Urban solid waste incineration process extreme abnormal flame image countermeasure generation method

Abstract

The adversarial generation method for extreme anomaly flame images in urban solid waste incineration belongs to the field of image control. The work is as follows: 1) For the first time, based on the physical location and imaging principles within the three-dimensional space of the incinerator, the theoretical combustion position in the furnace is accurately calibrated, thereby obtaining a pseudo-labeled sample set under extreme anomaly combustion conditions; 2) For the first time, a cyclic consistent adversarial network is introduced into the research on the generation of extreme anomaly flame samples in the MSWI process, realizing the generation of candidate missing samples; 3) For the first time, a two-level sample evaluation and selection method is proposed. The proposed algorithm can generate extreme anomaly flame images, and the obtained samples have good subjective visual effects. This method can effectively solve the problem of missing flame images in the MSWI process, providing strong support for the quantization of combustion lines and image-based control in the MSWI process.

Description

Technical Field

[0001] This invention belongs to the field of image control. Background Technology

[0002] The combustion line is one of the key parameters characterizing the combustion stability and operational safety of the Municipal Solid Waste Incineration (MSWI) process. Quantifying the combustion line can replace manual observation of the flame from a detection perspective, thereby improving the intelligence level of the MSWI process through real-time feedback. However, combustion line quantification requires a comprehensive flame image template library as a foundation, which can also provide desired settings for image-based control.

[0003] To address the issue of missing flame images in MSWI processes under extreme combustion conditions, this paper proposes a method for generating extreme anomalous flame images based on mechanistic knowledge and adversarial networks. First, based on the mechanistic mapping relationship between three-dimensional spatial locations within the incinerator and pixels, potential extreme anomalous flame samples are analyzed during the combustion process. Pseudo-labeled samples are obtained by translating, stitching, and combining normal sample pixels. Then, CycleGAN is used to transform the pseudo-labeled samples into extreme anomalous flame samples that conform to the distribution of real flame images. Finally, a two-level evaluation framework is proposed to evaluate and screen extreme anomalous flame samples.

[0004] Municipal solid waste (MSW) is influenced by factors such as residents' living habits, seasonal climate, and the degree of sorting, resulting in characteristics such as high impurities, high moisture content, and large fluctuations in calorific value. MSW incineration (MSWI), as a typical MSW treatment method widely used worldwide, offers advantages such as harmlessness, volume reduction, and resource recovery. Currently, MSWI technology still faces many challenges, the most prominent being the instability and volatility of combustion due to the arbitrariness and instability of manual operation, leading to large fluctuations in pollutant emissions. Furthermore, unstable combustion can easily cause coking, ash accumulation, and corrosion within the furnace, and in severe cases, even furnace explosions. Therefore, maintaining a stable combustion state is one of the keys to ensuring efficient operation and compliance with emission standards in the MSWI process.

[0005] Achieving stable combustion in the MSWI process requires timely adjustment of the operating parameters based on changes in the controlled variables. The main controlled variables involved in advanced international combustion control systems include: furnace temperature, flue gas oxygen content, steam flow rate, and the combustion line characterizing the MSWI combustion location (Ding Haixu et al., 2022). Currently, intelligent sensing and detection of the combustion line have not been achieved in China; methods such as…Figure 1 The method of "manual fire observation" shown is subjectively "quantified".

[0006] Figure 1 This indicates that operation experts identify the combustion location based on experience by observing flame images, and then adjust the "air and material distribution" strategy to ensure stable system operation. However, the experience-based method of determining the combustion location by operation experts is subjective and arbitrary, making it difficult to meet the current needs of MSWI process optimization. If a complete flame image template library can be constructed to achieve real-time quantification of the combustion line, it will provide strong support for guiding the intelligent control of the MSWI process. A complete flame template library includes normal flame images of MSW burning at the combustion grate, abnormal flame images of MSW burning at the rear end of the drying grate and the front end of the burnout grate, and extremely abnormal flame images of MSW burning at the front end of the drying grate. MSW burning at the rear end of the burnout grate is also considered an extremely abnormal flame, but it cannot be imaged due to camera angle issues. How to obtain images of extremely abnormal flames is the focus of this paper. Summary of the Invention

[0007] Therefore, to address the problem of missing extreme anomaly samples in the MSWI process, this paper proposes a method for generating extreme anomaly flame images by integrating mechanistic knowledge and adversarial networks. The main contributions of this paper are as follows: 1) For the first time, the theoretical combustion position within the furnace is accurately calibrated based on the physical location in the three-dimensional space of the incinerator and the imaging principles, thereby obtaining a pseudo-labeled sample set under extreme anomaly combustion conditions; 2) For the first time, Cycle-Consistent Adversarial Networks (CycleGAN) are introduced into the research on the generation of extreme anomaly flame samples in the MSWI process, realizing the generation of candidate missing samples; 3) For the first time, a two-level sample evaluation and selection method is proposed.

