Flame combustion state identification method and system based on image multi-threshold segmentation

Abstract

The invention discloses a flame combustion state recognition method and system based on image multi-threshold segmentation, and belongs to the technical field of image processing. Comprising the following steps: acquiring a flame image, taking an inter-class variance as a fitness function, and performing multi-threshold segmentation on the flame image by using a hybrid optimization strategy combining a genetic algorithm and a simulation degradation algorithm to generate a region segmentation image; performing feature extraction on the region segmentation image to obtain flame shape features and flame statistical features, and constructing a feature matrix; and processing the feature matrix through a trained flame combustion state recognition model to obtain a flame combustion state. Accurate identification and classification of flame combustion states can be realized through efficient image processing and deep learning methods; the problem that the generalization ability and the classification accuracy of flame combustion state recognition need to be improved in the prior art is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a flame combustion state recognition method and system based on image multi-threshold segmentation. Background Art

[0002] The statements in this section merely mention background art related to the present invention and do not necessarily constitute prior art.

[0003] The demand for waste disposal poses a severe challenge to public health, resource utilization and ecological balance. The content of organic matter in urban domestic waste is very high, and waste resources can be recycled through incineration. Through incineration and downstream energy conversion systems, waste incineration (MSWI) plants can generate various forms of energy, including heat, electricity and steam. Traditional MSWI plants rely on manual operation and monitoring, which are inefficient and difficult to achieve precise control. In order to improve efficiency and environmental performance, MSWI plants have gradually introduced automatic control systems. However, the reliability and effectiveness of traditional automatic control systems in monitoring the combustion process need to be improved, resulting in many deficiencies in handling complex combustion states, such as low efficiency and difficulty in achieving precise control.

[0004] In recent years, spectroscopy and image analysis techniques have provided rich diagnostic data for monitoring combustion processes. Multiple studies have explored different feature recognition and extraction methods, including color feature extraction, shape feature extraction, and statistical feature extraction. However, the reliability of these methods in complex situations such as illumination changes remains to be verified. In addition, commonly used support vector machines and deep learning methods have problems such as long computation time and sensitive parameter selection when processing large-scale data. Semi-supervised learning methods are effective when processing limited labeled data, but how to use unlabeled data to avoid model overfitting remains a major challenge. Summary of the invention

[0005] In order to address the deficiencies in the prior art, the present invention provides a flame combustion state recognition method, system, electronic device, computer-readable storage medium and computer program product based on image multi-threshold segmentation. The multi-threshold segmentation and comprehensive feature extraction method of the GA-SA hybrid algorithm significantly improves the recognition accuracy of the flame combustion state.

[0006] In a first aspect, the present invention provides a method for identifying flame combustion state based on image multi-threshold segmentation;

[0007] A flame combustion state recognition method based on image multi-threshold segmentation, comprising:

[0008] Acquire a flame image, use the inter-class variance as a fitness function, and use a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to perform multi-threshold segmentation on the flame image to generate a regional segmentation image;

[0009] Extracting features from the region segmented image, obtaining flame shape features and flame statistical features, and constructing a feature matrix;

[0010] The characteristic matrix is ​​processed by a trained flame combustion state recognition model to obtain the flame combustion state.

[0011] In some embodiments, the method of using the inter-class variance as a fitness function and utilizing a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to perform multi-threshold segmentation on the flame image comprises:

[0012] With the goal of maximizing the inter-class variance, a genetic algorithm is used to perform a global search on the flame image, and the population is optimized through selection, crossover and mutation operations; for the new individuals generated by the genetic algorithm, a simulated degradation algorithm is used to perform local optimization until the best segmentation threshold combination is obtained;

[0013] Dividing the flame image into regions according to the optimal segmentation threshold combination to generate a region segmentation image;

[0014] The region segmentation image includes a high temperature combustion region, an effective combustion region, an incomplete combustion region and a background region.

[0015] In some embodiments, feature extraction is performed on the region segmentation image to obtain flame shape features by calculating the flame effective areas of different combustion zones based on a comparison result of the grayscale values ​​in different combustion zones in the region segmentation image with a preset flame grayscale threshold, and calculating the flame effective area ratio of different combustion zones based on the length and height of the region segmentation image.

[0016] In some implementations, feature extraction is performed on the region segmented image to obtain flame statistical features specifically as follows:

[0017] According to the grayscale values ​​of all pixels in different burning areas in the region segmentation image, the average grayscale and grayscale standard deviation of different burning areas are calculated respectively;

[0018] Based on the region segmentation image, the Shannon entropy of the flame center position and the combustion region grayscale histogram is calculated. In some embodiments, the flame combustion state recognition model is a deep Q network combined with a pseudo label generation strategy.

[0019] In some embodiments, when training the flame combustion state recognition model, a joint loss function, a value function optimization loss function and a cross entropy loss function are introduced.

