AI-based flue gas emission monitoring methods and systems

Abstract

This application discloses an AI-based flue gas emission monitoring method and system, relating to the field of flue gas emission monitoring. It dynamically corrects the color and analyzes the physical characteristics of real-time flue gas videos, constructing a physical prediction baseline based on a Gaussian plume model in parallel. Based on this, an AI model driven by the macroscopic visual features of the plume is introduced. Visual feature vectors extracted by a deep network are used as key correction factors to dynamically compensate and finely refine the baseline concentration profile generated by the physical model, ultimately driving the generation of a high-precision pollutant concentration profile. In this way, by organically combining the universality of physical laws with the powerful learning capabilities of artificial intelligence, remote, non-contact, high-precision quantitative monitoring of flue gas emissions can be achieved, significantly improving the accuracy and robustness of visual monitoring solutions in quantitative pollutant concentration inversion in real industrial scenarios.

Description

Technical Field

[0001] This application relates to the field of flue gas emission monitoring, and more specifically, to an AI-based flue gas emission monitoring method and system. Background Technology

[0002] With the continuous advancement of global industrialization and the deepening of ecological civilization construction, the effective monitoring and precise treatment of industrial stationary pollution sources, especially flue gas emissions, has become a core issue in environmental protection. Traditional flue gas monitoring mainly relies on extraction or in-situ analytical instruments installed in the flue. While these methods can provide relatively accurate point-to-point data, they generally suffer from problems such as huge equipment investment, complex installation and maintenance, response delays, and limited monitoring point coverage, making it difficult to meet the modern demand for comprehensive monitoring of wide-area, continuous, and real-time emissions. Therefore, developing a low-cost, non-contact, and wide-coverage remote monitoring technology is of vital importance for improving environmental regulatory efficiency and implementing corporate environmental responsibility.

[0003] However, existing technologies primarily utilize cameras deployed around factory areas to automatically capture the visual characteristics of plumes through image recognition algorithms. These technologies largely achieve automatic identification and tracking of flue gas emissions, such as determining the presence or absence of flue gas, analyzing the diffusion trend of plumes, or making qualitative judgments or rough semi-quantitative classifications based on their color and opacity, referencing standards such as the Ringelmann blackness scale. However, these mainstream visual AI solutions face significant bottlenecks in the deep mining and application of data. The core challenge lies in how to accurately deduce the actual concentration distribution of specific pollutants (such as nitrogen oxides and sulfur dioxide) contained in flue gas from two-dimensional video images affected by lighting and the environment. Existing technological approaches often stop at the superficial analysis of flue gas appearance and cannot establish a precise mapping relationship between visual features and pollutant concentrations, failing to meet the urgent needs of refined environmental management for quantitative assessment and accurate source tracing of emission intensity.

[0004] Therefore, there is an urgent need for an optimized AI-based method and system for monitoring flue gas emissions. Summary of the Invention

[0005] This application is made in order to solve the above-mentioned technical problems.

[0006] According to one aspect of this application, an AI-based flue gas emission monitoring method is provided, comprising:

[0007] Based on the reference object ROI and the true color of the reference object, dynamic color correction is performed on the real-time smoke video frame to obtain the corrected smoke video frame.

[0008] Perform plume segmentation and center path extraction on the corrected flue gas video frames to obtain the plume mask and the pixel coordinates of the plume centerline;

[0009] Optical depth estimation is performed based on plume mask and plume centerline pixel coordinates to obtain optical depth profile;

[0010] Macroscopic visual features of the plume are extracted from the corrected smoke video frames to obtain the plume visual feature vector;

[0011] Baseline concentration profiles are obtained by performing baseline prediction based on a Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline.

[0012] AI residual correction is performed on the baseline concentration profile based on the visual feature vector of the smoke plume to obtain the corrected concentration profile.

[0013] According to another aspect of this application, an AI-based flue gas emission monitoring system is provided, comprising:

[0014] The dynamic color correction module is used to perform dynamic color correction on real-time smoke video frames based on the reference object ROI and the real color of the reference object to obtain the corrected smoke video frames.

[0015] The plume segmentation and center extraction module is used to segment the plumes and extract the center path from the corrected smoke video frames to obtain the plume mask and the pixel coordinates of the plume center line.

[0016] The optical depth estimation module is used to perform optical depth estimation based on the plume mask and the pixel coordinates of the plume centerline to obtain an optical depth profile.

[0017] The macroscopic visual feature extraction module is used to extract the macroscopic visual features of the plume from the corrected smoke video frames to obtain the plume visual feature vector.

[0018] The baseline prediction module is used to perform baseline prediction based on the Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline to obtain the baseline concentration profile.

[0019] The AI ​​residual correction module is used to perform AI residual correction on the baseline concentration profile based on the visual feature vector of the plume to obtain the corrected concentration profile.

[0020] Compared with existing technologies, this application provides an AI-based flue gas emission monitoring method and system. It dynamically corrects the color and analyzes the physical characteristics of real-time flue gas videos, constructing a physical prediction baseline based on a Gaussian plume model in parallel. Based on this, an AI model driven by the macroscopic visual features of the plume is introduced. Visual feature vectors extracted by a deep network are used as key correction factors to dynamically compensate and finely refine the baseline concentration profile generated by the physical model, ultimately driving the generation of a high-precision pollutant concentration profile. In this way, by organically combining the universality of physical laws with the powerful learning capabilities of artificial intelligence, remote, non-contact, high-precision quantitative monitoring of flue gas emissions can be achieved, significantly improving the accuracy and robustness of visual monitoring solutions in quantitative pollutant concentration retrieval in real industrial scenarios. Attached Figure Description

[0021] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0022] Figure 1 This is a flowchart of an AI-based flue gas emission monitoring method according to an embodiment of this application.

[0023] Figure 2 This is a data flow diagram of an AI-based flue gas emission monitoring method according to an embodiment of this application.

[0024] Figure 3 This is a flowchart of sub-step S1 of the AI-based flue gas emission monitoring method according to an embodiment of this application.

[0025] Figure 4 This is a flowchart of sub-step S2 of the AI-based flue gas emission monitoring method according to an embodiment of this application.

[0026] Figure 5 This is a flowchart of sub-step S3 of the AI-based flue gas emission monitoring method according to an embodiment of this application.

[0027] Figure 6 This is a flowchart of sub-step S35 of the AI-based flue gas emission monitoring method according to an embodiment of this application.

[0028] 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.

