Power plant boiler temperature field online monitoring method and system based on convolutional neural network and flame radiation image

Abstract

The present application relates to the technical field of power plant boiler temperature field monitoring, in particular to a power plant boiler temperature field online monitoring method and system based on convolutional neural network and flame radiation image, the monitoring method comprising: calibrating a detector with a blackbody furnace, establishing the relationship between image intensity and radiation intensity, and using the calibrated detector to collect flame images during the 20%-100% load change of the boiler to obtain original flame images; adding random obstructions to the flame area in the original flame images to obtain obstructed flame images; and simulating the situation where the detector lens is obstructed by slag. The present application introduces computer deep learning image repair technology into the field of combustion engineering, uses unobstructed and obstructed flame images as the learning set of the convolutional neural network model, trains the convolutional neural network model, so that the trained model can process the slag-free flame image output from the slagged flame image, solve the lens slagging interference problem, and accurately measure the flame temperature field.

Description

Technical Field

[0001] This invention relates to the field of power plant boiler temperature field monitoring technology, and in particular to a method and system for online monitoring of power plant boiler temperature field based on convolutional neural networks and flame radiation images. Background Technology

[0002] Accurate monitoring of the boiler temperature field is crucial for optimizing combustion efficiency, controlling pollutants, and ensuring safe equipment operation. This is especially true under deep peak-shaving conditions, where the temperature field distribution directly impacts the boiler's stability and economic efficiency. Currently, temperature monitoring methods for pulverized coal boilers in power plants primarily include thermocouples, acoustic wave methods, digital transducer laser absorption spectroscopy (TDLAS), and flame image thermometry.

[0003] Flame radiation image thermometry uses a calibrated camera to acquire flame images, converts the flame image intensity into monochromatic radiation intensity, and calculates the flame temperature distribution from the radiation intensity. The camera calibration process establishes the relationship between monochromatic radiation intensity, image intensity, and exposure time. This method not only achieves high spatial resolution temperature field monitoring but also reflects the combustion state in real time.

[0004] However, the applicant discovered that the flame radiation image thermometry method has problems with slag buildup or fly ash obstruction on the detector lens, resulting in slag in the captured flame images. This leads to deviations in the subsequent radiation intensity measurements, thus affecting the accuracy of the temperature field calculation. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose an online monitoring method and system for the temperature field of power plant boilers based on convolutional neural networks and flame radiation images, so as to solve the problem that the existing flame radiation image method cannot accurately reflect the flame radiation intensity due to the focusing and obstruction of the detector lens, thus resulting in the measured temperature field not accurately reflecting the combustion state.

[0006] To achieve the above objectives, this invention provides an online monitoring method for the temperature field of a power plant boiler based on a convolutional neural network and flame radiation images, specifically including the following steps:

[0007] S1. Using a blackbody furnace to calibrate the detector, establish the relationship between image intensity and radiation intensity, and use the calibrated detector to collect flame images during the boiler load variation period of 20%-100% to obtain the original flame image.

[0008] S2. Add random obstructions to the flame area in the original flame image to obtain an obstructed flame image; to simulate the situation where the detector lens is obstructed by slag.

[0009] S2 specifically includes the following steps:

[0010] S2.1 Flame area detection;

[0011] S2.2 The number and size of the obstructions are generated randomly;

[0012] S2.3, Obstruction shape generation;

[0013] S2.4 Place a shield in the flame area;

[0014] S3. Use the original flame image and the occluded flame image as the learning set of the CNN model to train the model, and use the trained model to reconstruct the flame image with slag captured on site to obtain the flame image without slag after repair.

[0015] S4. Calculate the temperature of the repaired flame image using the two-color method to obtain the flame temperature field.

[0016] Preferably, in step S1, the process of calibrating the detector on the blackbody furnace is as follows:

[0017] S1.1 Set the blackbody furnace temperature to 800℃-1440℃, and adjust the exposure time at the same temperature to gradually increase the maximum value of the image R channel to 250, and acquire ten images at each temperature.

[0018] S1.2 After image acquisition is completed, the obtained image data is processed and the calibration curves of the R channel and G channel are plotted.

