Computer vision-based monitoring of fugitive emissions in steelmaking and carbon flow analysis method

Abstract

The present application relates to the technical field of carbon emission monitoring, and discloses a steelmaking unorganized emission monitoring and carbon flow analysis method based on computer vision, which comprises the following steps: acquiring field video stream and spectral analysis data, extracting flue gas plume visual feature vectors, inputting the vectors into an inversion mapping model after space-time synchronization and fusion processing, calculating unorganized emission instantaneous concentration and acquiring emission source three-dimensional coordinates; inputting the instantaneous concentration and three-dimensional coordinates as boundary conditions into a workshop three-dimensional grid model to solve fluid mechanics, acquiring dynamic diffusion path and concentration gradient distribution; performing space volume integral operation in combination with the diffusion path and concentration gradient to output unorganized carbon emission estimation total amount within a time window; and finally determining abnormal state based on the instantaneous concentration and estimation total amount, generating hardware control instructions to adjust dust removal equipment and feeding back a business work order. The present application realizes quantitative accounting of steelmaking workshop unorganized carbon emission and closed-loop adjustment of abnormal emission state.

Description

Technical Field

[0001] This invention relates to the field of carbon emission monitoring technology, specifically to a computer vision-based method for monitoring fugitive emissions from steelmaking and analyzing carbon flow. Background Technology

[0002] Steel smelting processes generate waste gas containing carbon oxides. Fugitive emissions, particularly those from uncontrolled areas, are particularly challenging due to the lack of fixed exhaust ducts. The gas typically diffuses freely within the workshop and escapes into the atmosphere, making monitoring emission concentrations and calculating total carbon emissions difficult. Existing fugitive emission monitoring solutions primarily rely on deploying fixed-point gas sensors within the workshop. However, because steelmaking workshops are large and subject to thermal airflow disturbances, fixed sensors can only acquire gas composition data at specific local locations. They cannot capture the dynamic diffusion process of the flue gas plume throughout the entire three-dimensional space of the workshop, resulting in single-point monitoring data failing to represent the overall emission status of the workshop.

[0003] To address the limitations of large-space monitoring, some solutions have incorporated video surveillance technology to aid observation. Current visual monitoring largely focuses on two-dimensional image-level smoke morphology recognition and alarms, lacking methods for multi-source fusion of visual image features and gas physicochemical data, making it difficult to convert monitoring images into quantitative emission concentration values. After acquiring monitoring data, existing environmental protection systems typically only perform backend data display and recording. When abnormal fugitive emissions occur in the workshop, the system cannot automatically calculate the adjustment parameters of the underlying dust removal equipment based on the spatial diffusion flux and concentration gradient distribution of the flue gas. There is a lack of linkage between the monitoring network and the hardware control of on-site ventilation and dust removal equipment, requiring manual intervention based on experience to control the operating frequency of the dust removal and induced draft fans in the face of abnormal emissions. This approach struggles to achieve closed-loop adjustment of the underlying equipment based on actual fugitive concentration deviations, and due to the lack of a volume integral model for the spatial distribution of fugitive flue gas, it cannot quantify the total fugitive carbon emissions within a time window, resulting in missing data for the enterprise's overall carbon emission quota calculation. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking. This method solves the technical problems of existing technologies, such as the difficulty in quantitatively calculating the total carbon emissions in three-dimensional space and the lack of a closed-loop linkage adjustment mechanism between the monitoring system and the underlying dust removal equipment.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides a computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking, comprising the following steps: Acquire on-site video stream data and synchronously collected spectral analysis data of the target monitoring area; Image processing is performed on the on-site video stream data to extract the visual feature vector of the flue gas plume; The visual feature vector of the flue gas plume is fused with spectral analysis data, input into a preset inversion mapping model, and the instantaneous concentration value of unorganized emissions is calculated. The three-dimensional spatial coordinates of the emission source center are obtained by extracting the two-dimensional pixel coordinates of the centroid based on the on-site video stream data and combining them with the spatial transformation matrix. A three-dimensional mesh model of the workshop physical space is established. The instantaneous concentration values ​​of fugitive emissions and the three-dimensional spatial coordinates of the emission source center are used as the calculation boundary conditions to input the three-dimensional mesh model for fluid dynamics solution, and the dynamic diffusion path data and concentration gradient distribution data are output. By combining dynamic diffusion path data and concentration gradient distribution data, spatial volume integral calculation is performed to output the estimated total amount of fugitive carbon emissions within the accounting period. Anomalies are determined based on the instantaneous concentration of fugitive emissions and the estimated total amount of fugitive carbon emissions. If an anomaly is determined, hardware control instructions are generated, and business work order data is generated for feedback.

[0007] Furthermore, the visual feature vector of the flue gas plume includes geometric morphology parameters, diffusion velocity field parameters, opacity parameters, and texture parameters. During feature extraction, a deep learning visual feature extraction network outputs a binarized mask image of the flue gas plume from the live video stream data. The connected regions of the flue gas are located, and their contour coordinate sets are extracted. The aspect ratio, perimeter, and area of ​​the geometric bounding box of the connected regions are calculated, and tensor concatenation is performed to obtain the geometric morphology parameters. Simultaneously, pixel brightness data of the connected regions of the flue gas in two consecutive frames are acquired. Based on the dense optical flow algorithm, an optical flow constraint equation is constructed and solved under constant image brightness conditions to obtain the optical flow velocity vector distribution of the two-dimensional pixel coordinates. Its mean and variance matrices are then calculated as diffusion velocity field parameters.

[0008] In addition, a background reference model is established based on historical smokeless conditions. The opacity feature parameters are obtained by comparing the attenuation of the average pixel brightness in the current smoke area with that in the background area. The texture feature parameters of the connected regions are also extracted. Finally, the above parameters are input into the fully connected layer of the network for dimensionality reduction and splicing combination to output the visual feature vector of the smoke plume.

[0009] Before fusing the visual feature vectors of the flue gas plume with the spectral analysis data, spatiotemporal synchronization is performed. In the time dimension, a network time protocol is used for underlying clock synchronization, a global timestamp is assigned, and adjacent first and second spectral data timestamps in the spectral data sequence are found to perform linear interpolation calculations to obtain time-aligned interpolated spectral values.

[0010] In the spatial dimension, a spatial transformation matrix between the pixel coordinate system and the physical coordinate system is established by using on-site pre-set calibration references. The extracted two-dimensional pixel coordinates are mapped to three-dimensional physical coordinates, and intersection calculations are performed with the spatial straight line equation of the spectral detection optical path to verify the spatial overlap between the spectral detection path and the visual capture area.

[0011] During the data fusion stage, a maximum-minimum normalization algorithm is used to unify the data scale. The multi-source fusion feature vector is then input into the inversion mapping model of a multi-layer fully connected neural network architecture. In this model, the nonlinear activation function is a linear rectified function, which forces the negative values ​​calculated by the network to zero and retains the positive values ​​to meet the constraint that gas concentration values ​​cannot be negative in physical reality. Finally, the instantaneous concentration value of fugitive emissions is output.

[0012] The fluid mechanics solution process includes: The emission source release area is obtained by proportionally estimating the pixel area from the area of ​​the connected flue gas region in the binarized mask image using a spatial transformation matrix. The radius of the preset source term is then calculated using the equivalent circle area operation, thereby defining the set of mass source term units contained within the spherical spatial region. The fugitive flue gas mass release rate is obtained by multiplying the instantaneous fugitive emission concentration, the initial escape velocity of the flue gas plume, and the emission source release area. The fluid dynamics model defines the natural ventilation doors and windows and roof ventilators of the workshop as pressure inlet boundaries, the dust collection hood opening of the secondary dust removal fan as the velocity outlet boundary based on the flow rate calculated according to the real-time operating frequency, and sets the physical properties of the ambient air to follow the incompressible ideal gas law.

[0013] When solving the mass conservation equation, momentum conservation equation, and component transport equation, the standard k-value is used. A two-equation model is used to calculate the turbulent viscosity coefficient, and the density difference combined with the gravitational acceleration vector generates an upward thermal buoyancy driving term in the momentum conservation equation. A semi-implicit pressure-velocity coupling solution is used to solve the pressure and velocity problems. When the residual values ​​of mass, momentum, and components all decrease to below a preset residual threshold, the physical flow field calculation for the current time step is considered to have converged, generating a time series record of the fluid velocity vector and concentration gradient distribution data.

[0014] The system performs spatial volume integral calculations to output the estimated total amount of fugitive carbon emissions within the accounting period. This includes extracting two-dimensional mesh patches corresponding to the sections connected to the external atmospheric environment in the three-dimensional computational grid and stitching them together to generate flux integral boundaries. By comparing the direction determination vectors generated from the two-dimensional mesh patches and the centroid coordinates of adjacent internal mesh cells, the spatial unit normal vectors are uniformly corrected to point towards the external atmospheric environment.

[0015] The fluid velocity vector on the two-dimensional mesh patch is multiplied by the corrected spatial unit normal vector. A flux truncation function is used to include only values ​​with a positive value in the multiplication result in the flux calculation. The truncated multiplication result is then multiplied by the fluid density, component mass concentration, physical area, and discrete time step to obtain the total escaping mass within the calculation period as the boundary escaping flux.

[0016] The fluid density and component mass concentration in each three-dimensional computational grid are extracted simultaneously and multiplied with the corresponding physical volume to obtain the carbon emission mass in a single grid. The carbon emission masses of all three-dimensional computational grids are added together to obtain the spatial retention accumulation. The two results are combined and the background carbon emission mass at the initial moment is subtracted to obtain the total amount of unorganized carbon emissions.

[0017] Anomaly assessment is based on the instantaneous concentration of fugitive emissions and the estimated total amount of fugitive carbon emissions, including: A sliding window is established on the time axis. The average fugitive mass within the window is calculated and converted into the real-time fugitive rate. Simultaneously, the maximum component mass concentration within the calculation grid of the extended area surrounding the dust collector hood is extracted as the highest local concentration. The real-time fugitive rate and the highest local concentration are dimensionless and assigned flux and concentration weighting coefficients to construct an abnormal fugitive emission evaluation index. When this index is greater than the judgment benchmark value, the current state is determined to be abnormal fugitive emission, and an abnormal trigger signal is generated in the system logic controller. An incremental proportional-integral control algorithm is adopted, and the deviation between the abnormal evaluation index and the judgment benchmark value is used as the error input of the current system state deviating from the normal state. The target induced draft fan operating frequency is calculated by combining the time accumulation, proportional adjustment coefficient, and integral adjustment coefficient.

[0018] After the calculation frequency is limited to not exceeding the rated operating frequency of the induced draft fan, it is converted into a digital analog quantity setpoint and issued, along with simultaneous on / off commands to shut down the roof ventilator and start the spray dust suppression machine. Before generating business work order data for feedback, the total organized carbon emissions within the same period of the accounting cycle are obtained. The total organized carbon emissions are algebraically summed with the total unorganized carbon emissions to obtain the true total carbon emission load. The carbon emission quota early warning index is then calculated by combining the total carbon emission quota within the accounting cycle with the time schedule.

