Flue gas emission monitoring method and system based on AI

Abstract

The invention discloses an AI-based flue gas emission monitoring method and system, and relates to the field of flue gas emission monitoring, and the method comprises the steps: carrying out the dynamic color correction and physical characteristic analysis of a real-time flue gas video, building a physical prediction baseline based on a Gaussian plume model, introducing an AI model driven by the macroscopic visual features of the plume, and carrying out the real-time color correction and physical characteristic analysis of the real-time flue gas video. And taking a visual feature vector extracted by the deep network as a key correction factor, performing dynamic compensation and fine correction on a baseline concentration profile generated by the physical model, and finally driving to generate a high-precision pollutant concentration profile. Thus, by organically combining the universality of the physical law and the powerful learning ability of artificial intelligence, remote and non-contact high-precision quantitative monitoring of flue gas emission can be realized, so that the accuracy and robustness of pollutant concentration quantitative inversion in a real industrial scene by a visual monitoring scheme are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of flue gas emission monitoring, and more particularly, to an AI-based flue gas emission monitoring method and system. BACKGROUND

[0002] With the continuous advancement of global industrialization and the deepening of ecological civilization construction, effective monitoring and precise management of industrial fixed pollution sources, especially flue gas emissions, have become the core issue of environmental protection work. Traditional flue gas monitoring mainly relies on extraction or in-situ analysis instruments installed in the flue. Although such methods can provide relatively accurate point data, they generally have problems such as huge equipment investment, complex installation and maintenance, delayed response, and limited coverage of monitoring points, making it difficult to meet the modern needs of comprehensive supervision of wide-area, continuous, and real-time emission conditions. Therefore, developing a low-cost, non-contact, and widely-covered remote monitoring technology is of great significance for improving environmental supervision efficiency and implementing the environmental protection responsibilities of enterprises.

[0003] However, the existing technology mainly uses cameras deployed around the factory to automatically capture the visual features of the plume through image recognition algorithms. These technologies have largely achieved automatic identification and tracking of flue gas emission behavior, such as determining the presence or absence of flue gas, analyzing the diffusion trend of the plume, or making qualitative judgments or rough semi-quantitative classification based on its color and opacity, referring to the Lignellmann blackness scale and other standards. However, these mainstream visual AI solutions have significant bottlenecks in data mining and application. The core challenge is how to accurately reverse the actual concentration distribution of specific pollutants (such as nitrogen oxides and sulfur dioxide) in the flue gas from two-dimensional video images affected by light and environment. Existing technical paths often stop at the apparent analysis of the appearance of flue gas, and cannot establish an accurate mapping relationship between visual features and pollutant concentrations, making it difficult to meet the urgent needs of fine environmental management for quantitative assessment and precise tracing of emission intensity.

[0004] Therefore, there is an urgent need for an optimized AI-based flue gas emission monitoring method and system. SUMMARY

[0005] To solve the above technical problems, the present application is proposed.

[0006] According to an aspect of the present application, an AI-based flue gas emission monitoring method is provided, which includes: based on the reference object ROI and the reference object true color, performing dynamic color correction on the real-time flue gas video frame to obtain a corrected flue gas video frame; performing plume segmentation and center path extraction on the corrected flue gas video frame to obtain a plume mask and a plume center line pixel coordinate; optical depth estimation based on the plume mask and the plume centerline pixel coordinates to obtain an optical depth profile; extracting macro visual features of the plume from the corrected flue gas video frame to obtain a plume visual feature vector; baseline prediction based on a Gaussian plume model on the optical depth profile and the plume centerline pixel coordinates to obtain a baseline concentration profile; AI residual correction based on the plume visual feature vector on the baseline concentration profile to obtain a corrected concentration profile.

[0007] According to another aspect of the present application, an AI-based flue gas emission monitoring system is provided, which comprises: a dynamic color correction module configured to perform dynamic color correction on a real-time flue gas video frame based on a reference object ROI and a reference object true color to obtain a corrected flue gas video frame; a plume segmentation center extraction module configured to perform plume segmentation and center path extraction on the corrected flue gas video frame to obtain a plume mask and plume centerline pixel coordinates; an optical depth estimation module configured to perform optical depth estimation based on the plume mask and the plume centerline pixel coordinates to obtain an optical depth profile; a macro visual feature extraction module configured to extract macro visual features of the plume from the corrected flue gas video frame to obtain a plume visual feature vector; a baseline prediction module configured to perform baseline prediction based on a Gaussian plume model on the optical depth profile and the plume centerline pixel coordinates to obtain a baseline concentration profile; an AI residual correction module configured to perform AI residual correction based on the plume visual feature vector on the baseline concentration profile to obtain a corrected concentration profile.

[0008] Compared with the prior art, the AI-based flue gas emission monitoring method and system provided by the present application can realize remote and non-contact high-precision quantitative monitoring of flue gas emission by dynamically correcting the color and analyzing the physical properties of real-time flue gas video, constructing a physical prediction baseline based on a Gaussian plume model in parallel, introducing an AI model driven by plume macro visual features, using the visual feature vector extracted by the depth network as a key correction factor to dynamically compensate and finely correct the baseline concentration profile generated by the physical model, and finally generating a high-precision pollutant concentration profile. In this way, by organically combining the universality of physical laws and the strong learning ability of artificial intelligence, the accuracy and robustness of the visual monitoring scheme in quantitatively inverting the pollutant concentration in real industrial scenarios can be significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0009] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description thereof taken in conjunction with the accompanying drawings, in which: The accompanying drawings provide exemplary embodiments of the application and serve as an aid in understanding the application. They are a part of and constitute procedures for carrying out the application. They do not limit the scope of the application. In the drawings, like reference numerals refer to like elements or steps throughout.

[0010] Figure 1 Flowchart of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0011] Figure 2 Data flow diagram of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0012] Figure 3 Flowchart of sub-step S1 of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0013] Figure 4 Flowchart of sub-step S2 of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0014] Figure 5 Flowchart of sub-step S3 of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0015] Figure 6 Flowchart of sub-step S35 of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0016] Figure 7 Flowchart of sub-step S5 of an AI-based flue gas emission monitoring method according to an embodiment of the present application.

[0017] Figure 8 Block diagram of an AI-based flue gas emission monitoring system according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] Embodiments of the present disclosure will be described more fully hereinafter with reference to the accompanying drawings, in which some embodiments of the present disclosure are shown. This present disclosure may, however, be embodied in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and fully convey the scope of the present disclosure to those skilled in the art.

[0019] To address the problems in the above background art, the present application proposes an AI-based flue gas emission monitoring method. Figure 1 Flowchart of an AI-based flue gas emission monitoring method according to an embodiment of the present application.Figure 2 A data flow diagram of the AI-based flue gas emission monitoring method according to an embodiment of the present application. As shown in Figure 1 and Figure 2 The AI-based flue gas emission monitoring method includes the steps of: S1, performing dynamic color correction on real-time flue gas video frames based on a reference object ROI and a reference object true color to obtain corrected flue gas video frames; S2, performing plume segmentation and center path extraction on the corrected flue gas video frames to obtain a plume mask and a plume center line pixel coordinate; S3, performing optical depth estimation based on the plume mask and the plume center line pixel coordinate to obtain an optical depth profile; S4, extracting macroscopic visual features of the plume from the corrected flue gas video frames to obtain a plume visual feature vector; S5, performing baseline prediction based on a Gaussian plume model on the optical depth profile and the plume center line pixel coordinate to obtain a baseline concentration profile; and S6, performing AI residual correction on the baseline concentration profile based on the plume visual feature vector to obtain a corrected concentration profile.