[0008] In summary, this section proposes the following... Figure 2 The fusion mechanism knowledge and GAN network MSWI process extreme abnormal flame image generation strategy shown includes: pseudo-labeled sample generation module, candidate extreme abnormal flame sample generation module, and extreme abnormal flame sample evaluation and selection module. Figure 2 In the middle, G Real_to_False and G False_to_Real For two generators, D Real and D False For two discriminators, X Real For the real sample set, X False For the pseudo-labeled sample set, X Generated_False =G Real_to_False (X real ) and X Generated_Real =G False_to_Real (Xfalse ) represents the generated image set; X Reconstrute_Real =G False_to_Real (X Generated_False ) and X Reconstrute_False =G Real_to_False (X Generated_Real X is the reconstructed image set. False [.] represents a pseudo-labeled sample in the sample set.

[0009] Figure 2 The functions of different modules are described below:

[0010] 1) Pseudo-label sample acquisition module: Knowledge acquisition is achieved by proportional modeling of the grate, calibration of combustion position based on camera imaging principle, and mapping of three-dimensional spatial coordinates in the furnace to pixels. Pseudo-label samples of three-dimensional spatial position in the furnace are acquired by using precise translation, splicing, and combination of pixels.

[0011] 2) Candidate extreme anomaly flame sample generation module: Its input is pseudo-labeled samples with location information and unlabeled real samples, and its output is extreme anomaly samples. It transforms pseudo-labeled flame images into candidate extreme anomaly flame images through adversarial networks and cycle consistency methods.

[0012] 3) Extreme Anomaly Flame Sample Evaluation and Selection Module: Its inputs are the extreme anomaly samples, pseudo-labeled samples and real samples generated in the previous module, and its output is qualified extreme anomaly samples. The desired sample set is obtained through two-level evaluation and selection of the generated network parameters and generated images.

[0013] 2.1 Pseudo-labeled sample acquisition module

[0014] First, the positional information of the grate and camera is calculated through proportional modeling. Then, the combustion position is calculated and calibrated based on the positional information and the camera imaging principle. Next, the mapping relationship between the three-dimensional spatial position inside the furnace and the pixel is calculated by combining the marking results and the actual camera channel resolution. Finally, normal samples are translated and stitched together based on the mapping relationship and the correction for the thickness deviation of the material layer during the combustion process to obtain pseudo-marked samples. This module includes sub-modules for proportional grate modeling, combustion position calibration based on camera imaging principle, mapping of three-dimensional spatial position to pixel, and pseudo-marked sample acquisition.

[0015] 2.1.1 Grate Proportional Modeling Submodule

[0016] Based on the known grate length and camera position, construct a model proportional to the actual grate dimensions, and calculate the positional information between the grate and the camera, such as... Figure 3 As shown.

[0017] Figure 3The dashed lines in the diagram are auxiliary lines, and points A, B, C, D, E, F, G, H, I, J, K, G, H, I, J, K, L, and Z are important node markers; AB l BC l CD l DE l EF l BF l FG l HG l HI l AK and l HJ These represent the lengths of line segments AB, BC, CD, DE, EF, BF, FG, HG, HI, AK, and HJ, respectively, which include the lengths of each grate, each step, and the position information of the camera. ▲AED, ▲AGF, ▲JHA, ▲AFE, and ▲AHG contain the position information of each grate and camera. The calculation steps for each side length are as follows:

[0018] 1) ▲The AED is a triangular segment of burnt-out section, with three sides of length l. AD l AE and l ED Given ∠ACD = 90°, l AB l DC and l BC Then we can find Given l EC =l ED +l DC Then we can find

[0019] 2) ▲AGF is a triangular combustion segment with three sides of length l. GF l AF and l AG Given ∠ACD = 90°, l GF l AB l DC and l BF Then we can find Given l BG =l BF +l GF Then we can find

[0020] 3) ▲JHA is a triangular drying section with three sides of length l. JH l AH and l AJ Given l HI l BC l JH l AB and l DC Then we can find l KH and l KAThen we can calculate Next, we can seek... Given ∠JHL, we can find ∠JHA = 180° - ∠KAH - ∠JHL; then we can find...