[0019] Preferably, the detector consists of a CCD camera, a high-temperature resistant lens, and a stainless steel housing. Measuring the actual temperature of the flame using the calibrated detector in S1 includes the following steps:

[0020] S1.3, From Planck's blackbody radiation formula,

[0021]

[0022] In the formula, It is the blackbody radiation intensity; It is the blackbody temperature; It is the wavelength of radiation; It is the first radiation constant, with a value of ; It is the second radiation constant, with a value of ;

[0023] S1.4 Since the temperature during the calibration process and the flame temperature measurement process is between 1073-1873K, and the wavelength range is in the visible light band, Planck's blackbody radiation formula can be approximately replaced by Wien's displacement law:

[0024]

[0025] For a blackbody furnace, the radiation intensities of its R-channel and G-channel are:

[0026]

[0027]

[0028] A dual bandpass filter is set at the front of the CCD camera, with response wavelengths of 550nm and 650nm respectively, and a half-bandwidth of 10nm. Therefore, the formula... Pick ; Pick ;

[0029] S1.5 During temperature measurement, the radiation intensity of the actual object's R-channel and G-channel is obtained from the calibration curve. The radiation intensity of the R-channel and G-channel is:

[0030]

[0031]

[0032] In the formula, For the emissivity of the actual object, The temperature of the actual object;

[0033] S1.6. According to the gray body hypothesis, at the same temperature, the emissivity of a gray body is the same at different radiation wavelengths, that is... From the above formula, the actual temperature of the object can be obtained:

[0034] .

[0035] Preferably, step S2.1 includes the following steps:

[0036] S2.11 First, convert the original flame image into a grayscale image;

[0037] S2.12, The conversion of an RGB image to a grayscale image is achieved by the following formula:

[0038]

[0039] in, R represents the intensity value of the grayscale image; R, G, and B are the intensities of the three channels of the original image, respectively. These are pixel coordinates;

[0040] S2.13. Detect bright flame regions in the grayscale image and construct a binary flame mask as shown in the following formula.

[0041]

[0042] in, The grayscale brightness threshold is used; pixel areas exceeding this value are defined as flame areas.

[0043] In S2.2, the number of obstructions and the size of the obstruction Satisfy the following formula:

[0044]

[0045] .

[0046] Preferably, S2.3 includes the following steps:

[0047] S2.31, Number of vertices of the occluding polygon The vertex angles are evenly distributed in Above, the following formula is satisfied:

[0048]

[0049] Vertex radius , which follows a uniform integer distribution;

[0050] Vertex coordinates We obtain it from the following formula:

[0051]

[0052] ;

[0053] Use the inpolygon function to generate a mask. The polygon vertices are translated to the center of the mask, and it is determined whether the point is inside the polygon.

[0054] ;

[0055] S2.32. The mask is subjected to expansion and etching operations, and the expansion and etching process is represented by the following formula:

[0056]

[0057]

[0058] In the formula, Indicates the result of the expansion; This represents the expansion operation; The structure elements are randomly generated; Indicates the corrosion result; This represents the erosion operation;

[0059] S2.33, Generate random rotation angle The mask is rotated, and the rotation process is represented by the following formula:

[0060]

[0061] In the formula, Indicates the result of rotation; Indicates rotation operation;

[0062] S2.34. Apply Gaussian blur to the mask. The process can be represented by the following formula:

[0063]

[0064] In the formula, Indicates the Gaussian blur result; For Gaussian kernel, The Gaussian blur parameters are randomly generated; Indicates Gaussian blur operation;

[0065] In S2.4, the mask in the flame area Randomly select pixel position The obstruction completely covers the flame area, and the RGB values ​​of the obstruction area are generated random brightness values. This is consistent with the actual coking situation.

[0066] Preferably, the development process for CNN models in S3 is as follows:

[0067] S3.1 Data preparation includes image reading and alignment, normalization processing, and data partitioning;

[0068] S3.11 Image Reading and Alignment: Store the original flame image and the occluded flame image in two folders respectively. The filenames and quantities correspond one-to-one. The two sets can be represented by the following formula:

[0069]

[0070] ;

[0071] S3.12 Normalization Process: The pixel value of each image is linearly normalized from [0,255] to [0,1]. This process can be represented by the following formula.

[0072] ;

[0073] S3.13, Dataset Partitioning: Divide the input images into training and test sets in a 9:1 ratio;

[0074] S3.2 Model Construction: The convolutional neural network consists of convolutional layers, activation function layers, pooling layers, and fully connected layers. Each flame image has 300*403 pixels. The convolutional layers are set with 3*3 convolutional kernels and 64 filters.