[0029] Figure 8This is a block diagram of an AI-based flue gas emission monitoring system according to an embodiment of this application. Detailed Implementation

[0030] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0031] To address the problems mentioned above in the background technology, this application proposes an AI-based flue gas emission monitoring method. Figure 1 This is a flowchart of an AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 2 This is a data flow diagram of an AI-based flue gas emission monitoring method according to an embodiment of this application. (See diagram below.) Figure 1 and Figure 2 As shown, the AI-based flue gas emission monitoring method includes the following steps: S1, performing dynamic color correction on real-time flue gas video frames based on the reference object ROI and the true color of the reference object to obtain corrected flue gas video frames; S2, performing plume segmentation and center path extraction on the corrected flue gas video frames to obtain plume masks and plume centerline pixel coordinates; S3, performing optical depth estimation based on the plume masks and plume centerline pixel coordinates to obtain optical depth profiles; S4, extracting macroscopic visual features of the plumes from the corrected flue gas video frames to obtain plume visual feature vectors; S5, performing baseline prediction based on a Gaussian plume model on the optical depth profile and plume centerline pixel coordinates to obtain baseline concentration profiles; S6, performing AI residual correction on the baseline concentration profiles based on the plume visual feature vectors to obtain corrected concentration profiles.

[0032] In the aforementioned AI-based flue gas emission monitoring method, step S1 involves dynamically color-correcting the real-time flue gas video frame based on the reference object ROI and its true color to obtain a corrected flue gas video frame. It should be understood that real-time flue gas video frames are susceptible to interference from changes in sunlight intensity, weather conditions (such as cloudy days or evening backlighting), and ambient light color, causing the RGB pixel values ​​of the plume region to deviate from their true colors. Directly using these values ​​for subsequent plume segmentation and concentration inversion would introduce systematic errors. Therefore, this application further performs dynamic color calibration on the real-time video frame based on a preset reference object ROI in the scene to eliminate ambient light interference and restore the true color characteristics of the plume. This provides color-accurate video data for subsequent core steps such as plume mask extraction and optical depth calculation, avoiding misjudgments in plume identification or deviations in concentration inversion due to color distortion, ensuring the quantization accuracy of the entire monitoring system, and ensuring consistent plume visual characteristics are obtained under different time periods and meteorological conditions.

[0033] In particular, in one specific embodiment, Figure 3 This is a flowchart of sub-step S1 of the AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 3 As shown, step S1 includes: S11, extracting the RGB values ​​of all pixels in the region defined by the reference object ROI from the real-time smoke video frame; S12, calculating the average value of the RGB values ​​of all pixels in the region to obtain the average observed color vector; S13, calculating the color correction gain factor between the real color of the reference object and the average observed color vector; S14, applying the color correction gain factor to the real-time smoke video frame to obtain the corrected smoke video frame.

[0034] Specifically, in step S11, the RGB values ​​of all pixels within the region defined by the reference object ROI are extracted from the real-time smoke video frame. It should be understood that, in order to accurately quantify the impact of ambient light on video color, this application further locks the region defined by the reference object ROI in the real-time smoke video frame and extracts the RGB values ​​of all pixels within that region to obtain the actual color data of the reference object under ambient light interference. This provides accurate raw data support for subsequent color deviation calculations, avoids errors in judgment due to interference caused by selecting unstable regions, ensures the representativeness and reliability of the data source for subsequent color correction, and lays the foundation for the entire dynamic color correction process.

[0035] Specifically, in one possible embodiment, step S11 is implemented as follows: First, using an image annotation tool, an object with a constant color and unaffected by smoke is selected within the fixed field of view of the monitoring camera, such as a fixed white sign in the factory area. Its rectangular Region of Interest (ROI) is delineated, and the coordinate range of the ROI (including the pixel coordinates of the upper left and lower right corners) is stored in the system database. Second, after the system acquires the real-time smoke video stream, coordinate positioning is performed on each video frame. Based on the pre-stored ROI coordinate range, the area defined by the reference object ROI is precisely selected in the current video frame. Finally, the pixel reading module is invoked to traverse each pixel within the ROI area, sequentially extracting the R, G, and B channel values ​​of each pixel. These RGB values ​​are temporarily stored in the data buffer to ensure that the extracted pixels cover all positions within the ROI area without omissions or duplicate extractions.

[0036] Specifically, in step S12, the average RGB values ​​of all pixels within the region are calculated to obtain the average observed color vector. It should be understood that individual pixels within the ROI region of the reference object may contain minor noise, such as color anomalies in individual pixels caused by sensor errors. Directly using the RGB values ​​of a single pixel as the observation benchmark would introduce local errors. Therefore, this application further averages the RGB values ​​of all pixels within the defined region of the reference object ROI to obtain the average observed color vector. This eliminates the influence of individual pixel noise on the observation benchmark, obtaining an observed value that represents the current overall color state of the reference object. This allows the observed color to more closely match the true color performance of the reference object under the current ambient light, providing a stable and reliable observation benchmark for subsequent comparison with the true color of the reference object and calculation of correction factors, reducing the interference of local abnormal pixels on correction accuracy.

[0037] Specifically, in one possible embodiment, step S12 is implemented as follows: First, the RGB values ​​of all pixels within the extracted ROI region of the reference object 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 summed to obtain the total R channel values, and then this sum is divided by the total number of pixels to obtain the average R channel value. Similarly, the sum of the G channel values ​​and the sum of the B channel values ​​are calculated and divided by the total number of pixels to obtain the average G channel value and the average B channel value. Finally, the calculated average R channel value, average G channel value, and average B channel value are combined sequentially to form an average observed color vector. The three components of this vector correspond to the average observed color of the reference object in the current environment for the R, G, and B channels, respectively, and are stored in a temporary data file for later use.

[0038] Specifically, step S13 involves calculating the color correction gain factor between the true color of the reference object and the average observed color vector. It should be understood that due to ambient light interference, the average observed color vector of the reference object deviates from its real-time true color. This deviation leads to color distortion throughout the smoke video frame, and qualitative judgment alone cannot quantify the degree of this deviation, thus hindering accurate correction. Therefore, this application further calculates the color correction gain factor by comparing the true color of the reference object with the average observed color vector, thereby quantifying the deviation ratio of each color channel and determining the specific coefficient used to correct color distortion. In a specific example of this application, step S13 includes: calculating the color correction gain factor between the true color of the reference object and the average observed color vector using the following formula: ;in, and The average observed color vector, and This serves as the reference for the true color of the object. This provides a precise quantitative basis for subsequent overall color correction of the video frames, ensuring that deviations in each color channel can be specifically corrected, allowing the corrected video frame colors to return to their true state. This lays a precise color foundation for subsequent steps such as plume segmentation and density inversion.

[0039] Specifically, in step S14, the color correction gain factor is applied to the real-time flue gas video frame to obtain the corrected flue gas video frame. It should be understood that, to avoid the color characteristics of the plume region deviating from the true state, thus affecting the accuracy of subsequent plume segmentation and concentration calculation, this application applies the calculated color correction gain factor to each pixel of the real-time flue gas video frame to globally correct the color of the entire video frame, eliminating the interference of ambient light on all pixels. This allows the color of the plume region in the corrected video frame to be restored to a state close to that under real lighting, ensuring the accuracy and reliability of the plume's visual characteristics (such as color depth and boundary clarity), providing high-quality image data for subsequent core steps such as plume mask extraction and optical depth estimation, and ensuring the quantization accuracy of the entire flue gas monitoring system.

