Industrial smoke emission monitoring method and related apparatus

Abstract

The application discloses an industrial smoke emission monitoring method and related device, and relates to the technical field of computers. Firstly, the application fuses visible light and infrared thermal imaging multi-modal data, and introduces meteorological parameters and a physical constraint of the advection-diffusion equation when generating a concentration field in reverse diffusion, so that the reconstructed smoke concentration space-time field matches multi-modal visual features and conforms to smoke diffusion rules, avoids non-physical distortion such as local mutation of a pure data-driven model, and significantly improves the concentration field accuracy. Secondly, a physical information neural network is trained by taking the concentration field as data constraint and taking a physical equation as constraint, spatial peak values are extracted from a physically consistent source field as emission source coordinates and intensity by jointly inverting a source field, a wind speed field and a diffusion coefficient field, and compared with traditional numerical fitting, noise interference and parameter error are significantly reduced. Therefore, the scheme can significantly improve the emission concentration monitoring accuracy and the emission source identification accuracy.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method and related apparatus for monitoring industrial dust emissions. Background Technology

[0002] Industrial dust is a major air pollutant in industrial production settings, and monitoring its emissions is a core aspect of ecological and environmental supervision and industrial pollution control. Industrial dust monitoring includes emission concentration monitoring and emission source identification.

[0003] Current traditional industrial dust emission monitoring methods include industrial dust monitoring schemes based on single-point gas sensors and vision-based industrial dust monitoring schemes. However, traditional industrial dust emission monitoring methods are generally in the stage of "single-modal monitoring and simple numerical analysis", with low accuracy in emission concentration monitoring and emission source identification.

[0004] Therefore, how to provide an industrial dust emission monitoring solution to improve the accuracy of emission concentration monitoring and emission source identification has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, this application provides a method and related device for monitoring industrial dust emissions, so as to improve the accuracy of emission concentration monitoring and emission source identification. The specific solution is as follows:

[0006] The first aspect of this application provides a method for monitoring industrial dust emissions, including:

[0007] Acquire visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series of the target industrial area;

[0008] Based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, iterative denoising processing is performed according to the reverse diffusion process to obtain the smoke and dust concentration field at each time moment. The smoke and dust concentration fields at each time moment are then arranged in chronological order to generate the spatiotemporal field of smoke and dust concentration.

[0009] The spatiotemporal coordinates of the target to be monitored within the target industrial area are input into the trained physical information neural network to obtain the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field. The physical information neural network is trained with the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints.

[0010] The spatial peak position and peak intensity are extracted from the target predicted source term field. The spatial peak position is used as the emission source coordinates, and the peak intensity is used as the emission source intensity.

[0011] In one possible implementation, the step of performing iterative denoising processing based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, according to a reverse diffusion process, to obtain the smoke and dust concentration field at each time point includes:

[0012] The visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series are standardized to obtain a standardized visible light tensor sequence, a standardized infrared tensor sequence, and a meteorological parameter field sequence.

[0013] Cross-modal spatiotemporal feature fusion processing is performed on the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time step;

[0014] Using the spatiotemporal fusion feature map at each time point as input, and based on the Gaussian noise field and the meteorological parameter field sequence, iterative denoising processing is performed according to the reverse diffusion process to obtain the dust concentration field at each time point.

[0015] In one possible implementation, the cross-modal spatiotemporal feature fusion processing of the normalized visible light tensor sequence and the normalized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time step includes:

[0016] Based on the standardized visible light tensor sequence and the standardized infrared tensor sequence, a visible light block embedding sequence and an infrared block embedding sequence are constructed.

[0017] Based on the cross-modal temporal attention mechanism, the visible light block embedding sequence and the infrared block embedding sequence are fused to obtain the spatiotemporal fusion feature map at each time step.

[0018] In one possible implementation, constructing the visible light block embedding sequence and the infrared block embedding sequence based on the normalized visible light tensor sequence and the normalized infrared tensor sequence includes:

[0019] The standardized visible light tensor sequence is divided into multiple non-overlapping first local image blocks according to a preset size, and the standardized infrared tensor sequence is divided into multiple non-overlapping second local image blocks according to a preset size;

[0020] The first local image block and the second local image block are respectively subjected to feature flattening processing to obtain the corresponding one-dimensional image block features. Then, the one-dimensional image block features corresponding to the first local image block and the second local image block are respectively mapped to a preset high-dimensional feature space through a linear projection matrix to obtain the initial visible light block features and the initial infrared block features.

[0021] Based on the coordinate information of the first local image patch and the second local image patch in a unified spatial grid, a spatial location coding feature matching the unified spatial grid is constructed, and the spatial location coding feature is fused with the initial visible light patch feature and the initial infrared patch feature respectively to generate the visible light patch embedding sequence and the infrared patch embedding sequence.

[0022] In one possible implementation, the fusion of the visible light block embedding sequence and the infrared block embedding sequence based on the cross-modal temporal attention mechanism to obtain the spatiotemporal fusion feature map at each time step includes:

[0023] Perform a linear transformation on the visible light block embedding sequence at the current moment to generate the corresponding query matrix;

[0024] The visible light block embedding sequence at the current time and multiple historical time points are concatenated one by one with the corresponding infrared block embedding sequence to obtain multiple concatenated feature sequences. Linear transformation processing is performed on each of the concatenated feature sequences to generate the corresponding key matrix and value matrix.

[0025] Calculate the product of the query matrix and the transpose of the key matrix to obtain the product result, and normalize the product result to obtain the attention weight matrix;

[0026] The value matrix is ​​weighted and summed using the attention weight matrix to obtain the fusion block embedding feature. The fusion block embedding feature is then reshaped into a two-dimensional spatial feature. Upsampling is performed on the two-dimensional spatial feature to make its resolution the same as that of the unified spatial grid, thus obtaining the spatiotemporal fusion feature map at each time step.

[0027] In one possible implementation, the process of using the spatiotemporal fusion feature map at each time moment as conditional input, and performing iterative denoising processing based on the Gaussian noise field and the meteorological parameter field sequence according to the reverse diffusion process, yields the smoke and dust concentration field at each time moment, including:

[0028] Obtain a conditional constraint inverse diffusion network, which includes a feature encoding module, an iterative update module, and a physical constraint module;

[0029] Initialize a random noise field that conforms to a Gaussian distribution to obtain a Gaussian noise field, and obtain a preset feature length that matches the dust diffusion characteristics of the target industrial area. Based on the wind speed field in the meteorological parameter field sequence and the preset feature length, calculate the empirical diffusion coefficient.

[0030] Using the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, the spatiotemporal fusion feature map at each time moment is traversed, and the reverse diffusion iterative denoising process at the corresponding time moment is performed sequentially to obtain the dust concentration field at each time moment.

[0031] In one possible implementation, the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, traverses the spatiotemporal fusion feature map at each time step, and sequentially performs iterative reverse diffusion denoising processing at the corresponding time step to obtain the smoke and dust concentration field at each time step, including:

[0032] The spatiotemporal fusion feature maps at each time point are input into the feature encoding module for feature mapping processing to obtain the conditional guidance vectors at each time point.

[0033] The conditional guiding vector, the empirical diffusion coefficient, the wind speed field and the Gaussian noise field in the meteorological parameter field sequence are input into the iterative update module to perform initial iterative calculations and output the initial noise field estimate.

[0034] Based on the initial noise field estimate, a reverse diffusion iteration is performed. In each iteration, the noise field estimate of the current iteration step is called through the physical constraint module. The time difference term, spatial gradient term, and spatial divergence term are discretized and calculated on a unified spatial grid. They are combined in the form of the advection-diffusion equation to obtain the residual term. The sum of squares of the residual term is calculated to obtain the physical constraint loss value. In the first iteration, the noise field estimate of the current iteration step is the initial noise field estimate.

[0035] The physical constraint module calculates the gradient based on the physical constraint loss value and the noise field estimate of the current iteration step to obtain the physical guidance gradient; the physical guidance gradient is multiplied by the dynamic coefficient to obtain the physical constraint correction amount; wherein the dynamic coefficient is a coefficient that decreases linearly with the increase of the iteration step.

[0036] The iterative update module obtains the noise standard deviation of the current iteration step. Based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, the data-driven update amount is calculated using a preset reverse diffusion iterative formula. The physical constraint correction amount is subtracted from the data-driven update amount to obtain the corrected noise field estimate. The corrected noise field estimate is used as the noise field estimate of the current iteration step in the next iteration.

[0037] The reverse diffusion iteration process is repeated until the number of iterations reaches the preset maximum number of iterations. The iterative update module is used to denoise the corrected noise field estimate obtained in the last iteration to obtain the dust concentration field at the corresponding time.

[0038] In one possible implementation, the training method of the physical information neural network includes:

[0039] Construct a physical information neural network based on a fundamental physical information neural network;

[0040] Based on preset weighting coefficients, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are fused to construct a composite loss function;

[0041] During training, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are calculated. Gradient descent optimization is performed with the composite loss function as the objective function to iteratively update the network parameters of the physical information neural network, thereby obtaining the trained physical information neural network.

[0042] In one possible implementation, the data fitting loss is calculated as follows:

[0043] Based on the data constraints, multiple spatiotemporal sampling points are randomly sampled from the spatiotemporal field of the smoke and dust concentration, and the coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point are extracted.

[0044] The coordinates of each of the spatiotemporal sampling points are input into the physical information neural network to obtain the predicted concentration corresponding to each of the spatiotemporal sampling points;

[0045] The mean square error between the predicted concentration and the corresponding actual concentration value at each of the aforementioned spatiotemporal sampling points is calculated to obtain the data fitting loss.

[0046] In one possible implementation, the residual loss of the advection-diffusion equation is calculated as follows:

[0047] A spatiotemporal domain is constructed based on a unified spatial grid. Multiple configuration points are randomly sampled within the spatiotemporal domain, and the coordinates of each configuration point are extracted.

[0048] The coordinates of each of the configuration points are input into the physical information neural network to obtain the configuration prediction concentration, configuration prediction wind speed field and configuration prediction diffusion coefficient field corresponding to each of the configuration points.

[0049] The partial derivatives of the configured predicted concentration are calculated by automatic differentiation. The partial derivatives, the configured predicted wind speed field, and the configured predicted diffusion coefficient field are substituted into the advection-diffusion equation to obtain the equation residuals.

[0050] The advection-diffusion equation residual loss is calculated based on the equation residuals at each of the aforementioned configuration points.

[0051] In one possible implementation, the source term sparsity regularization loss is calculated in the following ways:

[0052] A spatial domain is constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within the spatial domain, and the coordinates of each source term monitoring point are extracted.

[0053] The coordinates of each source term monitoring point are input into the physical information neural network to obtain the predicted source term corresponding to each source term monitoring point;

[0054] The mean absolute value of the predicted source terms corresponding to each of the source term monitoring points is calculated to obtain the source term sparsity regularization loss.

[0055] In one possible implementation, extracting the spatial peak location and peak intensity from the target predicted source term field, using the spatial peak location as emission source coordinates, and the peak intensity as emission source intensity, includes:

[0056] Spatial nonmaximum suppression is performed on the target prediction source term field to remove discrete interference values ​​in the target prediction source term field, resulting in the processed target prediction source term field;

[0057] In the processed target prediction source term field, each spatial grid point is traversed, and the coordinates of the spatial grid point with the largest source term value in the processed target prediction source term field are taken as the spatial peak position.