[0020] In the above AI-based flue gas emission monitoring method, the step S1, based on the reference object ROI and the reference object true color, performs dynamic color correction on real-time flue gas video frames to obtain corrected flue gas video frames. It can be understood that due to the real-time flue gas video frames being susceptible to changes in sunlight intensity, weather conditions (such as overcast, evening backlight), and environmental light color interference, the RGB pixel values of the plume area deviate from the true color, and if directly used for subsequent plume segmentation and concentration inversion, systematic errors will be introduced. Therefore, the present application further performs dynamic color calibration on the real-time video frames based on the pre-set reference object ROI in the scene, so as to eliminate environmental light interference and restore the true color features of the plume. In this way, color-accurate video data can be provided for subsequent core steps such as plume mask extraction and optical depth calculation, avoiding plume recognition misjudgment or concentration inversion deviation caused by color distortion, ensuring the quantitative accuracy of the entire monitoring system, and ensuring consistent plume visual features under different time periods and different weather conditions.

[0021] In particular, in one specific embodiment, Figure 3 A flowchart of sub-step S1 of the AI-based flue gas emission monitoring method according to an embodiment of the present application. As shown in Figure 3 The step S1 includes: S11, extracting the RGB values of all pixels in the area defined by the reference object ROI from the real-time flue gas video frames; S12, calculating the average value of the RGB values of all pixels in the area to obtain an average observed color vector; S13, calculating a color correction gain factor between the reference object true color and the average observed color vector; and S14, applying the color correction gain factor to the real-time flue gas video frames to obtain the corrected flue gas video frames.

[0022] Specifically, the step S11 extracts the RGB values of all pixels in the region defined by the reference object ROI from the real-time flue gas video frame. It should be understood that, in order to accurately quantify the influence of ambient light on the color of the video, the present application further extracts the RGB values of all pixels in the region defined by the reference object ROI from the real-time flue gas video frame to obtain the actual color data of the reference object under the interference of ambient light. In this way, accurate original data support can be provided for subsequent calculation of color deviation, avoiding judgment errors caused by selecting unstable regions, ensuring that the data source for subsequent color correction has representativeness and reliability, and laying a foundation for the entire dynamic color correction process.

[0023] In particular, in one possible embodiment, the implementation process of the step S11 is as follows: first, by using an image labeling tool, an object with constant color and not affected by flue gas, such as a fixed white signboard in the factory area, is selected in the fixed field of view of the monitoring camera, a rectangular ROI region of the object is defined, and the coordinate range (including the upper left corner and the lower right corner pixel coordinates) of the ROI is stored in the system database. Second, when the system obtains the real-time flue gas video stream, the coordinates of each video frame are located, the region defined by the reference object ROI is accurately framed in the current video frame according to the pre-stored ROI coordinate range. Finally, a pixel reading module is called to traverse each pixel in the ROI region, and the R channel, G channel and B channel values of each pixel are extracted in turn, and these RGB values are temporarily stored in a data buffer area to ensure that the extracted pixels cover all positions in the ROI region without omission or repeated extraction.

[0024] Specifically, the step S12 calculates the average value of the RGB values of all pixels in the region to obtain an average observation color vector. It should be understood that, since individual pixels in the ROI region of the reference object may have slight noise, such as individual pixel color abnormalities caused by sensor errors, if the RGB values of individual pixels are directly used as observation criteria, local errors will be introduced. Therefore, the present application further averages the RGB values of all pixels in the region defined by the reference object ROI to obtain an average observation color vector, so as to eliminate the influence of individual pixel noise on the observation criteria and obtain an observation value that can represent the current overall color state of the reference object. In this way, the observation color can be more consistent with the true color performance of the reference object under the current ambient light, providing a stable and reliable observation criterion for subsequent comparison with the true color of the reference object and calculation of correction factors, and reducing the interference of local abnormal pixels on the correction accuracy.

[0025] In particular, in one possible embodiment, the step S12 is implemented as follows: first, the RGB values of all pixels in the extracted reference object ROI region are retrieved from the data buffer, and the total number of pixels in the region is counted. Second, the R channel values of all pixels are added up respectively to obtain the total sum of R channel values, and the total sum is divided by the total number of pixels to obtain the average value of the R channel. In the same way, the total sum of G channel values and the total sum of B channel values are calculated respectively, and each is divided by the total number of pixels to obtain the average value of the G channel and the average value of the B channel. Finally, the calculated average values of the R channel, the G channel and the B channel are combined in order to form an average observed color vector, and the three components of the vector correspond to the average observed colors of the reference object in the R, G and B channels in the current environment respectively. The average observed color vector is stored in a temporary data file for subsequent use.

[0026] Specifically, the step S13 calculates the color correction gain factor between the reference object true color and the average observed color vector. It should be understood that due to the interference of ambient light, the average observed color vector of the reference object deviates from its real-time true color, and this deviation will cause color distortion of the entire video frame of smoke, and qualitative judgment cannot quantify the degree of deviation and cannot achieve accurate correction. Therefore, the present application further calculates the color correction gain factor by comparing the reference object true color and the average observed color vector, so as to quantify the deviation proportion of each color channel and determine the specific coefficient for correcting the color distortion. In one specific example of the present application, the step S13 includes calculating the color correction gain factor between the reference object true color and the average observed color vector by the following formula: ; wherein, and is the average observed color vector, and is the reference object true color. In this way, accurate quantitative basis can be provided for subsequent overall color correction of the video frame, ensuring that the deviation of each color channel can be corrected specifically, so that the color of the corrected video frame returns to the true state, laying a precise color foundation for subsequent steps such as plume segmentation and concentration inversion.

[0027] Specifically, the step S14 applies the color correction gain factor to the real-time flue gas video frame to obtain the corrected flue gas video frame. It should be understood that, in order to avoid the color features of the plume area deviating from the true state, thereby affecting the accuracy of subsequent plume segmentation and concentration calculation, the application will apply the calculated color correction gain factor to each pixel of the real-time flue gas video frame respectively, so as to globally correct the color of the entire video frame, and eliminate the interference of ambient light on all pixels. In this way, the color of the plume area in the corrected video frame can be restored to a state close to the true illumination, ensuring that the visual features (such as color depth, boundary definition) of the plume are accurate and reliable, providing high-quality image data for subsequent core steps such as plume mask extraction and optical depth estimation, and ensuring the quantitative accuracy of the entire flue gas monitoring system.

[0028] In particular, in one possible embodiment, the implementation process of the step S14 is as follows: first, the color correction gain factors of the R, G and B channels are retrieved from the correction parameter cache area, and the real-time flue gas video frame to be corrected is obtained. Second, each pixel of the video frame is traversed, and for each pixel, its original R channel value, G channel value and B channel value are read respectively. Then, the original R channel value of the pixel is multiplied by the R channel correction gain factor to obtain the corrected R channel value, and the corrected G channel value and B channel value are calculated in the same way. During the calculation process, the corrected value of each channel is range-limited to ensure that it is within the effective interval of pixel values of 0-255, and if the calculation result exceeds this interval, it is truncated to 0 or 255. Finally, the corrected R, G and B channel values of each pixel are recombined to generate the corrected color data of the pixel, and after all the pixels are processed, a complete corrected flue gas video frame is formed and transmitted to the subsequent plume segmentation processing module.