[0021] 4) ▲AFE is a triangle formed by the angle between combustion and burning, with three sides of length l. AF l AE and l EF Given l BF and l ED Then we can find l FZ =l FB -l EC Given l EZ =l BC Then we can find

[0022] 5)▲AGH is a triangle formed by the drying and combustion sections, with the lengths of its three sides being l. AH l AG and l HG Given ∠HIG = 90°, l IG and l HI Then we can find After obtaining the location information inside the furnace, it is necessary to determine its imaging position on the image. Since the imaging principle of the camera is similar to "pinhole imaging," meaning light travels in a straight line, the result is as follows... Figure 4 As shown.

[0023] Figure 4 Chinese: l MN l NO l OP l PQ and l QR The image lengths of the burnout grate, the burnout grate and the combustion grate step, the combustion grate, the combustion grate and the drying grate step, and the drying grate, respectively, are represented by DW, which is perpendicular to DE and intersects AE, AF, AG, AH, and AJ at S, T, U, V, and W, respectively. DS l ST l TU l UV and l VW Let represent the lengths of line segments DS, ST, TU, UV, and VW, respectively. The calibration process is as follows: First, solve for l using the law of cosines at ▲DAS, ▲SAT, ▲TAU, ▲UAV, and ▲VAW. DS l ST l TU l UV and l VW Then, from similar triangles, we can obtain l MN :l NO :lOP :l PQ :l QR =l DS :l ST :l TU : l UV :l VW .

[0024] Depend on Figure 4 and Figure 5 It can be seen that the theoretical imaging results are consistent with the actual imaging results.

[0025] Taking a certain MSWI power plant in Beijing as an example: the resolution of the video captured by the camera is 720*576, that is, the width of the image is 576 pixels; since the drying grate cannot be fully imaged by the camera, the following method is used to calculate the mapping relationship from three-dimensional space to pixel points:

[0026] 1) Let Figure 5 Regions 1, 2, 3, 4, 5, 6, 7, and 8 occupy n1, n2, n3, n4, n5, n6, n7, and n8 pixels respectively, which can be obtained from the measurement. because Given this, we can calculate n6, which is the pixel point mapped to the combustion grate corresponding to PO;

[0027] 2) The pixel point corresponding to the actual position inside the furnace can be obtained from RQ:QP:PO:ON:NM.

[0028] Taking an image with a pixel width of 576 as an example, the calculation results of the mapping relationship between the three-dimensional spatial position and the pixel point without considering the thickness of the material layer are shown in Table 1.

[0029] Table 1. Mapping relationship from 3D space to pixel points

[0030]

[0031] 2.1.4 Pseudo-labeled sample acquisition submodule

[0032] Considering the influence of the material layer thickness, the imaging ratio needs to be corrected by 5% to 10%. MSW combustion at the combustion grate is the desired state, with a quantified combustion line position value of 60% ± 5%. MSW combustion at the front end of the drying grate is an extreme abnormal state, with a combustion line position value of approximately 45%.

[0033] To address the issue of missing images due to extreme anomalies in combustion locations, this paper employs a pixel-level translation and stitching method to obtain pseudo-labeled samples. Let the number of pixels in the image captured by the camera be n. high *n wide The sample set with clear and obvious MSW combustion locations was selected as X', and the generated pseudo-labeled sample set was X'. FalseIf the image is shifted up by n pixels, the following formula applies.

[0034] X false =X'[:,n:,:,:]+X'[:,n wide -n:,:,:] (1)

[0035] The value of n ranges from approximately 20 to 30 pixels.