[0019] When the warning index exceeds the set warning threshold, a structured data message containing the location of the occurrence, the current carbon emission concentration, and the warning index is pushed to the external business system. This message is used to trigger the production load downgrade procedure of the manufacturing execution system and the dispatch of environmental equipment abnormality inspection work orders in the equipment asset management system.

[0020] This invention provides a computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking. It offers the following advantages: 1. This invention acquires on-site video stream data and spectral analysis data, performs spatiotemporal synchronization processing, and then fuses the visual feature vectors and spectral data. The fusion is then input into an inversion mapping model that uses a linear rectified function as the activation function for calculation. This multi-source data fusion method overcomes the limitations of single visual monitoring being susceptible to background light interference and single spectral monitoring having limited coverage. Furthermore, by forcing negative values ​​to zero during model calculation, the output instantaneous concentration of fugitive emissions conforms to the physical constraint that gas concentration cannot be negative, ensuring the accuracy of the concentration calculation results.

[0021] 2. This invention uses the acquired instantaneous concentration values ​​and emission source center coordinates as boundary conditions input into a three-dimensional mesh model for hydrodynamic solution, generating flow field diffusion paths and concentration gradient distributions. It then combines spatial normal vector correction to extract flux integral boundaries, and uses spatial volume integral calculations to obtain the boundary fugitive flux and spatial retention accumulation within the time window. This method extends static monitoring at a single point to dynamic carbon flow distribution calculation in a three-dimensional workshop space, enabling quantitative calculation of total fugitive emissions and providing data support for enterprises' carbon emission quota consumption assessment.

[0022] 3. This invention constructs an anomaly evaluation index based on real-time emission rate and the highest concentration in a local space. When the evaluation index is determined to be greater than a benchmark value, an incremental proportional-integral control algorithm is used to calculate the operating frequency of the target induced draft fan, generating control commands that are sent to the programmable logic controller to adjust the secondary dust removal induced draft fan, and simultaneously linking the roof ventilator and dust suppression equipment. This processing method achieves closed-loop regulation from system data judgment to underlying hardware control when abnormal emission conditions are detected. It can automatically adjust the operating status of on-site environmental protection equipment according to concentration deviations, suppressing the diffusion of fugitive flue gas into the external environment. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a comparison chart of the instantaneous concentration prediction accuracy of the present invention; Figure 4 This is the abnormal evaluation index and linkage control response diagram of the present invention; Figure 5 This is a comparison chart of the overall energy-saving and emission-reduction effects of the present invention. Detailed Implementation

[0024] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Please see the appendix Figure 1 The present invention provides a computer vision-based system for monitoring fugitive emissions and analyzing carbon flow in steelmaking, which may include: a sensing and inversion module, a flow field simulation module, a carbon emission quantification module, and a linkage control module.

[0026] The sensing and inversion module is deployed at the dust-generating nodes in the smelting workshop. It connects to the flow field simulation module via a communication network. The flow field simulation module connects to both the sensing and inversion module and the carbon emission quantification module for data transmission. The carbon emission quantification module connects to the linkage control module. The linkage control module communicates with the actuators in the workshop.

[0027] Please see the appendix Figure 2 This invention provides a computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking, comprising the following steps: S10, the perception inversion module acquires the on-site video stream data of the target monitoring area, performs image processing on the on-site video stream data, and extracts the visual feature vector of the smoke plume, which includes geometric shape feature parameters, diffusion velocity field parameters, opacity feature parameters and texture feature parameters. S20, the perception inversion module acquires the spectral analysis data collected synchronously on site, fuses the visual feature vector with the spectral analysis data, inputs the preset inversion mapping model, calculates the instantaneous concentration value of fugitive emissions, and records the three-dimensional spatial coordinates of the emission source. S30, the flow field simulation module establishes a three-dimensional mesh model of the workshop physical space, receives the instantaneous concentration value of unorganized emissions and three-dimensional spatial coordinates and inputs them into the three-dimensional mesh model as calculation boundary conditions, uses fluid dynamics algorithms to solve the mass conservation equation and momentum conservation equation, and outputs the dynamic diffusion path data and concentration gradient distribution data of flue gas inside the workshop. S40, the carbon emission quantification module defines the three-dimensional control volume for calculation in the workshop coordinate system, obtains dynamic diffusion path data and concentration gradient distribution data, combines the airflow velocity field data of the corresponding spatial nodes, performs spatial volume integral calculation on the three-dimensional control volume within a set time period, and outputs the estimated total amount of fugitive carbon emissions within a specific time window. The S50 linkage control module acquires the instantaneous concentration value of fugitive emissions and the estimated total amount of fugitive carbon emissions. It compares the instantaneous concentration value with a preset threshold. If the instantaneous concentration value exceeds the preset threshold, it generates hardware control instructions and sends them to the field programmable logic controller to adjust the operating frequency of the secondary dust removal fan or the spray parameters of the dry fog dust suppression device. At the same time, it generates business work order data containing spatial coordinates, occurrence time and estimated total amount of fugitive carbon emissions and sends it to the central control room server.

[0028] The perception inversion module includes industrial vision acquisition equipment and spectral analysis equipment. To obtain continuous on-site video streams and spectral characteristics, corresponding acquisition equipment needs to be deployed at dust-generating nodes such as converter charging ports, electric furnace dust-generating areas, continuous casting cutting areas, and material storage yards in the smelting workshop.

[0029] Industrial visual acquisition equipment and spectral analysis equipment are deployed in the target monitoring area. The industrial visual acquisition equipment specifically consists of explosion-proof high-definition cameras or infrared thermal imagers. The smelting workshop is characterized by high temperature and high dust levels; therefore, hardware protection components are configured to maintain the normal operation of the industrial visual acquisition equipment. These components include a water-cooled protective housing and a compressed air purging device. The water-cooled protective housing isolates external high-temperature radiation, while the compressed air purging device, located at the front of the lens, prevents environmental dust from adhering to the lens surface through continuous airflow.

[0030] In terms of spatial deployment strategy, a multi-camera, cross-view deployment scheme is adopted based on the physical structure of the dust-generating nodes and the expected flue gas dispersion trajectory. At least two industrial vision acquisition devices are set up around the same dust-generating node, and their yaw and pitch angles are adjusted to ensure that the field of view of multiple devices forms a three-dimensional spatial coverage of the same fugitive emission source, avoiding blind spots caused by a single viewpoint. The spectral analysis equipment specifically employs a differential optical absorption spectrometer or a Fourier transform infrared spectrometer. The light source transmitter and signal receiver of the spectral analysis equipment are located on opposite sides of the dust-generating area. During installation and calibration, the central axis of the detection optical path of the spectral analysis equipment and the central line of sight of the industrial vision acquisition equipment are kept physically intersecting or parallel to each other to ensure that the observed targets acquired by the two devices have spatial overlap.

[0031] Industrial vision acquisition equipment acquires on-site video stream data and performs image preprocessing. The equipment continuously captures on-site footage at a set frame rate, outputting raw on-site video stream data. Due to complex lighting conditions and dust interference in industrial settings, basic processing of the raw video stream data is necessary. For image denoising, white balance adjustment, and lens distortion correction of the on-site video stream data, those skilled in the art can employ Gaussian filtering algorithms and camera intrinsic parameter matrix correction algorithms. This image preprocessing process is well-known in the field and will not be elaborated upon here.

[0032] The spectroscopic analysis equipment simultaneously acquires absorption spectral data of the target area. The light source emits a probe beam of a preset wavelength, which passes through the escaping plume of flue gas. Particulate matter and gas molecules in the flue gas absorb and scatter the beam. The receiver records the attenuated spectral signal and outputs a sequence of spectral data characterizing the current gas composition and concentration.

[0033] The on-site video stream data and spectral data are processed for spatiotemporal synchronization. Due to the different hardware sampling frequencies of the industrial vision acquisition equipment and the spectral analysis equipment, a unified spatiotemporal reference is established within the perception inversion module. In the time dimension, the perception inversion module uses the Network Time Protocol (NTP) to perform low-level clock synchronization between the industrial vision acquisition equipment and the spectral analysis equipment, assigning a global timestamp to each frame of video image and each spectral data line. For discrete data with timestamp discrepancies, the perception inversion module uses a linear interpolation algorithm to align the timestamps, generating time-aligned data pairs. Based on the timestamp of the visual data acquired by the industrial vision acquisition equipment, the perception inversion module searches for adjacent first and second spectral data timestamps in the spectral data sequence. The perception inversion module calculates the interpolated spectral values ​​using a linear interpolation algorithm; the interpolation formula is as follows: ; In the formula, These are interpolated spectral values; The first spectral value corresponding to the timestamp of the first spectral data; The second spectral value corresponding to the second spectral data timestamp; For visual data timestamps; This is the timestamp for the first spectral data; This is the timestamp for the second spectral data. (Visual data timestamp) The value range is set to be greater than or equal to the timestamp of the first spectral data. And less than or equal to the second spectral data timestamp By performing time alignment calculations, a correspondence between visual images and spectral data at time nodes is established.

[0034] In the spatial dimension, using pre-set calibration references on-site, the spatial transformation matrix between the two-dimensional image pixel coordinate system, the spectrometer optical path coordinate system, and the workshop's three-dimensional physical coordinate system is calculated. The spatial transformation matrix is ​​determined by establishing a mapping relationship between the pixel coordinates and physical coordinates of the corner points of a standard calibration board deployed on-site. The perception inversion module performs coordinate transformation using the spatial transformation matrix; the transformation formula is: ; In the formula, The workshop's three-dimensional physical coordinates; This is the space transformation matrix; The coordinates are two-dimensional pixel coordinates. The perception inversion module maps the two-dimensional pixel coordinates extracted from the video image to the three-dimensional physical coordinates of the workshop, and performs an intersection calculation with the spatial straight line equation of the spectral detection optical path. This spatial intersection calculation verifies whether the optical path of the current spectral data detection passes through the flue gas plume region captured by the video image, thus completing the spatial matching verification between visual and spectral data. The data after spatiotemporal synchronization processing serves as the input source for subsequent visual feature extraction and concentration inversion calculations.

[0035] A deep learning visual feature extraction network is constructed within the perception inversion module to process spatiotemporally synchronized live video stream data to obtain visual feature vectors of the smoke plume. The network's basic architecture employs a target detection and instance segmentation model incorporating a feature pyramid structure. Due to the blurred edges and irregular shapes of unorganized smoke emissions, deformable convolutional kernels are embedded in the backbone of the deep learning visual feature extraction network. These deformable convolutional kernels dynamically adjust the sampling position based on the target shape by adding a two-dimensional pixel coordinate offset to the conventional convolution sampling points, enabling the receptive field to conform to the dynamic diffusion edge of the smoke and outputting a high-dimensional feature map. For basic convolution operations and model backpropagation parameter updates, those skilled in the art can use the backpropagation algorithm combined with a gradient descent optimizer; the model training process is well-known in the field and will not be elaborated upon here.