[0058] Based on a preset neighborhood range of the spatial peak location, the numerical value of the target predicted source term field is integrated to obtain the integration result. The integration result is used as the peak intensity, and the peak intensity is converted into the corresponding emission rate. The emission rate is used as the emission source intensity.

[0059] A second aspect of this application provides an industrial dust emission monitoring device, comprising:

[0060] The acquisition unit is used to acquire visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series of the target industrial area.

[0061] The dust concentration spatiotemporal field generation unit is used to perform iterative denoising processing according to the reverse diffusion process based on the visible light image sequence, the infrared thermal imaging sequence and the meteorological parameter time series to obtain the dust concentration field at each time moment, and arrange the dust concentration fields at each time moment in chronological order to generate the dust concentration spatiotemporal field.

[0062] The physical information neural network prediction unit is used to input the spatiotemporal coordinates of the target to be monitored within the target industrial area into the trained physical information neural network to obtain the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field. The physical information neural network is trained with the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints.

[0063] The emission source identification unit is used to extract the spatial peak position and peak intensity from the target predicted source term field, and use the spatial peak position as the emission source coordinates and the peak intensity as the emission source intensity.

[0064] In one possible implementation, the smoke concentration spatiotemporal field generation unit includes:

[0065] The standardization processing unit is used to standardize the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series to obtain a standardized visible light tensor sequence, a standardized infrared tensor sequence, and a meteorological parameter field sequence.

[0066] A cross-modal spatiotemporal feature fusion processing unit is used to perform cross-modal spatiotemporal feature fusion processing on the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time point;

[0067] The iterative denoising processing unit is used to take the spatiotemporal fusion feature map at each time as a conditional input, and perform iterative denoising processing based on the Gaussian noise field and the meteorological parameter field sequence according to the reverse diffusion process to obtain the smoke and dust concentration field at each time.

[0068] In one possible implementation, the cross-modal spatiotemporal feature fusion processing unit includes:

[0069] An embedding sequence construction unit is used to construct visible light block embedding sequences and infrared block embedding sequences based on the standardized visible light tensor sequence and the standardized infrared tensor sequence;

[0070] A cross-modal temporal attention mechanism fusion unit is used to fuse the visible light block embedding sequence and the infrared block embedding sequence based on the cross-modal temporal attention mechanism to obtain a spatiotemporal fusion feature map at each time step.

[0071] In one possible implementation, the embedded sequence building unit is specifically used for:

[0072] The standardized visible light tensor sequence is divided into multiple non-overlapping first local image blocks according to a preset size, and the standardized infrared tensor sequence is divided into multiple non-overlapping second local image blocks according to a preset size;

[0073] The first local image block and the second local image block are respectively subjected to feature flattening processing to obtain the corresponding one-dimensional image block features. Then, the one-dimensional image block features corresponding to the first local image block and the second local image block are respectively mapped to a preset high-dimensional feature space through a linear projection matrix to obtain the initial visible light block features and the initial infrared block features.

[0074] Based on the coordinate information of the first local image patch and the second local image patch in a unified spatial grid, a spatial location coding feature matching the unified spatial grid is constructed, and the spatial location coding feature is fused with the initial visible light patch feature and the initial infrared patch feature respectively to generate the visible light patch embedding sequence and the infrared patch embedding sequence.

[0075] In one possible implementation, the cross-modal temporal attention mechanism fusion unit is specifically used for:

[0076] Perform a linear transformation on the visible light block embedding sequence at the current moment to generate the corresponding query matrix;

[0077] The visible light block embedding sequence at the current time and multiple historical time points are concatenated one by one with the corresponding infrared block embedding sequence to obtain multiple concatenated feature sequences. Linear transformation processing is performed on each of the concatenated feature sequences to generate the corresponding key matrix and value matrix.

[0078] Calculate the product of the query matrix and the transpose of the key matrix to obtain the product result, and normalize the product result to obtain the attention weight matrix;

[0079] The value matrix is ​​weighted and summed using the attention weight matrix to obtain the fusion block embedding feature. The fusion block embedding feature is then reshaped into a two-dimensional spatial feature. Upsampling is performed on the two-dimensional spatial feature to make its resolution the same as that of the unified spatial grid, thus obtaining the spatiotemporal fusion feature map at each time step.

[0080] In one possible implementation, the iterative denoising unit includes:

[0081] A conditional constraint reverse diffusion network acquisition unit is used to acquire a conditional constraint reverse diffusion network, wherein the conditional constraint reverse diffusion network includes a feature encoding module, an iterative update module, and a physical constraint module;

[0082] The parameter acquisition unit is used to initialize a random noise field that conforms to a Gaussian distribution, obtain a Gaussian noise field, and obtain a preset feature length that matches the dust diffusion characteristics of the target industrial area. Based on the wind speed field in the meteorological parameter field sequence and the preset feature length, the empirical diffusion coefficient is calculated.

[0083] The iterative denoising subunit is used to traverse the spatiotemporal fusion feature map at each time step through the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, and sequentially perform reverse diffusion iterative denoising processing at the corresponding time step to obtain the dust concentration field at each time step.

[0084] In one possible implementation, the iterative denoising subunit is specifically used for:

[0085] The spatiotemporal fusion feature maps at each time point are input into the feature encoding module for feature mapping processing to obtain the conditional guidance vectors at each time point.

[0086] The conditional guiding vector, the empirical diffusion coefficient, the wind speed field and the Gaussian noise field in the meteorological parameter field sequence are input into the iterative update module to perform initial iterative calculations and output the initial noise field estimate.

[0087] Based on the initial noise field estimate, a reverse diffusion iteration is performed. In each iteration, the noise field estimate of the current iteration step is called through the physical constraint module. The time difference term, spatial gradient term, and spatial divergence term are discretized and calculated on a unified spatial grid. They are combined in the form of the advection-diffusion equation to obtain the residual term. The sum of squares of the residual term is calculated to obtain the physical constraint loss value. In the first iteration, the noise field estimate of the current iteration step is the initial noise field estimate.

[0088] The physical constraint module calculates the gradient based on the physical constraint loss value and the noise field estimate of the current iteration step to obtain the physical guidance gradient; the physical guidance gradient is multiplied by the dynamic coefficient to obtain the physical constraint correction amount; wherein the dynamic coefficient is a coefficient that decreases linearly with the increase of the iteration step.

[0089] The iterative update module obtains the noise standard deviation of the current iteration step. Based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, the data-driven update amount is calculated using a preset reverse diffusion iterative formula. The physical constraint correction amount is subtracted from the data-driven update amount to obtain the corrected noise field estimate. The corrected noise field estimate is used as the noise field estimate of the current iteration step in the next iteration.

[0090] The reverse diffusion iteration process is repeated until the number of iterations reaches the preset maximum number of iterations. The iterative update module is used to denoise the corrected noise field estimate obtained in the last iteration to obtain the dust concentration field at the corresponding time.

[0091] In one possible implementation, the device further includes: a physical information neural network training unit, specifically used for:

[0092] Construct a physical information neural network based on a fundamental physical information neural network;

[0093] Based on preset weighting coefficients, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are fused to construct a composite loss function;

[0094] During training, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are calculated. Gradient descent optimization is performed with the composite loss function as the objective function to iteratively update the network parameters of the physical information neural network, thereby obtaining the trained physical information neural network.

[0095] In one possible implementation, the physical information neural network training unit includes: a data fitting loss calculation unit, specifically used for:

[0096] Based on the data constraints, multiple spatiotemporal sampling points are randomly sampled from the spatiotemporal field of the smoke and dust concentration, and the coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point are extracted.

[0097] The coordinates of each of the spatiotemporal sampling points are input into the physical information neural network to obtain the predicted concentration corresponding to each of the spatiotemporal sampling points;

[0098] The mean square error between the predicted concentration and the corresponding actual concentration value at each of the aforementioned spatiotemporal sampling points is calculated to obtain the data fitting loss.

[0099] In one possible implementation, the physical information neural network training unit includes: an advection-diffusion equation residual loss calculation unit, specifically used for:

[0100] A spatiotemporal domain is constructed based on a unified spatial grid. Multiple configuration points are randomly sampled within the spatiotemporal domain, and the coordinates of each configuration point are extracted.

[0101] The coordinates of each of the configuration points are input into the physical information neural network to obtain the configuration prediction concentration, configuration prediction wind speed field and configuration prediction diffusion coefficient field corresponding to each of the configuration points.

[0102] The partial derivatives of the configured predicted concentration are calculated by automatic differentiation. The partial derivatives, the configured predicted wind speed field, and the configured predicted diffusion coefficient field are substituted into the advection-diffusion equation to obtain the equation residuals.

[0103] The advection-diffusion equation residual loss is calculated based on the equation residuals at each of the aforementioned configuration points.

[0104] In one possible implementation, the physical information neural network training unit includes: a source term sparsity regularization loss calculation unit, specifically used for:

[0105] A spatial domain is constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within the spatial domain, and the coordinates of each source term monitoring point are extracted.

[0106] The coordinates of each source term monitoring point are input into the physical information neural network to obtain the predicted source term corresponding to each source term monitoring point;

[0107] The mean absolute value of the predicted source terms corresponding to each of the source term monitoring points is calculated to obtain the source term sparsity regularization loss.

[0108] In one possible implementation, the emission source identification unit is specifically used for:

[0109] Spatial nonmaximum suppression is performed on the target prediction source term field to remove discrete interference values ​​in the target prediction source term field, resulting in the processed target prediction source term field;

[0110] In the processed target prediction source term field, each spatial grid point is traversed, and the coordinates of the spatial grid point with the largest source term value in the processed target prediction source term field are taken as the spatial peak position.

[0111] Based on a preset neighborhood range of the spatial peak location, the numerical value of the target predicted source term field is integrated to obtain the integration result. The integration result is used as the peak intensity, and the peak intensity is converted into the corresponding emission rate. The emission rate is used as the emission source intensity.

[0112] A third aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the industrial dust emission monitoring method of the first aspect or any implementation thereof.

[0113] A fourth aspect of this application provides an electronic device, including at least one processor and a memory connected to the processor, wherein:

[0114] The memory is used to store computer programs;

[0115] The processor is used to execute the computer program so that the electronic device can implement the industrial dust emission monitoring method of the first aspect or any implementation thereof.

[0116] The fifth aspect of this application provides a computer-readable storage medium carrying one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the industrial dust emission monitoring method described in the first aspect or any implementation thereof.

[0117] By employing the aforementioned technical solution, the industrial dust emission monitoring method and related device provided in this application firstly fuse visible light and infrared thermal imaging multimodal data, and introduce meteorological parameter field sequences and advection-diffusion physical constraints during the reverse diffusion generation of the concentration field. This ensures that the generated dust concentration spatiotemporal field accurately matches multimodal visual features while strictly adhering to the physical laws of dust diffusion, avoiding non-physical distortions such as local concentration abrupt changes common in purely data-driven models, thereby significantly improving the reconstruction accuracy and spatiotemporal continuity of the concentration field. Furthermore, using this concentration spatiotemporal field as data constraints and the advection-diffusion equation as physical constraints, a physical information neural network is trained to achieve the conversion from a discrete field to a continuous spatiotemporal field and support accurate prediction of arbitrary spatiotemporal coordinates. This neural network, through joint inversion of the source term field, wind speed field, and diffusion coefficient field, can directly extract the spatial peak position and intensity from the physically consistent source term field as the emission source coordinates and intensity. Compared to traditional numerical fitting or simple peak detection, this significantly reduces errors caused by discrete noise and empirical parameters. Therefore, this solution can significantly improve the accuracy of emission concentration monitoring and emission source identification. Attached Figure Description

[0118] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0119] Figure 1 A schematic flowchart illustrating an industrial dust emission monitoring method provided in this application embodiment;

[0120] Figure 2 This is a schematic diagram of the structure of an industrial dust emission monitoring device provided in an embodiment of this application;

[0121] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0122] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0123] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0124] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements, but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0125] Industrial dust is a major air pollutant in industrial production settings, and monitoring its emissions is a core aspect of ecological and environmental supervision and industrial pollution control. Industrial dust monitoring includes emission concentration monitoring and emission source identification.

