A method and system for quantitatively detecting a combustion state of a furnace flame

Abstract

The application discloses a kind of furnace flame combustion state quantitative detection method and system, it is related to combustion state detection technical field, the method includes: collecting the image data and process parameter of furnace flame in set time period;Image latent feature extraction model is obtained after the preprocessed image data is input to image latent feature matrix;Image latent feature extraction model is the trained DANet-SCAE model;Dimension reduction is carried out to each image latent feature matrix, and combustion state index is obtained;From process parameter, the feature parameter with the correlation coefficient greater than set value between combustion state index is screened;Combustion state index of each image data is linearly weighted combination with feature parameter, and flame state index is obtained;Flame state index is clustered, and each image data is divided into different flame combustion state according to clustering result.The application realizes the quantitative detection of furnace flame combustion state.

Description

Technical Field

[0001] This invention relates to the field of combustion state detection technology, and in particular to a method and system for quantitative detection of combustion state of furnace flame. Background Technology

[0002] Currently, the proportion of new energy power in China is continuously increasing, but due to its strong intermittency and volatility, thermal power generation remains the most mainstream power generation method. The combustion state of the furnace flame is a crucial indicator of the power grid system, significantly impacting its stable operation. Traditional methods for detecting the combustion state of furnace flames in China are mainly based on experiments and mechanistic models. However, due to the complexity of industrial combustion behavior and the influence of multiple variables on the combustion state, the accuracy of these models is often limited. Image-based flame combustion state detection methods are widely used because they do not require complex mechanistic models, directly extract feature information from images to determine the combustion state, and offer higher timeliness and accuracy.

[0003] Due to the rise and development of artificial intelligence, current methods for evaluating the state of flames based on images mainly focus on machine learning methods based on feature engineering and the more popular deep learning methods in recent years. However, machine learning methods rely heavily on prior knowledge and feature selection, making the process complex and resulting in poor generalization ability. In contrast, deep learning methods do not require manual feature selection and have superior performance. In current combustion state detection, most methods only perform qualitative classification, dividing the combustion state into stable and unstable states. The distinction and identification of critical states for different categories are usually quite difficult, and the accuracy of model detection is limited by manually labeled data. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for quantitatively detecting the combustion state of a furnace flame, thereby achieving quantitative detection of the combustion state of the furnace flame.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for quantitatively detecting the combustion state of a furnace flame, comprising:

[0007] Collect image data and process parameters of the furnace flame within a set time period;

[0008] The image data described above are preprocessed;

[0009] The preprocessed image data is input into the image latent feature extraction model to obtain the image latent feature matrix; the image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method, and the DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE.

[0010] The combustion state index of each image data is obtained by dimensionality reduction of the latent feature matrix of each image.

[0011] Based on the Pearson correlation coefficient, feature parameters with a correlation coefficient greater than a set value with the combustion state index are selected from the process parameters corresponding to the image data.

[0012] The flame state index of each image data is obtained by linearly weighting the combustion state index and the corresponding feature parameters.

[0013] Clustering is performed on the flame state indicators of each image data, and the image data are divided into different flame combustion states based on the clustering results.

[0014] Optionally, the process parameters include burner flame detection signal, load command, actual load, main steam pressure, main steam temperature, main steam flow rate, feedwater flow rate, and total fuel quantity.

[0015] Optionally, the image data is preprocessed, specifically including:

[0016] Median filtering is performed on each image data to convert them into images with a resolution of 960×576.

[0017] The resolution of a 960×576 image was compressed to 128×128 using the bilinear method.

[0018] The normalization operation is used to convert the pixel values ​​in the image with a resolution of 128×128 from 0 to 255 to 0 to 1, thus obtaining the preprocessed image data.

[0019] Optionally, by performing dimensionality reduction on the latent feature matrices of each image, combustion state indices for each image data are obtained, specifically including:

[0020] Principal component analysis is used to reduce the dimensionality of the latent feature matrix of the image, resulting in a dimension-reduced matrix.

[0021] The matrix after dimensionality reduction is smoothed using a single exponential smoothing method.

[0022] The smoothed matrix is ​​normalized to obtain the combustion state index.

[0023] Optionally, the DANet-SCAE model includes a first encoder, DANet, a second encoder, a third encoder, a fourth encoder, a flattening layer, a fifth encoder, a fifth decoder, a dimension reconstruction layer, a fourth decoder, a third decoder, a second decoder, and a first decoder connected in sequence.

