A combustion stability diagnosis method based on multi-source information fusion

Abstract

A combustion stability diagnosis method based on multi-source information fusion, an image acquisition device captures the furnace combustion flame, forms a digital image and transmits it to the server; the server receives the flame image and records the monitoring information of different sensors at the same time; the deep features of the flame image are extracted by establishing the adversarial denoising auto-encoding, and the delay time between the deep image features and the sensor monitoring information is corrected by designing the correlation coefficient; then, the deep image features and the sensor monitoring information are fused and sent to the combustion stability evaluation index to quantitatively estimate the combustion stability state. The combustion stability diagnosis method can quantitatively characterize the combustion stability and provide an important guidance basis for combustion state adjustment.

Description

Technical Field

[0001] This invention relates to the field of combustion monitoring technology, and in particular to a combustion stability diagnosis method based on multi-source information fusion. Background Technology

[0002] Combustion is a crucial method of utilizing fossil fuels, and establishing and maintaining a stable combustion state is an important operational objective, beneficial for improving combustion efficiency and reducing pollutant emissions. However, influenced by factors such as fuel quality and combustion methods, the combustion state is difficult to maintain in the long term, posing a threat to the safe and environmentally friendly operation of the entire combustion system. Therefore, timely and accurate monitoring of combustion stability is essential for effectively preventing abnormal combustion states and avoiding safety accidents.

[0003] Based on the light and heat energy released during combustion, combustion diagnostic technologies such as thermocouples and flame fault detectors have been developed. However, these technologies have significant limitations in practical applications: (1) large errors and susceptibility to interference from the monitoring environment; (2) poor functionality, providing only local fire extinguishing signals; and (3) frequent maintenance and short equipment lifespan. In contrast, visual combustion diagnostic technologies are receiving increasing attention due to their advantages such as high sensitivity, anti-interference, and long lifespan. However, traditional visualization technologies such as flame television only provide simple monitoring functions, and the interpretation of image data relies heavily on the operator's subjective experience, lacking objective evaluation standards. Thanks to advancements in digital image processing technology, traditional monitoring functions have been expanded, enabling online analysis of flame images.

[0004] While flame images contain rich information about combustion status, accurate analysis of pixel information presents challenges. Looking at various data-driven methods, combustion stability diagnosis inevitably involves two key steps: feature extraction and feature analysis. Feature extraction aims to uncover the essential information of the image, eliminate non-critical repetitive content, and reduce image dimensionality. Every flame image possesses unique features that distinguish it from other images, such as hue and saturation. However, existing feature extraction methods are often limited by low computational efficiency, poor robustness, and weak generalization ability. Feature analysis, on the other hand, outputs deterministic diagnostic results through image feature analysis. The difficulty in interpreting flame image data, particularly the uninterpretability of its deep features, further complicates feature analysis. Furthermore, the numerous and difficult-to-characterize potential inducing factors of combustion instability make accurate diagnosis of combustion status extremely challenging.

[0005] In summary, feature extraction and feature analysis methods still require further improvement. To address the shortcomings of existing technologies, this invention will develop a combustion stability diagnostic system integrating key technologies such as image acquisition, image feature extraction, and image feature analysis. This system will provide important theoretical basis and technical foundation for optimizing combustion system operation, thereby assisting in the safe and environmentally friendly operation of the entire combustion system. Furthermore, this application has significant research value for studying combustion mechanisms and exploring the dynamic changes in combustion. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by providing a combustion stability diagnosis method based on multi-source information fusion, thereby solving the problems of poor interpretability of deep features in flame images and difficulty in quantifying combustion stability.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A combustion stability diagnostic method based on multi-source information fusion includes the following steps:

[0009] S1: Based on image acquisition equipment, capture flame images under different combustion conditions;

[0010] S2: Preprocessed flame image;

[0011] S3: Establish a feature extraction network, namely an adversarial denoising autoencoder, and train it using preprocessed flame images to improve the accuracy of the adversarial denoising autoencoder in recognizing deep features in flame images.

[0012] S4: For newly acquired flame images that need to be analyzed, after preprocessing, deep features of the flame image are extracted based on the trained adversarial noise reduction autoencoder; at the same time, data monitoring information in the furnace chamber is collected based on sensor equipment under the combustion conditions corresponding to the flame image.