[0036] The perception inversion module processes the high-dimensional feature map through instance segmentation, outputting a binarized mask image of the flue gas plume. Based on the binarized mask image, the perception inversion module locates the connected regions of the flue gas and extracts the contour coordinate set of these regions. Using the contour coordinate set, the perception inversion module calculates the aspect ratio, perimeter, and area of ​​the geometric bounding box of these connected regions. Specifically, the perception inversion module calculates the area by counting the total number of pixels within the connected regions of the flue gas in the binarized mask image, calculates the perimeter by counting the number of pixels at the boundary of the connected regions, and constructs the minimum bounding box surrounding the connected regions. The aspect ratio is calculated using the ratio of the height to the width of the minimum bounding box. The perception inversion module then performs tensor concatenation of the aspect ratio, perimeter, and area values ​​to obtain the geometric morphological feature parameters.

[0037] The perception inversion module extracts diffusion velocity field parameters. It acquires pixel brightness data of the connected regions of smoke in two consecutive image frames. The module uses a dense optical flow algorithm to calculate pixel displacement between adjacent frames. Optical flow is a method for estimating the velocity field based on the motion information of target objects in an image sequence. Its basic theoretical assumption is that the brightness of pixels representing the same physical target remains constant between two consecutive image frames, and that adjacent pixels have similar motion trajectories. For two-dimensional pixel coordinates... Set its corresponding pixel brightness function to ,in The time variable is used. Based on the condition of constant image brightness, the perception inversion module constructs the optical flow constraint equation: ; In the formula, The image space gradient vector corresponding to the two-dimensional pixel coordinates; This is the optical flow velocity vector corresponding to the two-dimensional pixel coordinates; For image time gradient; Two-dimensional pixel coordinates; This is a pixel brightness function; The time variable is used. The perception inversion module obtains the optical flow velocity vector distribution of all two-dimensional pixel coordinates within the flue gas connectivity region by solving this constraint equation, calculates the mean and variance matrix of this distribution, and uses it as the diffusion velocity field parameter.

[0038] The perception inversion module extracts opacity feature parameters. It establishes a background reference model using historical video images of the workshop under smoke-free conditions, obtaining pixel brightness values ​​for the background region under a fixed field of view. Based on the binarized mask image, the module extracts the pixel brightness values ​​for the corresponding smoke region within the current frame. To obtain a representative brightness value for the entire region, the module averages the brightness of all pixels within the connected smoke region, obtaining the pixel brightness value for the smoke region. It also averages the brightness of pixels at the same coordinates in the background reference model, obtaining the pixel brightness value for the background region. The perception inversion module calculates opacity by comparing brightness attenuation; the calculation formula is as follows: ; In the formula, Opacity is a characteristic parameter; This represents the pixel brightness value for the smoke area. This represents the pixel brightness value of the background area. Opacity feature parameter. The value range is set between 0 and 1, with a larger value indicating a higher degree of optical obstruction of the background medium by the flue gas.

[0039] The perception inversion module inputs the acquired geometric morphology parameters, diffusion velocity field parameters, opacity parameters, and texture parameters into the fully connected layer of the network for dimensionality reduction and concatenation, outputting a one-dimensional visual feature vector of the flue gas plume. This visual feature vector of the flue gas plume serves as the basic data input for subsequent calculations of the instantaneous concentration of fugitive emissions.

[0040] The perception inversion module acquires the extracted visual feature vector of the flue gas plume and the interpolated spectral values ​​obtained through spatiotemporal synchronization processing. Before data fusion, the perception inversion module performs scale unification processing on the input data. Since the elements in the visual feature vector of the flue gas plume differ in physical dimensions from the interpolated spectral values, the perception inversion module uses a maximum-minimum normalization algorithm to map each data point in both the visual feature vector and the interpolated spectral values ​​to a numerical range of 0 to 1. Specifically, the perception inversion module acquires the maximum and minimum values ​​from a pre-established historical sample set of each data point, calculates the difference between the current input data and the minimum value, and divides this difference by the range between the maximum and minimum values ​​to complete the data normalization calculation. The perception inversion module then performs matrix concatenation of the normalized visual feature vector of the flue gas plume and the interpolated spectral values ​​according to a set dimension to generate a multi-source fusion feature vector.

[0041] At the physical level, the features acquired by the visual perception system reflect the macroscopic diffusion state and optical obstruction degree of the flue gas plume, and are significantly affected by changes in ambient light. The features acquired by the spectral analysis system are based on the microscopic absorption and attenuation characteristics of gas molecules for photon energy, conforming to the Beer-Lambert law, and are unaffected by visible light interference, but can only acquire data from a single detection optical path. The perception inversion module integrates these two systems, enabling the system to simultaneously possess macroscopic spatial perception capabilities and microscopic component detection capabilities, thus establishing a mapping relationship between data and physical concentration.

[0042] The perception inversion module establishes a pre-defined inversion mapping model. This model employs a multi-layer fully connected neural network architecture. The multi-layer fully connected neural network fits the non-linear relationship between the input multi-source fusion feature vector and the fugitive emission concentration through its internal neurons and non-linear activation functions. The perception inversion module inputs the multi-source fusion feature vector into the inversion mapping model, and the output of the model is calculated using the following formula: ; In the formula, This represents the instantaneous concentration value of unorganized emissions; This is a multi-source fusion feature vector; This is the input layer weight matrix; The input layer bias vector; For activation functions; This is the output layer weight matrix; This represents the output layer bias vector. The activation function in the perceptual inversion module. Specifically, a linear rectification function is used. The function of the linear rectification function is to force the negative values ​​obtained from the network calculation to zero and retain the positive values, so as to meet the physical constraint that the gas concentration value cannot be negative in the real world.

[0043] For the input layer weight matrix in the inversion mapping model Input layer bias vector Output layer weight matrix and output layer bias vector The values ​​of the weight matrix and bias vector are determined through training on a pre-constructed historical dataset. The historical dataset contains multiple sets of synchronously collected historical fusion feature vectors and their corresponding true concentration labels. The true concentration labels are obtained through on-site measurements using a portable gas analyzer deployed during the testing phase. The perception inversion module uses a backpropagation algorithm to update the values ​​of the aforementioned weight matrix and bias vector until the model's prediction error converges to a set range.

[0044] While acquiring instantaneous concentration values ​​of fugitive emissions, the sensing and inversion module records the spatial location parameters of the emission source. The module extracts the set of pixel coordinates for all connected regions of the flue gas in the aforementioned binarized mask image, calculates the arithmetic mean of this set along the horizontal and vertical axes, and obtains the centroid's two-dimensional pixel coordinates. The module then uses the aforementioned spatial transformation matrix to convert the centroid's two-dimensional pixel coordinates into three-dimensional coordinates within the workshop's physical environment. The coordinate transformation formula is: ; In the formula, The three-dimensional spatial coordinates of the emission source center; This is the space transformation matrix; The coordinates of the centroid are two-dimensional pixel coordinates.

[0045] The perception inversion module binds and packages the calculated instantaneous concentration values ​​of fugitive emissions with the three-dimensional spatial coordinates of the emission source center. The packaged data serves as the dynamic input boundary conditions for the subsequent flow field simulation module to perform diffusion extrapolation.

[0046] The flow field simulation module, based on computational fluid dynamics (CFD) theory, establishes a three-dimensional mesh model of the workshop's physical space to simulate the physical diffusion process of unorganized flue gas within the workshop environment. The physical principle of mesh generation is to divide the continuous gas-fluid domain within the workshop into a finite number of discrete control volumes, allowing partial differential equations in fluid dynamics to be solved iteratively at these discrete nodes using numerical methods. The flow field simulation module acquires the building structure data and on-site equipment layout drawings of the smelting workshop. Based on these data, the module constructs a three-dimensional geometric solid model including the workshop's exterior walls, ventilation doors and windows, smelting furnaces, and dust collection hoods. In this model, physical structures that obstruct or guide airflow are retained, while micro-pipeline components smaller than a preset solid size threshold are ignored to reduce the computational complexity of the subsequent fluid dynamics solution matrix. The preset solid size threshold is determined based on the minimum mesh size of the specified dense area. The flow field simulation module marks structural components with a maximum span smaller than the minimum mesh size of the dense area as micro-pipeline components and discards them.

[0047] Based on a three-dimensional geometric solid model, the flow field simulation module discretizes the spatial fluid domain within the workshop where gas can flow, generating a three-dimensional computational mesh. Due to the vast interior space of the smelting workshop and the relatively concentrated areas of unorganized emissions, the flow field simulation module employs a locally refined unstructured mesh generation strategy to balance the numerical accuracy of the airflow simulation with the consumption of computer resources. The unstructured mesh uses tetrahedral or polyhedral mesh elements to accommodate the irregular physical curved boundaries of the smelting furnace and dust hood.

[0048] The flow field simulation module acquires the pre-marked locations of dust-generating equipment and sets them as the central region for local mesh refinement. The module sets the base mesh size and the minimum mesh size for the refinement zone. The side length of the mesh increases gradually from the refinement zone outwards. The module establishes a spatial mesh size growth function, dynamically determining the mesh cell size of the region containing each mesh node based on its distance from the dust-generating equipment. The formula for calculating the spatial mesh size growth function is: ; In the formula, This refers to the spatial grid size; Minimum grid size for the encrypted area; This is the grid size growth factor; This represents the three-dimensional spatial distance from the mesh node to the geometric center of the dust-generating equipment. To prevent computational divergence caused by mesh size jumps, the mesh size growth factor is... The value range is set to between 1.05 and 1.20. Simultaneously, the flow field simulation module adjusts the spatial grid size during the calculation process. The calculation results are compared with the basic mesh size. When the spatial mesh size is... When the calculated result is greater than the base mesh size, its value is forcibly set to the base mesh size to ensure that the maximum size of the global mesh is controlled.

[0049] By using a spatial grid size growth function, the flow field simulation module generates a high-density three-dimensional computational grid in the localized region of initial unorganized flue gas diffusion to capture complex flow field vortices and concentration gradient changes; and generates a low-density three-dimensional computational grid in the gentle airflow region far from the pollution source. For the distortion rate check, aspect ratio evaluation, and orthogonality test of the three-dimensional computational grid, those skilled in the art can use a general finite volume method preprocessing algorithm. The grid quality inspection process is a well-known technique in the field and will not be elaborated upon here.

[0050] The three-dimensional computational mesh, after discretization and quality inspection, serves as the underlying physical space framework for the flow field simulation module to solve the fluid dynamics equations.

[0051] After establishing the three-dimensional computational mesh, the flow field simulation module configures boundary conditions for solving partial differential equations. In fluid dynamics calculations, boundary conditions define the physical state at the edge of the fluid domain and are the foundation for driving the internal flow field calculations. The physical environment of the smelting workshop is dynamically changing, and the flow field simulation module needs to convert the field data calculated by the aforementioned sensing and inversion module into a numerical format conforming to computational fluid dynamics specifications and map it to the corresponding nodes of the three-dimensional computational mesh.

[0052] The flow field simulation module constructs the pollution source boundary. It acquires the previously recorded three-dimensional spatial coordinates of the emission source center. Since actual fugitive emissions manifest as a plume of smoke with spatial volume, rather than a single geometric point, the flow field simulation module uses the three-dimensional spatial coordinates of the emission source center as the center and a preset source term radius to divide a spherical spatial region. The module marks all three-dimensional computational grid cells falling within this spherical spatial region as a set of mass source term cells. To maintain physical consistency with the actual pollution source area, the preset source term radius is calculated from the emission source release area using the equivalent circle area formula, which is as follows: ; In the formula, The preset source term radius; The emission source area; Pi is a constant.