[0040] Specifically, in one possible embodiment, step S14 is implemented as follows: First, the color correction gain factors for the R, G, and B channels are retrieved from the correction parameter buffer, and the real-time smoke video frame to be corrected is acquired simultaneously. 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. 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. The corrected G channel values ​​and B channel values ​​are calculated using the same method. During the calculation process, the range of the corrected value for each channel is limited to ensure it falls within the valid pixel value range of 0-255. If the calculation result exceeds this range, it is truncated to 0 or 255. Finally, the corrected R, G, and B channel values ​​for each pixel are recombine to generate the corrected color data for that pixel. After all pixels have been processed, a complete corrected smoke video frame is formed and transmitted to the subsequent plume segmentation processing module.

[0041] In the aforementioned AI-based flue gas emission monitoring method, step S2 involves performing plume segmentation and center path extraction on the corrected flue gas video frame to obtain the plume mask and the pixel coordinates of the plume centerline. It should be understood that since the plume region in the corrected flue gas video frame still visually overlaps with the background (such as the sky, factory walls, and surrounding vegetation), this application further performs plume segmentation and center path extraction operations on the corrected flue gas video frame to separate the plume region from the complex background and determine its core diffusion path. This allows for precise definition of the plume's spatial range through the plume mask, completely eliminating the interference of background pixels on subsequent calculations. Simultaneously, the core trajectory of plume diffusion is obtained through the pixel coordinates of the plume centerline, providing accurate spatial location information for subsequent calculations and predictions, ensuring the accuracy and reliability of the subsequent pollutant concentration profile inversion in the spatial dimension.

[0042] In particular, in one specific embodiment, Figure 4 This is a flowchart of sub-step S2 of the AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 4 As shown, step S2 includes: S21, inputting the corrected smoke video frame into an instance segmentation model finely tuned on the smoke dataset to obtain the plume mask; S22, performing morphological processing-based centerline extraction on the plume mask to obtain the pixel coordinates of the plume centerline.

[0043] Specifically, in step S21, the corrected smoke video frame is input into an instance segmentation model fine-tuned on a smoke dataset to obtain the plume mask. In a specific example of this application, the instance segmentation model fine-tuned on the smoke dataset is YOLOv8-seg. It should be understood that smoke plumes exhibit blurred edges and gradual grayscale changes due to turbulent diffusion. Traditional threshold segmentation or edge detection methods struggle to accurately distinguish between the plume and the background, easily leading to background pixels being mixed into the plume region or plume edge pixels being lost. Therefore, this application further employs a YOLOv8-seg instance segmentation model specifically fine-tuned on a smoke dataset to process the corrected video frame, thereby achieving pixel-level accurate segmentation of the plume region. This effectively extracts the complete spatial range of the plume, generates an accurate plume mask, completely eliminates the interference of background pixels on subsequent centerline extraction and optical depth estimation, provides an accurate plume region benchmark for concentration inversion based on a physical model, and ensures the reliability of spatial parameter calculations in subsequent steps.

[0044] Specifically, in one possible embodiment, step S21 is implemented as follows: First, a smoke dataset is constructed and preprocessed. Based on a general smoke dataset, emission scenarios from different industrial pollution sources such as thermal power and steel, as well as smoke video frame samples under different meteorological conditions such as sunny days, cloudy days, and backlighting, are added. Pixel-level annotations are performed on the plume regions in each frame sample to clarify the boundary between the plume and the background, forming a dedicated dataset adapted to industrial smoke scenarios. Next, model fine-tuning is carried out. The YOLOv8-seg basic model and pre-trained weights are loaded. The constructed dedicated dataset is divided into a training set and a validation set according to a preset ratio, such as 8:2. Based on the model's three-stage architecture of Backbone (CSPDarknet structure), Neck (PAN-FPN structure), and Head (coupled detection and segmentation design), the weight parameters of the convolutional layers in the Backbone are adjusted during training to enhance the ability to capture multi-scale features of the plume. At the same time, the loss function weights of the Head module are optimized to improve segmentation accuracy. After fine-tuning, a YOLOv8-seg model adapted to industrial smoke scenarios is obtained. Finally, the corrected smoke video frames are input frame by frame into the fine-tuned model. The model extracts features of different scales of the plume through the Backbone, performs feature fusion through the Neck, and then outputs the binarized mask corresponding to the plume region from the Head, thus achieving accurate segmentation of the plume region.

[0045] Specifically, in step S22, the centerline of the plume mask is extracted based on morphological processing to obtain the pixel coordinates of the plume centerline. It should be understood that since the plume mask can only define the overall spatial range of the plume and cannot provide core trajectory information of plume diffusion, this application further extracts the centerline of the plume mask based on morphological processing to obtain the pixel coordinates of the core path of plume diffusion. This provides a precise starting point for the calculation of the unit normal vector at each point and the detection of the intersection point between the ray and the plume boundary in subsequent optical depth calculations, ensuring the spatial accuracy of plume width measurement and optical depth profile generation. This lays a reliable spatial parameter foundation for baseline concentration prediction in the Gaussian plume model and avoids concentration inversion errors caused by inaccurate spatial positioning.

[0046] Specifically, in one possible embodiment, step S22 is implemented as follows: First, the plume mask is preprocessed using morphological opening operations (erosion followed by dilation) with 3×3 structuring elements to remove minor noise points remaining from segmentation while maintaining the integrity of the main plume region. Next, the Zhang-Suen skeletonization algorithm is used to process the preprocessed mask. By iteratively traversing edge pixels, it is determined whether a pixel is a non-endpoint, non-branch point, and whether deletion will not disrupt connectivity. Edge pixels are gradually deleted until the central skeleton is retained. Then, the central skeleton is post-processed using a 5×5 neighborhood smoothing algorithm to remove short branches and jagged protrusions, ensuring a continuous and smooth centerline. Finally, the smoothed central skeleton pixels are traversed, and the (u, v) coordinates of each pixel are recorded to form a list of plume centerline pixel coordinates, which is stored in the system data cache for subsequent optical depth estimation.