[0024] The first encoder, DANet, second encoder, second decoder, and first decoder constitute the first convolutional autoencoder; the third encoder, fourth encoder, fourth decoder, and third decoder constitute the second convolutional autoencoder; the flattening layer, fifth encoder, fifth decoder, and dimension reconstruction layer constitute the autoencoder;

[0025] The training process of the DANet-SCAE model includes:

[0026] Based on the flame training set, the first convolutional autoencoder is trained with the learning objective of minimizing the error between the sample flame image input to the first convolutional autoencoder and the reconstructed flame image output by the first convolutional autoencoder.

[0027] The second convolutional autoencoder is trained with the learning objective of minimizing the error between the input first-layer latent feature vector of the second convolutional autoencoder and the reconstructed first-layer latent feature vector of the output second convolutional autoencoder.

[0028] The autoencoder is trained with the learning objective of minimizing the error between the second-layer latent feature vector of the input autoencoder and the reconstructed second-layer latent feature vector of the output autoencoder.

[0029] Optionally, the setting value is 0.8.

[0030] Optionally, different flame combustion states include stable combustion states and unstable combustion states.

[0031] This invention discloses a quantitative detection system for furnace flame combustion state, comprising:

[0032] The data acquisition module is used to collect image data and process parameters of the furnace flame within a set time period;

[0033] The preprocessing module is used to preprocess the image data.

[0034] The feature extraction module is used to input the preprocessed image data into the image latent feature extraction model to obtain the image latent feature matrix. The image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method. The DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE.

[0035] The combustion state index determination module is used to obtain the combustion state index of each image data by performing dimensionality reduction processing on the potential feature matrix of each image.

[0036] The feature parameter determination module is used to filter feature parameters from the process parameters corresponding to the image data based on the Pearson correlation coefficient, and to select feature parameters whose correlation coefficient with the combustion state index is greater than a set value.

[0037] The flame state index determination module is used to linearly weight and combine the combustion state index of each image data with the corresponding feature parameters to obtain the flame state index of each image data.

[0038] The combustion state determination module is used to cluster the flame state indicators of each image data and classify each image data into different flame combustion states based on the clustering results.

[0039] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0040] This invention inputs preprocessed image data into an image latent feature extraction model to obtain an image latent feature matrix. The image latent feature extraction model is a DANet-SCAE model trained using an unsupervised learning method, avoiding the process of manually labeling data. Based on the latent feature matrices of each image, flame state indices of each image data are obtained. The flame state indices of each image data are clustered, and the image data are divided into different flame combustion states according to the clustering results, thereby realizing the quantitative detection of the furnace flame combustion state. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 A flowchart of a method for quantitatively detecting the combustion state of a furnace flame provided in an embodiment of the present invention;

[0043] Figure 2 A schematic diagram illustrating the principle of a method for quantitatively detecting the combustion state of a furnace flame, provided in an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the DANet-SCAE model structure provided in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] The purpose of this invention is to provide a method and system for quantitatively detecting the combustion state of a furnace flame, thereby achieving quantitative detection of the combustion state of the furnace flame.

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0048] Autoencoders (AEs) are a type of artificial neural network used in semi-supervised and unsupervised learning to learn representations of input information. Stacked AEs (SAEs) can learn deeper information representations from the original input data, resulting in better image reconstruction performance. Convolutional AEs (CAEs) better preserve the spatial information of images. Building upon this, combining SAEs and CAEs to create stacked convolutional AEs (SCAEs) can better extract latent features from flame images. The Dual Attention Network (DANet) combines positional and channel attention mechanisms. The positional attention module selectively aggregates features at each position by weighted summation of features across all positions, while the channel attention module selectively emphasizes interdependent channel mappings by integrating relevant features across all channel mappings. The sum of the outputs of the two attention modules further improves the feature representation. To reduce reliance on manual labeling and improve detection accuracy, this invention establishes a flame latent feature extraction network based on DANet-SCAE, defines a combustion state index, and calculates combustion stability results by combining furnace signals.

[0049] like Figure 1 As shown in the figure, the method for quantitatively detecting the combustion state of a furnace flame provided in this embodiment specifically includes the following steps.

[0050] Step 101: Collect image data and process parameters of the furnace flame within a set time period.

[0051] The time period is set for 24 hours a day.

[0052] Step 102: Preprocess the image data.