[0075] S3.3 Model training includes the following steps:

[0076] S3.31, Training settings: Adam optimizer, maximum number of training epochs: 20, batch size per training session: 32, initial learning rate: 0.0001;

[0077] S3.32. The mean squared error (MSE) is used as the loss function during the training process, as shown in the following formula:

[0078] ;

[0079] This represents the number of training batches for the model.

[0080] Preferably, the development of the CNN model in S3 also includes the following steps:

[0081] S3.4. Use SSIM and PSNR to evaluate the performance of the trained model;

[0082] S3.41, The definition of SSIM is shown in the following formula:

[0083]

[0084]

[0085]

[0086]

[0087] In the formula , , These are the brightness correlation function, contrast correlation function, and structure correlation function between the reference image and the image under test, respectively. The average brightness of the two images; , Standard deviation; For covariance; , , Use very small positive numbers to prevent instability caused by denominators being 0 or close to 0; , , All are positive numbers used to adjust the weights of brightness, contrast, and structural relevance; the SSIM value ranges from [-1, 1]. The closer SSIM is to 1, the stronger the similarity between the two images; when the two images are completely identical... hour, ;

[0088] S3.42, The definition of PSNR is shown in the following formula:

[0089]

[0090] In the formula, The image is to be evaluated; It is the original image; It is the maximum grayscale value of the image. These are the pixels of the image;

[0091] PSNR is measured in dB. The higher the PSNR value, the smaller the difference between the reconstructed image and the original image, and the higher the image quality. When PSNR > 28 dB, the difference between the two images is not obvious; when PSNR > 35 dB, the human eye can hardly see the difference between the two images.

[0092] S3.5 Save the trained model as a .mat file.

[0093] An online monitoring system for temperature field of power plant boiler based on convolutional neural network and flame radiation image, including blackbody furnace calibration system, combustion detection system and CNN model system;

[0094] The blackbody furnace calibration system consists of a blackbody furnace, a detector, a detector bracket, a PoE switch, a gigabit network cable, and a computer.

[0095] The blackbody furnace is a device used to generate standard blackbody radiation for calibrating detectors;

[0096] The detector transmits the images captured inside the blackbody furnace to the PoE switch, and then to the computer via a gigabit network cable. The computer processes the obtained image data and plots the calibration curves for the R and G channels.

[0097] The combustion detection system includes A-layer burners, B-layer burners and C-layer burners installed inside the boiler, and six detectors calibrated by a blackbody furnace are evenly spaced on the A-layer burners.

[0098] It also includes switches, electronic control cabinets, DCS, a large screen in the central control room, and new fans;

[0099] The detector is connected to the switch via a gigabit network cable, and the switch is connected to the electronic room control cabinet via an optical fiber. The electronic room control cabinet is connected to the central control room screen via a signal line and to the DCS via RS485. This allows the detector to collect flame image data, transmit it to the switch in the field control cabinet via a gigabit network cable, and then transmit the data to the electronic room server via an optical fiber for processing. Finally, the temperature field results are displayed on the central control room screen. At the same time, the server's calculation results are transmitted to the DCS system via an RS485 signal line to assist in the parameter adjustment of the boiler system.

[0100] The newly added fan cools the detector through a cooling duct at its outlet.

[0101] The CNN model system is used to process slag-laden flame images and output non-slag-laden flame images.

[0102] The beneficial effects of this invention are as follows: It introduces computer deep learning image restoration technology into the field of combustion engineering. Random occlusions are added to the flame region of the original flame image to simulate the situation where the lens is obstructed by focus, resulting in two types of flame images: unobstructed and obstructed. These images serve as the training set for a convolutional neural network model, enabling the model to process slag-covered flame images and output unslag-covered flame images. The trained model achieves SSIM=0.9989 and PSNR=50.55dB, demonstrating excellent restoration performance. The trained model is used to restore slag-covered images captured on-site, obtaining the restored image. The temperature is then calculated using a two-color method to obtain the flame temperature field, thereby solving the problem of lens slag interference.