[0126] Current traditional methods for monitoring industrial dust emissions include industrial dust monitoring solutions based on single-point gas sensors and vision-based industrial dust monitoring solutions, among which:

[0127] Industrial dust monitoring solutions based on single-point gas sensors rely on these sensors to collect concentration data, which can only detect concentrations at local points. They cannot obtain the overall distribution and spatiotemporal evolution characteristics of dust within a spatial range, and have significant limitations in locating and calculating the intensity of fugitive emission sources.

[0128] While vision-based industrial dust emission monitoring schemes can reconstruct dust concentration fields and thus obtain the overall distribution and spatiotemporal evolution characteristics of dust within a spatial range, they rely solely on a single visible light image mode. This makes them susceptible to interference from variations in industrial lighting, dust obstruction, and complex operating conditions, making it difficult to guarantee the stability and effectiveness of feature extraction in complex industrial environments. In the dust concentration field reconstruction stage, on the one hand, traditional advection-diffusion physics models are typically used, requiring the pre-setting of numerous idealized parameters, resulting in poor adaptability to actual meteorological conditions and terrain, and the concentration simulation accuracy fails to meet practical needs. On the other hand, purely data-driven deep learning models are often employed, but without embedding physical constraints on dust diffusion, the reconstructed concentration field is prone to distortions that contradict physical laws, leading to insufficient reliability. In the emission source identification stage, traditional numerical fitting and simple peak detection algorithms are frequently used, resulting in low identification accuracy.

[0129] It is evident that traditional industrial dust emission monitoring methods, in terms of emission concentration monitoring, rely solely on single-point sensors to provide discrete and sparse "point" data, failing to characterize the continuous distribution and dynamic diffusion gradient of dust in three-dimensional space. Existing visual reconstruction or physical models either suffer from feature extraction instability due to interference from lighting conditions such as illumination and occlusion, or their idealized parameter assumptions are out of sync with actual meteorological and topographical conditions, causing simulated concentration values ​​to deviate from the true field distribution. Regarding emission source identification, neither numerical fitting based on single-point data nor simple peak detection based on the visual field couples the source term release mechanism with multi-physics fields (wind speed field, diffusion coefficient field), making it difficult to separate the true spatial location and intensity of the source term under noise interference. Therefore, traditional industrial dust emission monitoring methods are generally at the stage of "single-modal monitoring and simple numerical analysis," lacking multi-modal fusion sensing, field reconstruction under physical information constraints, and joint inversion capabilities of source term and flow field, resulting in low accuracy in emission concentration monitoring and emission source identification.

[0130] To address the aforementioned problems, this application provides a method for monitoring industrial dust emissions. The method for monitoring industrial dust emissions according to this application will be described in detail below with reference to the accompanying drawings.

[0131] Reference Figure 1 , Figure 1 This is a schematic flowchart of an industrial dust emission monitoring method provided in an embodiment of this application, as shown below. Figure 1 As shown in the embodiments of this application, an industrial dust emission monitoring method may include the following steps, which are described in detail below.

[0132] S101: Acquire visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series of the target industrial area;

[0133] In this application, continuous frame visual data can be collected first by visible light imaging equipment and infrared thermal imaging equipment deployed in the target industrial area to form visible light image sequences and infrared thermal imaging sequences, and time series of meteorological parameters such as wind speed, temperature and air pressure can be obtained based on regional meteorological monitoring terminals.

[0134] S102: Based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, perform iterative denoising processing according to the reverse diffusion process to obtain the smoke and dust concentration field at each time moment, and arrange the smoke and dust concentration fields at each time moment in chronological order to generate the smoke and dust concentration spatiotemporal field.

[0135] In one possible implementation, spatiotemporal fusion feature maps for each moment can be generated based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series. Using the spatiotemporal fusion feature maps for each moment as input conditions, iterative denoising processing is performed based on the Gaussian noise field and the meteorological parameter field sequence, following a reverse diffusion process, to obtain the smoke and dust concentration field for each moment.

[0136] In this application, iterative denoising is performed according to the reverse diffusion process, and physical consistency concentration estimation is achieved with meteorological parameters as constraints, overcoming the shortcomings of purely data-driven models that violate physical laws. By arranging the smoke and dust concentration fields at each time point in chronological order, a spatiotemporal field of smoke and dust concentration can be formed, so that the reconstruction results not only match the characteristics of multimodal visual observations, but also follow the physical laws of smoke and dust diffusion, further improving the accuracy of the spatiotemporal distribution of concentration.

[0137] In this application, a data-driven generative model can be combined with meteorological and physical constraints to iteratively generate a smoke and dust concentration distribution that conforms to reality from random noise, thus solving the problem that traditional concentration field reconstruction is difficult to characterize spatiotemporal dynamic changes.

[0138] S103: Input the spatiotemporal coordinates of the target to be monitored within the target industrial area into the trained physical information neural network to obtain the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field; the physical information neural network is trained with the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints.

[0139] In this application, the target spatiotemporal coordinates are a set of spatiotemporal points for which the physical field of smoke and dust is to be predicted. They consist of two parts: spatial coordinates and time coordinates. Specifically, they correspond to the location to be predicted within the target industrial area and the target time to be predicted. The target spatiotemporal coordinates are a discrete set pre-constructed based on the spatial range of the target industrial area and the monitoring time requirements. The spatial range of the target spatiotemporal coordinates completely corresponds to the target industrial area. It is a discretized representation of the target industrial area in the spatiotemporal dimension, i.e., a certain coordinate within the target industrial area.

[0140] Considering that iterative denoising processing based on the reverse diffusion process is a purely data-driven generative modeling approach, the resulting discrete temporal concentration field only contains discrete spatial concentration values ​​generated by the diffusion model iteration, lacking continuous field characteristics and potentially exhibiting non-physical distortions (such as local concentration abrupt changes or values ​​exceeding reasonable ranges), and lacking physical equation constraints, this application trains a physical information neural network with both data and physical constraints. The data constraints include the discrete spatiotemporal coordinates and corresponding concentration values ​​of the dust concentration spatiotemporal field, ensuring that the predicted concentration field output by the network remains consistent with the measured concentration data, thus providing data supervision for network training. The physical constraints include advection-diffusion equation constraints. The physical information neural network represents a dual-constraint inversion modeling approach, supporting concentration field prediction for arbitrary target spatiotemporal coordinates. It can further realize the conversion from discrete data to a continuous spatiotemporal field, and the prediction results strictly conform to the physical laws of dust diffusion without physical distortion.

[0141] In this application, the physical information neural network takes spatiotemporal coordinates as input and can realize the mapping from spatiotemporal coordinates to multiple physical quantities through feature mapping and nonlinear transformation of multiple hidden layers.

[0142] In one possible implementation, the physical information neural network is trained by minimizing a composite loss function, using the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints. The composite loss function includes a data fitting term, an advection-diffusion equation residual term, and a source term sparsity term.

[0143] S104: Extract the spatial peak position and peak intensity from the target predicted source term field, use the spatial peak position as the emission source coordinates, and use the peak intensity as the emission source intensity.

[0144] In this application, the spatial peak position and intensity are extracted from the source term field, realizing high-precision positioning and quantitative identification of emission sources, which significantly improves the accuracy and reliability of fugitive emission monitoring.

[0145] The industrial dust emission monitoring method provided in this solution first integrates multimodal data from visible light and infrared thermal imaging. During the reverse diffusion generation of the concentration field, a meteorological parameter field sequence and advection-diffusion physical constraints are introduced. This ensures that the generated dust concentration spatiotemporal field accurately matches multimodal visual features while strictly adhering to the physical laws of dust diffusion, avoiding non-physical distortions such as local concentration abrupt changes common in purely data-driven models. This significantly improves the reconstruction accuracy and spatiotemporal continuity of the concentration field. Furthermore, using this concentration spatiotemporal field as data constraints and the advection-diffusion equation as physical constraints, a physical information neural network is trained to achieve the conversion from a discrete field to a continuous spatiotemporal field and support accurate prediction of arbitrary spatiotemporal coordinates. This neural network, by jointly inverting the source term field, wind speed field, and diffusion coefficient field, can directly extract the spatial peak position and intensity from the physically consistent source term field as the emission source coordinates and intensity. Compared to traditional numerical fitting or simple peak detection, this significantly reduces errors caused by discrete noise and empirical parameters. Therefore, this solution can significantly improve the accuracy of emission concentration monitoring and emission source identification.

[0146] In one possible implementation, the step of performing iterative denoising processing based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, according to a reverse diffusion process, to obtain the smoke and dust concentration field at each time point includes:

[0147] S201: The visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series are standardized to obtain a standardized visible light tensor sequence, a standardized infrared tensor sequence, and a meteorological parameter field sequence.

[0148] In this application, a unified spatial grid in Cartesian form with equal spacing can be constructed according to the spatial range of the target industrial area and the accuracy requirements of smoke and dust characterization. Based on this unified spatial grid, visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series are standardized to obtain standardized visible light tensor sequences, standardized infrared tensor sequences, and meteorological parameter field sequences.

[0149] In one possible implementation, bilinear interpolation can be used to unify the projected values ​​of visible light image sequences and infrared thermal imaging sequences onto a unified spatial grid, correcting spatial offset and distortion errors to obtain standardized visible light tensor sequences and standardized infrared tensor sequences with consistent dimensions and spatial alignment. Considering the characteristic of meteorological parameters as single-point time-series data, Kriging interpolation can also be used to extend the meteorological parameter time series to all nodes of the unified spatial grid, generating a meteorological parameter field sequence that matches the visual tensor space.

[0150] In this application, by standardizing visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series, the heterogeneous differences in spatial resolution and distribution of visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series can be eliminated, achieving strict spatial alignment of multi-source data. This establishes a unified data benchmark for subsequent cross-modal spatiotemporal feature fusion processing and spatiotemporal field modeling of smoke and dust concentration. In this way, feature fusion deviations caused by spatial heterogeneity can be avoided, and continuous spatial meteorological constraints can be provided for subsequent physical modeling of smoke and dust diffusion.

[0151] S202: Perform cross-modal spatiotemporal feature fusion processing on the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time point;

[0152] In this application, a cross-modal temporal attention mechanism can be used to perform cross-modal spatiotemporal feature fusion processing on the standardized visible light tensor sequence and the standardized infrared tensor sequence, dynamically capturing temporal correlations, making up for the shortcomings of traditional static weighted fusion which is susceptible to changes in illumination and working conditions, and significantly improving the stability of feature extraction in complex industrial environments.