[0029] In the above AI-based flue gas emission monitoring method, the step S2 performs plume segmentation and center path extraction on the corrected flue gas video frame to obtain a plume mask and a plume center line pixel coordinate. It should be understood that, since the plume area in the corrected flue gas video frame is still visually superimposed on the background (such as the sky, the factory wall, and the surrounding vegetation), the application further performs plume segmentation and center path extraction on the corrected flue gas video frame to separate the plume area from the complex background and determine its core diffusion path. In this way, the spatial range of the plume can be accurately defined by the plume mask, completely excluding the interference of background pixels on subsequent calculations, and the core trajectory of plume diffusion can be obtained by the plume center line pixel coordinate, providing accurate spatial position basis for subsequent calculations and predictions, and ensuring the accuracy and reliability of subsequent pollutant concentration profile inversion in the spatial dimension.

[0030] In particular, in one specific embodiment, Figure 4A flowchart of sub-step S2 of the AI-based flue gas emission monitoring method according to the embodiments of the present application. As shown in Figure 4 S2 includes: S21, inputting the corrected flue gas video frame into the instance segmentation model fine-tuned on the smoke data set to obtain the plume mask; and S22, performing center line extraction based on morphological processing on the plume mask to obtain the plume center line pixel coordinates.

[0031] Specifically, the step S21, inputting the corrected flue gas video frame into the instance segmentation model fine-tuned on the smoke data set to obtain the plume mask. In one specific example of the present application, the instance segmentation model fine-tuned on the smoke data set is YOLOv8-seg. It should be understood that the plume has the characteristics of edge blur and gradual change of gray scale due to turbulent diffusion, and it is difficult to accurately distinguish the plume from the background by using traditional threshold segmentation or edge detection methods, and the problem of background pixels mixing into the plume area or losing plume edge pixels is easy to occur. Therefore, the present application further adopts the YOLOv8-seg instance segmentation model specially fine-tuned on the smoke data set to process the corrected video frame, so as to realize the pixel-level accurate segmentation of the plume area. In this way, the complete spatial range of the plume can be effectively extracted, an accurate plume mask can be generated, and the interference of background pixels on subsequent center line extraction and optical depth estimation can be completely eliminated, thereby providing an accurate plume area benchmark for concentration inversion based on a physical model and ensuring the reliability of spatial parameter calculation in subsequent steps.

[0032] In particular, in one possible embodiment, the implementation process of the step S21 is as follows: first, the construction and preprocessing of the smoke data set are performed, on the basis of the general smoke data set, flue gas video frame samples under different industrial pollution source emission scenarios such as thermal power and steel, and different weather conditions such as sunny and cloudy days and backlight are supplemented, the plume area in each frame sample is pixel-level labeled to clearly define the boundary between the plume and the background, and a special data set adapted to the industrial flue gas scene is formed. Then, the model fine-tuning is carried out, the YOLOv8-seg base model and the pre-training weight are loaded, the constructed special data set is divided into a training set and a verification set according to a preset ratio, such as 8:2, the three-order architecture of the model's Backbone (CSPDarknet structure), Neck (PAN-FPN structure) and Head (coupled detection and segmentation design) is based on, the weight parameters of the convolution layer in the Backbone are adjusted to strengthen the multi-scale feature capturing ability of the plume during the training process, and the loss function weight of the Head module is optimized to improve the segmentation accuracy, and the YOLOv8-seg model adapted to the industrial flue gas scene is obtained after fine-tuning. Finally, the corrected flue gas video frame is input into the fine-tuned model frame by frame, the model extracts different scale features of the plume through the Backbone, the features are fused through the Neck, and the binary mask corresponding to the plume area is output by the Head, thereby realizing the accurate segmentation of the plume area.

[0033] Specifically, the step S22 performs center line extraction based on morphological processing on the plume mask to obtain the plume center line pixel coordinates. It should be understood that since the plume mask can only define the overall spatial range of the plume, it cannot provide the core trajectory information of the plume diffusion, therefore, the present application further performs center line extraction based on morphological processing on the plume mask to obtain the core path pixel coordinates of the plume diffusion. In this way, it can provide an accurate starting point for the unit normal vector calculation of each point and the intersection point detection of the ray and the plume boundary in the subsequent optical depth calculation, ensuring the spatial accuracy of the plume width measurement and the optical depth profile generation, laying a reliable spatial parameter foundation for the baseline concentration prediction of the Gaussian plume model, and avoiding the concentration inversion error caused by inaccurate spatial positioning.

[0034] In particular, in one possible embodiment, the implementation process of the step S22 is as follows: first, pre-process the plume mask by using morphological opening operation (erosion followed by dilation) with a 3x3 structure element to remove the small noise points in the mask due to segmentation while maintaining the integrity of the plume main area. Then, process the pre-processed mask by using Zhang-Suen skeletonization algorithm, judge whether the pixel is a non-endpoint, non-branch point and deleting it without destroying the connectivity, and gradually delete the edge pixels until the center skeleton is reserved. Then, post-process the center skeleton by using 5x5 neighborhood smoothing algorithm to remove short branches and jagged protrusions, ensuring the continuity and smoothness of the center line. Finally, traverse the smoothed center skeleton pixels, record the (u, v) coordinate values of each pixel, form the plume center line pixel coordinate list, and store it in the system data buffer area for subsequent optical depth estimation.

[0035] In the above AI-based flue gas emission monitoring method, the step S3 performs optical depth estimation based on the plume mask and the plume center line pixel coordinates to obtain the optical depth profile. It should be understood that since the plume mask can only define the spatial boundary of the plume, and the plume center line can only provide the core path of the plume diffusion, neither of them can directly quantify the optical properties of the plume. Based on this, the present application further combines the area definition function of the plume mask and the path guiding function of the plume center line to carry out optical depth estimation, so as to obtain the optical depth information of the plume at different positions and construct a complete profile. In this way, it can accurately capture the change of the optical properties of the plume from the dense area near the emission port to the thin area at the end of the diffusion, completely exclude the interference of the background area on the optical parameter calculation, provide continuous and reliable optical data basis for the subsequent baseline concentration prediction based on the Gaussian plume model, ensure the accuracy of the physical parameters in the concentration profile inversion process, avoid the concentration calculation deviation caused by the lack of optical property quantization, and ensure the accuracy of the quantitative inversion of the entire monitoring system.

[0036] In particular, in one specific embodiment,Figure 5 A flowchart of sub-step S3 of the AI-based flue gas emission monitoring method according to the embodiments of the present application. As shown in the figure, the step S3 comprises: S31, extracting a first point from the plume centerline pixel coordinates; S32, calculating the unit normal vector of the first point; S33, taking the first point as the starting point, performing ray stepping along the unit normal vector and the reverse normal vector direction thereof, and recording the Euclidean distance between the two intersection points of the ray and the boundary of the plume mask as the plume width of the first point; S34, obtaining the pixel value of the corrected flue gas video frame at the first point, and extracting the luminance value of the blue channel from the pixel value of the corrected flue gas video frame at the first point as the foreground luminance; S35, sampling window interception based on the unit normal vector and the first point in the corrected flue gas video frame to obtain a background sampling window; S36, taking the statistical characteristics of the blue channel luminance in the background sampling window as the background luminance; S37, calculating the optical depth of the first point based on the foreground luminance and the background luminance of the first point. Figure 5 Specifically, the step S31 extracts a first point from the plume centerline pixel coordinates. Specifically, the present application extracts a first point from the plume centerline pixel coordinates to clearly define the first reference position for optical depth calculation. In this way, a specific operation starting point can be provided for subsequent unit normal vector calculation, plume width measurement and other steps, ensuring that each parameter is developed around the same reference point, avoiding calculation confusion caused by ambiguous starting points, and establishing a reusable operation paradigm for subsequent calculation of other points, ensuring the continuity and accuracy of the optical depth profile construction.