[0053] Step 102 specifically includes:

[0054] The mean value of each image data was calculated using the pixel averaging method.

[0055] Based on the mean of each image data point, median filtering is performed on each image data point to convert them into images with a resolution of 960×576. The median filtering operation aims to improve model training efficiency, reduce the amount of training data and time, mitigate the impact of furnace noise, and reduce signal strength loss during data transmission. Pixel averaging is used to process 25 frames per second, followed by median filtering; this yields high-quality combustion flame images representing that specific second-level time period.

[0056] The resolution of a 960×576 image was compressed to 128×128 using the bilinear method.

[0057] The normalization operation is used to convert the pixel value range of the image compressed to 128×128 resolution from 0 to 255 to 0 to 1, that is, from [0,255] to [0,1], to obtain the preprocessed image data.

[0058] Step 103: Input the preprocessed image data into the image latent feature extraction model to obtain the image latent feature matrix; the image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method, and the DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE.

[0059] like Figure 2 As shown, the DANet-SCAE model includes a first encoder e1, DANet, a second encoder e2, a third encoder e3, a fourth encoder e4, a flatten layer f, a fifth encoder e5, a fifth decoder d5, a dimension reconstruction layer uf, a fourth decoder d4, a third decoder d3, a second decoder d2, and a first decoder d1, which are connected in sequence.

[0060] The first encoder, DANet, second encoder, second decoder, and first decoder constitute the first convolutional autoencoder. Figure 2 The first CAE layer in the middle; the third encoder, fourth encoder, fourth decoder and third decoder constitute the second convolutional autoencoder ( Figure 2 The second CAE layer in the middle; the flattening layer, the fifth encoder, the fifth decoder, and the dimension reconstruction layer constitute the autoencoder ( Figure 2 (The third AE layer in the middle).

[0061] Both the fifth encoder and the fifth decoder are linear layers, and the dimension reconstruction layer is an unflattened layer.

[0062] In step 103, the first layer of the convolutional autoencoder is the first convolutional autoencoder, where the DANet attention mechanism layer is added to the first convolutional pooling of the first convolutional autoencoder in the SCAE model.

[0063] The training process of the DANet-SCAE model includes:

[0064] The DANet-SCAE model is trained using a flame training set. The sample data in the flame training set consists of sample flame images.

[0065] The construction of the flame training set included collecting video data of the furnace flame under various combustion states and related process parameters from the distributed control system (DCS) of a coal-fired power plant boiler database. The video data was 3 hours long with a frame rate of 25, recording image data of the flame state from stable combustion to unstable combustion until flameout, and then back to a stable state under different given conditions.

[0066] The process parameters include burner flame detection signal, load command, actual load, main steam pressure, main steam temperature, main steam flow rate, feedwater flow rate, total fuel quantity, first stage pressure 1, first stage pressure 5, first stage pressure 6, separator A outlet temperature, separator B outlet temperature, separator C outlet temperature, separator D outlet temperature, water tank exhaust pressure 1, water tank exhaust pressure 2, water tank exhaust pressure 3, reheat steam temperature, economizer inlet pressure 1, economizer inlet pressure 2, economizer inlet temperature 1, economizer inlet temperature 2, economizer inlet temperature 3, and other variables.

[0067] The specific conditions under which different conditions were given were as follows: from 13:00:00 to 14:35:00, the coal feed rate remained relatively stable, then gradually decreased. After a plateau period of about 8 minutes at 30t / h, the coal feed rate continued to decrease, reaching 0t / h at 14:45:15. The coal feed rate remained in a extinguished state until 15:02:00. Subsequently, the coal feed rate gradually increased to about 60t / h and remained stable until 16:00:00.

[0068] Training the first layer (first convolutional autoencoder): The first convolutional autoencoder is trained with the learning objective of minimizing the error between the sample flame image input to the first convolutional autoencoder and the reconstructed flame image output by the first convolutional autoencoder.

[0069] Training the second layer (second convolutional autoencoder): The second convolutional autoencoder is trained with the learning objective of minimizing the error between the input latent feature vector of the first layer of the second convolutional autoencoder and the reconstructed latent feature vector of the first layer of the output second convolutional autoencoder.

[0070] Training the third layer (autoencoder): The autoencoder is trained with the learning objective of minimizing the error between the input autoencoder's second-layer latent feature vector and the reconstructed second-layer latent feature vector of the output autoencoder.