[0153] S203: Using the spatiotemporal fusion feature map of each time moment as input, based on the Gaussian noise field and the meteorological parameter field sequence, perform iterative denoising processing according to the reverse diffusion process to obtain the dust concentration field of each time moment, and arrange the dust concentration fields of each time moment in time sequence to generate the dust concentration spatiotemporal field.

[0154] In one possible implementation, a Gaussian noise field with a uniform spatial grid resolution can be initialized first. The spatiotemporal fusion features at each time step are encoded into a conditional guiding vector. An empirical diffusion coefficient suitable for dust diffusion is calculated by combining the wind speed field in the meteorological parameter field. Subsequently, using the conditional guiding vector, empirical diffusion coefficient, wind speed field, and Gaussian noise field as inputs, a reverse diffusion iteration is performed. During the iteration process, physical constraints of the advection-diffusion equation can be introduced to calculate the physical residual of the noise field and generate a gradient correction amount to correct the data-driven iteration results and avoid non-physical distortions in the concentration field. After the iteration is completed, the dust concentration field at each time step can be output.

[0155] In one possible implementation, the cross-modal spatiotemporal feature fusion processing of the normalized visible light tensor sequence and the normalized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time step includes:

[0156] S301: Based on the standardized visible light tensor sequence and the standardized infrared tensor sequence, a visible light block embedding sequence and an infrared block embedding sequence are constructed;

[0157] In this application, the standardized visible light tensor sequence and the standardized infrared tensor sequence can be divided into non-overlapping local image blocks according to a preset size. After feature flattening and linear projection mapping to a high-dimensional feature space, linear embedding is performed, and spatial position coding adapted to a unified spatial grid is incorporated to retain spatial position information, thereby obtaining the visible light block embedding sequence and the infrared block embedding sequence.

[0158] In one possible implementation, a first local image block and a second local image block can be constructed based on the standardized visible light tensor sequence and the standardized infrared tensor sequence, respectively, and the first local image block and the second local image block can be linearly embedded to obtain a visible light block embedding sequence and an infrared block embedding sequence.

[0159] S302: Based on the cross-modal temporal attention mechanism, the visible light block embedding sequence and the infrared block embedding sequence are fused to obtain the spatiotemporal fusion feature map at each time step.

[0160] In this application, the visible light block embedding sequence at the current time can be used as the query matrix, and a key matrix and a value matrix can be constructed based on the visible light block embedding sequence and the infrared block embedding sequence at multiple times. Cross-modal temporal attention calculation is then performed to obtain the spatiotemporal fusion feature map at each time.

[0161] In this application, the visible light block embedding sequence at the current moment is used as the query matrix. A key-value matrix can be constructed based on the multi-time bimodal block embedding sequence, cross-modal temporal attention calculation is performed, and the temporal correlation features of visible light and infrared modes are dynamically mined. The limitations of static fusion are eliminated, and the advantages of bimodal features and the temporal evolution law of smoke and dust are fused simultaneously. Finally, the spatiotemporal fusion feature map of each moment is obtained, which provides multimodal temporal feature support for concentration field reconstruction.

[0162] In one possible implementation, constructing the visible light block embedding sequence and the infrared block embedding sequence based on the normalized visible light tensor sequence and the normalized infrared tensor sequence includes:

[0163] S401: Divide the standardized visible light tensor sequence into multiple non-overlapping first local image blocks according to a preset size, and divide the standardized infrared tensor sequence into multiple non-overlapping second local image blocks according to a preset size;

[0164] In this application, a preset size can be set based on the minimum identifiable scale of smoke and dust and computational efficiency to ensure that the local block can completely contain the local features of smoke and dust without feature fragmentation due to excessively small size or the introduction of redundant background information due to excessively large size. Based on this size, a fixed-size partitioning method can decompose the high-dimensional visual tensor into uniformly granular local feature units. This can focus on the local texture features and infrared temperature change features corresponding to smoke and dust in the image, avoiding computational redundancy caused by global feature calculation, and ensuring that the feature units of visible light and infrared modes are consistent in spatial scale, laying the foundation for subsequent cross-modal feature alignment. Furthermore, the non-overlapping partitioning rule can avoid duplicate feature extraction.

[0165] S402: Perform feature flattening processing on the first local image block and the second local image block respectively to obtain the corresponding one-dimensional image block features, and map the one-dimensional image block features corresponding to the first local image block and the second local image block to a preset high-dimensional feature space through a linear projection matrix to obtain the initial visible light block features and the initial infrared block features.

[0166] In this application, feature flattening operations are performed on the first and second local image blocks respectively. This converts the two-dimensional local image block features into one-dimensional vectors in a fixed row-major or column-major order, yielding the corresponding one-dimensional image block features. This feature flattening operation eliminates the spatial dimensional constraints of the local image blocks, transforming local visual features into a vector format suitable for linear projection operations. Furthermore, the conversion process preserves all visual feature information within the local image block, preventing feature loss and distortion, and ensuring that subsequent feature mapping can be based on the complete local features.

[0167] Furthermore, a preset linear projection matrix can be used to map the one-dimensional image patch features corresponding to the first and second local image patches respectively, mapping the one-dimensional image patch features from the original low-dimensional feature space to a preset high-dimensional feature space. The calculation expression for the linear projection can be:

[0168] ;

[0169] in, This refers to the initial visible light block features or the initial infrared block features. For one-dimensional image patch features, The projection matrix is ​​a learnable linear projection matrix, and Xavier initialization is used to ensure the stability of the variance of the features during the projection process. This is the linear projection bias term, initialized as a zero vector.

[0170] Linear projection can enhance the expressive dimension and abstraction of local visual features, improve the ability of features to represent subtle changes in smoke and dust, and unify the feature dimensions of visible light and infrared modes, so that the features of the two modes are in the same feature space, providing a dimensional basis for subsequent cross-modal feature fusion.

[0171] S403: Based on the coordinate information of the first local image block and the second local image block in a unified spatial grid, construct a spatial position coding feature that matches the unified spatial grid, and fuse the spatial position coding feature with the initial visible light block feature and the initial infrared block feature respectively to generate the visible light block embedding sequence and the infrared block embedding sequence.

[0172] In this application, since visual features lose their original spatial location association information after block segmentation, and the spatial distribution of smoke and dust is strongly correlated with the emission location, the lack of location information will lead to feature space misalignment during subsequent cross-modal fusion. Therefore, spatial location encoding can be used to supplement and embed location information. For example, this embodiment can use sine and cosine location encoding to generate spatial location encoded features, the calculation expression of which is:

[0173] ;

[0174] in, Encoding features for spatial location, This is an index for the location of a local image patch in a unified spatial grid. This index is generated by serializing the row and column coordinates of the patch according to a preset rule. The dimension index for the location-encoded feature is used to distinguish the different dimensional components of the encoded feature. This is the total dimension encoding the spatial location features. This dimension is consistent with the feature dimension after subsequent linear projection, ensuring dimension matching during feature fusion. This encoding method can transform discrete spatial location information into continuous high-dimensional features and can adapt to a unified spatial grid of different scales, accurately representing the spatial relative relationships of each local image patch, and providing support for the spatial alignment of cross-modal features.

[0175] Specifically, an element-wise addition method with complete feature dimension matching can be used to fuse the generated spatial location encoded features with the initial visible light block features and the initial infrared block features, respectively, ultimately generating the corresponding visible light block embedding sequence and infrared block embedding sequence. This fusion method does not require the introduction of additional learnable parameters and can efficiently achieve a deep combination of spatial location information and visual abstract features. This allows the generated block embedding sequence to simultaneously possess both local visual features and spatial location attributes of smoke and dust, ensuring that subsequent cross-modal temporal attention calculations can accurately associate the feature spatial locations at different times and in different modalities, thus improving the accuracy of feature fusion.

[0176] In one possible implementation, the fusion of the visible light block embedding sequence and the infrared block embedding sequence based on the cross-modal temporal attention mechanism to obtain the spatiotemporal fusion feature map at each time step includes:

[0177] S501: Perform a linear transformation on the visible light block embedding sequence at the current moment to generate the corresponding query matrix;

[0178] Specifically, the visible light modality can clearly characterize the spatial contours and morphological features of smoke and dust, and this feature has strong stability and discriminability. Therefore, using the visible light patch embedding sequence as the basis for constructing the query matrix, the core correlation information in cross-modal temporal features can be accurately anchored during the query process. For example, a learnable query transformation matrix and bias term can be introduced to map the visible light patch embedding sequence at the current moment to a preset feature interaction space, so that the generated query matrix meets the matching requirements of subsequent attention calculation in terms of dimension, and the expression of the core features of smoke and dust in the visible light modality can be enhanced through linear transformation, thereby improving the accuracy of feature matching.

[0179] S502: The visible light block embedding sequence at the current time and multiple historical times are concatenated one by one with the corresponding infrared block embedding sequence to obtain multiple concatenated feature sequences. Linear transformation processing is performed on each of the concatenated feature sequences to generate the corresponding key matrix and value matrix.

[0180] Specifically, the visible light block embedding sequences from the current moment and multiple historical moments can be sequentially concatenated with the corresponding infrared block embedding sequences. This fully leverages the complementary information between the morphological representation advantages of the visible light mode and the temperature sensitivity of the infrared mode, incorporating historical moment features to capture the temporal dynamics of smoke diffusion. This avoids the limitation of single-moment features failing to reflect the spatiotemporal evolution of smoke, generating multiple cross-modal concatenated feature sequences. Each concatenated feature sequence can be subjected to an independent learnable linear transformation, generating corresponding key and value matrices. The key matrix is ​​used to calculate similarity with the query matrix to establish feature associations, while the value matrix stores the core information of the cross-modal temporal features. Independent linear transformations allow the key and value matrices to adapt to the needs of query matching and feature aggregation, improving the effectiveness of feature interaction.

[0181] S503: Calculate the product of the query matrix and the transpose of the key matrix to obtain the product result, and normalize the product result to obtain the attention weight matrix;

[0182] Specifically, the product of the query matrix and the transpose of the key matrix is ​​calculated. This product directly reflects the correlation strength between the core features represented by the query matrix at the current time step and the cross-modal features at each time step. The product result is then normalized using the Softmax function to obtain the attention weight matrix. This normalization operation ensures that the elements in the weight matrix range from [0,1] and sum to 1, guaranteeing the rationality of attention allocation. The calculation formula is as follows:

[0183] ;

[0184] in, This is the attention weight matrix. For querying the matrix, The key matrix, This represents the transpose of the key matrix. To query the feature dimensions of the key matrix and the query matrix, the square root of this dimension is used as a scaling factor to alleviate the gradient vanishing problem caused by excessively large product values ​​when the feature dimensions are high, thus ensuring the stability of attention calculation.

[0185] S504: The value matrix is ​​weighted and summed using the attention weight matrix to obtain the fusion block embedding feature, and the fusion block embedding feature is reshaped into a two-dimensional spatial feature. Upsampling processing is performed on the two-dimensional spatial feature to make the resolution of the two-dimensional spatial feature the same as the resolution of the unified spatial grid, thereby obtaining the spatiotemporal fusion feature map at each time step.

[0186] The attention weight matrix obtained above is used to perform a weighted summation operation on the value matrix. During the weighted summation process, the elements with larger values ​​in the attention weight matrix can strengthen the contribution of relevant features in the value matrix, while the elements with smaller values ​​can weaken the influence of irrelevant features. This adaptively aggregates the effective information in cross-modal features from the current and historical moments, realizes multimodal complementarity and temporal correlation fusion of smoke and dust features, improves the integrity and robustness of features in representing smoke and dust, and finally obtains the fusion block embedded features.