[0013] Design a correlation coefficient to correct the time delay between deep features in flame images and data detected by sensor devices;

[0014] S5: The deep features of the flame image are fused with the sensor monitoring information to form image sensing features. Then, the structural similarity index is used to perform cluster analysis on the image sensing features to obtain two independent clusters.

[0015] S6: Calculate the information entropy of the two clusters respectively, and define the cluster with information entropy less than the set threshold as a stable cluster and the cluster with information entropy greater than the threshold as an unstable cluster;

[0016] S7: Calculate the centroids of the stable clusters and the unstable clusters respectively, and establish combustion stability evaluation index based on this;

[0017] S8: Analyze the image sensing features using combustion stability evaluation indicators to quantitatively estimate the combustion stability of a single flame image.

[0018] To optimize the above technical solution, the specific measures also include:

[0019] Furthermore,

[0020] In step S1, the image acquisition device includes an optical sight and a CCD camera;

[0021] In step S2, the preprocessing includes compressing and normalizing the flame image.

[0022] Furthermore, the specific content of step S3, which involves training using the preprocessed flame image to improve the accuracy of the adversarial denoising autoencoder in recognizing deep features in the flame image, is as follows:

[0023] S3.1: The input flame image x is processed by noise reduction encoding n to form a noisy image x. n Its data dimensions are O×P×3;

[0024] S3.2: Noisy image x n The features are sent to the convolutional encoder e1 for processing, which includes: first, feature extraction is performed by convolutional filters C1 with Q1 c×c windows and q strides; then, feature activation is performed by the ReLU activation function; finally, feature dimensionality reduction is performed by max pooling layer P1 with o×o windows and l strides, and the output feature vector h1 has dimensions a1×b1×c1.

[0025] S3.3: The feature vector h1 is processed sequentially by the convolutional encoders e2, e3, e4, and e5. The data processing process is similar to step S3.2, and the final output is a deep feature vector h5 with dimensions a5×b5×c5.

[0026] S3.4: The deep feature vector h5 is sent to the convolutional decoder d1 for processing, which includes: first, the feature dimension is increased by the upsampling layer U1 with a window size of g×g, and then it is processed by the convolutional filter C6 and the ReLU activation function to obtain the feature vector h6;

[0027] S3.5: The feature vector h6 is processed sequentially by convolutional decoders d2, d3, d4, and d5. The data processing procedure is similar to step S3.4, except that the activation function of convolutional decoder d5 is different; it uses the Sigmoid function. The final output is a reconstructed image x with dimensions O×P×3. rec ;

[0028] S3.6: Input flame image x is converted to grayscale After processing, a grayscale image x is formed. gIts data dimension is O×P×1;

[0029] S3.7: For grayscale image x g Analysis yielded nine static characteristics L: average gray level L1, flame brightness L2, effective area L3, non-uniformity L4, gray level variance L5, gray level standard deviation L6, flicker frequency L7, skewness L8, and kurtosis L9.

[0030] S3.8: The static features L and deep feature vector h5 of the flame image are sent to the discriminator, which is processed by the unfolding operation F, the fully connected operation f1, and the fully connected operation f2 in sequence. The discriminator determines whether the input data is a static feature or a deep feature. Then, based on the static feature as the standard, it guides the identified deep feature data to gradually tend towards the static feature. Through multiple training sessions with flame images, the adversarial denoising autoencoder can more accurately identify deep features in flame images.

[0031] Furthermore, in step S3, when training using the preprocessed flame image, a denoising coding technique is introduced to improve the robustness of feature extraction. The denoising coding process is represented as follows: In the formula Indicates the noise ratio. Let x be a random variable that follows a normal distribution. n "x" refers to the noise image, and "x" refers to the flame image.

[0032] Furthermore, the sensor devices mentioned in step S4 include a flame detector sensor, a thermocouple sensor, a pressure sensor, and a gas analysis sensor; the collected data information corresponds to flame detection intensity, furnace temperature, furnace pressure, NOx concentration, and CO2 concentration.