[0037] Specifically, the step S32 calculates the unit normal vector of the first point. It should be understood that the measurement direction of the plume width needs to be perpendicular to the plume centerline, and the centerline changes dynamically with the diffusion process. If the direction perpendicular to the centerline is not clear, the plume width measurement will be distorted due to the deviation of the direction, which will affect the accuracy of the optical depth calculation. Therefore, the present application further calculates the unit normal vector of the first point to determine the standard direction perpendicular to the centerline direction of the point. In this way, a precise direction guide can be provided for subsequent ray stepping to measure the plume width, ensuring that the ray always moves perpendicular to the centerline, avoiding width measurement errors caused by direction deviation.

[0038]

[0039] ​In particular, in one possible embodiment, the step S32 is implemented as follows: first, the previous and next adjacent points of the first point in the center line coordinate sequence are obtained, and the tangent direction vector of the center line at the first point is calculated based on the two-point coordinates. Then, the normal direction is derived according to the tangent direction vector, if the tangent vector is (dx, dy), the normal vector is (-dy, dx) or (dy, -dx), and the direction pointing to the outside of the plume is selected as the effective normal direction. Finally, the effective normal direction vector is unitized, that is, the components of the vector are divided by the modulus of the vector, to obtain a unit normal vector with a length of 1, and the component values of the vector are recorded and passed to the subsequent plume width calculation step.

[0040] Specifically, the step S33 is implemented as follows: starting from the first point, ray stepping is performed along the unit normal vector and the reverse normal vector direction, and the Euclidean distance between the two intersection points of the ray and the boundary of the plume mask is recorded as the plume width of the first point. It should be understood that since the plume width is a core parameter for measuring the spatial span of the plume at the point, and the width needs to correspond to the actual span perpendicular to the center line, it is impossible to accurately obtain the span value by visual observation. Therefore, the present application further performs ray stepping along the unit normal and reverse normal directions starting from the first point, records the intersection points with the plume mask boundary and calculates the Euclidean distance, so as to obtain the actual width of the plume at the first point. In this way, the spatial span of the plume perpendicular to the diffusion direction at the point can be accurately measured, which provides a true spatial scale basis for subsequent optical depth calculation, avoids errors in optical depth calculation caused by width estimation deviation, and ensures that the width measurement range is strictly limited within the plume region by ray stepping and mask boundary judgment, thereby excluding background region interference.

[0041] In particular, in one possible embodiment, the step S33 is implemented as follows: first, the ray stepping rule is set, starting from the first point, moving the ray in a pixel-by-pixel manner along the unit normal direction and the reverse normal direction respectively, and every time the ray moves one pixel, it is judged whether the pixel is outside the plume mask, that is, the pixel value is 0. When it is detected that a certain pixel first changes from inside the plume mask (pixel value is 1) to outside the mask, the pixel is recorded as a boundary intersection point, and the coordinates of the boundary intersection points in the two directions are obtained. Finally, according to the pixel coordinates of the two intersection points, the straight-line distance between the two points is calculated by using the Euclidean distance calculation formula, and the distance is stored in the parameter buffer area as the plume width at the first point.

[0042] Specifically, the step S34, the pixel value of the corrected flue gas video frame at the first point is obtained, and the luminance value of the blue channel is extracted from the pixel value of the corrected flue gas video frame at the first point as the foreground luminance. It should be understood that, since the plume is the foreground region, its optical characteristics are mainly reflected in the absorption and scattering of light, and the luminance change of the industrial flue gas under the blue channel can better reflect the concentration difference than the red and green channels. Therefore, the pixel value of the first point in the corrected video frame is further obtained, and the luminance value of the blue channel is extracted as the foreground luminance, so as to accurately characterize the actual optical state of the plume at the point. In this way, the difference in optical characteristics of the plume can be highlighted, and the influence of the background interference information in other channels on the foreground luminance judgment can be avoided.

[0043] Specifically, in one possible embodiment, the implementation process of the step S34 is as follows: first, according to the pixel coordinates of the first point, the pixel data of the first point in the data matrix of the corrected flue gas video frame is located. Then, from the RGB three-channel numerical value of the pixel, the luminance value of the B channel (blue channel) is specially read, and it is stored as the foreground luminance value of the first point in the temporary data area for subsequent comparison calculation with the background luminance.

[0044] Specifically, the step S35, based on the unit normal vector and the first point, a background sampling window is obtained by sampling window interception in the corrected flue gas video frame. It should be understood that, since the background luminance needs to be selected from a pure region outside the plume region and close to the first point, if the sampling is performed in a region far from the first point or other interference (such as equipment, building), the background luminance value will deviate from the true ambient light luminance, thereby affecting the optical depth calculation accuracy. Therefore, the background sampling window is further intercepted in the corrected video frame based on the unit normal vector and the first point, so as to obtain a background region consistent with the ambient light conditions of the first point and free from flue gas interference. In this way, it can be ensured that the background sampling region is in the same lighting environment as the first point, while avoiding the plume and other irrelevant interference regions, to provide a pure and reliable region sample for subsequent background luminance calculation, and avoid the optical depth calculation deviation caused by improper background selection.

[0045] Specifically, in one specific embodiment, Figure 6 The flowchart of the sub-step S35 of the AI-based flue gas emission monitoring method according to the embodiment of the present application is shown in FIG. 6. Figure 6 As shown in FIG. 6, the step S35 includes: S351, moving a preset distance along the unit normal vector direction of the first point to obtain a right background sampling center; S352, moving a preset distance along the reverse normal vector direction of the first point to obtain a left background sampling center; and S353, defining a background sampling window with a size of 5x5 pixels with the right background sampling center and the left background sampling center as centers, respectively.

[0046] More specifically, the step S351, the first point is moved along the direction of its unit normal vector by a preset distance to obtain the right side background sampling center. It should be understood that, since the first point is located on the plume center line, the surrounding area thereof belongs to the plume coverage range, if the background is directly sampled based on the first point, the plume pixels will be mixed to cause the background brightness distortion, and the unit normal vector direction is perpendicular to the plume center line, and moving along the direction can quickly get rid of the plume area. Therefore, the present application further moves the first point along the direction of the unit normal vector by a preset distance to determine the core position of the right side background sampling. In this way, it can be ensured that the right side background sampling center is completely outside the plume mask, avoiding the interference of the plume on the background brightness sampling, and at the same time, through the preset distance control, the sampling center and the first point maintain a reasonable distance, which is far away from the plume and in the same lighting environment, laying a position foundation for subsequent acquisition of pure background brightness.

[0047] More specifically, the step S352, the first point is moved along the direction of its reverse normal vector by a preset distance to obtain the left side background sampling center. It should be understood that, since only the unit normal vector direction is set to the single side background sampling center, the background may not be pure due to the existence of interfering objects such as factory equipment and pipelines on the side, and the reverse normal vector direction is on the other side of the center line, which can provide more comprehensive background selection to avoid the influence of single side interference. Therefore, the present application further moves the first point along the direction of the reverse normal vector by a preset distance to determine the core position of the left side background sampling. In this way, it can realize double side background sampling, eliminate the error caused by single side interference object by comparing the consistency of the brightness of the two sides, and at the same time, the double side sampling can obtain more representative environmental light data, ensuring that the background brightness value can truly reflect the actual lighting conditions of the area where the first point is located, and improving the reliability of subsequent optical depth calculation.