[0020] In a second aspect, the present invention provides a flame combustion state recognition system based on image multi-threshold segmentation;

[0021] A flame combustion state recognition system based on image multi-threshold segmentation, comprising:

[0022] The image multi-threshold segmentation module is configured to: obtain a flame image, use the inter-class variance as a fitness function, and perform multi-threshold segmentation on the flame image using a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to generate a region segmentation image;

[0023] The feature extraction module is configured to: perform feature extraction on the region segmentation image, obtain flame shape features and flame statistical features, and construct a feature matrix;

[0024] The flame combustion state recognition module is configured to process the feature matrix through a trained flame combustion state recognition model to obtain the flame combustion state.

[0025] In a third aspect, the present invention provides an electronic device;

[0026] An electronic device comprises a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the flame combustion state recognition method based on image multi-threshold segmentation.

[0027] In a fourth aspect, the present invention provides a computer-readable storage medium;

[0028] A computer-readable storage medium stores a computer program / instruction, which, when executed by a processor, implements the steps of the flame combustion state recognition method based on image multi-threshold segmentation.

[0029] In a fifth aspect, the present invention provides a computer program product;

[0030] A computer program product includes a computer program / instruction, which implements the steps of the flame combustion state recognition method based on image multi-threshold segmentation when executed by a processor.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] 1. The technical solution provided by the present invention adopts a hybrid optimization strategy (GA-SA) combining a genetic algorithm (GA) and a simulated annealing algorithm (SA) for multi-threshold segmentation; by maximizing the inter-class variance as the fitness function, the segmentation of each combustion area is clear, and the problem that a single algorithm may fall into a local optimum is effectively solved.

[0033] 2. The technical solution provided by the present invention extracts shape features and statistical features from the image after multi-threshold segmentation, including the effective flame area, area ratio, average gray value, gray standard deviation, flame center position and Shannon entropy of the gray histogram of the combustion area of ​​each combustion area; these features can effectively describe the combustion state of the flame and provide accurate and comprehensive basic data for combustion state identification.

[0034] 3. The technical solution provided by the present invention combines L1 regularization, recursive feature elimination (RFE) and random forest methods for feature selection, and then uses principal component analysis (PCA) to further reduce the dimension of the feature matrix to 5 principal components, thereby retaining most of the information in the data set and reducing the complexity of the model.

[0035] 4. The technical solution provided by the present invention combines the deep Q network (DQN) with the pseudo-label generation strategy to establish a new hybrid algorithm model DQN-PL; this model can not only generate pseudo-labels to improve the learning efficiency and classification accuracy of the model when data labels are incomplete, but also optimize the decision-making process through reinforcement learning to improve the overall performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0037] Figure 1 A schematic flow chart of a flame combustion state recognition method based on image multi-threshold segmentation provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0038] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used in the present invention have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0039] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0040] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0041] Embodiment 1

[0042] The reliability and effectiveness of existing flame combustion state recognition need to be improved, resulting in low automatic control efficiency and difficulty in achieving precise control; therefore, the present invention provides a flame combustion state recognition method based on image multi-threshold segmentation, which realizes precise recognition and classification of flame combustion states based on image multi-threshold segmentation and DQN-PL model.

[0043] Next, combine Figure 1 , a flame combustion state recognition method based on image multi-threshold segmentation disclosed in this embodiment is described in detail. The flame combustion state recognition method based on image multi-threshold segmentation includes:

[0044] S1. Obtain flame image.

[0045] In this embodiment, the flame image is collected by high-resolution, high-temperature resistant cameras installed on the left and right sides of the rear wall of the high-temperature incinerator. The cameras can withstand the high temperature environment when the incinerator is working while maintaining clear image acquisition quality.

[0046] Before acquisition, set the camera parameters, including exposure time, gain, resolution, etc., to ensure that the acquired flame image has sufficient brightness and contrast to accurately reflect the characteristics of the flame; set a reasonable acquisition frequency, such as acquiring 5 frames of images per minute.

[0047] Furthermore, the combustion of garbage inside the furnace is divided into five combustion states, including:

[0048] (1) Normal combustion state. The flame in the combustion section is vigorous and uniform, the flame is bright, the flame height is high, and the flame area is large. This state is the normal combustion state. During the control process, the existing control parameters should be maintained at this time to keep the garbage burning in a good state.

[0049] (2) Flame biased combustion state. The flame is strong on one side of the grate, while the other side is black and flameless. It is generally left-biased or right-biased. At this time, the garbage is not fully burned and the burning condition is poor. During the control process, the grate movement speed should be reduced and the air volume in the burning section should be increased.

[0050] (3) The garbage is too thick to burn. The image shows continuous black, the flame is small, and the flame area is small. At this time, the garbage is not burned fully and the combustion state is poor. During the control process, the movement speed of the pusher and grate should be reduced and the primary air volume should be increased.