[0187] By embedding the fused block features according to the row and column distribution rules of a unified spatial grid, it can be reshaped into two-dimensional spatial features. However, the resolution of the two-dimensional spatial features is lower than that of the unified spatial grid. Therefore, upsampling can be performed on them. For example, a bilinear interpolation algorithm can be used for upsampling. This algorithm can perform smooth interpolation based on the feature values ​​of adjacent pixels, improving the feature resolution while avoiding the introduction of spurious features. Ultimately, the resolution of the two-dimensional spatial features is completely consistent with the resolution of the unified spatial grid, resulting in a spatiotemporal fused feature map at each time step. This feature map contains both multimodal complementary information and temporal dynamic information, while maintaining the same spatial scale as the original data, providing feature constraints for the subsequent conditional reverse diffusion generation of the dust concentration field.

[0188] In one possible implementation, the process of using the spatiotemporal fusion feature map at each time moment as conditional input, and performing iterative denoising processing based on the Gaussian noise field and the meteorological parameter field sequence according to the reverse diffusion process, yields the smoke and dust concentration field at each time moment, including:

[0189] S601: Obtain the conditional constraint reverse diffusion network, which includes a feature encoding module, an iterative update module, and a physical constraint module;

[0190] Specifically, the conditional constraint reverse diffusion network includes a feature encoding module, an iterative update module, and a physical constraint module. The feature encoding module is responsible for mapping high-dimensional spatiotemporal fusion features into low-dimensional conditional guidance vectors, providing data constraint directions for reverse diffusion. The iterative update module can achieve gradual denoising of the noise field and approximation of the concentration field through multiple iterations based on the reverse diffusion principle. The physical constraint module, by embedding the advection-diffusion equation of smoke and dust diffusion, can correct the iterative process by calculating the physical residuals to ensure that the output results conform to physical laws.

[0191] S602: Initialize a random noise field that conforms to a Gaussian distribution, obtain a Gaussian noise field, and acquire a preset feature length that matches the dust diffusion characteristics of the target industrial area. Based on the wind speed field in the meteorological parameter field sequence and the preset feature length, calculate the empirical diffusion coefficient.

[0192] Specifically, a random noise field conforming to a standard Gaussian distribution can be initialized first. The spatial resolution of this noise field is strictly consistent with the resolution of the unified spatial grid, so that the spatial scale of the subsequent reverse diffusion process is consistent with the spatial benchmark of the previous multi-source data standardization and cross-modal fusion, avoiding spatial distortion of the concentration field caused by spatial scale mismatch. Furthermore, the randomness of the Gaussian distribution can cover the potential distribution range of the smoke and dust concentration field, providing a comprehensive initial search space for iterative denoising.

[0193] Subsequently, a preset characteristic length matching the dust diffusion characteristics of the target industrial area can be obtained. This length is determined based on the dust particle size, emission height, and historical diffusion monitoring data within the area, and is used to characterize the natural diffusion scale of dust under no external force influence. Furthermore, an empirical diffusion coefficient can be calculated based on the wind speed field in the meteorological parameter field sequence and this preset characteristic length. For example, since wind speed is a core meteorological factor affecting dust diffusion rate, the empirical diffusion coefficient can be determined by the ratio of the wind speed value to the preset characteristic length; that is, the higher the wind speed, the larger the diffusion coefficient. By introducing the empirical diffusion coefficient, initial physical parameters that fit actual meteorological conditions are provided for subsequent physical constraints.

[0194] S603: Using the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, the spatiotemporal fusion feature map at each time moment is traversed, and the reverse diffusion iterative denoising process at the corresponding time moment is performed sequentially to obtain the dust concentration field at each time moment.

[0195] In one possible implementation, the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, traverses the spatiotemporal fusion feature map at each time step, and sequentially performs iterative reverse diffusion denoising processing at the corresponding time step to obtain the smoke and dust concentration field at each time step, including:

[0196] S6031: Input the spatiotemporal fusion feature map at each time step into the feature encoding module, perform feature mapping processing, and obtain the conditional guidance vector at each time step;

[0197] Specifically, the spatiotemporal fusion feature maps at each time point are input into the feature encoding module. This module can perform dimensional compression and feature extraction on the feature maps through a multilayer perceptron, mapping the two-dimensional spatial features into a one-dimensional conditional guidance vector. This vector contains multimodal temporal fusion smoke and dust feature information, which is used to guide the noise field to evolve in a direction that conforms to the actual smoke and dust distribution during the reverse diffusion process.

[0198] S6032: Input the conditional guidance vector, the empirical diffusion coefficient, the wind speed field and the Gaussian noise field in the meteorological parameter field sequence into the iterative update module, perform initial iterative calculation, and output the initial noise field estimate.

[0199] By inputting the conditional guidance vector, empirical diffusion coefficient, wind speed field and Gaussian noise field from the meteorological parameter field sequence into the iterative update module, this module can aggregate multi-source constraint information through the attention mechanism, perform initial iterative calculation, and output an initial noise field estimate, which is the result of preliminary denoising and feature guidance of the Gaussian noise field.

[0200] S6033: Based on the initial noise field estimate, a reverse diffusion iteration is performed. In each iteration, the noise field estimate of the current iteration step is called through the physical constraint module. The time difference term, spatial gradient term, and spatial divergence term are discretized and calculated on a unified spatial grid. They are combined in the form of the advection-diffusion equation to obtain the residual term. The sum of squares of the residual term is calculated to obtain the physical constraint loss value. In the first iteration, the noise field estimate of the current iteration step is the initial noise field estimate.

[0201] Based on the initial noise field estimate, a reverse diffusion iterative process can be initiated. In each iteration, the physical constraint module can first call the noise field estimate of the current iteration step and perform discrete calculations using the finite difference method on a unified spatial grid. The time difference term can be obtained by the ratio of the difference between noise field estimates in adjacent iteration steps to the iteration step size, characterizing the temporal rate of change of the noise field. The spatial gradient term can be calculated by the neighborhood difference of the current noise field estimate on the spatial grid, characterizing the spatial distribution trend. The spatial divergence term can be obtained by the second-order discrete difference of the spatial gradient, characterizing the flux change in the diffusion process. Combining these three terms according to the mathematical form of the advection-diffusion equation yields the residual term of the equation. Sum of squares of the residual term across the entire spatial grid yields the physical constraint loss value, which quantifies the degree of deviation between the current noise field estimate and the physical laws of smoke and dust diffusion. The expression for calculating the physical constraint loss value can be:

[0202] ;

[0203] in, This represents the physical constraint loss value. For the first Next iteration step, spatial coordinates Noise field estimate at location, For the iteration step index, To unify the global extent of the spatial grid, For time difference terms, Spatial coordinates The wind speed field value at that location. For spatial gradient terms, The empirical diffusion coefficient is... The spatial divergence term, this formula can provide a quantitative basis for physical constraint correction by quantifying the deviation between the noise field estimate and the advection-diffusion equation.

[0204] S6034: The physical constraint module calculates the gradient based on the physical constraint loss value and the noise field estimate of the current iteration step to obtain the physical guidance gradient; the physical guidance gradient is multiplied by the dynamic coefficient to obtain the physical constraint correction amount; wherein, the dynamic coefficient is a coefficient that decreases linearly with the increase of the iteration step.

[0205] Specifically, the physical constraint module can obtain the physical guiding gradient by taking the partial derivative of the physical constraint loss value with respect to the noise field estimate of the current iteration step. This gradient points in the direction that reduces the physical constraint loss. Therefore, the physical guiding gradient can be multiplied by a dynamic coefficient to obtain the physical constraint correction amount. The dynamic coefficient is designed to be a parameter that decreases linearly with the increase of the iteration step. For example, in the early stage of iteration, the noise field deviates significantly from the true concentration field, so the dynamic coefficient can be close to 1 to strengthen the corrective effect of the physical constraint on the iteration direction. In the later stage of iteration, the noise field gradually approaches the true concentration field, so the dynamic coefficient can be close to 0 to weaken the physical constraint and avoid over-correction that causes the concentration field to deviate from the multimodal features.

[0206] S6035: The iterative update module obtains the noise standard deviation of the current iteration step. Based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, the data-driven update amount is calculated using a preset reverse diffusion iterative formula. The physical constraint correction amount is subtracted from the data-driven update amount to obtain the corrected noise field estimate. The corrected noise field estimate is used as the noise field estimate of the current iteration step in the next iteration.

[0207] In the iterative update module, the corresponding noise standard deviation can be obtained first based on the progress of the current iteration step. This value decreases as the iteration step increases, conforming to the gradual decay of noise in the back diffusion process. Subsequently, based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, a preset back diffusion iterative formula can be used to calculate the data-driven update amount. This update amount mainly relies on multimodal feature guidance and noise decay laws to ensure that the concentration field closely matches the observed characteristics. For example, the mathematical expression of the preset back diffusion iterative formula can be:

[0208] ;

[0209] in, This is the noise field estimate for the next iteration. This is the noise field estimate for the current iteration step. This represents the noise standard deviation for the current iteration step. This is a noise prediction network (with an embedded iterative update module), whose inputs are the current noise field estimate and the conditional guidance vector at the current time step. and noise standard deviation The output is the predicted noise residual. The retained weighting coefficients are used for the current noise field estimate. This formula achieves progressive denoising and concentration field approximation of the noise field by retaining the effective features of the current iteration and incorporating condition-guided noise residual correction, ensuring that the data-driven update quantity fits the multimodal observation characteristics.

[0210] Subtracting the physical constraint correction from the data-driven update achieves a dynamic balance between data-driven and physical constraints, ensuring that the updated noise field conforms to both multimodal observation characteristics and the physical laws of smoke and dust diffusion. The resulting corrected noise field estimate can then be used as input for the next iteration.

[0211] S6036: Repeat the reverse diffusion iteration process until the number of iterations reaches the preset maximum number of iterations. Through the iteration update module, the corrected noise field estimate obtained in the last iteration is denoised to obtain the dust concentration field at the corresponding time.

[0212] Specifically, the above-mentioned reverse diffusion iterative process is repeated until the preset maximum number of iterations is reached. The corresponding noise field can be considered to have been sufficiently denoised and feature-guided. The iterative update module can perform Gaussian filtering denoising on the corrected noise field estimate obtained in the last iteration to eliminate residual minor noise and obtain the smoke and dust concentration field at the corresponding time. By traversing the spatiotemporal fusion feature maps at each time step through the conditional constraint reverse diffusion network, the reverse diffusion iterative denoising process at each time step is completed sequentially. Finally, a smoke and dust concentration field covering the entire time series can be obtained, providing complete spatiotemporal distribution data of smoke and dust concentration for subsequent physical information neural network modeling.

[0213] In this application, the training method of the physical information neural network includes:

[0214] S701: Constructing a physical information neural network based on a fundamental physical information neural network;

[0215] Specifically, for the basic physical information neural network, a multi-layer fully connected neural network with a sinusoidal activation function can be used as the core architecture. This sinusoidal activation function has a stronger representation ability for spatiotemporally changing high-frequency physical fields, and can accurately capture the spatiotemporal gradient changes and local detailed features of physical quantities such as dust concentration and wind speed, avoiding the gradient vanishing and feature saturation problems that are prone to occur in deep networks with traditional ReLU-type activation functions. By configuring four output branches at the end of the basic physical information neural network, the physical information neural network can be obtained. Each output branch consists of a single fully connected layer, where each output branch corresponds to the physical quantity output of the predicted concentration field, predicted source term field, predicted wind speed field, and predicted diffusion coefficient field, respectively. This ensures the independence of each physical quantity output, and also enables the mining of coupling information between multiple physical quantities through a shared front-end feature extraction layer, so that each output physical quantity maintains an inherent consistency in spatiotemporal distribution, which conforms to the mutual influence relationship between concentration, wind speed, diffusion coefficient, and source term during dust diffusion.