[0048] More specifically, the step S353, the right side background sampling center and the left side background sampling center are respectively taken as the center to define a background sampling window with a size of 5x5 pixels. It should be understood that, since the pixels of a single background sampling center may have local interference such as sensor noise and small dust reflection, using only the brightness value of a single pixel as the background brightness will cause a large random error, and the statistical characteristics of a small size pixel window can smooth the local interference. Therefore, the present application further defines a background sampling window with a size of 5x5 pixels based on the double side sampling center to obtain a sufficient number of background pixel samples. In this way, the error caused by single pixel noise can be eliminated through the brightness statistics (such as mean, median) of multiple pixels in the window, and at the same time, the size of 5x5 can ensure uniform lighting conditions in the window on the premise of ensuring the sample size, avoiding the lighting difference caused by too large window, making the background brightness value more stable and more representative, and providing an accurate background reference for optical depth calculation.

[0049] In particular, in one possible implementation, the steps S351, S352 and S353 are implemented as follows: first, a first point on the plume centerline is taken as a reference, and a preset distance is moved outward along the normal direction perpendicular to the plume centerline, and the distance is set to be sufficient to ensure that the moved position is completely out of the coverage range of the point plume, so as to determine the center position of the right background sampling. Then, the same moving operation is performed in the opposite normal direction to determine the center position of the left background sampling. After the bilateral sampling centers are determined, a background sampling window of a preset size is defined with the two centers as references. Finally, the system performs purity verification on the two sampling windows to check whether the entire window region is located in the pure background area outside the plume mask. If the verification finds that the plume pixels are still contained in the window, the system will automatically adjust the corresponding sampling center further outward along the normal direction, and redefine and verify the sampling window until the entire window region completely meets the pure background condition.

[0050] Specifically, in the step S36, the statistical feature of the blue channel brightness in the background sampling window is taken as the background brightness. It can be understood that, since individual pixels in the background sampling window can have noise such as sensor error and small dust interference, if the blue channel brightness of an individual pixel is directly used as the background brightness, random errors will exist in the background brightness value, which will affect the contrast accuracy of the foreground and the background. Therefore, the statistical feature of the blue channel brightness in the background sampling window is further calculated and taken as the background brightness, so as to eliminate the influence of individual pixel noise on the background brightness. In this way, the noise interference can be smoothed by statistical method, and a more stable and more representative background brightness value can be obtained, so as to ensure that the contrast result of the foreground brightness and the background brightness is closer to the real optical attenuation condition, and to provide a reliable background reference for subsequent optical depth calculation.

[0051] In particular, in one possible implementation, the step S36 is implemented as follows: first, all pixels in the background sampling window are traversed, and the blue channel brightness value of each pixel is extracted to form a blue channel brightness data set. Then, statistical analysis is performed on the data set to calculate the statistical feature, and the mean or median of the data set is usually selected as the core statistical feature in consideration of the need to avoid the influence of extreme outliers. Finally, the calculated mean or median is determined as the background brightness value corresponding to the first point, and the value forms a paired data with the foreground brightness value for subsequent optical depth calculation.

[0052] Specifically, in step S37, the optical depth of the first point is calculated based on the foreground and background brightness. It should be understood that optical depth reflects the degree of light attenuation by the plume, and this attenuation can be quantified by the relationship between foreground brightness (brightness attenuated by the plume) and background brightness (ambient light brightness without attenuation). Therefore, this application further calculates the optical depth of the first point based on the foreground and background brightness to establish a quantitative correlation between visual brightness and plume optical characteristics. This transforms abstract brightness differences into specific optical parameters that can be used for concentration inversion, providing core physical parameter support for subsequent baseline concentration prediction based on the Gaussian plume model, ensuring that the concentration inversion process can be based on real plume optical characteristics, and improving the accuracy of the final concentration profile.

[0053] Specifically, in one possible embodiment, step S37 is implemented as follows: First, the foreground brightness value corresponding to the first point is obtained. and background brightness value Next, the optical depth calculation formula is derived from the Beer-Lambert law. The ratio of foreground brightness to background brightness is substituted into the natural logarithm function for calculation. To ensure the validity of the calculation, it is necessary to... and Preprocessing is performed to avoid zero or Greater than In abnormal situations, for example, it can The value is limited to no more than Within the range. Finally, the calculated optical depth value The coordinates of the centerline of the first point are associated with the data point in the optical depth profile and stored in the profile database.

[0054] In the AI-based flue gas emission monitoring method described above, the step S4 extracts the macroscopic visual features of the plume from the corrected flue gas video frame to obtain a plume visual feature vector. It should be understood that, although the baseline concentration profile generated based on the Gaussian plume model follows the physical law, the macroscopic visual features of the plume, such as the diffusion shape, color depth, and edge blur, in the actual industrial scene will dynamically adjust with the type of pollution source, airflow condition, and meteorological changes. These dynamic differences cannot be completely covered by the physical model, resulting in a possible deviation between the baseline concentration profile and the actual concentration, making it difficult to accurately match the actual emission situation. Therefore, the present application further extracts the macroscopic visual features of the plume from the corrected flue gas video frame and converts them into a plume visual feature vector, in order to capture the dynamic characteristics and detailed differences of the plume that are not covered by the physical model. In this way, key feature inputs can be provided for the subsequent AI residual correction module, so that the AI model can accurately identify the deviation law between the baseline concentration profile and the actual concentration based on these visualized features, and then implement fine correction of the baseline, improve the accuracy of the final pollutant concentration profile, and adapt to complex industrial scenes, avoiding quantitative monitoring errors caused by the limitations of the physical model.

[0055] In particular, in one possible embodiment, the implementation process of the step S4 is as follows: first, based on the plume mask obtained in advance, the complete plume region is accurately framed and cropped in the corrected flue gas video frame, ensuring that the extraction range only contains plume pixels, and completely excluding the interference of background areas such as factory buildings, sky, and equipment on the features. Then, multi-dimensional macroscopic visual feature extraction is carried out, in which the spatial form feature extracts the overall area, length-width ratio, and form irregularity coefficient of the plume, reflecting the degree of asymmetry of the plume diffusion; the color feature extracts the mean, variance of the brightness of the blue channel of the plume region, and the brightness gradient between the core region and the edge region, reflecting the concentration distribution difference of the plume; the texture feature extracts the uniformity index of the internal gray scale distribution of the plume, reflecting the particle distribution state of the plume. Finally, the extracted features are standardized, the numerical value of each feature is converted to a unified interval, and then all standardized numerical values are arranged in order according to the predetermined feature priority order, forming a plume visual feature vector with fixed dimensions.

[0056] In the aforementioned AI-based flue gas emission monitoring method, step S5 involves performing baseline prediction based on a Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline to obtain a baseline concentration profile. It should be understood that since the optical depth profile only reflects the optical attenuation characteristics of the plume, and the pixel coordinates of the plume centerline only provide the spatial diffusion path of the plume, neither can be directly correlated to the actual concentration distribution of pollutants. Therefore, this application further combines the optical information of the optical depth profile with the spatial information of the plume centerline pixel coordinates, and performs baseline prediction using a Gaussian plume model to construct an initial concentration profile that conforms to the physical diffusion laws of the plume. This transforms optical characteristics and spatial location into a concrete pollutant concentration distribution, providing a physically reasonable benchmark framework for subsequent AI residual correction, avoiding deviations from actual diffusion laws that may occur when the AI ​​model learns alone, and ensuring that the baseline concentration profile covers the concentration change trend of the plume from the emission port to the diffusion end, laying a reliable foundation for the generation of the final high-precision concentration profile.