[0053] For a set of mass source units, the flow field simulation module calculates the mass release rate of the unorganized flue gas used as input. The fluid mass release rate is determined by three physical quantities: concentration, velocity, and area. The flow field simulation module calls the previously calculated instantaneous concentration value of the unorganized emissions and combines it with geometric morphology parameters and diffusion velocity field parameters to obtain the final release rate through multiplication. The calculation formula is as follows: ; In the formula, The mass release rate of unorganized flue gas; This represents the instantaneous concentration value of unorganized emissions; The initial velocity of the flue gas plume; This refers to the emission source release area. For each variable parameter in the formula, the emission source release area... The value is derived from the area of ​​the connected flue gas region in the aforementioned binarized mask image, and then proportionally calculated from the pixel area to the physical area using the spatial transformation matrix. (Instantaneous concentration of fugitive emissions) Provided by the output of the aforementioned inversion mapping model. Initial escape velocity of the flue gas plume. The value is based on the mean of the optical flow velocity vector distribution of all two-dimensional pixel coordinates in the aforementioned diffusion velocity field parameters. The flow field simulation module obtains the physical scaling factor for the distance dimension in the aforementioned spatial transformation matrix, multiplies the mean of the optical flow velocity vector distribution by the physical scaling factor, and divides it by the time interval between two adjacent frames, thereby converting the pixel-level displacement velocity into the initial emission velocity of the smoke plume in physical space. .

[0054] The flow field simulation module sets environmental boundary conditions. These conditions are used to simulate the atmospheric flow background within the workshop. The module defines the natural ventilation doors and windows and roof ventilators of the workshop as pressure inlet boundaries, with their relative static pressure set to zero, indicating connection to the external standard atmospheric pressure. The module defines the dust collection hood opening of the secondary dust collector as the velocity outlet boundary, with its velocity parameters calculated based on the real-time feedback from the on-site programmable logic controller (PLC) of the fan's operating frequency. Based on the proportional law between fan operating frequency and exhaust volume, the module uses the following formula to calculate the velocity parameters at the velocity outlet boundary: ; In the formula, For flow velocity parameters; This refers to the real-time operating frequency of the induced draft fan. The rated operating frequency of the induced draft fan; This refers to the rated exhaust volume of the induced draft fan at its rated frequency. This refers to the physical cross-sectional area of ​​the opening of the dust hood.

[0055] Furthermore, the smelting process is accompanied by intense thermal radiation, and the thermal buoyancy generated by the decrease in the density of high-temperature gases is the main driving force for the flue gas to escape towards the top of the workshop. The flow field simulation module defines the outer wall surface of the smelting furnace as the isothermal wall heat source boundary and sets the physical properties of the ambient air to follow the law of incompressible ideal gases in order to reproduce the thermal buoyancy driving effect in subsequent calculations.

[0056] The flow field simulation module establishes a dynamic update mechanism. The fluid dynamics simulation employs a transient solver, whose calculation process is divided into continuous discrete time steps on the time axis. The flow field simulation module interacts with the transient solver through a user-defined function script. Before proceeding to the next time step iteration, the user-defined function script extracts the data timestamp corresponding to the current physical time and writes the unorganized flue gas mass release rate and corresponding velocity outlet boundary parameters at that timestamp into the memory matrix of the three-dimensional computational grid. This time-step incremental writing mechanism enables the boundary conditions to be updated synchronously with the actual on-site conditions. For the numerical discretization of various boundary condition matrix equations in the finite volume method, those skilled in the art can use readily available discretization formats in general fluid dynamics solvers; the numerical discretization process is a well-known technique in the field and will not be elaborated upon here.

[0057] After completing the three-dimensional computational mesh generation and boundary condition mapping, the flow field simulation module performs transient numerical simulation calculations based on computational fluid dynamics theory to obtain the dynamic diffusion and settling distribution of flue gas inside the workshop. The flow field simulation module reads the aforementioned constructed three-dimensional computational mesh and the corresponding dynamic boundary conditions, including the unorganized flue gas mass release rate, the set of mass source term elements within the preset source term radius, and the velocity outlet boundary parameters.

[0058] Due to the highly turbulent flow field generated by the convergence of the thermally driven airflow and the suction airflow from the fans within the smelting workshop, the flow field simulation module invokes the closed-loop control equations of the turbulence model. The turbulence model adopts the standard k-axis. Two-equation model. In the principles of fluid mechanics, turbulent motion of a fluid increases internal momentum and mass exchange. Standard k... The two-equation model calculates the turbulent viscosity coefficient of each grid node in the flow field by introducing the turbulent kinetic energy equation and the turbulent dissipation rate equation, so as to simulate the physical influence of fluid micro-particle pulsation on gas momentum and mass transfer.

[0059] The flow field simulation module establishes mass and momentum conservation equations based on the fluid control volume, and employs a semi-implicit pressure-velocity coupling algorithm to solve for the pressure and velocity. The formula for calculating the momentum conservation equation is as follows: ; In the formula, For fluid density; For fluid velocity vector; It is a time variable; For pressure; Equivalent dynamic viscosity; Let be the gravitational acceleration vector. To ensure a closed-loop solution to the momentum conservation equation, the equivalent dynamic viscosity is... The value of is obtained by algebraically summing the laminar dynamic viscosity of the fluid itself and the turbulent viscosity coefficient output from the aforementioned turbulence model. Based on the aforementioned incompressible ideal gas law, the fluid density... Local differences arise due to variations in the spatial distribution of the workshop temperature field. These density differences are combined with the gravitational acceleration vector. An upward thermal buoyancy driving term is generated in the momentum conservation equation; simultaneously, the gravitational acceleration vector... It acts on unorganized flue gas containing dust particles and heavy components, driving them to physically settle in a space without an upward airflow.

[0060] After obtaining the fluid velocity vector distribution, the flow field simulation module establishes component transport equations to simulate the convective diffusion and mixing processes of fugitive carbon emissions in the workshop airflow field. The flow field simulation module inputs the previously calculated fugitive flue gas mass release rate as source term data into the component transport equations. The calculation formula for the component transport equations is as follows: ; In the formula, For fluid density; The component mass concentration within the grid cell; It is a time variable; For fluid velocity vector; The equivalent diffusion coefficient; For component mass source terms. Equivalent diffusion coefficient. The diffusion coefficient is calculated by superimposing the molecular diffusion coefficient and the turbulent diffusion coefficient of the fluid. The value of the turbulent diffusion coefficient is obtained by dividing the previously acquired turbulent viscosity coefficient by the turbulent Schmidt number of the fluid medium, thereby establishing the physical mapping relationship between airflow pulsation and component diffusion. Component mass source term. The value of is obtained by dividing the aforementioned unorganized flue gas mass release rate by the volume of the three-dimensional computational grid where the source term element is located.

[0061] The flow field simulation module performs transient iterative calculations on the aforementioned partial differential equations. It monitors the residual values ​​of each equation at each time step. When the residual values ​​for mass, momentum, and components all decrease below a preset residual threshold, the physical flow field calculation for the current time step is considered converged. The preset residual threshold is set to a range of 10. -4 ~10 -6This requires balancing the accuracy of numerical calculations with the consumption of computer computing resources. For the algebraic equation system after the discretization of partial differential equations, those skilled in the art can use the Gauss-Seidel iterative algorithm to solve it. The iterative solution process of the equation system is a well-known technique in this field and will not be described in detail here.

[0062] After convergence is determined, the flow field simulation module will convert the fluid velocity vectors on each 3D computational grid node. Time-series data is recorded and spatially linked to generate dynamic diffusion path data. The flow field simulation module extracts the component mass concentration at each grid node. The data summarizes and outputs the concentration gradient distribution data within the workshop space. This dynamic diffusion path data and concentration gradient distribution data serve as the data foundation for downstream spatial integration calculations of total carbon emissions.

[0063] The carbon emission quantification module acquires three-dimensional computational grid data, fluid velocity vector distribution data, and concentration gradient distribution data output by the flow field simulation module. To achieve quantitative accounting of total fugitive carbon emissions, the carbon emission quantification module constructs a control volume within the physical space of the workshop based on the law of conservation of mass. The physical principle of the control volume is to artificially delineate a closed geometric region in the flow field and, by calculating the mass difference between the outflow and inflow into this region, as well as the accumulated mass within it, to regulate the conservation of matter within that region.

[0064] The carbon emission quantification module extracts the closed geometric surface enclosing the fluid domain inside the workshop from the aforementioned 3D geometric solid model, and performs a spatial topological set of all 3D computational grid cells located inside this closed geometric surface. The carbon emission quantification module then performs a scalar summation operation on the physical volumes of all 3D computational grid cells within the set to obtain the global control volume value. The global control volume covers the entire spatial range where fugitive flue gas undergoes physical diffusion and settling within the workshop.

[0065] Based on the constructed control volume, the carbon emission quantification module defines the integral boundary used to calculate mass flux. Since fugitive carbon emissions primarily escape into the external natural environment through open openings in the workshop, the carbon emission quantification module extracts physical sections from the building structure data that are directly connected to the external atmosphere, including natural ventilation doors and windows, roof ventilators, and unsealed factory openings. The module then extracts the corresponding two-dimensional mesh patches from the three-dimensional computational mesh and stitches them together to generate the flux integral boundary.

[0066] For each two-dimensional grid patch on the flux integration boundary, the carbon emission quantization module calculates its spatial normal vector. To accurately distinguish between outflow and inflow states in subsequent integration calculations, the carbon emission quantization module establishes a normal vector direction correction mechanism. The module obtains the three-dimensional coordinates of the centroid of the two-dimensional grid patch and, combined with the three-dimensional coordinates of the centroids of adjacent three-dimensional grid cells within its control volume, generates a direction determination vector. Through dot product comparison, the carbon emission quantization module uniformly corrects the spatial unit normal vectors of all two-dimensional grid patches to a direction pointing from the inside of the control volume to the external atmospheric environment. The determination and correction logic of the spatial unit normal vector adopts the following constraints: ; In the formula, It is the unit normal vector in space; The three-dimensional coordinates of the centroid of the two-dimensional mesh patch; These are the three-dimensional coordinates of the centroids of adjacent three-dimensional mesh cells. Satisfying this constraint indicates that the normal vector of the two-dimensional mesh patch is pointing towards the external atmospheric environment. When the dot product result is less than or equal to zero, it means that the spatial unit normal vector of the current two-dimensional mesh patch points into the control volume. The carbon emission quantification module multiplies the spatial unit normal vector by negative one for inverse processing, thereby achieving consistent correction of the outward orientation of the spatial unit normal vectors of all two-dimensional mesh patches on the flux integration boundary.