[0103] Furthermore, the model has wide applicability: different field flames require different learning sets to relearn the model; it has good adaptability to changing operating conditions: the learning set already contains images of the boiler under different loads. Attached Figure Description

[0104] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0105] Figure 1 This is a schematic diagram of the test of the blackbody furnace calibration system of the present invention;

[0106] Figure 2 This is a schematic diagram of the real convolutional neural network of the present invention;

[0107] Figure 3 This is a schematic diagram of the combustion detection system of the present invention;

[0108] Figure 4 This is a schematic diagram of the flame image with slag before and after the repair according to the present invention;

[0109] Figure 5 This is a schematic diagram of the temperature field of the flame image without slag formation and the flame image field with slag formation after repair according to the present invention. Detailed Implementation

[0110] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments and accompanying drawings.

[0111] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this invention should have the ordinary meaning understood by those skilled in the art to which this invention pertains. The terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0112] like Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 As shown, the online monitoring method for the temperature field of a power plant boiler based on convolutional neural networks and flame radiation images specifically includes the following steps:

[0113] S1. Using a blackbody furnace to calibrate the detector, establish the relationship between image intensity and radiation intensity, and use the calibrated detector to collect flame images during the boiler load variation period of 20%-100% to obtain the original flame image.

[0114] In step S1, the process of calibrating the detector on the blackbody furnace is as follows:

[0115] S1.1 Set the blackbody furnace temperature to 800℃-1440℃, and adjust the exposure time at the same temperature to gradually increase the maximum value of the image R channel to 250, and acquire ten images at each temperature.

[0116] S1.2 After image acquisition is completed, the obtained image data is processed and the calibration curves of the R channel and G channel are plotted.

[0117] The detector consists of a CCD camera, a high-temperature resistant lens, and a stainless steel housing. Measuring the actual temperature of the flame in S1 using the calibrated detector involves the following steps:

[0118] S1.3, From Planck's blackbody radiation formula,

[0119]

[0120] In the formula, It is the blackbody radiation intensity; It is the blackbody temperature; It is the wavelength of radiation; It is the first radiation constant, with a value of ; It is the second radiation constant, with a value of ;

[0121] S1.4 Since the temperature during the calibration process and the flame temperature measurement process is between 1073-1873K, and the wavelength range is in the visible light band, Planck's blackbody radiation formula can be approximately replaced by Wien's displacement law:

[0122]

[0123] For a blackbody furnace, the radiation intensities of its R-channel and G-channel are:

[0124]

[0125]

[0126] A dual bandpass filter is set at the front of the CCD camera, with response wavelengths of 550nm and 650nm respectively, and a half-bandwidth of 10nm. Therefore, the formula... Pick ; Pick ;

[0127] S1.5 During temperature measurement, the radiation intensity of the actual object's R-channel and G-channel is obtained from the calibration curve. The radiation intensity of the R-channel and G-channel is:

[0128]

[0129]

[0130] In the formula, For the emissivity of the actual object, The temperature of the actual object;

[0131] S1.6. According to the gray body hypothesis, at the same temperature, the emissivity of a gray body is the same at different radiation wavelengths, that is... From the above formula, the actual temperature of the object can be obtained:

[0132] .

[0133] S2. Add random obstructions to the flame area in the original flame image to obtain an obstructed flame image; to simulate the situation where the detector lens is obstructed by slag.

[0134] S2 specifically includes the following steps:

[0135] S2.1 Flame area detection;

[0136] S2.1 includes the following steps:

[0137] S2.11 First, convert the original flame image into a grayscale image;

[0138] S2.12, The conversion of an RGB image to a grayscale image is achieved by the following formula:

[0139]

[0140] in, R represents the intensity value of the grayscale image; R, G, and B are the intensities of the three channels of the original image, respectively. These are pixel coordinates;

[0141] S2.13. Detect bright flame regions in the grayscale image and construct a binary flame mask as shown in the following formula.

[0142]

[0143] in, The grayscale brightness threshold is used; pixel areas exceeding this value are defined as flame areas.

[0144] In S2.2, the number of obstructions and the size of the obstruction Satisfy the following formula:

[0145]

[0146] .

[0147] S2.2 The number and size of the obstructions are generated randomly;

[0148] In S2.2, the number N and size S of the obstructions satisfy the following formula:

[0149]

[0150] .