[0057] In particular, in one specific embodiment, Figure 7 This is a flowchart of sub-step S5 of the AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 7 As shown, step S5 includes: S51, performing coordinate system transformation on the pixel coordinates of the plume centerline to obtain a world coordinate path profile; S52, performing path integral density transformation on the optical depth profile to obtain an observed path integral density profile; S53, performing baseline prediction based on a Gaussian plume model on the world coordinate path profile and the observed path integral density profile to obtain the baseline density profile.

[0058] Specifically, in step S51, the pixel coordinates of the plume centerline are transformed to obtain a world coordinate path profile. It should be understood that since the pixel coordinates of the plume centerline only reflect the two-dimensional image position within the corrected flue gas video frame, they cannot correspond to the real three-dimensional spatial position in an industrial scenario. Therefore, this application further performs a coordinate transformation on the pixel coordinates of the plume centerline, converting the pixel coordinates into real-world coordinates to establish a correspondence between the plume diffusion path and the actual space of the industrial scenario. This ensures that the spatial parameters (such as plume diffusion distance and relative position to the emission source) relied upon by the Gaussian plume model during prediction conform to the real scenario, avoiding concentration prediction deviations caused by coordinate dimension mismatches. Simultaneously, it provides a unified spatial benchmark for subsequent path integral concentrations based on observations, ensuring the spatial accuracy of the baseline concentration profile.

[0059] In particular, in one possible embodiment, the step S51 is implemented as follows: first, the internal and external parameters of the monitoring camera are obtained, wherein the internal parameters include camera focal length, pixel size, etc., which are determined by factory calibration data and on-site calibration, and the external parameters include camera installation height, horizontal shooting angle, horizontal distance from the emission source, etc., which are obtained by on-site measurement tools. Then, based on the pinhole camera model in computer vision, the smoke plume centerline pixel coordinates are substituted into the conversion formula, and through the calculation of pixel coordinates and camera parameters, the world coordinates corresponding to each pixel coordinate are obtained, including horizontal distance, vertical height, etc. dimensions, and are arranged in order along the smoke plume diffusion path to form a world coordinate path profile. Finally, the result is verified, fixed landmarks with known true coordinates in the industrial scene are selected, such as the edges of the emission port and the corners of specific equipment, and their pixel coordinates are substituted into the same conversion process. If the error between the obtained world coordinates and the actual measured landmark coordinates is within the preset range, it is confirmed that the world coordinate path profile is valid, otherwise the camera parameters are recalibrated and converted again.

[0060] Specifically, the step S52, the optical depth profile is path-integrated concentration converted to obtain the observed path-integrated concentration profile. It should be understood that since the optical depth profile only represents the degree of attenuation of the smoke plume to the light, it belongs to the optical characteristic parameter and cannot be directly used by the Gaussian smoke plume model for concentration calculation. Therefore, the optical depth profile is further path-integrated concentration converted in the present application, the optical attenuation characteristic is converted into path-integrated data related to pollutant concentration, so as to establish the quantitative correlation between optical information and concentration information. In this way, the Gaussian smoke plume model can be provided with directly usable concentration observation basis, so that the model can optimize the concentration prediction result by comparing the observed path-integrated concentration with the theoretically calculated value, avoid the problem that the model cannot effectively integrate the optical information due to the mismatch of parameter types, and at the same time ensure that the converted concentration data can reflect the cumulative concentration characteristics of the smoke plume along the light propagation path, meeting the input requirements of the model.

[0061] In particular, in one possible embodiment, the step S52 is implemented as follows: first, a conversion relationship between optical depth and path-integrated concentration is determined, which is obtained through a laboratory calibration experiment: in a controllable environment, industrial flue gas samples with known concentrations are prepared, and the optical depth of different concentration samples at different propagation paths is measured to establish a quantitative mapping model between the two. Then the optical depth value of each reference point in the optical depth profile is substituted into the mapping model one by one, and the corresponding path-integrated concentration value of each reference point is calculated. These values are arranged in order along the centerline path of the plume to form the observed path-integrated concentration profile. Finally, the result is calibrated, and a period with known emission concentration in the industrial scene is selected, such as the standard emission concentration during stable operation of the equipment. The observed path-integrated concentration corresponding to this period is compared with the theoretically calculated path-integrated concentration. If the error is within the allowable range, the profile is confirmed to be valid, otherwise the mapping model parameters are adjusted and the conversion is performed again.

[0062] Specifically, the step S53, the world coordinate path profile and the observed path-integrated concentration profile are subjected to baseline prediction based on the Gaussian plume model to obtain the baseline concentration profile. It should be understood that since the world coordinate path profile only provides the real spatial path of plume diffusion, and the observed path-integrated concentration profile only provides the cumulative concentration observation along the path, neither of them can be used alone to generate a three-dimensional concentration distribution consistent with the physical diffusion law of the plume. Therefore, the two profiles are further input into the Gaussian plume model to integrate the spatial path and the observed concentration through the model, and the baseline prediction is carried out to generate an initial concentration profile consistent with the physical diffusion law. In this way, the mature physical description ability of the Gaussian plume model for plume diffusion can be utilized to integrate the spatial parameters and the observation data, and the baseline concentration profile covering the entire diffusion path of the plume and consistent with the actual concentration variation law is obtained, which provides a reliable physical benchmark for subsequent AI residual correction and avoids the concentration prediction deviation that may occur when the AI model is learned alone.

[0063] In particular, in one possible embodiment, the step S53 is implemented as follows: first, the Gaussian plume model parameters are adjusted according to the actual conditions of the industrial scene: the diffusion coefficient of the model is determined according to the real-time meteorological data (such as wind speed and direction), including the horizontal diffusion coefficient and the vertical diffusion coefficient, and the initial value of the source strength parameter of the model is determined according to the actual height and caliber of the emission source. Then, the world coordinate path profile and the observed path integral concentration profile are preprocessed to ensure one-to-one correspondence of the spatial sampling points, if the sampling densities are different, the interpolation method is used to unify the sampling interval, and the preprocessed data are input in the format required by the model. Then, the Gaussian plume model is run, the model calculates the theoretical path integral concentration based on the input spatial path, and through the comparison between the theoretical value and the observed path integral concentration, the source strength, diffusion coefficient and other parameters are iteratively optimized until the error between the theoretical value and the observed value is minimized. Finally, the generated baseline concentration profile is verified, the concentration value near the emission port in the profile is compared with the design emission concentration range of the industrial equipment, if it is in the reasonable interval, the baseline concentration profile is confirmed to be effective, otherwise the model parameters are adjusted again and the prediction is run again.

[0064] In the above AI-based flue gas emission monitoring method, the step S6 is to correct the baseline concentration profile based on the plume visual feature vector to obtain a corrected concentration profile. It should be understood that, in order to avoid the deviation between the baseline concentration profile and the true concentration distribution affecting the accuracy of the quantitative monitoring of the pollutant concentration, the present application further relies on the dynamic information such as macroscopic morphology, color gradient and texture uniformity contained in the plume visual feature vector, and corrects the baseline concentration profile through the AI model to accurately identify and correct the deviation between the baseline and the true concentration. In this way, the learning ability of the AI model for complex dynamic characteristics can be fully utilized, the limitations of the physical model in dealing with non-ideal industrial scenes can be compensated for, the corrected concentration profile not only conforms to the physical diffusion law, but also fits the actual emission state of the plume, significantly improves the adaptability to complex diffusion conditions in industrial scenes, meets the high-precision and high-robustness requirements of environmental protection monitoring for quantitative inversion of pollutant concentration, and provides reliable data support for subsequent emission tracing and compliance judgment.