[0033] Furthermore, the specific details of the delay time between the deep features of the flame image and the data information detected by the sensor device, as described in step S4, are as follows:

[0034] Given any dataset {h(i),R(j)|i=1,2,…,m,j=1,2,…,n}, where h(i) represents the deep features of the i-th flame image, R(j) represents the monitoring information of the j-th sensor, m represents the number of flame images, and n represents the number of sensors, the correlation coefficient I(h(i),R(j)) is defined as:

[0035]

[0036] In the formula, p(h(i)) represents the probability density of h(i), p(R(j)) represents the probability density of R(j), and p(h(i),R(j)) represents the joint probability density of h(i) and R(j). When the correlation coefficient I(h(i),R(j)) is greater than 0.8, it indicates that h(i) and R(j) are highly correlated. By observing the correlation coefficient between h(i) and R(j), the time delay correction of these two variables can be achieved.

[0037] Furthermore, the clustering analysis of image sensing features using structural similarity indices described in step S5, resulting in two independent clusters, is detailed below:

[0038] The structural similarity index ψ(a,b) has a numerical range of [0,1], and is specifically defined as follows:

[0039]

[0040] In the formula, a and b represent any two image sensing feature vectors; μ a and μ b σ represents the average value of the corresponding image sensing feature vectors; a and σ b σ represents the variance of the corresponding image sensing feature vector; ab θ1 and θ2 represent the covariance of two corresponding image sensing feature vectors; both θ1 and θ2 are constants less than 1.

[0041] Cluster analysis of image sensing features using structural similarity indices yielded two independent clusters, including:

[0042] S5.1: Randomly select two image sensing features as the initial cluster centroids of the two clusters;

[0043] S5.2: Calculate the structural similarity index between all image sensing features and these two centroids, and divide all image sensing features into two clusters according to the size of the structural similarity index;

[0044] S5.3: Redetermine the centroids of the two clusters;

[0045] S5.4: Repeat steps S5.2 and S5.3 until the centroid no longer changes or the limit number of iterations is reached, thereby completing the clustering of image sensing features and obtaining two independent clusters.

[0046] Furthermore, the specific content of calculating the information entropy ψ(ξ) in step S6 is as follows:

[0047]

[0048] In the formula, ξ(t) represents the t-th image sensing feature; p(ξ(t)) represents the probability that other data in the cluster are the same as ξ(t); and T represents the total number of image sensing features.

[0049] Further, in step S7

[0050] The centroid refers to the average of all image sensing features in the same cluster across all dimensions.

[0051] The established combustion stability evaluation index δ(t) reflects the degree of combustion stability of each image sensing feature. The specific calculation process is as follows:

[0052]

[0053] In the formula, δ(t) represents the degree of combustion stability of the t-th image sensing feature; ξ κ (t) represents the numerical value of the image sensing feature κ at the location; This represents the numerical value of the stable cluster center κ. This represents the numerical value of the center κ of the unstable cluster.

[0054] Furthermore, in step S8, the specific content of analyzing the image sensing features using combustion stability evaluation indicators and quantitatively estimating the combustion stability of a single flame image is as follows:

[0055] The numerical range of δ(t) is 0 to 100%, where 0 represents complete stability, 100% represents extreme instability, and 50% is the boundary between stable and unstable states.

[0056] The beneficial effects of this invention are:

[0057] 1. Develop a coupling mechanism that can interpret static features and dimensionless deep features to effectively improve the training performance of feature extraction networks and improve the accuracy of deep feature recognition.

[0058] 2. To address the issue of asynchronous flame images and sensor monitoring information, a delay correction method based on correlation coefficients was developed. Simultaneously, a structural similarity index was designed for image sensing feature clustering, and information entropy was used to determine stable and unstable clusters. A combustion stability evaluation index was established to overcome the difficulty in quantifying combustion stability. Attached Figure Description

[0059] Figure 1 This is a schematic diagram of the overall technical solution of the present invention.

[0060] Figure 2 This is a schematic diagram of the combustion stability diagnostic device of the present invention.

[0061] Figure 3This is a schematic diagram of the anti-noise reduction self-encoding structure of the present invention.

[0062] Figure 4 This is a schematic diagram of five heavy oil combustion conditions according to the present invention.

[0063] Figure 5 This is a schematic diagram of the image sensing feature clustering results of the present invention.