[0051] (4) Smoke and dust combustion state. Smoke and dust, the combustion flame is unclear. This is usually caused by the slag being unloaded during the movement of the grate in the burning section, which causes a large amount of garbage dust to be blown up. The raised dust obscures the camera, resulting in no obvious flame being observed in the image. In order to prevent misdiagnosis, the image recognition is suspended at this time and is not used as auxiliary combustion control information.

[0052] (5) Fire extinguishing state. No obvious flames can be observed in the image, and the grayscale of the entire image is also very low. This is the fire extinguishing state, which is a state that should be avoided as much as possible during the operation of the waste incinerator. At this time, oil should be added to assist combustion in order to restore the combustion to a good state.

[0053] S2. Preprocess the flame image and convert the preprocessed flame image into a grayscale image.

[0054] Specifically include:

[0055] S201, performing defogging processing on the flame image.

[0056] Since the environment inside a high-temperature incinerator may contain a large amount of smoke and steam, the quality of the collected images will be reduced; dark channel prior dehazing processing can improve image quality by enhancing the contrast and clarity of the image.

[0057] Exemplarily, the data processing flow of S201 is specifically as follows:

[0058] I(x)=J(x)+A×t(x);

[0059]

[0060] Where I(x) represents the flame image, J(x) represents the clear image, A represents the atmospheric light value, t(x) represents the transmittance, and I 0.95 (x,λ) represents the dark channel image.

[0061] S202, performing denoising processing on the flame image after the defogging processing.

[0062] Specifically, the median filter method is used to process the image to reduce or eliminate noise; the median filter sorts the pixel values ​​in the filter window and then selects the middle value as the pixel value after filtering, thereby reducing or eliminating noise.

[0063] S203, performing HSV conversion on the flame image after denoising.

[0064] Specifically, first, read the R, G, and B color components in the flame image after denoising; then, convert the three color components into grayscale values, expressed as:

[0065] g(x i ,yj )=0.2989×R(x i ,y j )+0.5870×G(x i ,y j )+0.1140×B(x i ,y j );

[0066] In the formula, g(x i ,y j ) means located at (x i ,y j ), R(x i ,y j ) means located at (x i ,y j ), G(x i ,y j ) means located at (x i ,y j ), B(x i ,y j ) means located at (x i ,y j ) is the blue component value at .

[0067] S3. Taking the inter-class variance as the fitness function, a hybrid optimization strategy combining genetic algorithm and simulated degradation algorithm is used to perform multi-threshold segmentation on the converted flame image to generate a regional segmentation image.

[0068] Here, the region segmentation image includes a high temperature combustion region, an effective combustion region, an incomplete combustion region, and a background region.

[0069] In this embodiment, a new GA-SA hybrid algorithm is proposed. On the basis of the GA algorithm, the local search mechanism of the SA algorithm is introduced, which can better balance the global search and local search, reduce the probability of the algorithm falling into the local optimum, and improve the convergence speed and result quality. Based on the idea of ​​the Otsu method, the inter-class variance is used as the fitness function. The larger the inter-class variance, the more significant the grayscale difference of the segmented sub-regions, thereby more effectively distinguishing the flame from the background or different parts of the flame, and the better the effect of threshold segmentation.

[0070] As an implementation method, S3 specifically includes:

[0071] S301, initialize the population, each individual in the population is a potential multi-threshold combination, define the fitness function to maximize the inter-class variance as the goal, so as to evaluate the quality of the segmentation effect of each multi-threshold combination.

[0072] The history of fitness values ​​can help track the progress of the algorithm and evaluate the performance of the algorithm by analyzing the evolution of the fitness value. In this embodiment, the fitness function is expressed as:

[0073]

[0074] In the formula, represents the inter-class variance, i.e., the fitness function; q i represents the number of pixels in the i-th category, μ i represents the average gray value of the i-th category, μ T Represents the average gray value of the entire image.

[0075] S302. New individuals are selected and generated through the selection, crossover and mutation operations in the genetic algorithm. These operations ensure the transmission of excellent genes and the exploration of new solutions.

[0076] In this embodiment, the selection operation can adopt a roulette wheel or tournament selection strategy to select excellent individuals; the crossover operation can be a single-point or multi-point crossover operation to combine the genes (threshold settings) of two excellent individuals to generate new individuals; the mutation operation can be a random adjustment of the threshold of the new individual.

[0077] S303, using simulated annealing algorithm to locally optimize the results of genetic algorithm, setting the initial temperature and gradually reducing it, so as to control the randomness and locality of the search by temperature; and applying Metropolis criterion to decide whether to accept new solutions, so as to avoid local optimality and expand the search space of solutions.