[0047] In the aforementioned AI-based flue gas emission monitoring method, step S3 involves optical depth estimation based on the plume mask and the pixel coordinates of the plume centerline to obtain an optical depth profile. It should be understood that since the plume mask only defines the spatial boundary of the plume, and the plume centerline only provides the core path of plume diffusion, neither can directly quantify the optical characteristics of the plume. Therefore, this application further combines the region-defining function of the plume mask with the path-guiding role of the plume centerline to perform optical depth estimation, thereby obtaining optical depth information at different locations of the plume and constructing a complete profile. This allows for accurate capture of the changes in optical characteristics of the plume from the dense region at the emission port to the sparse region at the diffusion end, completely eliminating the interference of the background region on the calculation of optical parameters. This provides a continuous and reliable optical data foundation for subsequent baseline concentration prediction based on the Gaussian plume model, ensuring the accuracy of physical parameters during concentration profile inversion, avoiding concentration calculation deviations caused by the lack of optical characteristic quantification, and guaranteeing the accuracy of the quantitative inversion of the entire monitoring system.

[0048] In particular, in one specific embodiment, Figure 5This is a flowchart of sub-step S3 of the AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 5 As shown, step S3 includes: S31, extracting a first point from the pixel coordinates of the plume centerline; S32, calculating the unit normal vector of the first point; S33, taking the first point as the starting point, performing ray stepping along its unit normal vector and reverse normal vector directions, 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 smoke video frame at the first point, and extracting the brightness value of the blue channel from the pixel value of the corrected smoke video frame at the first point as the foreground brightness; S35, performing sampling window truncation in the corrected smoke video frame based on the unit normal vector and the first point to obtain a background sampling window; S36, using the statistical characteristics of the blue channel brightness within the background sampling window as the background brightness; S37, calculating the optical depth of the first point based on the foreground brightness and background brightness of the first point.

[0049] Specifically, in step S31, a first point is extracted from the pixel coordinates of the plume centerline. In particular, this application defines the first reference position for optical depth calculation by extracting the first point from the pixel coordinates of the plume centerline. This provides a specific starting point for subsequent steps such as unit normal vector calculation and plume width measurement, ensuring that all parameters are developed around the same reference point, avoiding calculation confusion caused by an ambiguous starting point, and establishing a reusable operational paradigm for the calculation of other points, thus ensuring the continuity and accuracy of the optical depth profile construction.

[0050] Specifically, in step S32, the unit normal vector of the first point is calculated. It should be understood that since the measurement direction of the plume width must be perpendicular to the plume centerline, and the direction of the centerline changes dynamically with the diffusion process, if the direction perpendicular to the centerline is not clearly defined, the plume width measurement will be distorted due to directional deviation, thus affecting the accuracy of optical depth calculation. Therefore, this application further calculates the unit normal vector of the first point to determine the standard direction perpendicular to the direction of the centerline at that point. This provides accurate directional guidance for subsequent ray stepping measurements of the plume width, ensuring that the ray always moves along a direction perpendicular to the centerline, avoiding width measurement errors caused by directional deviation.

[0051] Specifically, in one possible embodiment, step S32 is implemented as follows: First, obtain the preceding and following adjacent points of the first point in the centerline coordinate sequence, and calculate the tangent direction vector of the centerline at the first point using the coordinates of the two points. Next, derive the normal direction based on the tangent direction vector. If the tangent vector is (dx, dy), then the normal vector is (-dy, dx) or (dy, -dx), and select the direction pointing outward of the plume as the effective normal direction. Finally, normalize the effective normal direction vector by dividing each component of the vector by the magnitude of the vector to obtain a unit normal vector of length 1. Record the component values ​​of this vector and pass them into the subsequent plume width calculation step.

[0052] Specifically, in step S33, starting from the first point, a ray is traversed along its unit normal vector and anti-normal vector, 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 at the first point. It should be understood that since plume width is a core parameter for measuring the spatial span of the plume at that point, and the width must correspond to the actual span perpendicular to the centerline, this span value cannot be accurately obtained solely through visual observation. Therefore, this application further uses the first point as the starting point, ray traversed along the unit normal and anti-normal vectors, records the intersection points with the plume mask boundary, and calculates the Euclidean distance to obtain the actual width of the plume at the first point. This allows for accurate measurement of the spatial span perpendicular to the plume diffusion direction at that point, providing a true spatial scale basis for subsequent optical depth calculations, avoiding optical depth calculation errors caused by width estimation deviations. Simultaneously, through ray traversal and mask boundary judgment, it ensures that the width measurement range is strictly limited to the plume area, eliminating background interference.

[0053] Specifically, in one possible embodiment, step S33 is implemented as follows: First, a ray stepping rule is set. Starting from the first point, the ray moves pixel by pixel along both the unit normal direction and the reverse normal direction. Each pixel moved determines whether the pixel is outside the plume mask (i.e., the pixel value is 0). When a pixel is detected to first change from inside the plume mask (pixel value 1) to outside the mask, this pixel is recorded as a boundary intersection point, and the coordinates of the boundary intersection point in both directions are obtained. Finally, based on the pixel coordinates of the two intersection points, the Euclidean distance formula is used to calculate the straight-line distance between the two points. This distance is the plume width at the first point, and it is stored in the parameter buffer.

[0054] Specifically, in step S34, the pixel value of the corrected flue gas video frame at the first point is obtained, and the brightness 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 brightness. 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 brightness change of industrial flue gas in the blue channel reflects its concentration difference more effectively than in the red and green channels. Therefore, this application further obtains the pixel value of the first point in the corrected video frame and extracts the brightness value of the blue channel as the foreground brightness to accurately characterize the actual optical state of the plume at that point. This highlights the differences in the optical characteristics of the plume and avoids the influence of background interference information in other channels on the foreground brightness judgment.

[0055] Specifically, in one possible embodiment, step S34 is implemented as follows: First, based on the pixel coordinates of the first point, the pixel data of that point is located in the data matrix of the corrected smoke video frame. Then, the brightness value of the B channel (blue channel) is specifically read from the RGB three-channel values ​​of that pixel and used as the foreground brightness value of the first point, stored in a temporary data area for subsequent comparison calculation with the background brightness.

[0056] Specifically, in step S35, a sampling window is extracted from the corrected smoke video frame based on the unit normal vector and the first point to obtain a background sampling window. It should be understood that since the background brightness needs to be selected from a clean area outside the smoke plume region and close to the first point, sampling in areas far from the first point or with other interference (such as equipment or buildings) will cause the background brightness value to deviate from the actual ambient light brightness, thus affecting the accuracy of optical depth calculation. Therefore, this application further extracts a background sampling window from the corrected video frame based on the unit normal vector and the first point to obtain a background area with the same ambient light conditions as the first point and free from smoke interference. This ensures that the background sampling area is under the same lighting environment as the first point, while avoiding smoke plumes and other irrelevant interference areas, providing a clean and reliable area sample for subsequent background brightness calculation, and avoiding optical depth calculation deviations caused by improper background selection.