[0064] Figure 6 This is a schematic diagram of the quantitative diagnosis of combustion stability according to the present invention. Detailed Implementation

[0065] The invention will now be described in further detail with reference to the accompanying drawings.

[0066] The technical solution of this application is as follows:

[0067] 1. A combustion stability diagnostic technology route based on multi-source information fusion, such as... Figure 1 As shown.

[0068] 2. A combustion stability diagnostic device based on multi-source information fusion has the following structure: Figure 2 As shown, it consists of image acquisition equipment (optical sight glass, CCD camera), flame detection sensor, thermocouple sensor, pressure sensor, gas analysis sensor, and server. The optical sight glass employs both air-cooling and water-cooling for temperature protection, and its front end features a corner design for easy installation.

[0069] 3. A combustion stability diagnosis method based on multi-source information fusion, the implementation steps of which are as follows: Figure 1 As shown, the details are as follows:

[0070] Step 1: The server sends instructions to the image acquisition device to capture flame images under different combustion conditions;

[0071] Step 2: Preprocess the flame image, including size compression and normalization;

[0072] Step 3: Establish a feature extraction network, namely an adversarial noise reduction autoencoder, and perform unsupervised training using preprocessed images;

[0073] Step 4: For the newly acquired flame images, after preprocessing, deep image features are extracted using a trained adversarial noise reduction autoencoder. Simultaneously, monitoring information from flame detectors, thermocouples, pressure sensors, and gas analysis sensors is used. A correlation coefficient is designed to correct the time delay between the deep image features and sensor monitoring information such as flame intensity, furnace temperature, furnace pressure, NOx concentration, and CO2 concentration.

[0074] Step 5: Fuse deep image features with sensor monitoring information to form image sensing features, and then use structural similarity index to perform cluster analysis on the image sensing features to obtain two independent clusters;

[0075] Step 6: Calculate the information entropy of the two clusters respectively, and define the cluster with the smaller information entropy as a stable cluster and the cluster with the larger information entropy as an unstable cluster;

[0076] Step 7: Calculate the centroids of the stable clusters and the unstable clusters respectively, and establish combustion stability evaluation indexes based on this.

[0077] Step 8: Analyze the image sensing features using combustion stability evaluation indicators to quantitatively estimate the combustion stability of a single flame image.

[0078] Furthermore, interpretable static features and dimensionless deep features are coupled. The structure of the adversarial denoising autoencoder is as follows: Figure 3 As shown, its hyperparameters are summarized in Table 2, and the specific implementation steps are as follows:

[0079] Step 3.1: After the input image x is processed by noise reduction encoding n, a noisy image x is formed. n Its data dimensions are O×P×3;

[0080] Step 3.2, Noisy image x n The data is sent to the convolutional encoder e1 for processing. First, features are extracted by convolutional filters C1 (Q1@c×c+q) with a window size of Q1 and a stride of q. Then, features are activated by the ReLU activation function [y(λ)=max(0,λ), where λ represents the input neuron]. Finally, feature dimensionality reduction is performed by a max pooling layer P1 (o×o+l) with a window size of o×o and a stride of l. The output feature vector h1 has a dimension of a1×b1×c1.

[0081] Step 3.3: The feature vector h1 is processed sequentially by convolutional encoders e2, e3, e4, and e5. The data processing process is similar to that in step 2.2, and the final output is a deep feature vector h5 with dimensions a5×b5×c5.

[0082] Step 3.4: The deep feature vector h5 is sent to the convolutional decoder d1 for processing. First, the feature is upsampled by the upsampling layer U1 (g×g) with a window size of g×g. Then, it is processed by the convolutional filter C6 (Q6@c×c+q) and the ReLU activation function to obtain the feature vector h6.

[0083] Step 3.5: The feature vector h6 is processed sequentially by convolutional decoders d2, d3, d4, and d5. The data processing procedure is similar to that in step 2.4, except that the activation function of convolutional decoder d5 is different, namely, the Sigmoid function [y(λ)=1 / 1+exp(-λ), where λ represents the input neuron] is used. The final output is a reconstructed image x with dimensions O×P×3. rec ;

[0084] Step 3.6: Input image x is converted to grayscale. After processing, a grayscale image x is formed. g Its data dimension is O×P×1;

[0085] Step 3.7, grayscale image x g After processing by the empirical formulas shown in Table 1, nine static characteristics are obtained, including average gray level L1, flame brightness L2, effective area L3, non-uniformity L4, gray level variance L5, gray level standard deviation L6, flicker frequency L7, skewness L8, and kurtosis L9.