[0065] In particular, in one possible embodiment, the step S6 is implemented as follows: first, an AI model for residual correction is constructed, which adopts an architecture combining a shallow convolutional neural network and a fully connected layer, the input layer is set to the dimension of the plume visual feature vector, the hidden layer extracts feature correlation information through convolution operation, and the output layer outputs a residual correction amount consistent with the spatial dimension of the baseline concentration profile. Then, training samples are prepared, and flue gas monitoring data under different industrial scenarios such as thermal power and steel, and different meteorological conditions such as sunny and cloudy days, and backlight are collected, each group of samples contains a plume visual feature vector, a corresponding baseline concentration profile, and a real-time concentration profile obtained by a high-precision in-situ analysis instrument, the difference between the baseline concentration and the real concentration is calculated as the residual label to form a complete sample set. Subsequently, model training is performed, the sample set is divided into a training set and a validation set according to a predetermined proportion, the mean square error of the residual prediction value and the real residual label is taken as the loss function, the gradient descent algorithm is used to iteratively optimize the model parameters until the prediction error of the model on the validation set is stable within the predetermined threshold. Then, residual correction is carried out, the plume visual feature vector to be processed is input into the trained AI model to obtain a point-by-point residual correction amount, which is added to the concentration value of the corresponding spatial position in the baseline concentration profile to obtain a preliminary corrected concentration profile. Finally, the result is verified, the corrected concentration profile is compared with the real-time concentration data monitored in-situ point by point, if the overall error is within the industry allowed range of industrial flue gas monitoring, the correction is confirmed to be effective and the final corrected concentration profile is output, otherwise the model hidden layer structure is adjusted or the samples are supplemented and the model is trained again until the correction result meets the accuracy requirement.

[0066] In particular, in another possible preferred embodiment, the step S6 includes: constructing a local feature vector containing the baseline concentration and local features of each point on the baseline concentration profile to obtain a local feature vector sequence; generating a correction embedding vector fused with global context information for each point's local feature vector based on the local feature vector sequence and the plume visual feature vector to obtain a correction embedding vector sequence; applying a multi-head attention mechanism based on the multi-head attention matrix, the correction embedding vector sequence, and a learnable residual prediction matrix to obtain a multi-head attention matrix; generating a residual correction vector based on the multi-head attention matrix, the correction embedding vector sequence, and a learnable residual prediction matrix, and correcting the baseline concentration profile using the residual correction vector to obtain a corrected concentration profile.

[0067] Specifically, in the baseline concentration profile obtained by performing baseline prediction on the world coordinate path profile and the observed path-integrated concentration profile based on the Gaussian plume model, the baseline concentration of each point corresponds to the global nature of the plume visual feature vector. That is, the plume is a continuous fluid, and the concentration, shape, and diffusion state between adjacent points on the profile are highly correlated, that is, have local relevance. The correction value (residual) at a point is likely to be affected by the states of the upstream and downstream points, for example, if the upstream point has experienced severe turbulent diffusion (which can be reflected in the visual features), the concentration decay pattern of the downstream point will deviate further from the standard Gaussian model.

[0068] On the other hand, since the plume visual feature vector is an average description of the entire plume, when the profile is very long and the plume shape differs greatly at different positions, for example, stable laminar flow at the root and turbulent diffusion at the tail, the global average feature distribution of the plume visual feature vector has a large amount of position-specific information relative to each point of the baseline concentration profile, which needs to be fully utilized.

[0069] First, a local feature vector containing the baseline concentration of each point on the baseline concentration profile and the local features is constructed to obtain a sequence of local feature vectors. Specifically, for the baseline concentration profile, the optical depth profile, the world coordinate path profile, and optional other local features such as local width, normal direction, etc., the corresponding baseline concentration CL_basei, optical depth Di, downwind distance xi, and height zi, etc. corresponding to each point i on the profile are obtained to form a local feature vector , and a feature sequence is obtained based on the local feature vectors of all points.

[0070] Then, for each local feature vector , and the plume visual feature vector , first, based on the sequence of local feature vectors and the plume visual feature vector, a correction embedding vector that fuses global context information is generated for each point's local feature vector to obtain a sequence of correction embedding vectors: ; wherein, denotes matrix multiplication, denotes the transpose symbol, denotes the plume visual feature vector, denotes the correction embedding vector, and the vectors here are row vectors, that is, for each point's local feature vector a correction embedding vector that fuses the context information on the entire profile is generated In this way, each local vector not only retains its own details, but also contains the overall state information of the plume. Whether the plume is asymmetrically diffused due to the influence of crosswinds or color gradient differences caused by changes in environmental light, it can be integrated into the local feature vector to avoid correction bias caused by the lack of a global perspective.

[0071] Then, based on the sequence of the correction embedding vectors and the plume visual feature vector, a multi-head attention mechanism is applied to calculate a multi-head attention matrix: ; wherein, represents vector subtraction, represents element-wise multiplication, represents the plume visual feature vector, represents element-wise reciprocal, that is, the reciprocal operation is performed on each element of the vector, represents the element-wise exponential function, that is, the exponential operation is performed on each element of the matrix, represents the multi-head attention matrix.

[0072] That is, in the case of local feature vector fused with global profile context information, the multi-head correlation based on residual prediction of global attention is performed through point-by-point attention residual correspondence, so as to accurately depict the correlation strength between different features and distinguish key correlations from secondary correlations.

[0073] Finally, based on the multi-head attention matrix, the sequence of the correction embedding vectors, and a learnable residual prediction matrix, a residual correction vector is generated: ; wherein, represents the learnable residual prediction matrix, represents the concatenation function, represents the residual correction vector.

[0074] That is, by focusing on the entire plume shape and state based on multi-head attention when correcting the concentration of a point, a more comprehensive and reasonable residual prediction is made. For each point on the profile, dynamically and non-locally focus on the information of all other points on the profile, and explicitly model the long-distance dependence between profile points. For example, it can be learned that if the geometric shape features at the end of the profile show strong downdraft, the baseline concentration in the middle of the profile needs to be corrected negatively even if it looks normal, because the physical model does not consider the rapid dilution caused by downdraft. Finally, the baseline concentration profile is corrected using the residual correction vector to obtain the corrected concentration profile. Specifically, first, the elements of the residual correction vector are one-to-one corresponding to the points on the baseline concentration profile in the order of the emission port, the middle of the diffusion, and the end of the diffusion. Then, the baseline concentration value is numerically superimposed with the corresponding correction amount point by point to obtain the preliminary corrected concentration. The preliminary corrected concentration less than 0 is set to 0 (consistent with the physical meaning of non-negative concentration). Finally, the corrected concentration values are reorganized in the original order to generate the corrected concentration profile, which takes into account the physical laws and feature correlation correction of the Gaussian plume model.

[0075] In summary, the AI-based flue gas emission monitoring method based on the embodiments of the present application is illustrated, which dynamically color corrects and analyzes the physical properties of real-time flue gas video, and constructs a physical prediction baseline based on the Gaussian plume model in parallel. Based on this, an AI model driven by the macro visual features of the plume is introduced, the visual feature vector extracted by the deep network is used as a key correction factor, the baseline concentration profile generated by the physical model is dynamically compensated and finely corrected, and finally a high-precision pollutant concentration profile is generated. In this way, by organically combining the universality of physical laws and the strong learning ability of artificial intelligence, remote, non-contact and high-precision quantitative monitoring of flue gas emission can be realized, thereby significantly improving the accuracy and robustness of the visual monitoring scheme in real industrial scenarios for quantitative inversion of pollutant concentration.