[0151] S2.3, Obstruction shape generation;

[0152] S2.3 includes the following steps:

[0153] S2.31, Number of vertices of the occluding polygon The vertex angles are evenly distributed in Above, the following formula is satisfied:

[0154]

[0155] Vertex radius , which follows a uniform integer distribution;

[0156] Vertex coordinates We obtain it from the following formula:

[0157]

[0158] ;

[0159] Translate the polygon vertices to the center of the mask and use the inpolygon function to generate the mask. Determine whether a point is inside the polygon;

[0160] ;

[0161] S2.32. The mask is subjected to expansion and etching operations, and the expansion and etching process is represented by the following formula:

[0162]

[0163]

[0164] In the formula, Indicates the result of the expansion; This represents the expansion operation; The structure elements are randomly generated; Indicates the corrosion result; This represents the erosion operation;

[0165] S2.33, Generate random rotation angle The mask is rotated, and the rotation process is represented by the following formula:

[0166]

[0167] In the formula, Indicates the result of rotation; Indicates rotation operation;

[0168] S2.34. Apply Gaussian blur to the mask. The process can be represented by the following formula:

[0169]

[0170] In the formula, Indicates the Gaussian blur result; For Gaussian kernel, The Gaussian blur parameters are randomly generated; This represents Gaussian blur operation.

[0171] S2.4 Place a shield in the flame area;

[0172] In S2.4, the mask in the flame area Randomly select pixel position The obstruction completely covers the flame area, and the RGB values ​​of the obstruction area are generated random brightness values. This is consistent with the actual coking situation.

[0173] S3. Use the original flame image and the occluded flame image as the learning set of the CNN model to train the model, and use the trained model to reconstruct the flame image with slag captured on site to obtain the flame image without slag after repair.

[0174] The development process for CNN models in S3 is as follows:

[0175] S3.1 Data preparation includes image reading and alignment, normalization processing, and data partitioning;

[0176] S3.11 Image Reading and Alignment: Store the original flame image and the occluded flame image in two folders respectively. The filenames and quantities correspond one-to-one. The two sets can be represented by the following formula:

[0177]

[0178] ;

[0179] S3.12 Normalization Process: The pixel value of each image is linearly normalized from [0,255] to [0,1]. This process can be represented by the following formula.

[0180] ;

[0181] S3.13, Dataset Partitioning: Divide the input images into training and test sets in a 9:1 ratio;

[0182] S3.2 Model Construction: The convolutional neural network consists of convolutional layers, activation function layers, pooling layers, and fully connected layers. Each flame image has 300*403 pixels. The convolutional layers are set with 3*3 convolutional kernels and 64 filters.

[0183] S3.3 Model training includes the following steps:

[0184] S3.31, Training settings: Adam optimizer, maximum number of training epochs: 20, batch size per training session: 32, initial learning rate: 0.0001;

[0185] S3.32. The mean squared error (MSE) is used as the loss function during the training process, as shown in the following formula:

[0186] ;

[0187] This represents the number of training batches for the model.

[0188] S3.4. Use SSIM and PSNR to evaluate the performance of the trained model;

[0189] S3.41, The definition of SSIM is shown in the following formula:

[0190]

[0191]

[0192]

[0193]

[0194] In the formula , , These are the brightness correlation function, contrast correlation function, and structure correlation function between the reference image and the image under test, respectively. The average brightness of the two images; , Standard deviation; For covariance; , , Use very small positive numbers to prevent instability caused by denominators being 0 or close to 0; , , All are positive numbers used to adjust the weights of brightness, contrast, and structural relevance; the SSIM value ranges from [-1, 1]. The closer SSIM is to 1, the stronger the similarity between the two images; when the two images are completely identical... hour, ;

[0195] S3.42, The definition of PSNR is shown in the following formula:

[0196]

[0197] In the formula, The image is to be evaluated; It is the original image; It is the maximum grayscale value of the image. These are the pixels of the image;

[0198] PSNR is measured in dB. The higher the PSNR value, the smaller the difference between the reconstructed image and the original image, and the higher the image quality. When PSNR > 28 dB, the difference between the two images is not obvious; when PSNR > 35 dB, the human eye can hardly see the difference between the two images.

[0199] S3.5 Save the trained model as a .mat file.

[0200] S4. Calculate the temperature of the repaired flame image using the two-color method to obtain the flame temperature field.

[0201] Two-color pyrometry is a non-contact temperature measurement technique that calculates the temperature of an object by measuring the ratio of its thermal radiation intensity at two different wavelengths (or bands). This method is based on the blackbody radiation law and assumes that the object being measured is a gray body (whose emissivity is independent of wavelength or has a known wavelength correlation), thus eliminating the influence of emissivity and simplifying the temperature inversion process.