[0216] It should be noted that the dimensions of each output physical quantity are matched with the resolution of the unified spatial grid to ensure consistency of spatial scale.

[0217] S702: Based on preset weighting coefficients, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are fused to construct a composite loss function;

[0218] In this application, the data fitting loss is the data fitting term of the composite loss function, the advection-diffusion equation residual loss is the advection-diffusion equation residual term of the composite loss function, and the source term sparsity regularization loss is the source term sparsity term of the composite loss function.

[0219] Specifically, a composite loss function can be constructed by weighting and fusing the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss based on preset weight coefficients. Its core expression can be:

[0220] ;

[0221] in, For composite loss function, , , These are the preset weighting coefficients for data fitting loss, advection-diffusion equation residual loss, and source term sparsity regularization loss, respectively. The weighting coefficients are determined based on the magnitude and importance of each loss term, and their values ​​are optimized through cross-validation to ensure that the three losses are balanced during training. For data fitting loss, The residual loss of the advection-diffusion equation, This is the source term sparsity regularization loss.

[0222] S703: During the training process, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are calculated, and gradient descent optimization is performed with the composite loss function as the objective function to iteratively update the network parameters of the physical information neural network, thereby obtaining the trained physical information neural network.

[0223] In one possible implementation, based on the data constraints, multiple spatiotemporal sampling points can be randomly sampled from the spatiotemporal field of the smoke and dust concentration. The coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point can be extracted. The coordinates of each spatiotemporal sampling point can be input into the physical information neural network to obtain the predicted concentration corresponding to each spatiotemporal sampling point. The mean square error between the predicted concentration and the corresponding actual concentration value of each spatiotemporal sampling point can be calculated to obtain the data fitting loss.

[0224] Specifically, based on the set data constraints, multiple spatiotemporal sampling points can be selected from the spatiotemporal field of smoke and dust concentration using a uniform random sampling method. The sampling process covers the entire spatiotemporal domain to avoid model overfitting caused by concentrated sampling, ensuring that the sampling points comprehensively reflect the spatiotemporal distribution characteristics of smoke and dust concentration. By extracting the spatiotemporal coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point, the spatiotemporal coordinates can be used as input vectors to the constructed physical information neural network. The predicted concentration corresponding to each spatiotemporal sampling point can then be obtained through forward propagation calculations between the network's front-end feature extraction layer and the concentration field output branch.

[0225] Specifically, the mean square error between the predicted concentration and the actual concentration value at each spatiotemporal sampling point is calculated. The mean square error of all sampling points can be averaged to obtain the data fitting loss. This loss term is used to constrain the deviation between the network prediction results and the observed data, ensuring the accuracy of the concentration field prediction.

[0226] In one possible implementation, a spatiotemporal domain can be constructed based on a unified spatial grid. Multiple configuration points are randomly sampled within the spatiotemporal domain, and the coordinates of each configuration point are extracted. The coordinates of each configuration point are input into the physical information neural network to obtain the configuration prediction concentration, configuration prediction wind speed field, and configuration prediction diffusion coefficient field corresponding to each configuration point. The partial derivatives of the configuration prediction concentration are calculated by automatic differentiation. The partial derivatives, along with the configuration prediction wind speed field and the configuration prediction diffusion coefficient field, are substituted into the advection-diffusion equation to obtain the equation residuals. The advection-diffusion equation residual loss is calculated based on the equation residuals of each configuration point.

[0227] Specifically, a complete spatiotemporal domain can be constructed based on a unified spatial grid, which is completely consistent with the spatiotemporal range of the smoke and dust concentration spatiotemporal field. Subsequently, multiple configuration points can be randomly sampled within the spatiotemporal domain. The number of configuration points is determined according to the strengthening requirements of physical constraints and does not need to coincide with data sampling points, thereby achieving physical constraints on the entire spatiotemporal domain.

[0228] Specifically, by extracting the spatiotemporal coordinates of each configuration point and inputting them into a physical information neural network, the corresponding configuration predicted concentration, configuration predicted wind speed field, and configuration predicted diffusion coefficient field can be obtained. Based on the automatic differentiation mechanism of the deep learning framework, the first-order partial derivatives of the configuration predicted concentration with respect to time, and the first and second-order partial derivatives with respect to spatial coordinates, can be calculated. These partial derivatives correspond to the temporal rate of change and spatial gradient and curvature characteristics of the dust concentration, respectively. Substituting the partial derivatives, the configuration predicted wind speed field, and the configuration predicted diffusion coefficient field into the preset advection-diffusion equation, the expression of this equation can be:

[0229] ;

[0230] in, Spacetime coordinates The dust concentration field at that location. The first-order partial derivative of the concentration field with respect to time represents the rate of temporal change of concentration. Spacetime coordinates The wind speed field at that location, The spatial gradient of the concentration field is represented by the dot product of the two terms, which constitutes the advection term and characterizes the driving effect of wind speed on the migration of dust. Spacetime coordinates The diffusion coefficient field at that location, The Laplace operator for the concentration field and the product of the two form the diffusion term, which characterizes the spatial diffusion effect of smoke and dust. Spacetime coordinates The source term field at the location represents the intensity contribution of the smoke and dust emission source.

[0231] By substituting the partial derivatives of each order with the predicted wind speed field, the predicted diffusion coefficient field, and the predicted source term field into the above equation, the difference between the left and right sides of the equation can be calculated to obtain the equation residual. The sum of squares of the equation residuals for all configuration points and the mean value are taken to obtain the advection-diffusion equation residual loss. This loss term can force the network output to meet the physical laws of dust diffusion and avoid non-physical concentration field distribution.

[0232] In one possible implementation, a spatial domain can be constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within the spatial domain, and the coordinates of each source term monitoring point are extracted. The coordinates of each source term monitoring point are then input into the physical information neural network to obtain the predicted source terms corresponding to each source term monitoring point. The absolute mean of the predicted source terms corresponding to each source term monitoring point is calculated to obtain the source term sparsity regularization loss.

[0233] Specifically, a spatial domain can be constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within this domain, covering the potential emission source distribution area of ​​the target industrial region. By extracting the spatial coordinates of each source term monitoring point and inputting them into a physical information neural network, the predicted source terms for each monitoring point can be obtained through the source term field output branch. Calculating the absolute mean of the predicted source terms for all monitoring points yields the source term sparsity regularization loss. This loss term originates from the actual distribution characteristics of industrial emission sources; that is, industrial dust emissions are typically concentrated in specific equipment or areas, and the source term field exhibits a sparse distribution. Therefore, by constraining the network through sparsity regularization, scattered false emission sources can be avoided, ensuring that the source term field is consistent with the actual industrial emission situation.

[0234] It should be noted that, using the composite loss function as the objective function, an adaptive momentum estimation optimizer can be employed to perform gradient descent optimization. During optimization, the gradient of the composite loss function with respect to all learnable parameters of the network can be calculated using the backpropagation algorithm, and the network weights and biases are iteratively updated based on this gradient information. Batch processing can also be used to improve training efficiency. By setting reasonable iteration counts and convergence criteria, training can be stopped when the composite loss function value stabilizes or reaches the preset maximum number of iterations, resulting in the trained physical information neural network. This network can accurately output predicted concentration fields, predicted source term fields, predicted wind speed fields, and predicted diffusion coefficient fields, conforming to data constraints, physical laws, and industrial realities, using spatiotemporal coordinates as input.

[0235] In one possible implementation, extracting the spatial peak location and peak intensity from the target predicted source term field, using the spatial peak location as emission source coordinates, and the peak intensity as emission source intensity, includes:

[0236] S801: Perform spatial nonmaximum suppression processing on the target prediction source term field to remove discrete interference values ​​in the target prediction source term field and obtain the processed target prediction source term field.

[0237] Specifically, the spatiotemporal coordinates of the target are input into the trained physical information neural network. This network can simultaneously generate the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field through front-end feature extraction and back-end four-branch output mapping. Among them, the target predicted source term field directly represents the spatial distribution and intensity potential information of industrial dust emission sources. Performing spatial nonmaximum suppression processing on the target predicted source term field can eliminate discrete spurious peaks and isolated interference values ​​caused by noise interference in the source term field, and retain the concentrated strong source term regions corresponding to the emission sources. For example, based on the resolution of a unified spatial grid and the spatial scale characteristics of emission sources, a square neighborhood window of a fixed size can be set. Each spatial grid point of the target predicted source term field is traversed, and the source term value of the current grid point is compared with the source term values ​​of all other grid points in the neighborhood window one by one. Only the points whose current grid point value is greater than the values ​​of all other grid points in the neighborhood are retained, and the values ​​of the remaining grid points are set to zero. In this way, local false peaks are suppressed and the core source term regions corresponding to real emission sources are strengthened, resulting in a spatially concentrated target predicted source term field with thorough noise removal. This provides a reliable source term data foundation for subsequent peak location and intensity quantification.

[0238] S802: Traverse each spatial grid point in the processed target prediction source term field, and take the coordinates of the spatial grid point with the largest source term value in the processed target prediction source term field as the spatial peak position.

[0239] Specifically, all spatial grid points can be traversed in the processed target prediction source term field. The traversal process is performed sequentially according to the row and column index order of the unified spatial grid to ensure that no grid points are missed. Based on the physical characteristics of industrial dust emission sources, namely that industrial emissions are usually concentrated in specific equipment or emission outlets, manifesting as a single or a few concentrated high-intensity core areas in the source term field, the source term values ​​corresponding to each grid point can be compared and the maximum value recorded in real time. After the traversal is completed, the coordinates of the spatial grid point corresponding to the maximum recorded source term value can be determined as the spatial peak position, because the grid point with the largest source term value is the spatial core position of the emission source. This method can achieve precise spatial positioning of the emission source.

[0240] S803: Based on a preset neighborhood range of the spatial peak location, the numerical value of the target predicted source term field is integrated to obtain the integration result. The integration result is used as the peak intensity, and the peak intensity is converted into the corresponding emission rate. The emission rate is used as the emission source intensity.