[0057] In particular, in one specific embodiment, Figure 6 This is a flowchart of sub-step S35 of the AI-based flue gas emission monitoring method according to an embodiment of this application. Figure 6 As shown, step S35 includes: S351, moving a preset distance along the direction of its unit normal vector at the first point to obtain the right background sampling center; S352, moving a preset distance along the direction of its reverse normal vector at the first point to obtain the left background sampling center; S353, defining a background sampling window with a size of 5×5 pixels with the right background sampling center and the left background sampling center as the center respectively.

[0058] More specifically, in step S351, the first point is moved a preset distance along its unit normal vector direction to obtain the right-side background sampling center. It should be understood that since the first point is located on the plume centerline, its surrounding area is within the plume's coverage. If the background is sampled directly using the first point as a reference, plume pixels will be mixed in, causing background brightness distortion. The unit normal vector direction is perpendicular to the plume centerline, and moving along this direction allows for quick escape from the plume area. Therefore, this application further moves the first point a preset distance along the unit normal vector direction to determine the core position of the right-side background sampling. This ensures that the right-side background sampling center is completely outside the plume mask, avoiding interference from the plume on background brightness sampling. Simultaneously, by controlling the preset distance, the sampling center maintains a reasonable distance from the first point, both away from the plume and within the same lighting environment, laying the positional foundation for subsequently obtaining pure background brightness.

[0059] More specifically, in step S352, the first point is moved a preset distance along its reverse normal vector direction to obtain the left background sampling center. It should be understood that setting a single-sided background sampling center only along the unit normal vector direction may result in an impure background due to interference from objects such as factory equipment or pipes on that side. The reverse normal vector direction, being on the other side of the center line, provides a more comprehensive background selection and avoids the influence of unilateral interference. Therefore, this application further moves the first point a preset distance along the reverse normal vector direction to determine the core position of the left background sampling. This enables dual-sided background sampling, eliminating errors caused by unilateral interference by comparing the consistency of background brightness on both sides. Simultaneously, dual-sided sampling obtains more representative ambient light data, ensuring that the background brightness value accurately reflects the actual lighting conditions of the area where the first point is located, thus improving the reliability of subsequent optical depth calculations.

[0060] More specifically, in step S353, a 5×5 pixel background sampling window is defined, centered on both the right and left background sampling centers. It should be understood that since pixels at a single background sampling center may experience local interference such as sensor noise and reflections from tiny dust particles, using only the brightness value of a single pixel as the background brightness would lead to significant randomness errors in the data. The statistical characteristics of a small pixel window can smooth out local interference. Therefore, this application further defines a 5×5 pixel background sampling window based on both sampling centers to obtain a sufficient number of background pixel samples. This allows for the elimination of errors caused by single-pixel noise through the statistical analysis of the brightness of multiple pixels within the window (such as the mean and median). Simultaneously, the 5×5 size ensures uniform illumination conditions within the window while maintaining a sufficient sample size, avoiding illumination differences caused by an excessively large window. This makes the background brightness value more stable and representative, providing an accurate background benchmark for optical depth calculation.

[0061] Specifically, in one possible embodiment, steps S351, S352, and S353 are implemented as follows: First, using the first point on the plume centerline as a reference, the system moves outward a preset distance along the normal direction perpendicular to the plume centerline. This distance is set sufficiently to ensure that the moved position is completely outside the plume coverage area of ​​that point, thereby determining the center position of the right-side background sampling. Subsequently, the same movement operation is performed along its reverse normal direction to determine the center position of the left-side background sampling. After determining the sampling centers on both sides, a background sampling window of a preset size is defined based on these two centers respectively. Finally, the system performs purity verification on these two sampling windows, checking whether the entire window area is located in a clean background area outside the plume mask. If the verification finds that the window still contains plume pixels, the system automatically adjusts the corresponding sampling center further outward along its normal direction and redefines and verifies the sampling window until the entire window area completely meets the condition of a clean background.

[0062] Specifically, in step S36, the statistical characteristics of the blue channel brightness within the background sampling window are used as the background brightness. It should be understood that since individual pixels within the background sampling window may contain noise, such as sensor errors or minute dust interference, directly using the blue channel brightness of a single pixel as the background brightness would result in random errors in the background brightness value, affecting the contrast accuracy between the foreground and background. Therefore, this application further calculates the statistical characteristics of the blue channel brightness within the background sampling window and uses it as the background brightness to eliminate the influence of individual pixel noise on the background brightness. This allows for the smoothing of noise interference through statistical methods, resulting in a more stable and representative background brightness value. This ensures that the contrast between the foreground and background brightness is closer to the actual optical attenuation, providing a reliable background benchmark for subsequent optical depth calculations.

[0063] Specifically, in one possible embodiment, step S36 is implemented as follows: First, all pixels within the background sampling window are traversed, and the blue channel luminance value of each pixel is extracted to form a blue channel luminance dataset. Next, statistical analysis is performed on this dataset to calculate its statistical characteristics. Considering the need to avoid the influence of extreme outliers, the mean or median of the dataset is typically selected as the core statistical characteristic. Finally, the calculated mean or median is determined as the background luminance value corresponding to the first point. This value, along with the foreground luminance value, forms paired data for subsequent optical depth calculation.

[0064] 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.

[0065] 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.

[0066] In the aforementioned AI-based flue gas emission monitoring method, step S4 involves extracting macroscopic visual features of the plume from the corrected flue gas video frames to obtain a plume visual feature vector. It should be understood that while the baseline concentration profile generated based on the Gaussian plume model follows physical laws, the macroscopic visual features of the plume, such as diffusion morphology, color intensity, and edge blurring, dynamically adjust with pollution source type, airflow conditions, and meteorological changes in actual industrial scenarios. These dynamic differences cannot be fully covered by the physical model, leading to potential deviations between the baseline concentration profile and the actual concentration, making it difficult to accurately match actual emissions. Therefore, this application further extracts macroscopic visual features of the plume from the corrected flue gas video frames and converts them into a plume visual feature vector to capture the dynamic characteristics and detailed differences of the plume not covered by the physical model. This provides crucial feature input for the subsequent AI residual correction module, enabling the AI ​​model to accurately identify the deviation patterns between the baseline concentration profile and the actual concentration based on these concrete visual features. This allows for fine-tuning of the baseline, improving the accuracy of the final pollutant concentration profile and its adaptability to complex industrial scenarios, and avoiding quantitative monitoring errors caused by the limitations of the physical model.

[0067] Specifically, in one possible embodiment, step S4 is implemented as follows: First, based on the previously acquired plume mask, the complete plume region is precisely selected and cropped from the corrected smoke video frame to ensure that the extraction range only includes plume pixels, completely eliminating interference from background areas such as factory buildings, sky, and equipment. Next, multi-dimensional macroscopic visual feature extraction is performed. Spatial morphology features include the overall area, aspect ratio, and irregularity coefficient of the plume, reflecting the asymmetry of plume diffusion; color features include the mean and variance of the blue channel brightness in the plume region, and the brightness gradient between the core and edge regions, reflecting the differences in plume concentration distribution; texture features include the uniformity index of grayscale distribution within the plume, reflecting the distribution state of plume particles. Finally, the extracted features are standardized, converting each feature value to a uniform range, and then arranging all standardized values ​​sequentially according to a preset feature priority order to form a fixed-dimensional plume visual feature vector.