[0202] The two-color method measures temperature by comparing the radiant intensity of two wavelengths. Combined with calibration curves, it can quickly reconstruct the flame temperature field, making it a common method for combustion diagnostics and industrial furnace monitoring. In practical applications, it is necessary to verify the gray body assumption and optimize wavelength selection to balance accuracy and computational complexity.

[0203] The online monitoring system for the temperature field of a power plant boiler based on convolutional neural networks and flame radiation images is characterized by including a blackbody furnace calibration system, a combustion detection system, and a CNN model system.

[0204] The blackbody furnace calibration system consists of a blackbody furnace, a detector, a detector bracket, a PoE switch, a gigabit network cable, and a computer.

[0205] The blackbody furnace is a device used to generate standard blackbody radiation for calibrating detectors;

[0206] The detector transmits the images captured inside the blackbody furnace to the PoE switch, and then to the computer via a gigabit network cable. The computer processes the obtained image data and plots the calibration curves for the R and G channels.

[0207] The combustion detection system includes A-layer burners, B-layer burners and C-layer burners installed inside the boiler, and six detectors calibrated by a blackbody furnace are evenly spaced on the A-layer burners.

[0208] It also includes switches, electronic control cabinets, DCS, a large screen in the central control room, and new fans;

[0209] The detector is connected to the switch via a gigabit network cable, and the switch is connected to the electronic room control cabinet via an optical fiber. The electronic room control cabinet is connected to the central control room screen via a signal line and to the DCS via RS485. This allows the detector to collect flame image data, transmit it to the switch in the field control cabinet via a gigabit network cable, and then transmit the data to the electronic room server via an optical fiber for processing. Finally, the temperature field results are displayed on the central control room screen. At the same time, the server's calculation results are transmitted to the DCS system via an RS485 signal line to assist in the parameter adjustment of the boiler system.

[0210] The outlet of the newly added fan cools the detector through a cooling duct to ensure that the detector's long-term operating temperature is below 55 degrees Celsius.

[0211] The CNN model system is used to process slag-laden flame images and output non-slag-laden flame images.

[0212] like Figure 4 As shown, using a combustion detection system, 100,000 flame images were selected at uniform time intervals from all images acquired during the boiler's variable load period of 20%-100%. Then, random numbers and sizes of occluders were added to the bright areas of the flames (where R is greater than 40) in these 100,000 flame images to simulate the condition of camera obstruction. The original flame images and the obstructed flame images were used as a training set and input into a convolutional neural network for training. Each flame image had a pixel size of 300*403, with 3*3 convolutional kernels and 64 filters. The training settings used the Adam optimizer, a maximum of 20 training epochs, a batch size of 32 per training session, and an initial learning rate of 0.0001, enabling the model to process slag-laden flame images and output unslag-laden flame images. The trained model achieved an SSIM of 0.9989 and a PSNR of 50.55 dB, demonstrating excellent restoration performance. The trained model was then used to reconstruct the slag-laden images captured on-site, resulting in the restored images.

[0213] like Figure 5 As shown, the temperature fields of the unrepaired and repaired flame images with slag are different. The temperature field of the slag-covered flame exhibits abrupt changes in the flame center region, with the temperature at the slag-covered area being 1000K, which does not match the actual flame temperature. Furthermore, the highest temperature at the flame center reaches 2525K, while the repaired flame temperature field shows a maximum temperature of 1588K, an error of 59%, a significant discrepancy from the actual flame temperature field. Calculation results indicate that the temperature field calculated from flame images with slag obstruction cannot reflect the true situation and has a large deviation. Incorrect calculation results may lead operators to misjudge the boiler's operating status, resulting in incorrect adjustments to operating parameters and seriously affecting the safe and stable operation of the power plant. Images with slag, after being repaired using a well-learned model, can more realistically reflect the actual flame temperature field, helping operators to correctly judge the boiler combustion situation and contributing to the safe production of coal-fired power plants.

[0214] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of the different aspects of the invention as described above, which are not provided in detail for the sake of brevity.