[0071] In step 103, the latent feature matrix of the image is the output of the trained DANet-SCAE model, i.e., the output of the first decoder.

[0072] As a specific implementation method, the DANet-SCAE model training process includes the following:

[0073] Training the first layer: The preprocessed image is passed through the first encoder e1, DANet, and the first encoder e2 to obtain the first layer latent feature vector of size 32×32×32. Then, it is reconstructed by the second decoder d2 and the first decoder d1 to reconstruct the flame image with the same pixels as the original. The learning objective of the network at this stage is to minimize the error between the reconstructed flame image and the image input to the network after preprocessing.

[0074] Training the second layer: The first layer latent feature vector is processed by the third encoder e3 and the fourth encoder e4 to obtain the second layer latent feature vector of size 8×8×8. Then, it is reconstructed by the fourth decoder d4 and the third decoder d3 to reconstruct the first layer latent feature vector. The learning objective of the network in this stage is to minimize the error between the reconstructed first layer latent feature vector and the input first layer latent feature vector.

[0075] Training the third layer: The second-layer latent feature vector is passed through the data flattening layer f and the fifth encoder e5 to obtain the final 256-dimensional flame latent feature vector. Then, it is reconstructed through the fifth decoder d5 and the dimension reconstruction layer uf to reconstruct the second-layer latent feature vector. The learning objective of the network at this stage is to minimize the error between the reconstructed second-layer latent feature vector and the input second-layer latent feature vector.

[0076] Fine-tune the parameters of the DANet-SCAE feature extraction network (DANet-SCAE model): At this point, all layers are treated as a whole network model, and all weights, biases and other parameters of the network are readjusted through backpropagation and gradient descent algorithms.

[0077] The network configuration of the DANet-SCAE model is shown in Table 1.

[0078] Table 1 Network configuration of the DANet-SCAE model

[0079]

[0080] Step 104: Based on the dimensionality reduction of the latent feature matrix of each image, the combustion state index of each image data is obtained.

[0081] Step 104 specifically includes:

[0082] Principal component analysis (PCA) is used to reduce the dimensionality of the latent feature matrix of the image, resulting in a dimensionality-reduced matrix, which includes:

[0083] Let n = 256, and calculate the cumulative variance contribution rate for k = 1, 2, 3, 4, 5, 6... respectively. k1 is the minimum value that makes the cumulative variance contribution rate of the first k1 principal components greater than 99%. The specific calculation formula is as follows:

[0084]

[0085] Where m is the number of samples, Y and Y' are the sample matrix (image latent feature matrix) and the sample mean matrix, respectively. Y' is obtained by subtracting the mean of each dimension of the sample matrix Y. C is the variance matrix, and λ i Let η be the descending order of the eigenvalues ​​of C. k Let η be the variance contribution rate of the k-th principal component. t Let U be the cumulative variance contribution rate of the first t principal components, U be the eigenvectors corresponding to the k principal components, and Z be the matrix after dimensionality reduction.

[0086] The matrix after dimensionality reduction was smoothed using a single exponential smoothing method to reduce redundant information, with the smoothing factor set to 0.7.

[0087] The smoothed matrix is ​​normalized and scaled to [0,1] to obtain the combustion state index.

[0088] The formula for normalization is:

[0089]

[0090] In the formula, Z' is the smoothed matrix, Z min Z is the minimum value among the combustion state indicators. max Φ represents the maximum value among the combustion state indicators, and Φ is the final combustion state indicator.

[0091] Step 105: Based on the Pearson correlation coefficient, select feature parameters from the process parameters corresponding to the image data whose correlation coefficient with the combustion state index is greater than a set value.

[0092] The set value is 0.8.

[0093] Step 105 specifically includes:

[0094] Perform outlier screening, missing value supplementation, and normalization on the process parameters acquired by DCS.

[0095] For each image data, features (process parameters corresponding to each image data) are initially screened based on the Pearson correlation coefficient. Feature parameters with a correlation coefficient greater than 0.8 between the parameters (process parameters after outlier screening, missing value supplementation, and normalization) and combustion state indicators are selected.

[0096] Step 106: Perform a linear weighted combination of the combustion state index and corresponding feature parameters of each image data to obtain the flame state index of each image data.

[0097] The formula for calculating the linear weighted combination is: and

[0098] Where Q represents the flame state index, n represents the sum of the number of characteristic parameters and the number of combustion state indices, and y i w is the i-th value in the set consisting of combustion state indices and their corresponding characteristic parameters. i For y i The corresponding weighting coefficients.