[0036] This module consists of two generator networks and two discriminator networks. The structure of the generator network is as follows: Figure 6 As shown. By Figure 6 As can be seen, the generative network is mainly divided into downsampling, residual network, and upsampling modules to achieve feature extraction of flame images. The downsampling module consists of three stacks of zero-padding, convolutional layers, instance normalization, and 'ReLU' activation functions: the first zero-padding convolutional kernel is (3,3); the second and third convolutional kernels are (1,1); the first convolutional layer has 64 channels and a kernel of (7,7); the second convolutional layer has 128 channels and a kernel of (3,3); the third convolutional layer has 256 channels and a kernel of (3,3); the axis of the three instance normalization layers is 3; and the three activation functions are all 'ReLU'. In addition to the stacked zero-padding layers, convolutional layers, instance normalization layers, and 'ReLU' activation functions, some input data in the residual modules can be directly passed to the output layer: the kernels of the first and second zero-padding layers are (1,1); the first convolutional layer has 64 channels and a kernel of (7,7); the second convolutional layer has 256 channels and a kernel of (3,3); the axes of the two instance normalization layers are both 3; and all three activation functions are 'ReLU'. There are a total of 9 residual modules with the same structure in this network. The upsampling module consists of two stacked layers: an upsampling layer, a zero-padding layer, a convolutional layer, an instance normalization layer, and a 'ReLU' activation function. The final generated data is then obtained through the zero-padding layer, convolutional layer, instance normalization layer, and 'tanh' activation function. The first and second upsampling layers have convolutional kernels of (2,2); the first and second zero-padding layers have convolutional kernels of (1,1); the third zero-padding layer has convolutional kernels of (3,3); the first convolutional layer has 128 channels and a kernel of (3,3); the second convolutional layer has 64 channels and a kernel of (3,3); the third convolutional layer has 3 channels and a kernel of (7,7); the axis of all three instance normalization layers is 3; the first and second activation functions are 'relu'; the third activation function is 'tanh'.

[0037] The discriminant network consists of convolutional layers, the 'LeakyReLU' activation function, and instance normalization layers, such as... Figure 7 As shown.

[0038] Figure 7In the diagram, the first convolutional layer has 64 channels, a (4,4) kernel, a stride of 1, and 'same' padding; the second convolutional layer has 128 channels, a (4,4) kernel, a stride of 1, and 'same' padding; the third convolutional layer has 256 channels, a (4,4) kernel, a stride of 1, and 'same' padding; the fourth convolutional layer has 512 channels, a (4,4) kernel, a stride of 1, and 'same' padding; the fifth convolutional layer has 1 channel, a (3,3) kernel, a stride of 1, and 'same' padding; all four activation functions are 'LeakyReLU' with a slope alpha of 0.2.

[0039] Depend on Figure 7 As can be seen, the goal of the discriminative network is to obtain the probability that an image is real. The process is as follows: feature extraction is achieved by stacking convolutional layers. LeakyReLU is added between the convolutional layers to increase the nonlinearity of the network while ensuring the stability of the discriminative network in the game with the generator network. Instance normalization can calculate the mean and variance of pixels in each image separately, which avoids mutual influence between images compared to batch normalization. The convolutional layer is used as the output layer to determine the probability of each pixel being real.

[0040] 2.2.2 Network Game Process

[0041] The parameters are updated through a game between the generator network and the discriminator network. This paper uses two generators G. Real_to_False and G False_to_Real 2 discriminators D Real and D False .

[0042] First, obtain the relevant variables: the variable obtained directly is the real sample X. Real Pseudo-labeled sample X False The indirectly obtained variable is the generated image X. Generated_False =G Real_to_False (X Real ) and X Generated_Real =G False_to_Real (X False The reconstructed image is X. Reconstrute_Real =G False_to_Real (X Generated_False ) and X Reconstrute_False =G Real_to_False (X Generated_Real The image used for authentication is X. Real_id =G False_to_Real (X Real ), X False_id =G Real_to_False (X False ); The result of the discriminant network discrimination: YGenerated_False =D False (X Generated_False ), Y False =D False (X False ), Y Generated_Real =D Real (X Generated_Real ) and Y Real =D Real (X Real ); and These represent column vectors that are all 0 and all 1, respectively.