[0086] Step 3.8: The static feature L and deep feature h of the flame image are sent to the discriminator, which is processed by the unfolding operation F, the fully connected operation f1 and the fully connected operation f2 in sequence to determine whether the input data is static feature L (true) or deep feature h (false).

[0087] Furthermore, a denoising coding technique is introduced during unsupervised training to improve feature robustness. The denoising coding process is represented as follows: in Indicates the noise ratio. This represents a random variable that follows a normal distribution.

[0088] Furthermore, the correlation coefficient can be used to calculate the degree of correlation between deep image features and sensor monitoring information. Given any dataset {h(i),R(j)|i=1,2,…,m,j=1,2,…,n}, where h(i) represents the deep features of the i-th flame image, R(j) represents the monitoring information of the j-th sensor, m represents the number of flame images, and n represents the number of sensors, the correlation coefficient I(h(i),R(j)) is defined as:

[0089]

[0090] In the formula, p(h(i)) represents the probability density of h(i), p(R(j)) represents the probability density of R(j), and p(h(i),R(j)) represents the joint probability density of h(i) and R(j). When the correlation coefficient I(h(i),R(j)) is greater than 0.8, it indicates that h(i) and R(j) are highly correlated. By observing the correlation coefficient between h(i) and R(j), the time delay correction of these two variables is achieved.

[0091] Furthermore, the aforementioned structural similarity index can cluster image sensing features into two clusters in an unsupervised manner. The structural similarity index ψ(a,b) has a numerical range of [0,1], and is specifically defined as follows:

[0092]

[0093] In the formula, a and b represent any two image sensing feature vectors; μ a and μ b σ represents the average value of the eigenvectors. a and σ b σ represents the variance of the eigenvectors. ab θ1 and θ2 represent the covariance of the eigenvectors; θ1 and θ2 are constants less than 1.

[0094] Furthermore, the clustering analysis is implemented in the following steps: ① Randomly select two image sensing features as the initial cluster centroids; ② Calculate the structural similarity index between all image sensing features and these two centroids, and divide all image sensing features into two clusters according to the size of the structural similarity index; ③ Recalculate the centroids of the two clusters; ④ Repeat steps ② and ③ until the centroids no longer change or the limit number of iterations is reached.

[0095] Furthermore, the information entropy, which can quantify the complexity of data within a cluster, is specifically defined as:

[0096]

[0097] In the formula, ξ(t) represents the t-th image sensing feature, and p(ξ(t)) represents the probability that other data in the cluster are the same as ξ(t).

[0098] Furthermore, the centroid refers to the average of all image sensing features in the same cluster across all dimensions.

[0099] Furthermore, the aforementioned combustion stability evaluation index can quantitatively estimate combustion stability, and is specifically defined as follows:

[0100]

[0101] In the formula, δ(t) represents the degree of combustion stability of the t-th image sensing feature; ξ k (t) represents the numerical value of the image sensing feature κ at the location; This represents the numerical value of the stable cluster center κ. The value represents the position of the unstable cluster center κ. The value of δ(t) ranges from 0 to 100%, where 0 represents complete stability, 100% represents extreme instability, and 50% is the boundary between stable and unstable states.

[0102] The present invention will now be described in detail with reference to the accompanying drawings and tables:

[0103] 1. Build a heavy oil combustion test platform and install a combustion stability diagnostic device.

[0104] 2. Five combustion conditions were established, and the implementation steps for the combustion stability diagnosis method are as follows:

[0105] Step 1: Acquire 200 flame images. Example images are shown below. Figure 4 As shown, and record sensor monitoring information;

[0106] Step 2: Preprocess the flame image, including size compression and normalization;

[0107] Step 3: Establish a feature extraction network, namely an adversarial noise reduction autoencoder, and perform unsupervised training using preprocessed images;

[0108] Step 4: Acquire new flame images and sensor information. After preprocessing, use the correlation coefficient of formula (1) to correct the delay time of deep image features and sensor monitoring information, and fuse the deep image features and sensor monitoring information to form image sensing features.