[0067] The carbon emission quantification module establishes a data alignment and binding mechanism. It traverses all two-dimensional mesh patches on the flux integration boundary, mapping and interpolating the previously acquired fluid velocity vector distribution data and concentration gradient distribution data to the centroid positions of the corresponding two-dimensional mesh patches. The carbon emission quantification module also positions and binds the physical volume values ​​of all three-dimensional computational meshes within the global control volume to their corresponding component mass concentrations.

[0068] By dividing the control volume and defining the integration boundary, the carbon emission quantification module transforms the continuous three-dimensional physical space into a discrete dataset, providing boundary conditions and data structure support for subsequent carbon emission flux integration and spatial accumulation calculations on the time and spatial axes. For the calculation of the three-dimensional coordinates of the centroid of the two-dimensional mesh patches and the spatial mapping interpolation algorithm for the mesh node data, those skilled in the art can use inverse distance weighted interpolation or centroid averaging algorithms. The spatial geometric interpolation process is a well-known technique in this field and will not be elaborated upon here.

[0069] After defining the control volume and integral boundary, the carbon emission quantification module performs a quantitative calculation of the total fugitive carbon emissions based on the principle of mass conservation. At the theoretical analysis level, the total carbon emissions from fugitive processes are estimated by spatially integrating the material density distribution and airflow velocity field within the flow field. The carbon emission quantification module incorporates a theoretical estimation model for fugitive carbon emissions, with the calculation formula as follows: ; In the formula, This refers to the total amount of unorganized carbon emissions; To simulate the carbon dioxide density distribution in space; For airflow velocity field; The three-dimensional control volume for integral calculation; This is a differential spatial unit within the three-dimensional control volume. Based on the aforementioned theoretical estimation model, the carbon emission quantification module transforms this spatial calculus process into numerical accumulation operations within a discretized three-dimensional computational grid, comprehensively calculating the boundary escaping flux and the spatial retention accumulation.

[0070] The carbon emission quantification module calculates the boundary flux flux. It calls upon the fluid velocity vector data, concentration gradient distribution data, and spatial unit normal vector mapped to the flux integral boundary. The module performs a dot product operation between the fluid velocity vector and the spatial unit normal vector on the two-dimensional grid patch. Since the spatial unit normal vector has been uniformly corrected to point towards the external atmosphere, a positive dot product result indicates carbon-containing gas flowing out of the workshop, while a negative result indicates external air flowing into the workshop. The module introduces a flux truncation function, including only values ​​with a dot product result greater than zero in the flux calculation. For dot product results less than or equal to zero, it indicates that there is reverse inflow of external airflow into the current grid patch, and no outward carbon emission loss will occur. The system forcibly sets the flux value generated by this item to zero to prevent the external clean airflow from offsetting the actual carbon emissions during mathematical accumulation. The module multiplies the truncated dot product result by the fluid density, the component mass concentration within the grid cell, the physical area of ​​the two-dimensional grid patch, and the discrete time step to calculate the carbon emission mass passing through the two-dimensional grid patch in a single time step. The calculation results for all time steps and all 2D mesh patches are summed to obtain the total escaping mass within the calculation cycle.

[0071] The carbon emission quantification module calculates the cumulative amount of carbon emissions retained in the space. It iterates through all three-dimensional computational grid cells within the previously acquired global control volume. At the end of the calculation cycle, the module extracts the fluid density and component mass concentration within each three-dimensional computational grid cell, multiplies them by their corresponding physical volume, and obtains the carbon emission mass within a single grid cell. The carbon emission masses of all three-dimensional computational grid cells are summed to obtain the total cumulative carbon emission mass retained within the workshop at the end of the calculation period.

[0072] The carbon emission quantification module determines the background carbon emission mass within the initial control volume based on the initial environmental conditions. By integrating boundary emission flux, spatially accumulated emissions, and background carbon emission mass, the carbon emission quantification module establishes the core formula for spatial calculus-based calculation: ; In the formula, The total amount of unorganized carbon emissions is accounted for. This represents the total number of time steps. This represents the total number of two-dimensional mesh patches on the flux integral boundary. For fluid density; This is the function for maximizing the value, used to perform the aforementioned flux truncation judgment; For fluid velocity vector; It is the unit normal vector in space; The component mass concentration within the grid cell; The physical area of ​​the two-dimensional mesh surface; The discrete time step; This represents the total number of 3D computational meshes within the global control volume. The physical volume of the three-dimensional computational mesh; To determine the background carbon emissions within the control volume at the initial moment.

[0073] For the values ​​of each parameter in the core formula, the background carbon emission mass within the control volume at the initial moment is determined. The value is obtained by multiplying the average environmental background concentration measured by a portable gas analyzer at the start of the accounting cycle by the global control volume value. For discrete time steps... The numerical setting and adaptive adjustment can be dynamically constrained by those skilled in the art based on the Courant number criterion in fluid dynamics calculations. The time discretization and step size configuration process is a well-known technology in this field and will not be elaborated here.

[0074] Using the aforementioned spatial calculus calculation algorithm, the carbon emission quantification module transforms the discrete grid data generated by flow field simulation into carbon emission entity mass values ​​over a continuous time period. After obtaining the total amount of fugitive carbon emissions, the system aggregates and stores it in the underlying database as data records for subsequent carbon emission quota management and corporate environmental review.

[0075] After acquiring the aforementioned flow field diffusion state and total carbon emission calculation results, the linkage control module establishes a logical assessment mechanism to determine whether fugitive emissions in the workshop exceed the standards. Since airflow disturbances during the smelting process are sporadic, the linkage control module needs to convert the fluid dynamics calculation data into environmental condition assessment indicators to determine whether to trigger subsequent physical treatment equipment intervention.

[0076] The linkage control module obtains the dynamic emission flux at the flux integration boundary and the component mass concentration of each three-dimensional computational grid cell within the workshop from the carbon emission quantification module. To eliminate numerical spikes that occur in a few time steps during transient flow field calculations, the linkage control module establishes a sliding window on the time axis. The length of the sliding window is set to include 5–15 consecutive discrete time steps. The linkage control module calculates the average mass of emissions to the external atmosphere for all time steps within the sliding window and converts it into a real-time emission rate. The linkage control module extracts all three-dimensional computational grid cells within the outer region of the dust collection hood at the top of the workshop. Based on the physical boundary of the dust collection hood, the linkage control module extends outward by a spatial distance of 1.0–3.0 meters, defining the three-dimensional computational grid cells falling within this distance range as the outer region of the dust collection hood. Among these grid cells, the linkage control module selects the component mass concentration with the highest value, taking it as the highest local spatial concentration.

[0077] The linkage control module establishes an evaluation system encompassing both external fugitive flux and internal local concentration to cover the external loss risk and internal retention risk of carbon emissions. In the principles of physics and environmental engineering, the fugitive flux to the external environment determines the overall carbon emission compliance, while the retention of high-concentration flue gas inside the workshop directly impacts the working environment and occupational health. The linkage control module performs dimensionless processing of real-time fugitive rates and the highest local concentration, assigns weighting coefficients, and constructs an evaluation index for anomaly in fugitive emissions. The formula for calculating the evaluation index for anomaly in fugitive emissions is: ; In the formula, An evaluation index for unorganized emissions anomalies; This is the flux weighting coefficient; Real-time dissipation rate; The preset escape rate threshold; This is the concentration weighting coefficient; This is the function for maximizing the value, used to perform the aforementioned flux truncation judgment; This refers to the collection of component mass concentrations within each grid unit in the area surrounding the dust collection hood at the top of the workshop. This is a preset concentration limit threshold.

[0078] For each constraint parameter in this formula, a preset exhaust rate threshold is used. Compared with the preset concentration limit threshold The value is determined based on national or local standards for fugitive emissions of air pollutants and the company's internal environmental control indicators. Flux weighting coefficient With concentration weighting coefficient The sum of these values ​​is fixed at 1. Under conditions where the focus is on controlling total carbon emissions, the linkage control module will adjust the flux weighting coefficient. The value range is set between 0.6 and 0.8; under the condition of focusing on controlling the occupational health environment inside the workshop, the linkage control module will use the concentration weighting coefficient. The value range is set to be between 0.6 and 0.8.

[0079] The linkage control module will calculate the abnormal evaluation index of fugitive emissions. The value is compared with the judgment benchmark. The judgment benchmark is set to 1.0. When the evaluation index for abnormal unorganized emissions... A value greater than 1.0 indicates that the current fugitive carbon emissions in the workshop have exceeded the set control targets in terms of flux or local concentration. The linkage control module determines that the current situation is an abnormal emission state and generates an abnormal trigger signal in the system logic controller. When the fugitive emission abnormality evaluation index is greater than 1.0, it indicates that the current fugitive carbon emissions have exceeded the set control targets in terms of flux or local concentration. When the value is less than or equal to 1.0, the linkage control module determines that the system is in a normal controlled state and continues to maintain the normal monitoring cycle.

[0080] By establishing an evaluation model that incorporates both rate and concentration indicators, the linkage control module achieves engineering-based risk identification of the fluid diffusion process. For the mean filtering algorithm within the sliding window and the normalization procedure for numerical dimensionlessness, those skilled in the art can implement them using general functions for time series data processing; the data smoothing and standardization processes are well-known techniques in the field and will not be elaborated upon here.

[0081] Upon generating an anomaly trigger signal, the linkage control module initiates the closed-loop control mechanism for fugitive carbon emissions. To control the continued spread of fugitive carbon emissions, the linkage control module generates adjustment commands for the physical treatment equipment within the workshop. This physical treatment equipment includes a secondary dust removal variable frequency induced draft fan, an electric roof ventilator, and a spray dust suppression system. Based on the aforementioned calculated fugitive emission anomaly evaluation index, the linkage control module calculates the target operating frequency of the variable frequency induced draft fan to suppress flue gas leakage by increasing the negative pressure suction capacity of the dust collection hood.

[0082] The linkage control module uses an incremental proportional-integral control algorithm to calculate the target induced draft fan operating frequency. In control engineering principles, this algorithm utilizes the error of the system's current state deviating from the normal state and the cumulative amount of the error over time to calculate the adjustment force required to correct this deviation. The linkage control module uses the deviation between the aforementioned fugitive emission anomaly evaluation index and the judgment benchmark value of 1.0 as input to calculate the frequency adjustment increment. The formula for calculating the target induced draft fan operating frequency is: ; In the formula, The target operating frequency of the induced draft fan; This refers to the real-time operating frequency of the induced draft fan. This is the proportional adjustment coefficient; An evaluation index for unorganized emissions anomalies; This is the integral adjustment coefficient; This refers to the time step number; To control the number of time steps within the cycle; is the discrete time step.

[0083] Regarding the parameter settings in this formula, the proportional adjustment coefficient... and integral adjustment coefficient The value of this parameter is determined by on-site engineers based on the power response characteristics and inertial load of the variable frequency induced draft fan motor. To balance system response time and prevent overshoot oscillation, the proportional adjustment coefficient... The value range is set between 1.5 and 5.0, and the integral adjustment coefficient is... The value range is set between 0.1 and 0.5. To protect the variable frequency induced draft fan motor from overload burnout, the linkage control module adjusts the calculated target induced draft fan operating frequency. Amplitude limiting is applied. The linkage control module calls the aforementioned rated operating frequency of the induced draft fan. When the target induced draft fan operates at a frequency Greater than the rated operating frequency of the induced draft fan At that time, the linkage control module will adjust the operating frequency of the target induced draft fan. Forced to be set to the rated operating frequency of the induced draft fan The value.

