; The input feature F is operated through the graph convolutional network to obtain the new channel feature F c =θ(A c FW c ), where A c represents the adjacency matrix, W c represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function; F c Perform global average pooling and global maximum pooling to obtain the feature vector F avg 、F max : F avg =GAP(F c ) F max =GMP(F c ) The eigenvector F avg 、F max Input shared multilayer perceptron, output channel weight M c : M c =θ(MLP(F avg )+MLP(F max )) The channel weight M c Multiply it with the original input feature F to get the weighted feature F channel =M c ·F.

5. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 3, characterized in that: The implementation steps of the spatial graph convolution module are as follows: The dimension of the input feature F is in is the height, is the width, is the channel dimension; Each pixel of the input feature F is regarded as a node in the graph, and only the height and width of the spatial dimension are concerned. Therefore, the number of spatial nodes is Calculate the adjacency matrix A based on the cosine similarity of pixels in space s , this matrix represents the relationship between nodes in the graph; The input feature F is operated by the graph convolutional network to obtain the new feature F s =θ(A s FW s ), where A s represents the adjacency matrix, W s represents the learnable convolution weight matrix, and θ represents the Sigmoid activation function; For the weighted feature F s Perform global average pooling and global maximum pooling on the channel dimension to generate two two-dimensional feature maps, namely These two 2D feature maps After splicing, a 1×1 convolution operation and a Sigmoid function are performed to obtain the spatial weights. The spatial weight M s Multiply with the original input feature F to generate the final weighted feature F spatial =M s ·F.

6. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 3, characterized in that: After performing channel weighting and spatial weighting on the input features, the results are fused, specifically including: The feature map output by the channel graph convolution module is F channel , the feature map output by the spatial graph convolution module is F spatial ; F channel and F spatial Add the corresponding position values ​​to get the fused feature, i.e. F fused =F channel +F spatial , its feature dimension is still 7. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: In step S5, the exponential smoothing method is used to reduce the noise of the core features. The expression is as follows: With' i =αz i +(1-α)z′ i-1 Where z′ i represents the core features after smoothing, z i represents the core feature, and α represents the smoothing factor.

8. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: The combustion state index λ in step S5 i Calculate according to the following formula: Where z max and z min Indicates the maximum and minimum values ​​of the core features after smoothing; The combustion performance index δ i Calculate according to the following formula: Where v(i) and μ(i) represent the i-th flame image z′ i The standard deviation and mean of Represents the weight parameter, satisfying 9. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: In step S6, the sliding time window is set to T, and the combustion state index within the time window is expressed as λ1, λ2, ..., λ T , calculate the variance σ of the combustion state index in each time window i , and the degree of combustion fluctuation is obtained: Where, represents the average value of CSI within the window, that is, 10. The method for quantitatively evaluating furnace combustion stability based on flame image and stacked convolutional autoencoder according to claim 1, characterized in that: The combustion stability comprehensive evaluation index S in step S7 is defined as follows: S=ω1·λ i +ω2·d i -ω2·s i Where λ i represents CSI, σ i represents CFD, ω1, ω2, and ω3 are weight parameters and satisfy ω1+ω2+ω3=1.

Images