[0068] 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.

[0069] 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.

[0070] 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.

[0071] Specifically, in one possible embodiment, step S51 is implemented as follows: First, the intrinsic and extrinsic parameters of the monitoring camera are acquired. The intrinsic parameters include the camera focal length and pixel size, which are determined by the camera's factory calibration data and prior on-site calibration. The extrinsic parameters include the camera's installation height, horizontal shooting angle, and horizontal distance from the emission source, which are obtained through on-site measurement tools. Next, based on the pinhole camera model in computer vision, the pixel coordinates of the plume centerline are substituted into the conversion formula. Through the calculation of pixel coordinates and camera parameters, the world coordinates corresponding to each pixel coordinate are obtained, including dimensions such as horizontal distance and vertical height, and arranged in the order of the plume diffusion path to form a world coordinate path profile. Finally, the results are verified. Fixed landmarks with known real coordinates in the industrial scene are selected, such as the edge of the emission outlet or the corner of a specific device. 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 a preset range, the world coordinate path profile is confirmed to be valid; otherwise, the camera parameters are recalibrated and the conversion is performed again.

[0072] Specifically, step S52 involves converting the optical depth profile into path integral concentration to obtain the observed path integral concentration profile. It should be understood that since the optical depth profile only characterizes the attenuation of light by the plume and is an optical characteristic parameter, it cannot be directly used by the Gaussian plume model for concentration calculation. Therefore, this application further converts the optical depth profile into path integral concentration, transforming the optical attenuation characteristics into path integral data related to pollutant concentration, thereby establishing a quantitative correlation between optical information and concentration information. This provides the Gaussian plume model with directly usable concentration observation data, allowing the model to optimize concentration prediction results by comparing the observed path integral concentration with theoretically calculated values. This avoids the problem of the model being unable to effectively integrate optical information due to parameter type mismatch, while ensuring that the converted concentration data reflects the cumulative concentration characteristics of the plume along the light propagation path, meeting the model's input requirements.

[0073] Specifically, in one possible embodiment, step S52 is implemented as follows: First, the conversion relationship between optical depth and path integral concentration is determined. This relationship is obtained through laboratory calibration experiments: In a controlled environment, industrial flue gas samples of known concentrations are prepared, and the optical depth of samples of different concentrations under different propagation paths is measured to establish a quantitative mapping model between the two. Next, the optical depth value of each reference point in the optical depth profile is substituted into the mapping model one by one to calculate the path integral concentration value corresponding to each reference point. These values ​​are arranged in the order of the plume centerline path to form the observed path integral concentration profile. Finally, the results are calibrated. A time period with known emission concentrations in the industrial scenario is selected, such as the standard emission concentration during stable equipment operation. The observed path integral concentration corresponding to this time period is compared with the theoretically calculated path integral 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 repeated.

[0074] Specifically, in step S53, baseline prediction based on a Gaussian plume model is performed on the world coordinate path profile and the observed path integral concentration profile to obtain the baseline concentration profile. It should be understood that since the world coordinate path profile only provides the actual spatial path of plume diffusion, and the observed path integral concentration profile only provides the cumulative concentration observation along the path, neither can generate a three-dimensional concentration distribution that conforms to the physical laws of plume diffusion when used alone. Therefore, this application further inputs the two profiles into a Gaussian plume model, integrates the spatial path and observed concentration through the model, and performs baseline prediction to generate an initial concentration profile that conforms to the physical laws of diffusion. In this way, the mature physical description capability of the Gaussian plume model for plume diffusion can be utilized to fuse spatial parameters with observational data, obtaining a baseline concentration profile that covers the entire plume diffusion path and whose concentration changes conform to actual laws. This provides a reliable physical benchmark for subsequent AI residual correction and avoids concentration prediction biases that may occur when the AI ​​model learns alone, deviating from physical laws.

[0075] Specifically, in one possible embodiment, step S53 is implemented as follows: First, the Gaussian plume model parameters are adjusted according to the actual conditions of the industrial scenario: the diffusion coefficient of the model, including the horizontal diffusion coefficient and the vertical diffusion coefficient, is determined based on real-time meteorological data (such as wind speed and wind direction), and the initial value of the source strength parameter of the model is determined based on the actual height and diameter of the emission source. Next, the world coordinate path profile and the observed path integral concentration profile are preprocessed to ensure that the spatial sampling points of the two correspond one-to-one. If the sampling density is different, an interpolation method is used to unify the sampling interval, and the preprocessed data is 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. By comparing the theoretical value with 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 by comparing the concentration values ​​near the emission outlet in the profile with the design emission concentration range of the industrial equipment. If it is within a reasonable range, the baseline concentration profile is confirmed to be valid; otherwise, the model parameters are readjusted and the prediction is run again.

[0076] In the aforementioned AI-based flue gas emission monitoring method, step S6 involves performing AI residual correction on the baseline concentration profile based on the visual feature vector of the plume to obtain a corrected concentration profile. It should be understood that, to avoid the deviation between the baseline concentration profile and the actual concentration distribution affecting the accuracy of quantitative pollutant concentration monitoring, this application further relies on the dynamic information contained in the visual feature vector of the plume, such as macroscopic morphology, color gradient, and texture uniformity, to perform residual correction on the baseline concentration profile using an AI model. This accurately identifies and corrects the deviation between the baseline and the actual concentration. This fully leverages the AI ​​model's learning ability for complex dynamic features, compensates for the limitations of physical models in dealing with non-ideal industrial scenarios, and ensures that the corrected concentration profile not only conforms to physical diffusion laws but also fits the actual emission state of the plume. This significantly improves the adaptability to complex diffusion conditions in industrial scenarios, meets the high-precision and high-robustness requirements of environmental monitoring for quantitative pollutant concentration inversion, and provides reliable data support for subsequent emission source tracing and compliance determination.