[0215] The embodiments of this invention are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for online monitoring of the temperature field of a power plant boiler based on convolutional neural networks and flame radiation images, characterized in that, Specifically, the following steps are included: S1. Using a blackbody furnace to calibrate the detector, establish the relationship between image intensity and radiation intensity, and use the calibrated detector to collect flame images during the boiler load variation period of 20%-100% to obtain the original flame image. S2. Add random obstructions to the flame area in the original flame image to obtain an obstructed flame image; to simulate the situation where the detector lens is obstructed by slag. S2 specifically includes the following steps: S2.1 Flame area detection; S2.2 The number and size of the obstructions are generated randomly; S2.3, Obstruction shape generation; S2.4 Place a shield in the flame area; S3. Use the original flame image and the occluded flame image as the learning set of the CNN model to train the model, and use the trained model to reconstruct the flame image with slag captured on site to obtain the flame image without slag after repair. S4. Calculate the temperature of the repaired flame image using the two-color method to obtain the flame temperature field.

2. The online monitoring method for power plant boiler temperature field based on convolutional neural network and flame radiation image according to claim 1, characterized in that, In step S1, the process of calibrating the detector on the blackbody furnace is as follows: S1.1 Set the blackbody furnace temperature to 800℃-1440℃, and adjust the exposure time at the same temperature to gradually increase the maximum value of the image R channel to 250, and acquire ten images at each temperature. S1.2 After image acquisition is completed, the obtained image data is processed and the calibration curves of the R channel and G channel are plotted.

3. The online monitoring method for power plant boiler temperature field based on convolutional neural network and flame radiation image according to claim 2, characterized in that, The detector consists of a CCD camera, a high-temperature resistant lens, and a stainless steel housing. Measuring the actual temperature of the flame in S1 using the calibrated detector involves the following steps: S1.3, From Planck's blackbody radiation formula, ； In the formula, It is the blackbody radiation intensity; It is the blackbody temperature; It is the wavelength of radiation; It is the first radiation constant, with a value of ; It is the second radiation constant, with a value of ; S1.4 Since the temperature during the calibration process and the flame temperature measurement process is between 1073-1873K, and the wavelength range is in the visible light band, Planck's blackbody radiation formula can be approximately replaced by Wien's displacement law: ； For a blackbody furnace, the radiation intensities of its R-channel and G-channel are: ； ； A dual bandpass filter is set at the front of the CCD camera, with response wavelengths of 550nm and 650nm respectively, and a half-bandwidth of 10nm. Therefore, the formula... Pick ; Pick ; S1.5 During temperature measurement, the radiation intensity of the actual object's R-channel and G-channel is obtained from the calibration curve. The radiation intensity of the R-channel and G-channel is: ； ； In the formula, For the emissivity of the actual object, The temperature of the actual object; S1.

6. According to the gray body hypothesis, at the same temperature, the emissivity of a gray body is the same at different radiation wavelengths, that is... From the above formula, the actual temperature of the object can be obtained: 。 4. The online monitoring method for power plant boiler temperature field based on convolutional neural network and flame radiation image according to claim 1, characterized in that, S2.1 includes the following steps: S2.11 First, convert the original flame image into a grayscale image; S2.12, The conversion of an RGB image to a grayscale image is achieved by the following formula: ； in, R represents the intensity value of the grayscale image; R, G, and B are the intensities of the three channels of the original image, respectively. These are pixel coordinates; S2.

13. Detect bright flame regions in the grayscale image and construct a binary flame mask as shown in the following formula. ； in, The grayscale brightness threshold is used; pixel areas exceeding this value are defined as flame areas. In S2.2, the number of obstructions and the size of the obstruction Satisfy the following formula: ； 。 5. The online monitoring method for power plant boiler temperature field based on convolutional neural network and flame radiation image according to claim 4, characterized in that, S2.3 includes the following steps: S2.31, Number of vertices of the occluding polygon The vertex angles are evenly distributed in Above, the following formula is satisfied: ； Vertex radius , which follows a uniform integer distribution; Vertex coordinates We obtain it from the following formula: ； ； Use the inpolygon function to generate a mask. The polygon vertices are translated to the center of the mask, and it is determined whether the point is inside the polygon. ； S2.

32. The mask is subjected to expansion and etching operations, and the expansion and etching process is represented by the following formula: ； ； In the formula, Indicates the result of the expansion; This represents the expansion operation; The structure elements are randomly generated; Indicates the corrosion result; This represents the erosion operation; S2.33, Generate random rotation angle The mask is rotated, and the rotation process is represented by the following formula: ； In the formula, Indicates the result of rotation; Indicates rotation operation; S2.