[0084] After calculating the target induced draft fan operating frequency, the linkage control module assesses the state of the workshop's natural ventilation boundary. Since the roof ventilators are located at the top of the workshop, driven by thermal buoyancy, their open state provides a physical channel for carbonaceous flue gas to escape into the external environment. During periods of abnormal triggering, the linkage control module generates on / off control commands to shut down the roof ventilators, cutting off the path of fugitive carbon emissions. Furthermore, the linkage control module sends activation commands to the spray dust suppression machines in the emission source area, utilizing the collision and condensation of water mist particles to intercept suspended carbonaceous dust in the air.

[0085] The linkage control module establishes an industrial communication connection with the field-programmable logic controller (FPGA). The linkage control module converts the target induced draft fan operating frequency into a digital-to-analog setpoint and packages the aforementioned switch control commands into a communication message, which is then sent to the FPGA via the industrial Ethernet protocol. Upon receiving the communication message, the FPGA drives the frequency converter, electric actuator, and corresponding relay contactors to perform physical actions.

[0086] The linkage control module acquires the device status word and real-time operating frequency data returned by the field programmable logic controller (PLC) and verifies the execution of control commands. After the physical equipment operates and changes the exhaust airflow field inside the workshop, the aforementioned flow field simulation module and carbon emission quantification module continue to perform transient iterative calculations based on the sensing and monitoring data of the next time period, re-evaluate, and output a new fugitive emission anomaly evaluation index. When the fugitive emission anomaly evaluation index falls back to 1.0 or below, the linkage control module releases the abnormal trigger signal and generates shutdown and reset commands. For the specific execution of the reset command, the linkage control module restores the set frequency of the variable frequency induced draft fan to the stable operating frequency at the moment before the abnormal trigger, sends a shutdown message to the spray dust suppression machine to cut off its power supply circuit, and simultaneously sends an open command to the electric roof ventilator to restore the natural ventilation boundary, so that each physical treatment device returns to its normal basic operating condition, completing the closed-loop feedback control of the treatment process. For the message encapsulation and parsing of the industrial Ethernet protocol and the low-level driver writing of the PLC, those skilled in the art can use a general industrial automation control specification library to implement it. The device communication and hardware driver process are well-known technologies in this field and will not be described in detail here.

[0087] After controlling the physical equipment on-site, the system constructs a data collaboration and management closed-loop mechanism at the business logic layer to achieve asset-based management of carbon emission data and cross-departmental information linkage. The system obtains the total unorganized carbon emission accounting amount calculated above and reads the total organized carbon emission amount within the same period from the exhaust stack online monitoring system. The system algebraically sums the two to obtain the enterprise's actual total carbon emission load within the accounting period, and stores it in a timestamp format in the underlying relational database to establish a carbon emission ledger for compliance review.

[0088] To prevent the risk of exceeding carbon emission limits and defaulting, the system establishes a dynamic early warning model for carbon emission quotas. In the principles of carbon asset management, enterprises need to extrapolate the total quota consumption at the end of the entire accounting cycle based on the current emission rate, thereby intervening in production load before over-emissions occur. The system calculates a carbon quota consumption early warning index based on the current total carbon emission load and the production time already consumed. The formula for calculating the carbon quota consumption early warning index is: ; In the formula, This serves as an early warning index for carbon quota consumption. The total amount of unorganized carbon emissions is accounted for. Total organized carbon emissions; This refers to the total carbon emission allowances within the accounting period; This is the total duration of the accounting cycle; This represents the current elapsed time.

[0089] Regarding parameter settings, the total carbon emission allowance within the accounting period... The values ​​are assigned from the verification documents issued by local environmental protection management departments. The calculated carbon quota consumption early warning index... When the emission level exceeds the set warning threshold, it indicates that the company will face the risk of quota overruns at the end of the accounting period, based on the current carbon emission rate. The system sets the warning threshold range between 0.9 and 1.0 to allow for a buffer period for management intervention.

[0090] Upon triggering a quota warning or receiving an abnormal trigger signal generated by the aforementioned linkage control module, the system initiates a cross-business system collaborative procedure. Simple automatic equipment adjustment may not be sufficient to handle continuous leakage caused by damaged dust collector bags or cracked pipes. Therefore, the system packages the anomaly type, location, current carbon emission concentration, and warning index into a structured data message. The system then invokes a message middleware, using a lightweight message queue publish-subscribe mechanism, to push this message to the enterprise's Manufacturing Execution System and Equipment Asset Management System.

[0091] When the Manufacturing Execution System (MES) receives a structured data message and parses it to show that the carbon allowance consumption warning index exceeds the warning threshold, it triggers a production load downgrade procedure. The MES adjusts the production scheduling of underlying equipment, extends the feeding interval of the smelting furnace, and issues a postponement instruction to the workshop dispatch console for high-emission processes. Through physical-level source reduction of raw materials and process delays, the MES curbs the rapid consumption of carbon emission allowances from the production end.

[0092] After receiving structured data messages, the equipment asset management system parses the message content in the background, generates an environmental protection equipment anomaly inspection work order containing the spatial coordinates of the anomaly source, and dispatches it to the portable mobile terminals of on-site maintenance personnel via the enterprise intranet. Maintenance personnel then proceed to the corresponding area according to the work order instructions to perform manual physical inspection and repair of vulnerable parts such as dust collector bags and dust hood flange sealing surfaces.

[0093] After maintenance personnel complete on-site operations and submit confirmation of fault elimination via a portable mobile terminal, the system receives the feedback status code. The system records the closed-loop timeline from the occurrence of the anomaly to its completion in the relational database and updates the work order status field to "completed." At this point, from the underlying flow field status perception and automatic hardware intervention to the top-level quota early warning assessment and manual dispatching of maintenance orders, the system completes the integrated hardware and software management closed loop for the control of fugitive carbon flows.

[0094] For high-concurrency data writing in relational databases and message push and subscription parsing of lightweight message queue protocols, those skilled in the art can use industrially common microservice architecture components for configuration and deployment. The data storage and network communication processes are well-known technologies in the field and will not be elaborated here.

[0095] To verify the system's operational logic in a real-world engineering environment, a scenario based on a non-ferrous metal smelter was configured. The smelter is a single-story, large-space steel structure workshop with a total length of 120.0m, a span of 40.0m, and a roof ridge height of 24.0m. Louvered sidewalls are located on the north and south sides of the workshop as air inlets for natural ventilation; an 80.0m long electrically powered roof ventilator is located in the center of the roof, serving as an exhaust vent driven by thermal pressure.

[0096] Inside this physical space, a tilting converter serves as both a carbon emission source and a heat source. During charging and tilting slag removal operations, unorganized fugitive flue gas may be generated due to inadequate sealing of the secondary dust collection hood. The system presets the converter's physical operating parameters based on on-site measured data. The initial temperature of the carbon-containing flue gas escaping from the furnace mouth is set to 180℃, and the outlet gas velocity is set to 2.5 m / s. The mixed mass concentration of carbon monoxide and carbon dioxide in the flue gas is set to 500.0 mg / m³. 3 These initial source terms constitute the source term boundary conditions used by the flow field simulation module for solving the problem.

[0097] The system deploys hardware for state awareness within the workshop. An array of anemometers is installed at 2.0m intervals along the vertical direction at the side wall louvers to collect the initial wind speed and direction of the external atmosphere flowing into the workshop. Two laser gas analyzers are mounted on the crossbeam structure at an elevation of 15.0m above the converter. These analyzers acquire spatial three-dimensional point cloud concentration data of fugitive gases by cross-scanning the airflow area above and around the dust collection hood. Regarding the configuration of treatment equipment, the workshop is equipped with facilities operating at a rated frequency... This includes a 50.0Hz variable frequency induced draft fan and a spray dust suppression system mounted on supports around the converter. Wherein, This refers to the rated operating frequency of the induced draft fan.

[0098] The system inputs the physical dimension data, the three-dimensional spatial coordinates of the sensors, and the rated operating parameters of the mechanical equipment from the above scenario into the underlying database. The system calls a three-dimensional solid modeling program to generate a geometric model proportional to the actual workshop based on the physical dimensions, and divides it into a total of 1.2 million three-dimensional computational mesh cells.

[0099] At the level of pre-setting computational parameters, the system defines the time discretization parameters and physical property parameters for flow field simulation. Discrete time step The value is set to 0.5s, and the total calculation cycle is set to 24h. The fluid medium is assumed to be an incompressible ideal gas mixture, and its dynamic viscosity and thermal conductivity are based on constant property data under standard atmospheric pressure. Where, is the discrete time step.

[0100] To support subsequent evaluation and coordinated control calculations, the system presets the determination thresholds and control parameters of each algorithm module. The system presets the escape rate threshold to 5.0kg / h, and presets the concentration limit threshold to 150.0mg / m 3 . The system sets the flux weight coefficient to 0.6, and the concentration weight coefficient to 0.4. In the control algorithm part, the system sets the proportional adjustment coefficient to 2.5, and the integral adjustment coefficient to 0.2. At the same time, according to the environmental protection verification document, the system presets the total carbon emission quota within the accounting period to 5000.0kg, and sets the early warning threshold for carbon quota consumption to 0.95. In the formula, is the preset escape rate threshold; is the preset concentration limit threshold; is the flux weight coefficient; is the concentration weight coefficient; is the proportional adjustment coefficient; is the integral adjustment coefficient; is the total carbon emission quota within the accounting period.

[0101] By completing the spatial geometric layout, sensor point binding, calculation parameter configuration and logic threshold setting, the system establishes the physical and digital mapping benchmark for the smelting scenario, and provides a complete initial state parameter set for subsequent calculation and deduction.

[0102] For the parameter configuration of the industrial sensor communication interface and the call of the data import and export interface of the 3D drawing software, those skilled in the art can implement it by adopting a general data bus standard, and the interface configuration process belongs to the well-known technology in the art, which will not be repeated here.

[0103] During the monitoring period of system operation, the No. 3 tilting converter performs molten iron pouring operation. As a large amount of black smoke overflows from the furnace mouth and cannot be completely captured by the secondary dedusting suction hood, part of the carbon-containing flue gas diffuses to the top space of the workshop. The industrial vision acquisition equipment, laser gas analyzer and anemometer array deployed on site collect the concentration point cloud and wind speed boundary data of this process in real time, and transmit them synchronously to the flow field simulation module. Based on the preset 1.2 million 3D calculation grid cells, the flow field simulation module performs physical solution and iterative calculation on the spatial diffusion trajectory of the flue gas.