[0241] Specifically, a preset neighborhood range can be defined based on the spatial peak location. The size of this neighborhood range is determined according to the typical influence radius of the emission source in the target industrial area, the resolution of the unified spatial grid, and historical emission monitoring data. This ensures that the neighborhood range can completely cover the core emission area of ​​the emission source and avoids introducing irrelevant background source terms over an excessively large area. The source term values ​​of all grid points within this neighborhood range in the target predicted source term field are integrated. For example, a numerical integration method adapted to the unified spatial grid can be used, where the integration is achieved by summing the product of the source term values ​​at grid points and the area of ​​the grid micro-element, yielding an initial integration result. This result initially reflects the total intensity of the emission source, but does not consider the dilution effect during the smoke and dust diffusion process, and needs to be further converted into an emission rate. In the process of converting this emission rate, the target predicted diffusion coefficient field output by the physical information neural network can be combined, and a diffusion correction mechanism can be introduced to restore the true emission intensity. The core calculation formula can be:

[0242] ;

[0243] in, The final smoke and dust emission rate serves as the core quantitative indicator of emission source intensity. For the spatial peak position The pre-defined neighborhood integration domain centered on the center; The target prediction source term field after spatial nonmaximum suppression processing is used to characterize the source term intensity of each grid point in the neighborhood. Predict the diffusion coefficient field in spacetime coordinates for the target The value at that location is synchronously output by the physical information neural network, reflecting the smoke and dust diffusion capacity at the corresponding location; Peak moment Below, the spatial average value of all grid points in the target prediction diffusion coefficient field within the unified spatial grid is used to normalize the local diffusion coefficient and avoid correction bias caused by differences in the overall diffusion level of the region; The preset diffusion correction coefficient is calibrated based on the diffusion characteristics of industrial dust and historical measured data, and its value range is [value range missing]. , used to regulate the intensity of diffusion correction; To unify the spatial micro-element area corresponding to the spatial grid and ensure the consistency of spatial scale in integral operations.

[0244] The above formula can incorporate the spatiotemporal distribution characteristics of the diffusion coefficient field into the source term integration process, i.e., the diffusion coefficient within the neighborhood. The larger the value, the stronger the smoke and dust diffusion and dilution effect at that location, and the higher the degree to which the source term value is diluted. The calculated relative diffusion coefficient quantifies the degree of local dilution, which can then be corrected using a correction factor. By adjusting the upward fluctuation range, the integral results can be dynamically corrected, ensuring that the calculated emission rate can offset the effects of diffusion and dilution, thus restoring the true emission intensity of the emission source. Using this emission rate as the emission source intensity achieves both intensity quantification and, through physical correction, ensures the accuracy and physical rationality of the quantification results, providing a reliable quantitative basis for the regulation of industrial dust emissions.

[0245] The above describes an industrial dust emission monitoring method provided by the embodiments of this application. The following describes the apparatus for performing the above-described industrial dust emission monitoring method.

[0246] Please see Figure 2 , Figure 2 This is a schematic diagram of an industrial dust emission monitoring device provided in an embodiment of this application. Figure 2 As shown, the industrial dust emission monitoring device includes:

[0247] Acquisition unit 11 is used to acquire visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series of the target industrial area;

[0248] The smoke and dust concentration spatiotemporal field generation unit 12 is used to perform iterative denoising processing according to the reverse diffusion process based on the visible light image sequence, the infrared thermal imaging sequence and the meteorological parameter time series to obtain the smoke and dust concentration field at each time moment, and arrange the smoke and dust concentration fields at each time moment in chronological order to generate the smoke and dust concentration spatiotemporal field.

[0249] The physical information neural network prediction unit 13 is used to input the spatiotemporal coordinates of the target to be monitored within the target industrial area into the trained physical information neural network to obtain the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field. The physical information neural network is trained with the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints.

[0250] The emission source identification unit 14 is used to extract the spatial peak position and peak intensity from the target predicted source term field, and use the spatial peak position as the emission source coordinates and the peak intensity as the emission source intensity.

[0251] In one possible implementation, the smoke concentration spatiotemporal field generation unit includes:

[0252] The standardization processing unit is used to standardize the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series to obtain a standardized visible light tensor sequence, a standardized infrared tensor sequence, and a meteorological parameter field sequence.

[0253] A cross-modal spatiotemporal feature fusion processing unit is used to perform cross-modal spatiotemporal feature fusion processing on the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time point;

[0254] The iterative denoising processing unit is used to take the spatiotemporal fusion feature map at each time as a conditional input, and perform iterative denoising processing based on the Gaussian noise field and the meteorological parameter field sequence according to the reverse diffusion process to obtain the smoke and dust concentration field at each time.

[0255] In one possible implementation, the cross-modal spatiotemporal feature fusion processing unit includes:

[0256] An embedding sequence construction unit is used to construct visible light block embedding sequences and infrared block embedding sequences based on the standardized visible light tensor sequence and the standardized infrared tensor sequence;

[0257] A cross-modal temporal attention mechanism fusion unit is used to fuse the visible light block embedding sequence and the infrared block embedding sequence based on the cross-modal temporal attention mechanism to obtain a spatiotemporal fusion feature map at each time step.

[0258] In one possible implementation, the embedded sequence building unit is specifically used for:

[0259] The standardized visible light tensor sequence is divided into multiple non-overlapping first local image blocks according to a preset size, and the standardized infrared tensor sequence is divided into multiple non-overlapping second local image blocks according to a preset size;

[0260] The first local image block and the second local image block are respectively subjected to feature flattening processing to obtain the corresponding one-dimensional image block features. Then, the one-dimensional image block features corresponding to the first local image block and the second local image block are respectively mapped to a preset high-dimensional feature space through a linear projection matrix to obtain the initial visible light block features and the initial infrared block features.

[0261] Based on the coordinate information of the first local image patch and the second local image patch in a unified spatial grid, a spatial location coding feature matching the unified spatial grid is constructed, and the spatial location coding feature is fused with the initial visible light patch feature and the initial infrared patch feature respectively to generate the visible light patch embedding sequence and the infrared patch embedding sequence.

[0262] In one possible implementation, the cross-modal temporal attention mechanism fusion unit is specifically used for:

[0263] Perform a linear transformation on the visible light block embedding sequence at the current moment to generate the corresponding query matrix;

[0264] The visible light block embedding sequence at the current time and multiple historical time points are concatenated one by one with the corresponding infrared block embedding sequence to obtain multiple concatenated feature sequences. Linear transformation processing is performed on each of the concatenated feature sequences to generate the corresponding key matrix and value matrix.

[0265] Calculate the product of the query matrix and the transpose of the key matrix to obtain the product result, and normalize the product result to obtain the attention weight matrix;

[0266] The value matrix is ​​weighted and summed using the attention weight matrix to obtain the fusion block embedding feature. The fusion block embedding feature is then reshaped into a two-dimensional spatial feature. Upsampling is performed on the two-dimensional spatial feature to make its resolution the same as that of the unified spatial grid, thus obtaining the spatiotemporal fusion feature map at each time step.

[0267] In one possible implementation, the iterative denoising unit includes:

[0268] A conditional constraint reverse diffusion network acquisition unit is used to acquire a conditional constraint reverse diffusion network, wherein the conditional constraint reverse diffusion network includes a feature encoding module, an iterative update module, and a physical constraint module;

[0269] The parameter acquisition unit is used to initialize a random noise field that conforms to a Gaussian distribution, obtain a Gaussian noise field, and obtain a preset feature length that matches the dust diffusion characteristics of the target industrial area. Based on the wind speed field in the meteorological parameter field sequence and the preset feature length, the empirical diffusion coefficient is calculated.

[0270] The iterative denoising subunit is used to traverse the spatiotemporal fusion feature map at each time step through the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, and sequentially perform reverse diffusion iterative denoising processing at the corresponding time step to obtain the dust concentration field at each time step.

[0271] In one possible implementation, the iterative denoising subunit is specifically used for:

[0272] The spatiotemporal fusion feature maps at each time point are input into the feature encoding module for feature mapping processing to obtain the conditional guidance vectors at each time point.

[0273] The conditional guiding vector, the empirical diffusion coefficient, the wind speed field and the Gaussian noise field in the meteorological parameter field sequence are input into the iterative update module to perform initial iterative calculations and output the initial noise field estimate.

[0274] Based on the initial noise field estimate, a reverse diffusion iteration is performed. In each iteration, the noise field estimate of the current iteration step is called through the physical constraint module. The time difference term, spatial gradient term, and spatial divergence term are discretized and calculated on a unified spatial grid. They are combined in the form of the advection-diffusion equation to obtain the residual term. The sum of squares of the residual term is calculated to obtain the physical constraint loss value. In the first iteration, the noise field estimate of the current iteration step is the initial noise field estimate.

[0275] The physical constraint module calculates the gradient based on the physical constraint loss value and the noise field estimate of the current iteration step to obtain the physical guidance gradient; the physical guidance gradient is multiplied by the dynamic coefficient to obtain the physical constraint correction amount; wherein the dynamic coefficient is a coefficient that decreases linearly with the increase of the iteration step.

[0276] The iterative update module obtains the noise standard deviation of the current iteration step. Based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, the data-driven update amount is calculated using a preset reverse diffusion iterative formula. The physical constraint correction amount is subtracted from the data-driven update amount to obtain the corrected noise field estimate. The corrected noise field estimate is used as the noise field estimate of the current iteration step in the next iteration.

[0277] The reverse diffusion iteration process is repeated until the number of iterations reaches the preset maximum number of iterations. The iterative update module is used to denoise the corrected noise field estimate obtained in the last iteration to obtain the dust concentration field at the corresponding time.

[0278] In one possible implementation, the device further includes: a physical information neural network training unit, specifically used for:

[0279] Construct a physical information neural network based on a fundamental physical information neural network;

[0280] Based on preset weighting coefficients, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are fused to construct a composite loss function;

[0281] During training, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are calculated. Gradient descent optimization is performed with the composite loss function as the objective function to iteratively update the network parameters of the physical information neural network, thereby obtaining the trained physical information neural network.

[0282] In one possible implementation, the physical information neural network training unit includes: a data fitting loss calculation unit, specifically used for:

[0283] Based on the data constraints, multiple spatiotemporal sampling points are randomly sampled from the spatiotemporal field of the smoke and dust concentration, and the coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point are extracted.

[0284] The coordinates of each of the spatiotemporal sampling points are input into the physical information neural network to obtain the predicted concentration corresponding to each of the spatiotemporal sampling points;

[0285] The mean square error between the predicted concentration and the corresponding actual concentration value at each of the aforementioned spatiotemporal sampling points is calculated to obtain the data fitting loss.

[0286] In one possible implementation, the physical information neural network training unit includes: an advection-diffusion equation residual loss calculation unit, specifically used for:

[0287] A spatiotemporal domain is constructed based on a unified spatial grid. Multiple configuration points are randomly sampled within the spatiotemporal domain, and the coordinates of each configuration point are extracted.

[0288] The coordinates of each of the configuration points are input into the physical information neural network to obtain the configuration prediction concentration, configuration prediction wind speed field and configuration prediction diffusion coefficient field corresponding to each of the configuration points.

[0289] The partial derivatives of the configured predicted concentration are calculated by automatic differentiation. The partial derivatives, the configured predicted wind speed field, and the configured predicted diffusion coefficient field are substituted into the advection-diffusion equation to obtain the equation residuals.

[0290] The advection-diffusion equation residual loss is calculated based on the equation residuals at each of the aforementioned configuration points.

[0291] In one possible implementation, the physical information neural network training unit includes: a source term sparsity regularization loss calculation unit, specifically used for:

[0292] A spatial domain is constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within the spatial domain, and the coordinates of each source term monitoring point are extracted.

[0293] The coordinates of each source term monitoring point are input into the physical information neural network to obtain the predicted source term corresponding to each source term monitoring point;

[0294] The mean absolute value of the predicted source terms corresponding to each of the source term monitoring points is calculated to obtain the source term sparsity regularization loss.

[0295] In one possible implementation, the emission source identification unit is specifically used for:

[0296] Spatial nonmaximum suppression is performed on the target prediction source term field to remove discrete interference values ​​in the target prediction source term field, resulting in the processed target prediction source term field;

[0297] In the processed target prediction source term field, each spatial grid point is traversed, and the coordinates of the spatial grid point with the largest source term value in the processed target prediction source term field are taken as the spatial peak position.