34. Apply Gaussian blur to the mask. The process can be represented by the following formula: ； In the formula, Indicates the Gaussian blur result; For Gaussian kernel, The Gaussian blur parameters are randomly generated; Indicates Gaussian blur operation; In S2.4, the mask in the flame area Randomly select pixel position The obstruction completely covers the flame area, and the RGB values ​​of the obstruction area are generated random brightness values. This is consistent with the actual coking situation.

6. The online monitoring method for power plant boiler temperature field based on convolutional neural network and flame radiation image according to claim 1, characterized in that, The development process for CNN models in S3 is as follows: S3.1 Data preparation includes image reading and alignment, normalization processing, and data partitioning; S3.11 Image Reading and Alignment: Store the original flame image and the occluded flame image in two folders respectively. The filenames and quantities correspond one-to-one. The two sets can be represented by the following formula: ； ； S3.12 Normalization Process: The pixel value of each image is linearly normalized from [0,255] to [0,1]. This process can be represented by the following formula. ； S3.13, Dataset Partitioning: Divide the input images into training and test sets in a 9:1 ratio; S3.2 Model Construction: The convolutional neural network consists of convolutional layers, activation function layers, pooling layers, and fully connected layers. Each flame image has 300*403 pixels. The convolutional layers are set with 3*3 convolutional kernels and 64 filters. S3.3 Model training includes the following steps: S3.31, Training settings: Adam optimizer, maximum number of training epochs: 20, batch size per training session: 32, initial learning rate: 0.0001; S3.

32. The mean squared error (MSE) is used as the loss function during the training process, as shown in the following formula: ； This represents the number of training batches for the model.

7. The online monitoring method for the temperature field of a power plant boiler based on convolutional neural network and flame radiation image according to claim 6, characterized in that, The development of CNN models in S3 also includes the following steps: S3.

4. Use SSIM and PSNR to evaluate the performance of the trained model; S3.41, The definition of SSIM is shown in the following formula: ； ； ； ； In the formula , , These are the brightness correlation function, contrast correlation function, and structure correlation function between the reference image and the image under test, respectively. The average brightness of the two images; , Standard deviation; For covariance; , , Use very small positive numbers to prevent instability caused by denominators being 0 or close to 0; , , All are positive numbers used to adjust the weights of brightness, contrast, and structural relevance; the value range of SSIM is [-1, 1]. The closer SSIM is to 1, the stronger the similarity between the two images; when the two images are completely identical... hour, ; S3.42, The definition of PSNR is shown in the following formula: ； In the formula, The image is to be evaluated; It is the original image; It is the maximum grayscale value of the image. These are the pixels of the image; PSNR is measured in dB. The higher the PSNR value, the smaller the difference between the reconstructed image and the original image, and the higher the image quality. When PSNR > 28 dB, the difference between the two images is not obvious; when PSNR > 35 dB, the human eye can hardly see the difference between the two images. S3.5 Save the trained model as a .mat file.

8. An online monitoring system for the temperature field of a power plant boiler based on convolutional neural networks and flame radiation images, characterized in that, This includes a blackbody furnace calibration system, a combustion detection system, and a CNN model system; The blackbody furnace calibration system consists of a blackbody furnace, a detector, a detector bracket, a PoE switch, a gigabit network cable, and a computer. The blackbody furnace is a device used to generate standard blackbody radiation for calibrating detectors; The detector transmits the images captured inside the blackbody furnace to the PoE switch, and then to the computer via a gigabit network cable. The computer processes the obtained image data and plots the calibration curves for the R and G channels. The combustion detection system includes A-layer burners, B-layer burners and C-layer burners installed inside the boiler, and six detectors calibrated by a blackbody furnace are evenly spaced on the A-layer burners. It also includes switches, electronic control cabinets, DCS, a large screen in the central control room, and new fans; The detector is connected to the switch via a gigabit network cable, and the switch is connected to the electronic room control cabinet via an optical fiber. The electronic room control cabinet is connected to the central control room screen via a signal line and to the DCS via RS485. This allows the detector to collect flame image data, transmit it to the switch in the field control cabinet via a gigabit network cable, and then transmit the data to the electronic room server via an optical fiber for processing. Finally, the temperature field results are displayed on the central control room screen. At the same time, the server's calculation results are transmitted to the DCS system via an RS485 signal line to assist in the parameter adjustment of the boiler system. The newly added fan cools the detector through a cooling duct at its outlet. The CNN model system is used to process slag-laden flame images and output non-slag-laden flame images.

Images