[0078] The two algorithms are executed alternately, with the genetic algorithm responsible for global search and the simulated annealing algorithm responsible for local fine search; until the preset number of iterations is met or the fitness is no longer significantly improved. Finally, the threshold combination with the highest current fitness is output as the best multi-threshold segmentation scheme, that is, the best segmentation threshold combination. In this embodiment, the acceptance probability of the simulated annealing algorithm is expressed as:

[0079]

[0080] Where P a represents the acceptance probability of the simulated annealing algorithm, represents the difference in the inter-class variance between the new and old solutions, and T represents the current temperature.

[0081] S304, according to the obtained optimal segmentation threshold combination, the flame image is segmented into four areas: a high-temperature combustion area, an effective combustion area, an incomplete combustion area and a background area.

[0082] S4. Extract features from the region segmented image to obtain flame shape features and flame statistical features.

[0083] In this embodiment, the flame shape characteristics include the flame effective areas of the effective combustion zone and the incomplete combustion zone in the left picture, the flame effective areas of the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the right picture, the flame effective area rate of the incomplete combustion zone in the left picture, the flame effective area rate of the incomplete combustion zone in the right picture, and the flame effective area rate of the high-temperature combustion zone in the right picture; the flame statistical characteristics include the average grayscale value of the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the left and right pictures, the grayscale standard deviation of the high-temperature combustion zone and the effective combustion zone in the left picture, the horizontal and vertical center positions of the flame in the left and right pictures, and the Shannon entropy of the grayscale histogram in the right picture.

[0084] Specifically, according to the total number of pixels whose grayscale values ​​in the high-temperature combustion zone are greater than the first flame grayscale threshold of the high-temperature combustion zone, the flame effective area of ​​the high-temperature combustion zone is calculated, and the area rate of the high-temperature combustion zone is calculated in combination with the length and height of the regional segmentation image; according to the total number of pixels whose grayscale values ​​in the effective combustion zone are greater than the flame grayscale threshold of the effective combustion zone, the flame effective area of ​​the effective combustion zone is calculated, and the area rate of the effective combustion zone is calculated in combination with the length and height of the regional segmentation image; according to the total number of pixels whose grayscale values ​​in the incomplete combustion zone are greater than the flame grayscale threshold of the incomplete combustion zone, the flame effective area of ​​the incomplete combustion zone is calculated, and the area rate of the incomplete combustion zone is calculated in combination with the length and height of the regional segmentation image. According to the grayscale values ​​of all pixels in the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone, the average grayscale and grayscale standard deviation are calculated respectively; based on the regional segmentation image, the Shannon entropy of the flame center position and the grayscale histogram of the combustion area is calculated.

[0085] The flame effective area of ​​the high-temperature combustion zone refers to the total number of pixels in the high-temperature combustion zone, that is, the total number of pixels in the image whose grayscale value is greater than the specified threshold. The flame effective area rate of the high-temperature combustion zone represents the proportion of the high-temperature combustion zone. The high-temperature zone usually corresponds to the full combustion of the fuel and is closely related to the combustion efficiency. The flame effective area A_H and area rate R_H of the high-temperature combustion zone are expressed as:

[0086]

[0087]

[0088] In the formula, g(x i ,y j ) means (x i ,y j ) is the gray value of the image at ; Indicates the flame grayscale threshold of the high-temperature combustion area; N and M represent the length and height of the entire flame image respectively; is a unit step function used to ensure that the pixels are within the required range.

[0089] The flame effective area rate R_E in the effective combustion zone refers to the ratio of the area of ​​the effective combustion zone to the total area, reflecting the proportion of the flame area actually participating in the combustion. The calculation is similar to the flame effective area rate in the high temperature zone, except that the gray threshold is different. The formula is as follows:

[0090]

[0091]

[0092] In the formula, The flame grayscale threshold that represents the effective combustion area.

[0093] The flame area of ​​the effective combustion zone can reflect the overall combustion of the garbage in the furnace on the combustion grate. The reaction on the combustion grate is the most important stage of the garbage combustion process. Generally speaking, when the garbage combustion condition is better, the area of ​​the effective zone is larger. When the combustion condition is unstable, the area of ​​the effective zone will fluctuate violently. Therefore, the flame area rate of the effective zone can be used to reflect the combustion condition.

[0094] The effective area rate of the incomplete combustion zone flame refers to the proportion of the combustion area that has not been fully burned in the total combustion area. The effective area A_N and area rate R_N of the incomplete combustion zone flame are expressed as:

[0095]

[0096] In the formula, The flame grayscale threshold indicating the incomplete combustion area.

[0097] The average grayscale of the high-temperature combustion area refers to the average grayscale value of all positions in the high-temperature area. The average grayscale value G_H of the high-temperature combustion area is expressed as:

[0098]

[0099] If the effective combustion area A_H of the high temperature region is 0, then G_H is specified to be 0. The average gray value of the high temperature combustion region can provide information on the temperature distribution and combustion intensity of the high temperature region during the combustion process.