[0298] Based on a preset neighborhood range of the spatial peak location, the numerical value of the target predicted source term field is integrated to obtain the integration result. The integration result is used as the peak intensity, and the peak intensity is converted into the corresponding emission rate. The emission rate is used as the emission source intensity.

[0299] Each unit in the aforementioned industrial dust emission monitoring device can be implemented entirely or partially through software, hardware, or a combination thereof. These units can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each unit.

[0300] This application also provides an electronic device in its embodiments. (See reference...) Figure 3 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 3 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0301] like Figure 3As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0302] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0303] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the industrial dust emission monitoring methods provided in this application.

[0304] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the industrial dust emission monitoring methods provided in this application.

[0305] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0306] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0307] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0308] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for monitoring industrial dust emissions, characterized in that, include: Acquire visible light image sequences, infrared thermal imaging sequences, and meteorological parameter time series of the target industrial area; Based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, iterative denoising processing is performed according to the reverse diffusion process to obtain the smoke and dust concentration field at each time moment. The smoke and dust concentration fields at each time moment are then arranged in chronological order to generate the spatiotemporal field of smoke and dust concentration. The spatiotemporal coordinates of the target to be monitored within the target industrial area are input into the trained physical information neural network to obtain the target predicted concentration field, the target predicted source term field, the target predicted wind speed field, and the target predicted diffusion coefficient field. The physical information neural network is trained with the discrete spatiotemporal coordinates of the dust concentration spatiotemporal field and the corresponding concentration values ​​as data constraints, and the advection-diffusion equation as physical constraints. The spatial peak position and peak intensity are extracted from the target predicted source term field. The spatial peak position is used as the emission source coordinates, and the peak intensity is used as the emission source intensity.

2. The method according to claim 1, characterized in that, The process of performing iterative denoising based on the visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series, following a reverse diffusion process, yields the smoke and dust concentration field at each time point, including: The visible light image sequence, the infrared thermal imaging sequence, and the meteorological parameter time series are standardized to obtain a standardized visible light tensor sequence, a standardized infrared tensor sequence, and a meteorological parameter field sequence. Cross-modal spatiotemporal feature fusion processing is performed on the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time step; Using the spatiotemporal fusion feature map at each time point as input, and based on the Gaussian noise field and the meteorological parameter field sequence, iterative denoising processing is performed according to the reverse diffusion process to obtain the dust concentration field at each time point.

3. The method according to claim 2, characterized in that, The cross-modal spatiotemporal feature fusion processing of the standardized visible light tensor sequence and the standardized infrared tensor sequence to obtain spatiotemporal fusion feature maps at each time step includes: Based on the standardized visible light tensor sequence and the standardized infrared tensor sequence, a visible light block embedding sequence and an infrared block embedding sequence are constructed. Based on the cross-modal temporal attention mechanism, the visible light block embedding sequence and the infrared block embedding sequence are fused to obtain the spatiotemporal fusion feature map at each time step.

4. The method according to claim 3, characterized in that, The construction of visible light block embedding sequences and infrared block embedding sequences based on the standardized visible light tensor sequence and the standardized infrared tensor sequence includes: The standardized visible light tensor sequence is divided into multiple non-overlapping first local image blocks according to a preset size, and the standardized infrared tensor sequence is divided into multiple non-overlapping second local image blocks according to a preset size; The first local image block and the second local image block are respectively subjected to feature flattening processing to obtain the corresponding one-dimensional image block features. Then, the one-dimensional image block features corresponding to the first local image block and the second local image block are respectively mapped to a preset high-dimensional feature space through a linear projection matrix to obtain the initial visible light block features and the initial infrared block features. Based on the coordinate information of the first local image patch and the second local image patch in a unified spatial grid, a spatial location coding feature matching the unified spatial grid is constructed, and the spatial location coding feature is fused with the initial visible light patch feature and the initial infrared patch feature respectively to generate the visible light patch embedding sequence and the infrared patch embedding sequence.

5. The method according to claim 3, characterized in that, The visible light block embedding sequence and the infrared block embedding sequence are fused based on a cross-modal temporal attention mechanism to obtain spatiotemporal fusion feature maps at each time step, including: Perform a linear transformation on the visible light block embedding sequence at the current moment to generate the corresponding query matrix; The visible light block embedding sequence at the current time and multiple historical time points are concatenated one by one with the corresponding infrared block embedding sequence to obtain multiple concatenated feature sequences. Linear transformation processing is performed on each of the concatenated feature sequences to generate the corresponding key matrix and value matrix. Calculate the product of the query matrix and the transpose of the key matrix to obtain the product result, and normalize the product result to obtain the attention weight matrix; The value matrix is ​​weighted and summed using the attention weight matrix to obtain the fusion block embedding feature. The fusion block embedding feature is then reshaped into a two-dimensional spatial feature. Upsampling is performed on the two-dimensional spatial feature to make its resolution the same as that of the unified spatial grid, thus obtaining the spatiotemporal fusion feature map at each time step.

6. The method according to claim 2, characterized in that, The process involves using the spatiotemporal fusion feature maps at each time point as input, and performing iterative denoising based on the Gaussian noise field and the meteorological parameter field sequence, following a reverse diffusion process, to obtain the smoke and dust concentration field at each time point, including: Obtain a conditional constraint inverse diffusion network, which includes a feature encoding module, an iterative update module, and a physical constraint module; Initialize a random noise field that conforms to a Gaussian distribution to obtain a Gaussian noise field, and obtain a preset feature length that matches the dust diffusion characteristics of the target industrial area. Based on the wind speed field in the meteorological parameter field sequence and the preset feature length, calculate the empirical diffusion coefficient. Using the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, the spatiotemporal fusion feature map at each time step is traversed, and the reverse diffusion iterative denoising process at the corresponding time step is performed sequentially to obtain the dust concentration field at each time step.

7. The method according to claim 6, characterized in that, The process involves using the conditionally constrained reverse diffusion network, based on the Gaussian noise field and the empirical diffusion coefficient, to traverse the spatiotemporal fusion feature maps at each time step, and sequentially perform iterative reverse diffusion denoising processing at the corresponding time step to obtain the smoke and dust concentration field at each time step, including: The spatiotemporal fusion feature maps at each time point are input into the feature encoding module for feature mapping processing to obtain the conditional guidance vectors at each time point. The conditional guiding vector, the empirical diffusion coefficient, the wind speed field and the Gaussian noise field in the meteorological parameter field sequence are input into the iterative update module to perform initial iterative calculations and output the initial noise field estimate. Based on the initial noise field estimate, a reverse diffusion iteration is performed. In each iteration, the noise field estimate of the current iteration step is called through the physical constraint module. The time difference term, spatial gradient term, and spatial divergence term are discretized and calculated on a unified spatial grid. They are combined in the form of the advection-diffusion equation to obtain the residual term. The sum of squares of the residual term is calculated to obtain the physical constraint loss value. In the first iteration, the noise field estimate of the current iteration step is the initial noise field estimate. The physical constraint module calculates the gradient based on the physical constraint loss value and the noise field estimate of the current iteration step to obtain the physical guidance gradient; the physical guidance gradient is multiplied by the dynamic coefficient to obtain the physical constraint correction amount; wherein the dynamic coefficient is a coefficient that decreases linearly with the increase of the iteration step. The iterative update module obtains the noise standard deviation of the current iteration step. Based on the noise field estimate, conditional guidance vector, and noise standard deviation of the current iteration step, the data-driven update amount is calculated using a preset reverse diffusion iterative formula. The physical constraint correction amount is subtracted from the data-driven update amount to obtain the corrected noise field estimate. The corrected noise field estimate is used as the noise field estimate of the current iteration step in the next iteration. The reverse diffusion iteration process is repeated until the number of iterations reaches the preset maximum number of iterations. The iterative update module is used to denoise the corrected noise field estimate obtained in the last iteration to obtain the dust concentration field at the corresponding time.

8. The method according to claim 1, characterized in that, The training methods for the physical information neural network include: Construct a physical information neural network based on a fundamental physical information neural network; Based on preset weighting coefficients, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are fused to construct a composite loss function; During training, the data fitting loss, the advection-diffusion equation residual loss, and the source term sparsity regularization loss are calculated. Gradient descent optimization is performed with the composite loss function as the objective function to iteratively update the network parameters of the physical information neural network, thereby obtaining the trained physical information neural network.

9. The method according to claim 8, characterized in that, The calculation methods for the data fitting loss include: Based on the data constraints, multiple spatiotemporal sampling points are randomly sampled from the spatiotemporal field of the smoke and dust concentration, and the coordinates and actual concentration values ​​corresponding to each spatiotemporal sampling point are extracted. The coordinates of each of the spatiotemporal sampling points are input into the physical information neural network to obtain the predicted concentration corresponding to each of the spatiotemporal sampling points; The mean square error between the predicted concentration and the corresponding actual concentration value at each of the aforementioned spatiotemporal sampling points is calculated to obtain the data fitting loss.

10. The method according to claim 8, characterized in that, The calculation method for the residual loss of the advection-diffusion equation includes: A spatiotemporal domain is constructed based on a unified spatial grid. Multiple configuration points are randomly sampled within the spatiotemporal domain, and the coordinates of each configuration point are extracted. The coordinates of each of the configuration points are input into the physical information neural network to obtain the configuration prediction concentration, configuration prediction wind speed field and configuration prediction diffusion coefficient field corresponding to each of the configuration points. The partial derivatives of the configured predicted concentration are calculated by automatic differentiation. The partial derivatives, the configured predicted wind speed field, and the configured predicted diffusion coefficient field are substituted into the advection-diffusion equation to obtain the equation residuals. The advection-diffusion equation residual loss is calculated based on the equation residuals at each of the aforementioned configuration points.

11. The method according to claim 8, characterized in that, The calculation methods for the source term sparsity regularization loss include: A spatial domain is constructed based on a unified spatial grid. Multiple source term monitoring points are randomly sampled within the spatial domain, and the coordinates of each source term monitoring point are extracted. The coordinates of each source term monitoring point are input into the physical information neural network to obtain the predicted source term corresponding to each source term monitoring point; The mean absolute value of the predicted source terms corresponding to each of the source term monitoring points is calculated to obtain the source term sparsity regularization loss.

12. The method according to claim 1, characterized in that, The step of extracting the spatial peak position and peak intensity from the target predicted source term field, using the spatial peak position as the emission source coordinates, and the peak intensity as the emission source intensity, includes: Spatial nonmaximum suppression is performed on the target prediction source term field to remove discrete interference values ​​in the target prediction source term field, resulting in the processed target prediction source term field; In the processed target prediction source term field, each spatial grid point is traversed, and the coordinates of the spatial grid point with the largest source term value in the processed target prediction source term field are taken as the spatial peak position. Based on a preset neighborhood range of the spatial peak location, the numerical value of the target predicted source term field is integrated to obtain the integration result. The integration result is used as the peak intensity, and the peak intensity is converted into the corresponding emission rate. The emission rate is used as the emission source intensity.

13. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the industrial dust emission monitoring method as described in any one of claims 1 to 12.

14. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the industrial dust emission monitoring method as described in any one of claims 1 to 12.

15. A computer-readable storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the industrial dust emission monitoring method as described in any one of claims 1 to 12.

Images