[0099] Weighting coefficient w i This was calculated using a genetic algorithm (GA).

[0100] Step 107: Cluster the flame state indicators of each image data, and classify each image data into different flame combustion states based on the clustering results.

[0101] In specific implementations, different flame combustion states include stable combustion states and unstable combustion states.

[0102] As another specific implementation method, the different flame combustion states include five state intervals.

[0103] The key to this invention lies in protecting the ability to perform quantitative stability analysis by integrating furnace information data to calculate the combustion state index and stability results when detecting the combustion process state of the furnace flame using an unsupervised learning method.

[0104] The beneficial effects of this invention are:

[0105] To ensure the stable operation of thermal power units, accurate detection and evaluation of the flame combustion status within the unit are necessary. This invention extracts flame feature information from video data of furnace flame combustion and uses this information to train a DANet-SCAE model, thereby achieving accurate detection and evaluation capabilities. This is of great significance for realizing the economical, efficient, and safe operation of thermal power generating units.

[0106] The experiment divided 10,800 processed flame images into training and test sets in an 8:2 ratio. Adaptive moment estimation (Adam) was used for training, with MSE as the loss function. The loss function value quickly ceased its significant decreasing trend with increasing iterations, and the loss function at each layer eventually converged, demonstrating the model's reliability.

[0107] Comparing the DANet-SCAE model with other autoencoders in terms of MSE, peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) using test data, it can be concluded that the DANet-SCAE feature extraction network exhibits superior performance, better noise reduction capability, and robustness, making it more suitable for extracting potential features of furnace flames. Analysis of the combustion stability results during the data acquisition experiment, combined with the furnace air-coal ratio and primary / secondary air ratio, demonstrates that combustion state indicators and combustion stability results can reflect the combustion state of the flame within the furnace.

[0108] like Figure 3 As shown, this embodiment provides a quantitative detection system for furnace flame combustion status, comprising:

[0109] The data acquisition module 201 is used to acquire image data and process parameters of the furnace flame within a set time period.

[0110] The preprocessing module 202 is used to preprocess each of the image data.

[0111] The feature extraction module 203 is used to input the preprocessed image data into the image latent feature extraction model to obtain the image latent feature matrix; the image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method, and the DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE.

[0112] The combustion state index determination module 204 is used to obtain the combustion state index of each image data by performing dimensionality reduction processing on the potential feature matrix of each image.

[0113] The feature parameter determination module 205 is used to filter feature parameters from the process parameters corresponding to the image data based on the Pearson correlation coefficient, and to select feature parameters whose correlation coefficient with the combustion state index is greater than a set value.

[0114] The flame state index determination module 206 is used to linearly weight and combine the combustion state index of each image data with the corresponding feature parameters to obtain the flame state index of each image data.

[0115] The combustion state determination module 207 is used to cluster the flame state indicators of each image data and classify each image data into different flame combustion states based on the clustering results.

[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0117] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for quantitatively detecting the combustion state of a furnace flame, characterized in that, include: Collect image data and process parameters of the furnace flame within a set time period; The image data described above are preprocessed; The preprocessed image data is input into the image latent feature extraction model to obtain the image latent feature matrix; the image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method, and the DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE. The combustion state index of each image data is obtained by dimensionality reduction of the latent feature matrix of each image. Based on the Pearson correlation coefficient, feature parameters with a correlation coefficient greater than a set value with the combustion state index are selected from the process parameters corresponding to the image data. The flame state index of each image data is obtained by linearly weighting the combustion state index and the corresponding feature parameters. Cluster the flame state indicators of each image data, and classify each image data into different flame combustion states based on the clustering results; The DANet-SCAE model includes a first encoder, DANet, a second encoder, a third encoder, a fourth encoder, a flattening layer, a fifth encoder, a fifth decoder, a dimension reconstruction layer, a fourth decoder, a third decoder, a second decoder, and a first decoder, all connected in sequence. The first encoder, DANet, second encoder, second decoder, and first decoder constitute the first convolutional autoencoder; The third encoder, the fourth encoder, the fourth decoder, and the third decoder constitute the second convolutional autoencoder; The flattening layer, the fifth encoder, the fifth decoder, and the dimension reconstruction layer constitute the autoencoder; The training process of the DANet-SCAE model includes: Based on the flame training set, the first convolutional autoencoder is trained with the learning objective of minimizing the error between the sample flame image input to the first convolutional autoencoder and the reconstructed flame image output by the first convolutional autoencoder. The second convolutional autoencoder is trained with the learning objective of minimizing the error between the input first-layer latent feature vector of the second convolutional autoencoder and the reconstructed first-layer latent feature vector of the output second convolutional autoencoder. The autoencoder is trained with the learning objective of minimizing the error between the second-layer latent feature vector of the input autoencoder and the reconstructed second-layer latent feature vector of the output autoencoder.