[0100] By calculating the horizontal and vertical center positions of the flame, the geometric center of the flame can be monitored to evaluate the combustion status or abnormal conditions. Generally, the ideal position of the flame is within a specific area. If the flame deviates too much, it indicates that there may be problems with biased burning or burner failure. It can be expressed as:

[0101]

[0102]

[0103] The average gray value of the effective combustion area refers to the average gray value of all positions in the effective area. The average gray value G_E of the effective combustion area is expressed as:

[0104]

[0105] If the effective combustion area A_E of the effective combustion zone = 0, then G_E = 0. The average gray value of the effective combustion zone reflects the temperature level in the furnace. As the temperature increases, G_E increases.

[0106] The average gray value of the incomplete combustion area refers to the average gray value of all positions in the incomplete combustion area. The average gray value G_N of the incomplete combustion area is expressed as:

[0107]

[0108] If the effective combustion area A_N = 0, then G_N = 0 is specified. A lower G_N value indicates incomplete combustion, lower temperature, unburned carbon particles and other incomplete combustion products. By monitoring and analyzing G_N, problems in the combustion process can be identified.

[0109] The grayscale standard deviation of the high-temperature combustion area refers to the standard deviation of the grayscale values ​​of all pixels in the high-temperature combustion area, which is an important indicator for evaluating the flame uniformity and stability in the high-temperature area. The grayscale standard deviation S_H of the high-temperature combustion area is expressed as:

[0110]

[0111] By calculating the horizontal and vertical center positions of the flame, the geometric center of the flame can be monitored to evaluate the combustion status or abnormal conditions. Generally, the ideal position of the flame is within a specific area. If the flame deviates too much, it indicates that there may be problems with biased burning or burner failure. It can be expressed as:

[0112]

[0113]

[0114] In the formula, x c Indicates the horizontal center position of the flame, y c Indicates the vertical center position of the flame, The flame grayscale threshold representing the effective combustion area.

[0115] The Shannon entropy of the grayscale histogram of the burning area is an important indicator to measure the complexity and randomness of the grayscale distribution in the area. A higher entropy value indicates that the combustion in the area is more intense and unstable, while a lower entropy value indicates that the combustion is more stable or uniform. It can be expressed as:

[0116]

[0117]

[0118] Where p(k) represents the probability distribution of gray level k in the high temperature and effective combustion region, and the gray level k ranges from 0 to 255; is the Dirac function, which is used to determine whether the grayscale value of a pixel is k; E represents the Shannon entropy, which reflects the complexity and randomness of the grayscale distribution.

[0119] S5. Perform dimensionality reduction processing on flame shape features and flame statistical features, construct a feature matrix, input the feature matrix into a trained flame combustion state recognition model, process the feature matrix through the trained flame combustion state recognition model, and obtain the flame combustion state recognition model.

[0120] In this embodiment, the flame combustion state recognition model is a deep Q network combined with a pseudo-label generation strategy, namely a DQN-PL model, which takes the feature matrix as input and outputs the flame combustion state recognition result.

[0121] In the training process of the deep Q network, unlabeled data is gradually introduced through the pseudo-label generation strategy to improve the generalization ability of the model. As an implementation method, the specific process of training the flame combustion state recognition model is as follows:

[0122] Step 1: Obtain multiple sets of combustion state data sets and construct training sets.

[0123] Step 2: Preprocess the images in the training set and convert them into grayscale images. Take the inter-class variance as the fitness function and use a hybrid optimization strategy combining genetic algorithm and simulated degradation algorithm to perform multi-threshold segmentation on the converted images to generate regional segmentation images.

[0124] Step 3: Extract features from the region segmented image to obtain flame shape features and flame statistical features.

[0125] Here, the flame shape features and the flame statistical features include all features in the left picture and the right picture.

[0126] Step 4: Perform feature selection and dimension reduction on the flame shape features and flame statistical features to construct a feature matrix, specifically including:

[0127] Step 401 , combining L1 regularization, recursive feature elimination (RFE) and random forest feature importance to perform feature selection on all flame shape features and flame statistical features, reducing the dimension of the feature matrix from 28 to 22.

[0128] Here, the 22 feature parameters obtained after feature selection include: the flame effective area of ​​the effective combustion zone and the incomplete combustion zone in the left picture, the flame effective area of ​​the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the right picture, the flame effective area rate of the incomplete combustion zone in the left picture, the flame effective area rate of the incomplete combustion zone in the right picture, the flame effective area rate of the high-temperature combustion zone in the right picture, the average grayscale value of the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the left and right pictures, the grayscale standard deviation of the high-temperature combustion zone and the effective combustion zone in the left picture, the horizontal and vertical center positions of the flame in the left and right pictures, and the Shannon entropy of the grayscale histogram in the right picture.

[0129] Furthermore, the specific process of step 401 is as follows:

[0130] (1) All flame shape features and flame statistical features are screened through L1 regularization.