[0043] Next, the generated network is updated as shown in the following formula: In the formula, L G and θ G The generator loss function and network parameters consist of two generators. and G Real_to_False and G False_to_Real The network parameters, L1, L2, L3, L4, L5, and L6, are all loss functions, defined as follows:

[0044]

[0045]

[0046] L3 = L mae (X Reconstrute_Real ,X Real (5)

[0047] L4 = L mae (X Reconstrute_False ,X False (6)

[0048] L5 = L mae (X Real_id ,X Real (7)

[0049] L6 = L mae (X False_id ,X False (8)

[0050] In the formula, L1 and L2 are the losses of a standard GAN, designed to make the generated images more realistic; L3 and L4 are cycle-consistent losses, used to ensure X... Generated_Real and X False and X Real and X Generated_FalseThe correlation; L5 and L6 are Identity losses, used to guarantee G. False_to_Real (X Real The generated samples are still from the real sample set, G Real_to_False (X False The generated samples are still from a pseudo-sample set;

[0051]

[0052]

[0053] In the formula, Y and There are m elements, y i and Representing Y and The i-th element.

[0054] Then, update the two discriminant networks as shown in the following equation:

[0055]

[0056]

[0057] Finally, repeat the above steps, saving the network parameters and the generated extreme anomaly flame images at the end of each loop.

[0058] 2.3 Extreme Anomaly Flame Sample Evaluation and Selection Module

[0059] The requirements for generating the desired extreme anomaly flame image are: (1) conforming to the distribution of the real samples; and (2) preserving the location information of the pseudo-labeled samples. Considering that it is contradictory for general evaluation methods to simultaneously meet the above requirements, this paper proposes a two-level evaluation method for extreme anomaly samples, wherein: the first level of evaluation filters the generator model parameters that can generate samples that conform to the sample distribution; the second level of evaluation filters images that are similar to the distribution of pseudo-labeled samples. The pseudocode is shown in Table 2:

[0060] Table 2. Pseudocode of the two-level evaluation algorithm

[0061]

[0062]

[0063] The input to the FID function is the real sample set X. r Generate a sample set X g The feature extractor outputs FID, and the process is as follows: First, the feature extractor is loaded to extract the feature matrix z of the two sample sets. r With z g Then, calculate the mean μ of the multivariate normal distribution of the feature matrix. r With μg and the covariance matrix Cov r With Cov g Next, the characteristic matrix z is calculated. r Trace T r Finally, calculate the FID score according to equation (13):

[0064]

[0065] In FID calculation, X r The value is X False and X Real X g The value is X Generated The feature extractor uses the Inception V3 model. In the second-level evaluation, FID is positively correlated with location features. Based on subjective visual judgment, when the FID threshold is set to 47, the image can retain the combustion line location features similar to those of the pseudo-labeled samples; therefore, the threshold is set to 47 in this paper.

[0066] Experimental results show that the proposed algorithm can generate extreme anomaly flame images, and the obtained samples have good subjective visual effects. The extreme anomaly flame image generation method that integrates mechanistic knowledge and adversarial networks proposed in this paper can effectively solve the problem of missing flame images in the MSWI process, providing strong support for the realization of combustion line quantization and image-based control in the MSWI process. Attached Figure Description

[0067] Figure 1 MSWI process diagram

[0068] Figure 2 Overall Strategy Diagram

[0069] Figure 3 Proportional Modeling Diagram

[0070] Figure 4 Proportional modeling of the image

[0071] Figure 5 Correspondence between camera images and furnace shutdown photos

[0072] Figure 6 Generative Network Structure

[0073] Figure 7 Determine network structure

[0074] Figure 8 FID between sample sets in Level 1 assessment

[0075] Figure 9 Evaluation results for each generated image in Level 2 evaluation Detailed Implementation

[0076] The data in this article comes from an MSWI power plant in Beijing. Flame videos captured by cameras installed on the rear wall of the furnace are transmitted via coaxial cable. Single-channel video is then acquired through a video capture card and corresponding software on an industrial control computer, and the flame images are stored minute by minute.

[0077] The data source X' for the pseudo-labeled sample acquisition module is selected from 228 images of MSW burning at the grate on July 21, 2021. During the training phase of the candidate extreme anomaly flame sample generation module: X Real Images selected from between 10:23 AM and 5:43 PM on December 6, 2021, totaling 561 images; X False A total of 546 images were generated, representing X' shifted up by 20 pixels and 30 pixels respectively. These were used in the candidate extreme anomaly flame sample generation module testing phase: X Real Images selected from those acquired on July 21, 2021, totaling 495 images; X False There are 228 images in total, each representing an image where X' is shifted up by 25 pixels.

[0078] 3.2 Experimental Results and Analysis

[0079] The pseudo-labeled samples were obtained by translating pixels in a typical combustion state.