[0077] Specifically, in one possible embodiment, step S6 is implemented as follows: First, an AI model for residual correction is constructed, employing 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 association information through convolution operations, and the output layer outputs the residual correction amount consistent with the spatial dimension of the baseline concentration profile. Next, training samples are prepared by collecting flue gas monitoring data from different industrial scenarios such as thermal power and steel, as well as under different meteorological conditions such as sunny days, cloudy days, and backlighting. Each sample set includes a plume visual feature vector, a corresponding baseline concentration profile, and a concurrent real concentration profile obtained through a high-precision in-situ analysis instrument. The difference between the baseline concentration and the real concentration is calculated as the residual label, forming a fully labeled sample set. Subsequently, model training is performed by dividing the sample set into a training set and a validation set according to a preset ratio. The mean squared error between the predicted residual value and the real residual label is used as the loss function, and the gradient descent algorithm is used to iteratively optimize the model parameters until the prediction error of the model on the validation set is stably within a preset threshold. Next, residual correction is performed. The visual feature vector of the plume to be processed is input into the trained AI model to obtain the point-by-point residual correction amount. This correction amount is then superimposed with the concentration value at the corresponding spatial location in the baseline concentration profile to obtain the preliminary corrected concentration profile. Finally, the results are verified by comparing the corrected concentration profile with the real concentration data from in-situ monitoring during the same period point by point. If the overall error is within the industry-acceptable range for industrial flue gas monitoring, the correction is confirmed to be effective and the final corrected concentration profile is output. Otherwise, the hidden layer structure of the model is readjusted or additional samples are added before retraining the model until the correction result meets the accuracy requirements.

[0078] In particular, in another possible preferred embodiment, step S6 includes: constructing a local feature vector containing the baseline concentration and local features for each point on the baseline concentration profile to obtain a local feature vector sequence; generating a correction embedding vector that integrates 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; calculating a multi-head attention matrix by applying a multi-head attention mechanism based on the correction embedding vector sequence and the plume visual feature vector; generating a residual correction vector based on the multi-head attention matrix, the correction embedding vector sequence, and a learnable residual prediction matrix, and using the residual correction vector to correct the baseline concentration profile to obtain a corrected concentration profile.

[0079] Specifically, in the baseline concentration profile obtained by baseline prediction based on a Gaussian plume model from the world coordinate path profile and the observed path integral concentration profile, the baseline concentration at each point corresponds globally to the visual feature vector of the plume. That is, as a continuous fluid, the concentration, morphology, and diffusion state between adjacent points on the plume profile are highly correlated, exhibiting local correlation. The correction value (residual) at a point is likely to be affected by the states of its upstream and downstream points. For example, if severe turbulent diffusion has occurred at the upstream point (which can be reflected in visual features), then the concentration decay pattern at the downstream point will deviate further from the standard Gaussian model.

[0080] On the other hand, since the visual feature vector of the plume is an average description of the entire plume, when the profile is very long and the plume morphology varies greatly at different locations, 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 lot of position-specific information relative to each point of the baseline concentration profile, which needs to be fully utilized.

[0081] First, a local feature vector containing the baseline concentration and local features at each point on the baseline concentration profile is constructed to obtain a sequence of local feature vectors. Specifically, for the baseline concentration profile, optical depth profile, world coordinate path profile, and other optional local features such as local width and normal direction, the corresponding baseline concentration CL_basei, optical depth Di, downwind distance xi, and height zi are obtained for each point i on the profile, forming a local feature vector. And obtain the feature sequence based on the local feature vectors of all points.

[0082] Then, for each local feature vector and smoke plume visual feature vectors First, based on the sequence of local feature vectors and the visual feature vectors of the plume, a corrected embedding vector that incorporates global contextual information is generated for each point's local feature vector, resulting in a sequence of corrected embedding vectors: ;in, Represents matrix multiplication. Indicates the transpose symbol. Represents the visual feature vector of smoke plumes. This represents the corrected embedding vector, where all vectors are row vectors, meaning they are the local feature vectors for each point. Generate a corrected embedding vector that incorporates contextual information from the entire profile. This ensures that each local vector retains its own details while also containing information about the overall state of the plume. Whether it is the asymmetric diffusion of the plume due to crosswinds or the color gradient differences caused by changes in ambient light, these can all be incorporated into the local feature vector, avoiding correction biases caused by a lack of a global perspective.

[0083] Then, based on the sequence of the corrected embedding vectors and the visual feature vectors of the plume, a multi-head attention matrix is ​​calculated using a multi-head attention mechanism: ;in, This represents vector subtraction. This represents element-wise multiplication. Represents the visual feature vector of smoke plumes. This indicates element-wise reciprocal calculation, meaning performing the reciprocal operation on each element of the vector. This represents an element-wise exponential function, which performs an exponential operation on each element of the matrix. This represents a multi-head attention matrix.

[0084] That is, in the local feature vector By incorporating global profile context information, multi-head association based on global attention and residual prediction is performed through point-by-point attention residual correspondence, thereby accurately characterizing the association strength between different features and distinguishing between key associations and secondary associations.

[0085] Finally, based on the multi-head attention matrix, the sequence of the corrected embedding vectors, and the learnable residual prediction matrix, a residual correction vector is generated: ;in, Represents the learnable residual prediction matrix. Represents a cascade function. This represents the residual correction vector.

[0086] In other words, by focusing on the morphology and state of the entire plume with multi-head attention when correcting the concentration at a single point, a more comprehensive and reasonable residual prediction is made. For each point on the profile, information from all other points on the entire profile is dynamically and non-locally considered, and long-distance dependencies between profile points are explicitly modeled. For example, it can be learned that if the geometry at the end of the profile shows strong downwashing, then the baseline concentration in the middle of the profile, even if it looks normal, requires negative correction because the physical model does not consider the rapid dilution caused by downwashing. Finally, the baseline concentration profile is corrected using the residual correction vector to obtain the corrected concentration profile. Specifically, firstly, a one-to-one correspondence is established between the elements of the residual correction vector and the points on the baseline concentration profile, according to the actual diffusion order of the plume from the emission port, the middle of the diffusion range, and the end of the diffusion range. Then, the baseline concentration value is numerically superimposed with the corresponding correction amount point by point to obtain the preliminary corrected concentration. Preliminary corrected concentrations less than 0 are set to 0 (consistent with the non-negative physical meaning of concentration). Finally, the corrected concentration values ​​are recombined in the original order to generate a corrected concentration profile, which takes into account both the physical laws and characteristic correlation corrections of the Gaussian plume model.

[0087] In summary, the AI-based flue gas emission monitoring method based on the embodiments of this application is explained. It dynamically corrects the color and analyzes the physical characteristics of real-time flue gas videos, and constructs a physical prediction baseline based on a Gaussian plume model in parallel. Based on this, an AI model driven by the macroscopic visual features of the plume is introduced. The visual feature vectors extracted by the deep network are used as key correction factors to dynamically compensate and finely correct the baseline concentration profile generated by the physical model, ultimately driving the generation of a high-precision pollutant concentration profile. In this way, by organically combining the universality of physical laws with the powerful learning ability of artificial intelligence, remote, non-contact, high-precision quantitative monitoring of flue gas emissions can be achieved, significantly improving the accuracy and robustness of visual monitoring solutions in quantitative inversion of pollutant concentrations in real industrial scenarios.