[0104] The linkage control module invokes the background algorithm to analyze and determine that the dust removal efficiency has decreased. The carbon emission quantification module uses the aforementioned flow field data and immediately calculates the additional carbon emissions from this overflow using spatial calculus, obtaining a current real-time emission rate of 6.2 kg / h. Simultaneously, it extracts the component mass concentration set from each grid cell within the area surrounding the dust collection hood at the top of the workshop, selecting the maximum value in this set as 210.0 mg / m³. 3 The linkage control module obtains the above values ​​and, based on the preset flux weighting coefficient of 0.6 and concentration weighting coefficient of 0.4, calculates the current fugitive emission anomaly evaluation index as 1.304. Because this calculated value is greater than the judgment benchmark value of 1.0, the linkage control module determines that the fugitive emissions exceed the control target and sets an anomaly trigger flag in the system memory.

[0105] At the on-site treatment equipment coordination level, the linkage control module initiates control logic based on the abnormal trigger flag. The system reads the current real-time operating frequency of the induced draft fan as 40.0Hz. Based on a proportional adjustment coefficient of 2.5 and an integral adjustment coefficient of 0.2, the linkage control module uses an incremental proportional-integral algorithm to substitute the deviation value of the abnormal evaluation index into the calculation, obtaining a target induced draft fan operating frequency of 40.8Hz. Since this frequency value does not exceed the rated operating frequency of 50.0Hz for the induced draft fan, the linkage control module sends it as a digital-analog setpoint to the field programmable logic controller (PLC). The PLC drives the frequency converter to increase the induced draft fan speed, increasing the negative pressure suction volume inside the dust collection hood. Simultaneously, the linkage control module sends a closing command to the electric roof ventilator to cut off the physical natural escape channel at the top, and starts the spray dust suppression machine around the converter, using water mist particles to physically intercept carbonaceous dust.

[0106] At the enterprise business management collaboration level, the system reads and accumulates the current total unorganized carbon emissions and total organized carbon emissions, and substitutes them into the dynamic early warning model for carbon emission quotas. At this moment, the system reads that the current consumed time is 12 hours. The underlying database records the current total unorganized carbon emissions as 800.0 kg and the total organized carbon emissions as 1625.0 kg. The system divides the sum of the above emission data (2425.0 kg) by the preset quota total of 5000.0 kg and multiplies it by the time ratio coefficient. Through internal system deduction, the current carbon quota consumption early warning index reaches 0.97. Because 0.97 is greater than the set early warning threshold of 0.95, the system calls the message middleware to directly push the rectification work order data message containing the spatial coordinates of the abnormal state, the additional carbon emissions, and the early warning index to the central control room server.

[0107] After receiving and parsing the message, the Manufacturing Execution System (MES) triggers a production load downgrade procedure, delaying the next batch feed time of the converter by fifteen minutes via the production scheduling interface to reduce carbon flow generation at the source of the smelting process. Simultaneously, the Equipment Asset Management System extracts the spatial coordinate information from the message, generates an environmental protection equipment anomaly inspection work order, and dispatches it to the portable mobile terminals of on-site maintenance personnel. Based on this, maintenance personnel arrive at the designated physical area, physically seal the gaps in the sealing baffles on the outer edge of the dust hood, and submit a confirmation receipt on their mobile terminals confirming the fault completion.

[0108] With the physical leak point sealed and the induced draft fan's suction volume continuously adjusted, the flow field simulation module's calculations at subsequent time steps showed a decrease in the concentration of stagnant flue gas inside the workshop. The linkage control module recalculated the anomaly evaluation index for fugitive emissions. When this index value dropped back to 0.85, the linkage control module determined that the environmental state had returned to normal and removed the anomaly trigger flag. The linkage control module issued a reset command, restoring the variable frequency induced draft fan to its pre-anomaly trigger frequency of 40.0Hz, the electric roof ventilator's push rod actuated to reopen the natural ventilation channel, and the main power supply circuit of the spray dust suppression machine was cut off and spraying stopped. Thus, the system completed the closed-loop management of the entire process of data perception, automatic intervention, and business scheduling for this fugitive carbon emission loss event.

[0109] For data packet encryption and encapsulation in cross-module communication at the system's underlying level and transaction rollback processing in relational databases, those skilled in the art can use information security standards and database engine mechanisms commonly used in industrial software for configuration. The data security assurance and transaction integrity verification process is a well-known technology in this field and will not be elaborated here.

[0110] Specific application examples: To verify the application effect of the computer vision-based steelmaking fugitive emission monitoring and carbon flow analysis system and method provided in the embodiments of the present invention, a prototype system was deployed at a steelmaking plant and a 72-hour on-site comparative verification experiment was conducted.

[0111] The experiment was conducted by comparing two operating modes: Benchmarking Group (Traditional Model): Employs single-visual opacity monitoring and fixed threshold alarm; The secondary dust removal induced draft fan operates at a constant frequency (50.0Hz) at full load without incorporating flow field fluid dynamics simulation and dynamic intervention.

[0112] Experimental group (patent of this invention): The system employs multi-source sensing fusion inversion, carbon flow spatial calculus quantization based on flow field grids, and an evaluation index based on fugitive emission anomalies. Incremental PI closed-loop linkage control.

[0113] Multi-source data fusion and concentration inversion accuracy verification: Please see the appendix Figure 3 During the experiment, portable infrared gas analyzers were placed at equal intervals along the edge of the converter dust hood, and their measurement data were used as the ground truth concentration acquisition benchmark. Figure 3 The study demonstrates the instantaneous concentration values ​​of fugitive emissions calculated by different algorithms over the entire cycle (60 minutes) from feeding to tilting and dumping slag. Due to dust obstruction and varying background lighting at the smelting site, traditional single-vision algorithms (relying solely on opacity feature parameters) are insufficient. (Estimation) Due to illumination interference, the calculated values ​​were too high during the strong light arc phase from the 15th to the 25th minute; single-spectral analysis failed to detect the flue gas concentration when the flue gas deviated from the detection optical path from the 40th to the 50th minute (the core anomaly point was the 45th minute). This invention addresses this by fusing feature vectors from multiple sources. Input a multilayer fully connected inversion mapping model and output the calculation formula. This invention combines the macroscopic spatial anti-deviation characteristics of vision with the microscopic anti-illuminance characteristics of the spectrum. The results show that the output of this invention... Pearson correlation coefficient between the curve and the true value measured by the portable instrument ( The mean absolute error (MAE) reached 0.965, and the mean absolute error (MAE) decreased to 8.4 mg / m³. 3 This improves data stability under various environmental interferences.

[0114] Verification of carbon flow spatial calculus and dynamic linkage control: Please see the appendix Figure 4 In the figure, the left vertical axis represents the anomaly evaluation index, and the right vertical axis represents the target operating frequency of the induced draft fan. At the 15th minute of the experimental cycle, carbon-containing flue gas is generated during the converter's molten iron addition. The flow field simulation module, based on the component transport equation and flux cutoff function, outputs the dynamic concentration gradient distribution. The carbon emission quantification module uses the core formula of spatial calculus. The rise in boundary escaping flux was calculated.

[0115] like Figure 4 As shown, the abnormal evaluation index of fugitive emissions calculated by the linkage control module (Based on normalized weighted average of flux and local concentration) The threshold of 1.0 was exceeded at approximately 16 minutes, and peaked at approximately 1.50 at 20 minutes. The system then triggered a linkage mechanism based on the formula. Equipment control is implemented. The target operating frequency of the variable frequency induced draft fan is set. The frequency is gradually increased from a base of 40.0Hz to a maximum of 50.0Hz. With the increase in suction volume and the automatic shutdown of the electric roof ventilator, the escaping smoke is collected and treated. By the 35th minute... The frequency of the induced draft fan dropped back to the baseline safe range of around 0.6, and under algorithmic control, the frequency decreased to a basic operating state. This process demonstrates that the control logic of this invention can control carbon emissions and reduce the frequent start-ups and shutdowns of mechanical equipment.

[0116] Comparison of energy conservation and emission reduction data: Please see the appendix Figure 5 The left vertical axis of the graph represents the total cumulative loss of fugitive carbon emissions, and the right vertical axis represents the cumulative power consumption of the variable frequency induced draft fan.

[0117] By comparing and verifying the cumulative data over 72 hours using calculus: In terms of carbon emission control, the benchmark group, due to the lack of flow field control, had a total loss of 1450.0 kg of fugitive carbon emissions. Through flow field simulation and anomaly index linkage control, the experimental group reduced the total cumulative fugitive carbon emissions to 320.0 kg, thus lowering the carbon quota consumption warning index. The risk of exceeding the limit.

[0118] In terms of energy consumption, the benchmark group's induced draft fan maintained a frequency of 50.0 Hz, with a cumulative power consumption of 2150.0 kW. The experimental group adjusted the operating frequency using a PI control algorithm. When the average operating frequency was 42.3Hz and the cumulative power consumption was 1380.0kW, the power consumption was reduced by about 35.8% compared with the benchmark group.

[0119] The above results demonstrate that the method provided in this embodiment of the invention takes into account both dust control and carbon emission management.

Claims

1. A computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking, characterized in that, Includes the following steps: Acquire on-site video stream data and synchronously collected spectral analysis data of the target monitoring area; Image processing is performed on the on-site video stream data to extract the visual feature vector of the flue gas plume; The visual feature vector of the flue gas plume is fused with the spectral analysis data, input into a preset inversion mapping model, and the instantaneous concentration value of unorganized emissions is calculated. Based on the on-site video stream data, the centroid two-dimensional pixel coordinates are extracted and combined with the spatial transformation matrix to obtain the three-dimensional spatial coordinates of the emission source center. A three-dimensional mesh model of the workshop physical space is established. The instantaneous concentration value of the fugitive emissions and the three-dimensional spatial coordinates of the emission source center are used as the calculation boundary conditions and input into the three-dimensional mesh model for fluid dynamics solution. The dynamic diffusion path data and concentration gradient distribution data are output. By combining the dynamic diffusion path data and the concentration gradient distribution data, a spatial volume integral operation is performed to output the estimated total amount of fugitive carbon emissions within the accounting period; An anomaly is determined based on the instantaneous concentration of the unorganized emissions and the estimated total amount of unorganized carbon emissions. If an anomaly is determined, a hardware control command is generated, and a business work order is generated for feedback.

2. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 1, characterized in that, The visual feature vector of the flue gas plume includes geometric shape feature parameters, diffusion velocity field parameters, opacity feature parameters, and texture feature parameters; The step of image processing of the on-site video stream data to extract visual feature vectors of the smoke plume includes: The deep learning visual feature extraction network outputs a binarized mask image of the flue gas plume in the on-site video stream data, locates the flue gas connected region, extracts the contour coordinate set of the flue gas connected region, uses the contour coordinate set to calculate the aspect ratio, perimeter and area of ​​the geometric bounding box of the flue gas connected region, and performs tensor concatenation of the aspect ratio, perimeter and area of ​​the geometric bounding box to obtain geometric morphological feature parameters. The pixel brightness data of the smoke-connected region in two consecutive frames of the on-site video stream data is obtained. The pixel position displacement between adjacent frames is calculated based on the dense optical flow algorithm. The optical flow constraint equation is constructed and solved. The optical flow velocity vector distribution of all two-dimensional pixel coordinates in the smoke-connected region is obtained. The mean and variance matrix of the optical flow velocity vector distribution are calculated as diffusion velocity field parameters.

3. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 2, characterized in that, The step of image processing of the on-site video stream data to extract visual feature vectors of the flue gas plume also includes: Acquire workshop video images under historical smoke-free conditions, establish a background reference model based on the workshop video images, extract the pixel brightness values ​​of the corresponding smoke area in the current frame, calculate the average brightness of all pixels in the smoke-connected area to obtain the pixel brightness values ​​of the smoke area, and calculate the average brightness of pixels at the same coordinates in the background reference model to obtain the pixel brightness values ​​of the background area. Calculate the opacity by comparing brightness attenuation to obtain the opacity feature parameters. Extract the texture feature parameters of the flue gas connected region; The geometric shape feature parameters, the diffusion velocity field parameters, the opacity feature parameters, and the texture feature parameters are input into the fully connected layer of the network for dimensionality reduction and splicing combination, and the visual feature vector of the smoke plume is output.

4. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 1, characterized in that, Before fusing the visual feature vector of the flue gas plume with the spectral analysis data, the process further includes performing spatiotemporal synchronization processing on the on-site video stream data and the spectral analysis data: In the time dimension, the underlying clock is synchronized using the network time protocol. A global timestamp is assigned to each frame of video image and each spectral data. Based on the visual data timestamp when the on-site video stream data is acquired, the adjacent first spectral data timestamp and second spectral data timestamp are found in the spectral data sequence corresponding to the spectral analysis data. The interpolated spectral value is calculated by a linear interpolation algorithm. In the spatial dimension, the spatial transformation matrix is ​​calculated using the pre-set calibration reference on site, and the two-dimensional pixel coordinates extracted from the video image are mapped to the three-dimensional physical coordinates of the workshop. The intersection calculation is performed with the spatial straight line equation of the spectral detection optical path. The spatial intersection calculation is used to verify whether the optical path of the current spectral data detection passes through the flue gas plume region captured by the video image. If it is determined that it passes through, the corresponding spectral analysis data is included in the subsequent fusion processing.

5. The computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking, as described in claim 4, is characterized in that... The process of fusing the visual feature vector of the flue gas plume with the spectral analysis data, inputting it into a preset inversion mapping model, and calculating the instantaneous concentration value of fugitive emissions includes: The maximum-minimum normalization algorithm is used to map the visual feature vector of the flue gas plume and the interpolated spectral values ​​to the numerical range of 0 to 1 respectively, and then perform matrix splicing according to the set dimensions to generate a multi-source fusion feature vector. The multi-source fusion feature vector is input into the inversion mapping model of a multi-layer fully connected neural network architecture, and the nonlinear relationship between the input multi-source fusion feature vector and the fugitive emission concentration is fitted by a nonlinear activation function. The nonlinear activation function employs a linear rectification function to force negative results calculated by the network to zero and retain positive results, thereby ensuring that the output results meet the physical constraint that gas concentration values ​​cannot be negative in the real world, and outputting the instantaneous concentration value of the unorganized emissions.

6. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 2, characterized in that, The process of establishing a three-dimensional mesh model of the workshop's physical space, and inputting the instantaneous concentration values ​​of the fugitive emissions and the three-dimensional spatial coordinates of the emission source center as computational boundary conditions into the three-dimensional mesh model for fluid dynamics solution, includes: The emission source release area is obtained by combining the area value of the flue gas connected region in the binarized mask image with the spatial transformation matrix to proportionally calculate the physical area from the pixel area; Using the three-dimensional spatial coordinates of the emission source center as the center of a sphere, a spherical spatial region is divided according to a preset source term radius. All three-dimensional computational grid cells falling into the spherical spatial region are marked as a mass source term cell set. The value of the preset source term radius is obtained by converting the emission source release area using the equivalent circle area formula; The unorganized flue gas mass release rate is obtained by multiplying the instantaneous concentration value of the unorganized emissions, the initial escape velocity of the flue gas plume obtained by converting the diffusion velocity field parameters, and the release area of ​​the emission source. The natural ventilation doors and windows and roof ventilators of the workshop are defined as the pressure inlet boundary, and the dust collection hood opening of the secondary dust removal induced draft fan in the workshop is defined as the velocity outlet boundary. The flow velocity parameters of the velocity outlet boundary are calculated based on the real-time operating frequency of the induced draft fan fed back by the field programmable logic controller. The outer wall surface of the smelting furnace is defined as the boundary of the isothermal wall heat source, and the physical properties of the ambient air are set to follow the law of incompressible ideal gases.

7. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 6, characterized in that, The partial differential equations involved in the fluid mechanics solution include the mass conservation equation, the momentum conservation equation, and the component transport equation. The output dynamic diffusion path data and concentration gradient distribution data include: Using standard k The turbulent viscosity coefficient of each grid node in the flow field is calculated by the two-equation model. The upward thermal buoyancy driving term is generated in the momentum conservation equation by combining the density difference with the gravitational acceleration vector. The momentum conservation equation is established and the pressure and velocity coupling solution is performed by the semi-implicit pressure relationship equation algorithm to obtain the fluid velocity vector distribution. The mass release rate of the unorganized flue gas is input as source term data into the component transport equation, and transient iterative calculations are performed on the mass conservation equation, the momentum conservation equation, and the component transport equation. When the residual values ​​of mass, momentum, and components all drop below the preset residual threshold, the physical flow field calculation of the current time step is determined to be converged. The fluid velocity vectors on each three-dimensional computational grid node are recorded in time series and spatially connected to generate dynamic diffusion path data. The preset residual threshold value range is set to 10. -4 ~10 -6 between; Extract the component mass concentration at each grid node and summarize the concentration gradient distribution data within the workshop space.

8. The computer vision-based method for monitoring fugitive emissions and analyzing carbon flow in steelmaking, as described in claim 7, is characterized in that... The step of combining the dynamic diffusion path data and the concentration gradient distribution data to perform spatial volume integral calculations and outputting the estimated total amount of fugitive carbon emissions within the accounting period includes: Extract the two-dimensional mesh patches corresponding to the physical cross-sections that are directly connected to the external atmospheric environment in the three-dimensional computational grid, and stitch them together to generate the flux integral boundary; Calculate the spatial unit normal vector of the two-dimensional mesh patch, and combine it with the three-dimensional coordinates of the centroid of the adjacent three-dimensional mesh cells inside the control volume where the two-dimensional mesh patch is located to generate a direction determination vector. By comparing the dot product, the spatial unit normal vector is uniformly corrected to the direction from the inside of the control volume to the external atmospheric environment. The fluid velocity vector on the two-dimensional grid patch is multiplied by the corrected spatial unit normal vector. A flux truncation function is used to include only the values ​​with a product greater than zero in the flux calculation. The total escaping mass within the calculation period is obtained by multiplying the fluid density, the component mass concentration in the grid cell corresponding to the concentration gradient distribution data, the physical area of ​​the two-dimensional grid patch, and the discrete time step corresponding to the fluid dynamics solution. This is used as the boundary escaping flux. Extract the fluid density and the component mass concentration in each three-dimensional computational grid, and multiply them with the physical volume corresponding to the three-dimensional computational grid to obtain the carbon emission mass in a single grid. Add up the carbon emission masses of all three-dimensional computational grids to obtain the total cumulative carbon emission mass retained in the workshop at the end of the accounting period as the spatial retention accumulation. The boundary escaping flux is added to the spatial stagnation accumulation, and the background carbon emissions within the control volume at the initial moment are subtracted to obtain the total amount of fugitive carbon emissions, which is then used as the estimated total amount of fugitive carbon emissions.

9. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 8, characterized in that, The abnormal state determination is performed based on the instantaneous concentration value of the fugitive emissions and the estimated total amount of fugitive carbon emissions. If an abnormality is determined, hardware control instructions are generated, including: A sliding window is established on the time axis, and the average value of the mass of the external atmospheric environment emitted at all time steps within the sliding window is calculated. The average value is then converted into the real-time emission rate. Extract all three-dimensional computational mesh cells within the outer area of ​​the dust collection hood at the top of the workshop. The outer area of ​​the dust collection hood is: Using the physical boundary of the dust collection hood as a reference, extend outward by a spatial distance of 1.0 to 3.0 meters to cover the area. In the three-dimensional computational grid cell, the item with the largest component mass concentration is selected as the highest concentration in the local space; The real-time emission rate and the highest concentration in the local space are dimensionless, and weight coefficients are assigned to the dimensionless real-time emission rate and the highest concentration in the local space respectively to construct an evaluation index for unorganized emissions anomalies, wherein the weight coefficients include flux weight coefficients and concentration weight coefficients. When the evaluation index of unorganized emissions is greater than the judgment benchmark value, it is determined that the current state is abnormal and an abnormal trigger signal is generated in the system logic controller to trigger the generation of the hardware control command. The value of the determination benchmark is set to 1.0; The generation of hardware control instructions includes: The digital-analog setpoint corresponding to the target induced draft fan operating frequency is generated and sent to the field programmable logic controller to adjust the secondary dust removal variable frequency induced draft fan. At the same time, a switch control command is generated to shut down the roof ventilator and a start command is sent to the spray dust suppression machine in the emission source area. The specific process of adjusting the secondary dust removal variable frequency induced draft fan includes: An incremental proportional-integral control algorithm is adopted, and the deviation between the unorganized emission anomaly evaluation index and the judgment benchmark value is used as the error input of the current state of the system deviating from the normal state. The target induced draft fan operating frequency is calculated by combining the cumulative amount of the error over time, the proportional adjustment coefficient and the integral adjustment coefficient. The operating frequency of the target induced draft fan is limited to ensure that it does not exceed the rated operating frequency. The operating frequency of the target induced draft fan is converted into the digital-analog set value, and the switch control command is packaged into a communication message and sent through the industrial Ethernet protocol.

10. The method for monitoring fugitive emissions and analyzing carbon flow in steelmaking based on computer vision according to claim 1, characterized in that, The business work order data includes the three-dimensional spatial coordinates of the emission source center, the time of occurrence, and the estimated total amount of fugitive carbon emissions. The generated business work order data is fed back, including: The business work order data is sent to the central control room server; Before sending the business work order data, the following is also included: Obtain the total organized carbon emissions within the same period as the accounting period, and algebraically sum the total organized carbon emissions with the estimated total unorganized carbon emissions to obtain the enterprise's actual total carbon emission load within the accounting period. The carbon quota consumption early warning index is calculated by dividing the actual total carbon emission load by the total carbon emission quota within the accounting period and multiplying it by the ratio of the total duration of the accounting period to the current duration already consumed. When the carbon quota consumption warning index exceeds the set warning threshold, the abnormality type, location of occurrence, current carbon emission concentration and warning index are packaged into a structured data message and pushed to the enterprise's manufacturing execution system and equipment asset management system through the message middleware. This is used to trigger the manufacturing execution system to execute the production load downgrade procedure, adjust the production schedule of the underlying equipment, extend the feeding interval of the smelting furnace, and issue a temporary suspension instruction for high emission processes to the workshop dispatch console. And to instruct the equipment asset management system to parse the message content in the background after receiving the structured data message and generate an environmental protection equipment anomaly inspection work order containing the spatial coordinate information of the anomaly source; The warning threshold is set to a value between 0.9 and 1.0.

Images