[0076] Figure 8 The block diagram of the AI-based flue gas emission monitoring system according to the embodiments of the present application is shown in FIG. 1. As shown in FIG. 1, the system comprises a video acquisition module 101, a video preprocessing module 102, a video feature extraction module 103, a physical prediction baseline construction module 104, an AI model construction module 105, a concentration profile correction module 106, and a concentration profile output module 107. Figure 8As shown, the AI-based flue gas emission monitoring system 100 according to the embodiments of the present application comprises: a dynamic color correction module 110 configured to perform dynamic color correction on a real-time flue gas video frame based on a reference object ROI and a reference object true color to obtain a corrected flue gas video frame; a plume segmentation center extraction module 120 configured to perform plume segmentation and center path extraction on the corrected flue gas video frame to obtain a plume mask and a plume center line pixel coordinate; an optical depth estimation module 130 configured to perform optical depth estimation based on the plume mask and the plume center line pixel coordinate to obtain an optical depth profile; a macroscopic visual feature extraction module 140 configured to extract a macroscopic visual feature of the plume from the corrected flue gas video frame to obtain a plume visual feature vector; a baseline prediction module 150 configured to perform baseline prediction on the optical depth profile and the plume center line pixel coordinate based on a Gaussian plume model to obtain a baseline concentration profile; and an AI residual correction module 160 configured to perform AI residual correction on the baseline concentration profile based on the plume visual feature vector to obtain a corrected concentration profile.

[0077] As described above, the AI-based flue gas emission monitoring system 100 according to the embodiments of the present application can be implemented in various wireless terminals, such as a server with AI-based flue gas emission monitoring algorithm, etc. In one possible implementation, the AI-based flue gas emission monitoring system 100 according to the embodiments of the present application can be integrated into a wireless terminal as a software module and / or a hardware module. For example, the AI-based flue gas emission monitoring system 100 can be a software module in the operating system of the wireless terminal, or can be an application program developed for the wireless terminal; of course, the AI-based flue gas emission monitoring system 100 can also be one of the many hardware modules of the wireless terminal.

[0078] Alternatively, in another example, the AI-based flue gas emission monitoring system 100 and the wireless terminal can also be separate devices, and the AI-based flue gas emission monitoring system 100 can be connected to the wireless terminal through a wired and / or wireless network, and transmit interactive information in an agreed data format.

[0079] Here, those skilled in the art can understand that the specific operations of each step in the above AI-based flue gas emission monitoring system have been described in detail above with reference to the description of the AI-based flue gas emission monitoring method of Figures 1 to 7 , and therefore, the repeated description thereof will be omitted.

Claims

1. An AI-based flue gas emission monitoring method, characterized by, The method comprises the following steps: based on the reference object ROI and the reference object true color, performing dynamic color correction on the real-time smoke video frame to obtain a corrected smoke video frame; performing plume segmentation and center path extraction on the corrected smoke video frame to obtain a plume mask and plume center line pixel coordinates; based on the plume mask and the plume center line pixel coordinates, performing optical depth estimation to obtain an optical depth profile; extracting macroscopic visual features of the plume from the corrected smoke video frame to obtain a plume visual feature vector; based on the Gaussian plume model, performing baseline prediction on the optical depth profile and the plume center line pixel coordinates to obtain a baseline concentration profile; based on the plume visual feature vector, performing AI residual correction on the baseline concentration profile to obtain a corrected concentration profile. 2.The AI-based flue gas emission monitoring method of claim 1, wherein, Based on the reference object ROI and the reference object true color, the real-time smoke video frame is dynamically color corrected to obtain a corrected smoke video frame, comprising: extracting the RGB values of all pixels in the region defined by the reference object ROI from the real-time smoke video frame; calculate the average value of the RGB values of all pixels in the region to obtain the average observed color vector; calculate the color correction gain factor between the reference object true color and the average observed color vector; apply the color correction gain factor to the real-time smoke video frame to obtain the corrected smoke video frame. 3.The AI-based flue gas emission monitoring method of claim 2, wherein, computing a color correction gain factor between the reference object true color and the average observed color vector, comprising: computing a color correction gain factor between the reference object true color and the average observed color vector with a formula as follows: ; wherein, and is the average observed color vector, and is the reference object true color. 4.The AI-based flue gas emission monitoring method of claim 1, wherein, The corrected smoke video frame is segmented and the center path is extracted to obtain the plume mask and the plume center line pixel coordinates, comprising: input the corrected smoke video frame into the instance segmentation model fine-tuned on the smoke dataset to obtain the plume mask; perform center line extraction based on morphological processing on the plume mask to obtain the plume center line pixel coordinates. 5.The AI-based flue gas emission monitoring method of claim 4, wherein, The instance segmentation model fine-tuned on the smoke dataset is YOLOv8-seg. 6.The AI-based flue gas emission monitoring method of claim 1, wherein, Based on the plume mask and the plume center line pixel coordinates, optical depth estimation is performed to obtain an optical depth profile, comprising: extracting a first point from the plume center line pixel coordinates; calculate the unit normal vector of the first point; starting from the first point, perform ray stepping along its unit normal vector and inverse normal vector direction, and record the Euclidean distance between the two intersection points of the ray and the boundary of the plume mask as the plume width of the first point; obtain the pixel value of the corrected smoke video frame at the first point, and extract the luminance value of the blue channel from the pixel value of the corrected smoke video frame at the first point as the foreground luminance; based on the unit normal vector and the first point, sample window cutting is performed in the corrected smoke video frame to obtain a background sampling window; the statistical characteristics of the blue channel luminance in the background sampling window are taken as the background luminance; based on the foreground luminance and the background luminance of the first point, the optical depth of the first point is calculated. 7.The AI-based flue gas emission monitoring method of claim 6, wherein, Based on the unit normal vector and the first point, sample window cutting is performed in the corrected smoke video frame to obtain a background sampling window, comprising: move a preset distance from the first point along its unit normal vector direction to obtain a right background sampling center; move a preset distance from the first point along its inverse normal vector direction to obtain a left background sampling center; A background sampling window of size 5x5 pixels is defined with the right and left background sampling center as the center, respectively. 8.The AI-based flue gas emission monitoring method of claim 1, wherein, Baseline prediction based on the Gaussian plume model is performed on the optical depth profile and the plume centerline pixel coordinates to obtain a baseline concentration profile, including: Coordinate system conversion is performed on the plume centerline pixel coordinates to obtain a world coordinate path profile; Path integral concentration conversion is performed on the optical depth profile to obtain an observed path integral concentration profile; Baseline prediction based on the Gaussian plume model is performed on the world coordinate path profile and the observed path integral concentration profile to obtain the baseline concentration profile.

9. An AI-based flue gas emission monitoring system characterized in that, It includes: A dynamic color correction module for performing dynamic color correction on real-time smoke video frames based on a reference object ROI and a reference object true color to obtain corrected smoke video frames; A plume segmentation center extraction module for performing plume segmentation and center path extraction on the corrected smoke video frames to obtain a plume mask and plume centerline pixel coordinates; An optical depth estimation module for performing optical depth estimation based on the plume mask and the plume centerline pixel coordinates to obtain an optical depth profile; A macroscopic visual feature extraction module for extracting macroscopic visual features of the plume from the corrected smoke video frames to obtain a plume visual feature vector; A baseline prediction module for performing baseline prediction based on the Gaussian plume model on the optical depth profile and the plume centerline pixel coordinates to obtain a baseline concentration profile; An AI residual correction module for performing AI residual correction on the baseline concentration profile based on the plume visual feature vector to obtain a corrected concentration profile.

Images