[0080] The reasons for the successful generation of candidate extreme anomaly flame samples are as follows: 1) X Generated_Real X was retained False The location information, which is obtained based on knowledge translation operations, is missing in the real sample set; 2) X Generated_Real The image conforms to the causal relationship between features of the furnace flame image, but the pseudo-labeled sample X... False Not satisfied, such as: X Generated_Real The brightness of the flame is significantly lower than that of X. False The brightness of the flame varies with the degree of combustion of the MSW and its distance from the camera. If the MSW is burning at the front of the drying grate, the brightness of this extremely abnormal flame image should be significantly lower than that of the normal combustion position from a causal perspective. However, shifting the position of the flame does not satisfy the above causal relationship; 3) Due to the shift X False The existence of unreasonable feature information in X Generated_Real The above is eliminated, such as: X False There are obvious horizontal lines left by the flipping, while X Generated_Real However, the characteristic information of this error does not exist.

[0081] The first level of evaluation measures the similarity between the generated sample set and the real sample set, and between the generated sample set and the pseudo-labeled sample set, thereby obtaining the parameters for training the model, such as... Figure 8As shown. The second-level evaluation, based on the first-level evaluation, assesses the images in the generated sample set and the images in the pseudo-labeled samples. This paper uses a batch size of 10, and the results are shown below. Figure 9 As shown:

[0082] Samples with an FID less than 47 between the generated sample and the pseudo-labeled sample are selected as qualified samples.

[0083] This paper proposes a novel method for generating extreme anomaly flame images by integrating mechanistic knowledge and adversarial networks. Its innovations include: 1) For the first time, it accurately calibrates the combustion position within the furnace based on the physical location in three-dimensional space and imaging principles; 2) It is the first method to integrate mechanistic knowledge and adversarial networks to generate extreme anomaly flame images, compensating for missing samples in the MSWI process; 3) It proposes a two-level evaluation method to address the difficulty in evaluating and selecting missing extreme anomaly flame samples. Experimental results show that this method can generate extreme anomaly flame images, and the acquired samples have good subjective visual effects.