2. The method for quantitatively detecting the combustion state of a furnace flame according to claim 1, characterized in that, The process parameters include burner flame detection signal, load command, actual load, main steam pressure, main steam temperature, main steam flow rate, feedwater flow rate, and total fuel quantity.

3. The method for quantitatively detecting the combustion state of a furnace flame according to claim 1, characterized in that, Preprocessing of the image data includes: Median filtering is performed on each image data to convert them into images with a resolution of 960×576. The resolution of a 960×576 image was compressed to 128×128 using a bilinear method. The normalization operation is used to convert the pixel values ​​in the image with a resolution of 128×128 from 0 to 255 to 0 to 1, thus obtaining the preprocessed image data.

4. The method for quantitatively detecting the combustion state of a furnace flame according to claim 1, characterized in that, By performing dimensionality reduction on the latent feature matrices of each image, the combustion state indices for each image data are obtained, specifically including: Principal component analysis is used to reduce the dimensionality of the latent feature matrix of the image, resulting in a dimension-reduced matrix. The matrix after dimensionality reduction is smoothed using a single exponential smoothing method. The smoothed matrix is ​​normalized to obtain the combustion state index.

5. The method for quantitatively detecting the combustion state of a furnace flame according to claim 1, characterized in that, The set value is 0.

8.

6. The method for quantitatively detecting the combustion state of a furnace flame according to claim 1, characterized in that, Different flame combustion states include stable combustion state and unstable combustion state.

7. A quantitative detection system for furnace flame combustion state, characterized in that, include: The data acquisition module is used to collect image data and process parameters of the furnace flame within a set time period; The preprocessing module is used to preprocess the image data. The feature extraction module is used to input the preprocessed image data into the image latent feature extraction model to obtain the image latent feature matrix. The image latent feature extraction model is a DANet-SCAE model trained by an unsupervised learning method. The DANet-SCAE model is the model after the first convolutional pooling of the convolutional autoencoder with DANet added to the first layer of SCAE. The combustion state index determination module is used to obtain the combustion state index of each image data by performing dimensionality reduction processing on the potential feature matrix of each image. The feature parameter determination module is used to filter feature parameters from the process parameters corresponding to the image data based on the Pearson correlation coefficient, and to select feature parameters whose correlation coefficient with the combustion state index is greater than a set value. The flame state index determination module is used to linearly weight and combine the combustion state index of each image data with the corresponding feature parameters to obtain the flame state index of each image data. The combustion state determination module is used to cluster the flame state indicators of each image data and classify each image data into different flame combustion states based on the clustering results. The DANet-SCAE model includes a first encoder, DANet, a second encoder, a third encoder, a fourth encoder, a flattening layer, a fifth encoder, a fifth decoder, a dimension reconstruction layer, a fourth decoder, a third decoder, a second decoder, and a first decoder, all connected in sequence. The first encoder, DANet, second encoder, second decoder, and first decoder constitute the first convolutional autoencoder; The third encoder, the fourth encoder, the fourth decoder, and the third decoder constitute the second convolutional autoencoder; The flattening layer, the fifth encoder, the fifth decoder, and the dimension reconstruction layer constitute the autoencoder; The training process of the DANet-SCAE model includes: Based on the flame training set, the first convolutional autoencoder is trained with the learning objective of minimizing the error between the sample flame image input to the first convolutional autoencoder and the reconstructed flame image output by the first convolutional autoencoder. The second convolutional autoencoder is trained with the learning objective of minimizing the error between the input first-layer latent feature vector of the second convolutional autoencoder and the reconstructed first-layer latent feature vector of the output second convolutional autoencoder. The autoencoder is trained with the learning objective of minimizing the error between the second-layer latent feature vector of the input autoencoder and the reconstructed second-layer latent feature vector of the output autoencoder.

Images