[0109] Step 5: Use the structural similarity index of equation (2) to perform cluster analysis on the image sensing features to obtain two independent clusters, and use the information entropy of equation (3) to determine the stability category of the two clusters. The results are as follows. Figure 5 As shown;

[0110] Step 6: Calculate the centroids of the two clusters and establish the combustion stability evaluation index as shown in equation (4);

[0111] Step 7: Repeat the acquisition of new flame images and sensor monitoring information, then repeat step 4 to obtain image sensing features, and finally use the combustion stability evaluation index in step 6 to identify the image sensing features and quantitatively output the combustion stability diagnosis results.

[0112] 3. Figure 6 The combustion stability diagnostic results for five combustion conditions are presented, which are 36%, 73%, 89%, 14%, and 11%, respectively. It can be seen that the combustion stability of condition 3 is the worst, while the combustion stability of condition 5 is the best.

[0113] The following needs further explanation:

[0114] Table 1

[0115]

[0116] Table 2

[0117]

[0118] It should be noted that the terms such as "upper", "lower", "left", "right", "front", and "back" used in the invention are only for clarity of description and are not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention.

[0119] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A combustion stability diagnostic method based on multi-source information fusion, characterized in that, Includes the following steps: S1: Based on image acquisition equipment, capture flame images under different combustion conditions; S2: Preprocessed flame image; S3: Establish a feature extraction network, namely an adversarial denoising autoencoder, and train it using preprocessed flame images to improve the accuracy of the adversarial denoising autoencoder in recognizing deep features in flame images. S4: For newly acquired flame images that need to be analyzed, after preprocessing, deep features of the flame image are extracted based on the trained adversarial noise reduction autoencoder; at the same time, data monitoring information in the furnace chamber is collected based on sensor equipment under the combustion conditions corresponding to the flame image. Design a correlation coefficient to correct the time delay between deep features in flame images and data detected by sensor devices; The specific details of the design correlation coefficient correction for the delay time between the deep features of the flame image and the data information detected by the sensor device in step S4 are as follows: Given any dataset {h(i),R(j)|i=1,2,…,m,j=1,2,…,n}, where h(i) represents the deep features of the i-th flame image, R(j) represents the monitoring information of the j-th sensor, m represents the number of flame images, and n represents the number of sensors, the correlation coefficient I(h(i),R(j)) is defined as: In the formula, p(h(i)) represents the probability density of h(i), p(R(j)) represents the probability density of R(j), and p(h(i),R(j)) represents the joint probability density of h(i) and R(j). When the correlation coefficient I(h(i),R(j)) is greater than 0.8, it indicates that h(i) and R(j) are highly correlated. By observing the correlation coefficient between h(i) and R(j), the time delay correction of these two variables can be achieved. S5: The deep features of the flame image are fused with the sensor monitoring information to form image sensing features. Then, the structural similarity index is used to perform cluster analysis on the image sensing features to obtain two independent clusters. S6: Calculate the information entropy of the two clusters respectively, and define the cluster with information entropy less than the set threshold as a stable cluster and the cluster with information entropy greater than the threshold as an unstable cluster; S7: Calculate the centroids of the stable clusters and the unstable clusters respectively, and establish combustion stability evaluation index based on this; S8: Analyze the image sensing features using combustion stability evaluation indicators to quantitatively estimate the combustion stability of a single flame image.

2. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, In step S1, the image acquisition device includes an optical sight and a CCD camera; In step S2, the preprocessing includes size compression and normalization of the flame image.

3. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, The specific content of step S3, which involves training using the preprocessed flame image to improve the accuracy of the adversarial denoising autoencoder in recognizing deep features in the flame image, is as follows: S3.1: Input flame image x after noise reduction encoding After processing, a noisy image x is formed. n Its data dimensions are O×P×3; S3.2: Noisy image x n The features are sent to the convolutional encoder e1 for processing, which includes: first, feature extraction is performed by convolutional filters C1 with Q1 c×c windows and q strides; then, feature activation is performed by the ReLU activation function; finally, feature dimensionality reduction is performed by max pooling layer P1 with o×o windows and l strides, and the output feature vector h1 has dimensions a1×b1×c1. S3.3: The feature vector h1 is processed sequentially by the convolutional encoders e2, e3, e4, and e5. The data processing process is similar to step S3.2, and the final output is a deep feature vector h5 with dimensions a5×b5×c5. S3.4: Input flame image x converted to grayscale After processing, a grayscale image x is formed. g Its data dimension is O×P×1; S3.5: For grayscale image x g Analysis yielded nine static characteristics L: average gray level L1, flame brightness L2, effective area L3, non-uniformity L4, gray level variance L5, gray level standard deviation L6, flicker frequency L7, skewness L8, and kurtosis L9. S3.6: The static features L and deep feature vector h5 of the flame image are sent to the discriminator, which is processed by the unfolding operation F, the fully connected operation f1, and the fully connected operation f2 in sequence. The discriminator determines whether the input data is a static feature or a deep feature. Then, based on the static feature as the standard, it guides the identified deep feature data to gradually tend towards the static feature. Through multiple training sessions with flame images, the adversarial denoising autoencoder can more accurately identify deep features in flame images.

4. The combustion stability diagnosis method based on multi-source information fusion according to claim 3, characterized in that, In step S3, when training using the preprocessed flame image, a denoising coding technique is introduced to improve the robustness of feature extraction. The denoising coding process is represented as follows: In the formula Indicates the noise ratio. Let x be a random variable that follows a normal distribution. n "x" refers to the noise image, and "x" refers to the flame image.

5. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, The sensor devices mentioned in step S4 include a flame detector sensor, a thermocouple sensor, a pressure sensor, and a gas analysis sensor; the collected data information corresponds to flame detection intensity, furnace temperature, furnace pressure, NOx concentration, and CO2 concentration.

6. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, Step S5 describes using structural similarity indices to perform cluster analysis on image sensing features, resulting in two independent clusters, the details of which are as follows: The structural similarity index ψ(a,b) has a numerical range of [0,1], and is specifically defined as follows: In the formula, a and b represent any two image sensing feature vectors; μ a and μ b σ represents the average value of the corresponding image sensing feature vectors; a and σ b σ represents the variance of the corresponding image sensing feature vector; ab θ1 and θ2 represent the covariance of two corresponding image sensing feature vectors; both θ1 and θ2 are constants less than 1. Cluster analysis of image sensing features using structural similarity indices yielded two independent clusters, including: S5.1: Randomly select two image sensing features as the initial cluster centroids of the two clusters; S5.2: Calculate the structural similarity index between all image sensing features and these two centroids, and divide all image sensing features into two clusters according to the size of the structural similarity index; S5.3: Redetermine the centroids of the two clusters; S5.4: Repeat steps S5.2 and S5.3 until the centroid no longer changes or the limit number of iterations is reached, thereby completing the clustering of image sensing features and obtaining two independent clusters.

7. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, The specific content of calculating the information entropy ψ(ξ) in step S6 is as follows: In the formula, ξ(t) represents the t-th image sensing feature; p(ξ(t)) represents the probability that other data in the cluster are the same as ξ(t); and T represents the total number of image sensing features.

8. The combustion stability diagnosis method based on multi-source information fusion according to claim 1, characterized in that, In step S7 The centroid refers to the average of all image sensing features in the same cluster across all dimensions. The established combustion stability evaluation index δ(t) reflects the degree of combustion stability of each image sensing feature. The specific calculation process is as follows: In the formula, δ(t) represents the degree of combustion stability of the t-th image sensing feature; ξ κ (t) represents the numerical value of the image sensing feature κ at the location; This represents the numerical value of the stable cluster center κ. This represents the numerical value of the center κ of the unstable cluster.

9. The combustion stability diagnosis method based on multi-source information fusion according to claim 8, characterized in that, In step S8, the specific content of analyzing the image sensing features using combustion stability evaluation indicators and quantitatively estimating the combustion stability of a single flame image is as follows: The numerical range of δ(t) is 0 to 100%, where 0 represents complete stability, 100% represents extreme instability, and 50% is the boundary between stable and unstable states.

Images