Claims

1. A method for generating adversarial images of extreme anomalies in urban solid waste incineration processes, characterized by: and For two generators, and There are two discriminators. For the real sample set, For pseudo-labeled sample sets, and For the generated image set; and For the reconstructed image set, This represents a pseudo-labeled sample in the sample set; It includes: a pseudo-labeled sample acquisition module, a candidate extreme anomaly flame sample generation module, and an extreme anomaly flame sample evaluation and selection module; The functions of the different modules are described below: 1) Pseudo-label sample acquisition module: Knowledge acquisition is achieved by proportional modeling of the grate, combustion position calibration based on camera imaging principle, and mapping of three-dimensional spatial coordinates in the furnace to pixels. Pseudo-label samples of three-dimensional spatial positions in the furnace are acquired by accurately translating, splicing, and combining pixels. 2) Candidate extreme anomaly flame sample generation module: Its input is pseudo-labeled samples with location information and unlabeled real samples, and its output is extreme anomaly samples. It transforms pseudo-labeled flame images into candidate extreme anomaly flame images through adversarial networks and cycle consistency methods. 3) Extreme Anomaly Flame Sample Evaluation and Selection Module: Its inputs are the extreme anomaly samples, pseudo-labeled samples and real samples generated in the previous module, and its output is qualified extreme anomaly samples. The desired sample set is obtained through two-level evaluation and selection of the generated network parameters and the generated image. Pseudo-labeled sample acquisition module: First, the position information of the grate and the camera is calculated through proportional modeling. Then, the combustion position is calculated and calibrated based on the position information and the camera imaging principle. Next, the mapping relationship between the three-dimensional spatial position in the furnace and the pixel point is calculated by combining the marking results and the actual camera channel resolution. Finally, normal samples are translated and stitched together based on the mapping relationship and the correction of the material layer thickness deviation during the combustion process to obtain pseudo-marked samples. This module includes sub-modules for proportional grate modeling, combustion position calibration based on camera imaging principle, mapping of three-dimensional spatial position to pixel point, and pseudo-marked sample acquisition. The grate proportional modeling submodule constructs a model proportional to the size of the grate on site based on the known grate length and camera position, and calculates the positional information between the grate and the camera. The pseudo-label sample acquisition submodule: Considering the influence of the material layer thickness, the imaging ratio needs to be corrected by 5%~10%; the MSW combustion at the combustion grate is in the desired state, and the quantized combustion line position value is 60%±5%; Let the pixel of the camera collecting the image be , select the MSW burning position clear sample set as , and the generated pseudo-labeled sample set is , then the image is moved n pixels as shown in the following formula; (1) Where n ranges from 20 to 30 pixels; This module consists of two generator networks and two discriminator networks. The generator network is divided into downsampling, residual, and upsampling modules to extract features from flame images. The downsampling module consists of three stacks of zero-padding, convolutional layers, instance normalization, and ReLU activation functions: the first zero-padding convolutional kernel is (3, 3); the second and third convolutional kernels are (1, 1); the first convolutional layer has 64 channels and a kernel of (7, 7); the second convolutional layer has 128 channels and a kernel of (3, 3); the third convolutional layer has 256 channels and a kernel of (3, 3); the three instance normalization layers all have an axis of 3; and all three activation functions are ReLU. In addition to the stacked zero-padding, convolutional, instance normalization, and ReLU activation functions, the residual module allows some input data to be directly passed to the output layer: the first and second zero-padding layers have a kernel of (1, 1); the first convolutional layer has 64 channels and a kernel of (7, 7). 7); The second convolutional layer has 256 channels and a kernel of (3, 3); Both instance normalization layers have an axis of 3; All three activation functions are ReLU; There are 9 residual modules with the same structure in this network; The upsampling module consists of two stacks of upsampling layer, zero-padding layer, convolutional layer, instance normalization layer and ReLU activation function, and then obtains the final generated data through zero-padding layer, convolutional layer, instance normalization layer and 'tanh' activation function; The convolutional kernels of the first and second upsampling layers are (2, 2); The convolutional kernels of the first and second zero-padding layers are (1, 1); The convolutional kernel of the third zero-padding layer is (3, 3); The first convolutional layer has 128 channels and a kernel of (3, 3); The second convolutional layer has 64 channels and a kernel of (3, 3); The third convolutional layer has 3 channels and a kernel of (7, 3). 7); The axis of the three instance normalization layers is 3; the first and second activation functions are ReLU; the third activation function is 'tanh'; The discriminant network consists of convolutional layers, LeakyReLu activation functions, and instance normalization layers. The first convolutional layer has 64 channels, a (4, 4) kernel, a stride of 1, and the same padding. The second convolutional layer has 128 channels, a (4, 4) kernel, a stride of 1, and the same padding. The third convolutional layer has 256 channels, a (4, 4) kernel, a stride of 1, and the same padding. The fourth convolutional layer has 512 channels, a (4, 4) kernel, a stride of 1, and the same padding. The fifth convolutional layer has 1 channel, a (3, 3) kernel, a stride of 1, and the same padding. All four activation functions are LeakyReLu with a slope alpha of 0.

2. The goal of the discriminative network is to obtain the probability that an image is real. The process is as follows: feature extraction is achieved by stacking convolutional layers. LeakyReLU is added between the convolutional layers to ensure the stability of the discriminative network in the game with the generator network. Instance normalization can calculate the mean and variance of pixels in each image separately, which avoids the mutual influence between images compared with batch normalization. The convolutional layer is used as the output layer to determine the probability of each pixel being real. There are two generators and , two discriminators and ; First, obtain the relevant variables: the variables obtained directly are the real samples. pseudo-labeled samples The variables obtained indirectly include the generated image. and The reconstructed image is and The image used for authentication is , The results of the discrimination network: , , and ; and These represent column vectors that are all 0 and all 1, respectively. Next, the generated network is updated as shown in the following formula: (2) In the formula, and The generator loss function and network parameters consist of two generators. and They represent and Network parameters, Both are loss functions, defined as follows: (3) (4) (5) (6) (7) (8) In the formula, and It is the loss of a regular GAN, and its purpose is to make the generated images more realistic; and It is a cycle-consistent loss, used to guarantee and as well as and The correlation; and It is the Identity loss, used to guarantee The generated samples are still from the real sample set. The generated samples are still from a pseudo-sample set; (9) (10) In the formula, and There are m elements. and They represent and The i-th element; Then, update the two discriminant networks as shown in the following equation: (11) (12) Repeat the above steps, saving the network parameters and the generated extreme anomaly flame image at the end of each loop.

Images