[0131] A deep neural network (DNN) classification model is used to evaluate the effect of the model before and after feature selection. The sparsity of the weights is controlled by adjusting the regularization parameter λ to screen out the most important feature parameters. The loss function is as follows:

[0132]

[0133] In the formula, y i is the actual value, is the predicted value, w j is the weight of the feature.

[0134] Here, there are 14 characteristic parameters obtained, namely: the flame effective area of ​​the effective combustion zone and the incomplete combustion zone in the left picture, the effective area of ​​the high-temperature combustion zone and the effective combustion flame in the right picture, the flame effective area rate of the effective combustion zone in the left picture, the flame effective area rate of the high-temperature combustion zone and the effective combustion flame in the right picture, the average grayscale value of the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the left and right pictures, and the horizontal center position of the flame in the left picture.

[0135] (2) All feature parameters are processed through recursive feature elimination method.

[0136] First, all feature parameters are processed by a deep neural network (DNN) classification model to obtain the importance scores of the feature parameters; then, the least important feature parameters are recursively deleted and the model is retrained until the predetermined number of features is reached.

[0137] Here, there are five characteristic parameters obtained through screening, namely: the flame effective area and average gray value of the incomplete combustion area in the left picture, the gray standard deviation of the high temperature combustion area in the left picture, the horizontal center position of the flame in the left picture, and the Shannon entropy of the gray histogram in the right picture.

[0138] (3) Random Forest.

[0139] The contribution of each feature parameter is determined by calculating its importance in the decision tree, which is expressed as:

[0140]

[0141] In the formula, y i is the actual value, is the predicted value, w j is the weight of the feature.

[0142] Here, there are 20 characteristic parameters obtained, namely: the effective area of ​​the flame in the effective combustion zone in the left picture, the effective areas of the flame in the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the right picture, the effective area rate of the flame in the incomplete combustion zone in the left picture, the effective area rate of the flame in the incomplete combustion zone in the right picture, the effective area rate of the flame in the high-temperature combustion zone in the right picture, the average grayscale value of the high-temperature combustion zone, the effective combustion zone and the incomplete combustion zone in the left and right pictures, the grayscale standard deviation of the high-temperature combustion zone and the effective combustion zone in the left picture, and the horizontal and vertical center positions of the flames in the left and right pictures.

[0143] Through the above three feature selection methods, the 28 feature parameters of the data set are reduced to 22 feature parameters.

[0144] Step 402: Use principal component analysis (PCA) to reduce the dimension of the feature matrix after feature selection, and further reduce the feature dimension to 5 principal components; use a deep neural network (DNN) classification model to evaluate the prediction effect of the model before and after dimensionality reduction.

[0145] Principal component analysis (PCA) is a data dimensionality reduction method. The core idea is to select the eigenvectors (principal components) with the largest variance in the data set. These principal components are the eigenvectors of the covariance matrix, which can maximize the retention of the characteristics of the original data. Assuming that m n-dimensional eigenvectors have been extracted, the vector w is defined as the mapping vector of the low-dimensional space. The formula for maximizing the variance of the data projected onto this vector is as follows:

[0146]

[0147] In the formula, x i is the feature vector of the i-th sample; is the mean vector of all samples; w is the direction vector we want to find.

[0148] W is a matrix consisting of column vectors containing all feature map vectors. This matrix can be transformed into an optimized objective function through linear transformation, which can minimize the spread of data points in the new projection space (the trace of the covariance matrix), thereby finding the most important component (the direction of maximum variance). The output of PCA is Y = W'X. The optimal W is composed of the eigenvectors corresponding to the first k largest eigenvalues ​​of the data covariance matrix as column vectors, thereby reducing the original dimension to k dimensions.

[0149]

[0150]

[0151] Where, the projection matrix W is composed of multiple projection direction vectors w; tr represents the trace of the matrix; I is the identity matrix, ensuring that the column vectors in W are orthogonal to each other; A is the covariance matrix of the data.

[0152] Assuming that the final data set has n samples and each sample has m features, the new data set feature matrix is ​​as follows:

[0153]

[0154] Where X is an n×m feature matrix, x ij represents the eigenvalue of the i-th row and j-th column.

[0155] Step 5: Input the reduced feature matrix into the deep Q network. During the model training process, the unlabeled data is gradually introduced through the pseudo-label generation strategy to improve the generalization ability of the model.

[0156] The DQN (Deep Q Network) algorithm is a value-based deep reinforcement learning algorithm that combines a deep neural network with a Q-Learning algorithm. The deep neural network DNN is used as a function approximator to estimate the value function Q(s,a), and the optimal action is determined through the state space s and the action space a. The goal is to maximize the cumulative reward. The experience replay mechanism selects actions from the current state using an ε-greedy strategy, and obtains rewards and the next state from the environment. The experience (state s, action a, reward r, next state s) is stored in a fixed-size memory pool D for experience replay during subsequent training. Two deep neural networks with the same weights but different structures are used as Q networks. During training, the model extracts a batch of data from the memory pool D, uses the above-mentioned Q-learning update rule to update the weight w of the neural network, calculates the current Q value, uses gradient descent and back propagation to update the parameters of the current Q network, adjusts the difference between the predicted Q value and the target Q value, and regularly copies the network parameters to the target Q network to ensure stable training.