[0088] Figure 8 This is a block diagram of an AI-based flue gas emission monitoring system according to an embodiment of this application. Figure 8As shown, the AI-based flue gas emission monitoring system 100 according to an embodiment of this application includes: a dynamic color correction module 110, used to perform dynamic color correction on real-time flue gas video frames based on the reference object ROI and the real color of the reference object to obtain corrected flue gas video frames; a plume segmentation center extraction module 120, used to perform plume segmentation and center path extraction on the corrected flue gas video frames to obtain a plume mask and plume centerline pixel coordinates; an optical depth estimation module 130, used to perform optical depth estimation based on the plume mask and plume centerline pixel coordinates to obtain an optical depth profile; a macroscopic visual feature extraction module 140, used to extract macroscopic visual features of the plume from the corrected flue gas video frames to obtain a plume visual feature vector; a baseline prediction module 150, used to perform baseline prediction based on a Gaussian plume model on the optical depth profile and plume centerline pixel coordinates to obtain a baseline concentration profile; and an AI residual correction module 160, used to perform AI residual correction on the baseline concentration profile based on the plume visual feature vector to obtain a corrected concentration profile.

[0089] As described above, the AI-based flue gas emission monitoring system 100 according to the embodiments of this application can be implemented in various wireless terminals, such as servers with AI-based flue gas emission monitoring algorithms. In one possible implementation, the AI-based flue gas emission monitoring system 100 according to the embodiments of this application can be integrated into the wireless terminal as a software module and / or 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 it can be an application developed for the wireless terminal; of course, the AI-based flue gas emission monitoring system 100 can also be one of many hardware modules of the wireless terminal.

[0090] 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 connect to the wireless terminal via wired and / or wireless networks and transmit interactive information in accordance with an agreed data format.

[0091] Here, those skilled in the art will understand that the specific operations of each step in the AI-based flue gas emission monitoring system described above have been referenced. Figures 1 to 7 The AI-based flue gas emission monitoring method is described in detail here, and therefore, its repeated description will be omitted.

Claims

1. An AI-based flue gas emission monitoring method, characterized by, include: Based on the reference object ROI and the true color of the reference object, dynamic color correction is performed on the real-time smoke video frame to obtain the corrected smoke video frame. Perform plume segmentation and center path extraction on the corrected flue gas video frames to obtain the plume mask and the pixel coordinates of the plume centerline; Optical depth estimation based on plume mask and plume centerline pixel coordinates to obtain an optical depth profile includes: extracting a first point from the plume centerline pixel coordinates; calculating the unit normal vector of the first point; taking the first point as the starting point, performing ray stepping along its unit normal vector and reverse normal vector directions, 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; obtaining the pixel value of the corrected smoke video frame at the first point, and extracting 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; performing sampling window truncation in the corrected smoke video frame based on the unit normal vector and the first point to obtain a background sampling window; using the statistical characteristics of the blue channel luminance within the background sampling window as the background luminance; and calculating the optical depth of the first point based on the foreground luminance and background luminance. Macroscopic visual features of the plume are extracted from the corrected smoke video frames to obtain the plume visual feature vector; Baseline concentration profiles are obtained by performing baseline prediction based on a Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline. AI residual correction is performed on the baseline concentration profile based on the visual feature vector of the smoke plume to obtain the 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's true color, dynamic color correction is performed on real-time smoke video frames to obtain corrected smoke video frames, including: Extract the RGB values ​​of all pixels within the region defined by the reference object ROI from the real-time smoke video frame; Calculate the average RGB values ​​of all pixels within the region to obtain the average observed color vector; Calculate the color correction gain factor between the true color of the reference object and the average observed color vector; The color correction gain factor is applied 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, Calculating the color correction gain factor between the true color of the reference object and the average observed color vector includes: calculating the color correction gain factor between the true color of the reference object and the average observed color vector using the following formula: wherein , and are the average observed color vectors, , and are the true colors of the reference objects. 4.The AI-based flue gas emission monitoring method of claim 1, wherein, The corrected smoke video frames are segmented and their center paths are extracted to obtain the smoke plume mask and the pixel coordinates of the smoke plume centerline, including: The corrected smoke video frames are input into a finely tuned instance segmentation model on the smoke dataset to obtain the plume mask. The centerline of the plume is extracted by morphological processing to obtain the pixel coordinates of the plume centerline. 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 according to claim 5, characterized in that, Based on the unit normal vector and the first point, a sampling window is extracted from the corrected smoke video frame to obtain the background sampling window, including: Move a preset distance along the direction of its unit normal vector from the first point to obtain the right background sampling center; Move a preset distance along the direction of its reverse normal vector from the first point to obtain the sampling center of the left background. Define background sampling windows with a size of 5×5 pixels, centered on the right background sampling center and the left background sampling center respectively.

7. The AI-based flue gas emission monitoring method according to claim 1, characterized in that, Baseline concentration profiles are obtained by performing baseline prediction based on a Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline, including: Perform coordinate system transformation on the pixel coordinates of the plume centerline to obtain the world coordinate path profile; The optical depth profile is converted into path integral density to obtain the observed path integral density profile. The baseline concentration profile is obtained by performing baseline prediction based on a Gaussian plume model on the world coordinate path profile and the observed path integral concentration profile.

8. An AI-based flue gas emission monitoring system, characterized in that, include: The dynamic color correction module is used to perform dynamic color correction on real-time smoke video frames based on the reference object ROI and the real color of the reference object to obtain the corrected smoke video frames. The plume segmentation and center extraction module is used to segment the plumes and extract the center path from the corrected smoke video frames to obtain the plume mask and the pixel coordinates of the plume center line. An optical depth estimation module is used to perform optical depth estimation based on a plume mask and the pixel coordinates of the plume centerline to obtain an optical depth profile. The module includes: extracting a first point from the pixel coordinates of the plume centerline; calculating the unit normal vector of the first point; taking the first point as the starting point, performing ray stepping along the directions of its unit normal vector and reverse normal vector, 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; obtaining the pixel value of the corrected smoke video frame at the first point, and extracting 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; performing sampling window truncation in the corrected smoke video frame based on the unit normal vector and the first point to obtain a background sampling window; using the statistical characteristics of the blue channel luminance within the background sampling window as the background luminance; and calculating the optical depth of the first point based on the foreground luminance and background luminance. The macroscopic visual feature extraction module is used to extract the macroscopic visual features of the plume from the corrected smoke video frames to obtain the plume visual feature vector. The baseline prediction module is used to perform baseline prediction based on the Gaussian plume model on the optical depth profile and the pixel coordinates of the plume centerline to obtain the baseline concentration profile. The AI ​​residual correction module is used to perform AI residual correction on the baseline concentration profile based on the visual feature vector of the plume to obtain the corrected concentration profile.

Images