[0157] The Q value update formula is as follows:

[0158]

[0159]

[0160] In the formula, α is the learning rate, γ is the discount factor, and r t is the reward value.

[0161] The pseudo-label generation strategy is combined with the DQN model to establish the DQN-PL model, which is used to deal with the situation where the classification labels are missing in the data set and improve the generalization ability of the model. unlabeled Predict the action corresponding to the maximum Q value (i.e., the classification label), and generate the pseudo label y pseudo With labeled data y labeled Combine to form a new dataset X combined , as the input of the final classification model.

[0162]

[0163] In the DQN-PL model, DQN is not only used to generate pseudo labels, but also serves as the core component of the classification network, affecting the final classification prediction. DQN-PL introduces a joint loss function L total , the Q-learning update rule is used to optimize the value function of DQN, and the loss function is L DQN , the classification task uses the standard cross entropy loss function L classfy The two are combined in a weighted manner and share part of the network structure. During the optimization process, the Q value update of DQN will affect the weight of the neural network. The pseudo labels generated by DQN update the weight of the neural network through supervised learning, thereby improving the quality of the model.

[0164] L total =λ1L DQN +λ2L classfy ;

[0165]

[0166]

[0167] In the formula, y i is a true label or a pseudo label, is the predicted label; λ1 and λ2 are weight coefficients used to balance the impact of the two parts of the loss.

[0168] The hyperparameter settings are shown in Table 1. An upper limit of 2000 training steps is set in the experience replay memory pool, and the discount factor is set to 0.95 to ensure that future rewards can be reasonably discounted to current values. The learning rate of the model is set to 0.001 to help stabilize the training process and avoid excessive weight updates. In addition, the exploration rate of the model is initially 1.2, and a decay rate of 0.99 is set, which eventually drops to 0.01. In the pseudo-label generation module, the confidence threshold of the pseudo-label is set to 80%. The network structure of its deep neural network DNN part is shown in Table 2. The input layer is responsible for receiving the feature matrix of the processed flame image. The hidden layer contains two dense layers, each of which is configured with 32 neurons, and the ReLU activation function is used to enhance the model's ability to process nonlinear data.

[0169] Table 1 Hyperparameter settings

[0170]

[0171] Table 2 Network structure parameters

[0172]

[0173] Furthermore, after training is completed, the performance of the model is evaluated through k-fold cross validation and test sets, including accuracy, cross entropy loss, confusion matrix and other related indicators. The details are as follows:

[0174] (1) k-fold cross validation.

[0175] The training set data is divided into k non-overlapping subsets (k=5). Then, each subset is used as a validation set in turn, and the remaining k-1 subsets are used as training sets to train and validate the model. Finally, the average performance index of k validations is calculated.

[0176] (2) Performance indicator evaluation.

[0177] Use the trained model to make predictions on an independent test set and calculate its performance metrics.

[0178] (1) Accuracy. This measures the proportion of correct classifications by the model. It is calculated as follows:

[0179]

[0180] In the formula, TP refers to the number of samples that are actually positive and predicted as positive, TN refers to the number of samples that are actually negative and predicted as negative, FP refers to the number of samples that are actually negative but predicted as positive, and FN refers to the number of samples that are actually positive but predicted as negative.

[0181] (2) Cross entropy loss. The smaller the cross entropy loss value, the closer the probability distribution predicted by the model is to the actual probability distribution, and the better the performance of the model.

[0182]

[0183] In the formula, N represents the number of samples, y ij is the actual label, p ij is the probability distribution predicted by the model.

[0184] (3) Confusion matrix. In the problem of flame combustion state recognition, the confusion matrix is ​​a 5×5 table. It is used to show the comparison between the classification results of the model and the actual labels. Each row represents the true category and each column represents the predicted category. Through the confusion matrix, a variety of performance indicators can be calculated, such as precision, recall, F1 score, etc.

[0185] Table 3 Confusion matrix

[0186]

[0187] Precision: refers to the ratio of the number of correct samples predicted by the model to the number of samples predicted to be in a certain category.

[0188]

[0189] Recall refers to the ratio of the number of correct samples that are actually in a certain category to the number of samples that are actually in that category.

[0190]

[0191] The F1 score is the harmonic mean of precision and recall, which comprehensively measures the performance of the model in a certain category.

[0192]

[0193] Embodiment 2

[0194] This embodiment discloses a flame combustion state recognition system based on image multi-threshold segmentation, comprising:

[0195] The image multi-threshold segmentation module is configured to: obtain a flame image, use the inter-class variance as a fitness function, and perform multi-threshold segmentation on the flame image using a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to generate a region segmentation image;

[0196] The feature extraction module is configured to: perform feature extraction on the region segmentation image, obtain flame shape features and flame statistical features, and construct a feature matrix;

[0197] The flame combustion state recognition module is configured to process the feature matrix through a trained flame combustion state recognition model to obtain the flame combustion state.

[0198] It should be noted that the above-mentioned image multi-threshold segmentation module, feature extraction module and flame combustion state recognition module correspond to the steps in Embodiment 1, and the examples and application scenarios implemented by the above-mentioned modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above-mentioned Embodiment 1. It should be noted that the above-mentioned modules as part of the system can be executed in a computer system such as a set of computer executable instructions.

[0199] Embodiment 3

[0200] Embodiment 3 of the present invention provides an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor. When the computer instructions are run by the processor, the steps of the flame combustion state recognition method based on multi-threshold image segmentation are completed.

[0201] Embodiment 4

[0202] Embodiment 4 of the present invention provides a computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, the steps of the flame combustion state recognition method based on multi-threshold image segmentation are completed.

[0203] Embodiment 5

[0204] Embodiment 5 of the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps of the flame combustion state recognition method based on image multi-threshold segmentation.

[0205] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0206] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0207] These computer program instructions can also be loaded onto a computer or other programmable data processing device, and a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0208] The description of each embodiment in the above embodiments has different emphases. For parts not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0209] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A flame combustion state recognition method based on image multi-threshold segmentation, characterized in that: include: Acquire a flame image, use the inter-class variance as a fitness function, and use a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to perform multi-threshold segmentation on the flame image to generate a regional segmentation image; Extracting features from the region segmented image, obtaining flame shape features and flame statistical features, and constructing a feature matrix; The characteristic matrix is ​​processed by a trained flame combustion state recognition model to obtain the flame combustion state.

2. The flame combustion state recognition method based on image multi-threshold segmentation according to claim 1 is characterized in that: The method of using the inter-class variance as the fitness function and utilizing a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to perform multi-threshold segmentation on the flame image includes: With the goal of maximizing the inter-class variance, a genetic algorithm is used to perform a global search on the flame image, and the population is optimized through selection, crossover and mutation operations; for the new individuals generated by the genetic algorithm, a simulated degradation algorithm is used to perform local optimization until the best segmentation threshold combination is obtained; Dividing the flame image into regions according to the optimal segmentation threshold combination to generate a region segmentation image; The region segmentation image includes a high temperature combustion region, an effective combustion region, an incomplete combustion region and a background region.

3. The flame combustion state recognition method based on image multi-threshold segmentation according to claim 1 is characterized in that: The feature extraction of the region segmentation image is performed to obtain the flame shape feature as follows: based on the comparison result of the grayscale value in different combustion zones in the region segmentation image and the preset flame grayscale threshold, the flame effective area of ​​different combustion zones is calculated, and the flame effective area rate of different combustion zones is calculated in combination with the length and height of the region segmentation image.

4. The flame combustion state recognition method based on image multi-threshold segmentation according to claim 1 is characterized in that: The feature extraction is performed on the region segmentation image to obtain the flame statistical features as follows: According to the grayscale values ​​of all pixels in different burning areas in the region segmentation image, the average grayscale and grayscale standard deviation of different burning areas are calculated respectively; Based on the region segmentation image, the Shannon entropy of the flame center position and the grayscale histogram of the burning area are calculated.

5. The flame combustion state recognition method based on image multi-threshold segmentation according to claim 1 is characterized in that: The flame combustion state recognition model is a deep Q network combined with a pseudo-label generation strategy.

6. The flame combustion state recognition method based on image multi-threshold segmentation according to claim 1 is characterized in that: When training the flame combustion state recognition model, a joint loss function, a value function optimization loss function and a cross entropy loss function are introduced.

7. The flame combustion state recognition system based on image multi-threshold segmentation is characterized by: include: The image multi-threshold segmentation module is configured to: obtain a flame image, use the inter-class variance as a fitness function, and perform multi-threshold segmentation on the flame image using a hybrid optimization strategy combining a genetic algorithm and a simulated degradation algorithm to generate a region segmentation image; The feature extraction module is configured to: extract features from the region segmentation image, obtain flame shape features and flame statistical features, and construct a feature matrix; The flame combustion state recognition module is configured to process the feature matrix through a trained flame combustion state recognition model to obtain the flame combustion state.

8. An electronic device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the flame combustion state recognition method based on image multi-threshold segmentation according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program / instruction stored thereon, characterized in that: When the computer program / instruction is executed by a processor, the steps of the flame combustion state recognition method based on image multi-threshold segmentation as described in any one of claims 1 to 6 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instruction is executed by a processor, the steps of the flame combustion state recognition method based on image multi-threshold segmentation as described in any one of claims 1 to 6 are implemented.

Images