Lensless diffraction on-line detection and distribution positioning method of airborne particulate matter

Abstract

The application discloses an airborne particulate matter lens-free diffraction online detection and distribution positioning method, belongs to the technical field of intelligent environment monitoring, collects aerosol diffraction images and meteorological data, constructs a diffraction spectrum meteorological space-time alignment set through timestamp alignment; extracts frequency spectrum energy distribution characteristics and ring characteristics to input a deep inversion model, introduces humidity, wind speed and temperature, combines a preset threshold correction weight, inverts particulate matter equivalent diameter distribution and concentration, generates a particulate matter basic parameter set, deduces a high-probability source area based on a concentration gradient and a meteorological wind field direction, generates an initial source area estimation map; combined with constraint conditions, a constraint type reverse trajectory tracing model is constructed to deduce a particulate matter propagation path; a concentration and position error feedback mechanism is adopted to construct a joint loss function to iteratively optimize path initial weight and source area intensity distribution, correct the propagation path and positioning, output particulate matter online detection results and high-resolution spatial distribution positioning maps, and realize accurate tracing and monitoring.

Description

Technical Field

[0001] This invention relates to a lensless online detection and distribution location method for airborne particulate matter, belonging to the field of intelligent environmental monitoring technology. Background Technology

[0002] Inside livestock sheds, airborne particulate matter carries pollutants such as harmful microorganisms and ammonia, which not only damage the ecological environment but also irritate the respiratory mucosa of livestock, reducing their immunity and productivity. Accurate detection and location of airborne particulate matter in livestock sheds are crucial for tracing pollution sources, assessing environmental risks, and developing scientific environmental management strategies. Traditional particulate matter detection methods suffer from limitations such as low resolution, slow detection speed, and difficulty in real-time location within the complex environments of livestock farming. Lensless diffraction imaging technology, with its advantages of high resolution, wide field of view, label-free operation, and high throughput, has become a research hotspot for rapid, online, and real-time detection of airborne particulate matter in livestock sheds.

[0003] Chinese patent application CN119845904A discloses a method and system for locating and detecting viral infection based on diffraction optics. This method utilizes diffraction imaging principles to acquire diffraction fingerprints corresponding to all cells within an imaging area. The texture features of each diffraction fingerprint in the imaging area are extracted using the gray-level co-occurrence matrix method. The extracted texture features are compared with pre-stored standard texture features in a model relating cell morphological changes to diffraction fingerprints, thereby determining the cell state of each cell in the imaging area and locating and identifying the viral infection status of each cell. Applying this efficient optical diffraction spectroscopy system to viral infection detection eliminates the need for specialized personnel and expensive, complex on-site testing equipment. It meets the requirements for continuous, high-throughput assessment of cell status after viral infection, laying the foundation for the widespread application of point-of-care viral infection detection platforms.

[0004] Although existing virus infection detection technologies based on diffraction optics can locate virus infections at the cellular level, they are limited to the microscopic imaging area and do not integrate meteorological data to track the trajectory of particulate matter in the atmosphere, making it difficult to achieve accurate positioning of airborne viruses and other particulate matter in complex spaces. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the present invention aims to provide a lensless online detection and distribution location method for airborne particulate matter. By spatiotemporal alignment and fusion of aerosol diffraction images and meteorological data, dynamic weight correction, multi-constraint inverse trajectory tracing, and dual-error optimization, the method achieves accurate inversion of the equivalent diameter distribution and concentration of particulate matter, efficient source region location, and clear path tracing, thereby improving the reliability and practicality of online detection and distribution location of airborne particulate matter.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A lensless online detection and distribution localization method for airborne particulate matter, including:

[0008] Aerosol diffraction images and meteorological data were collected, preprocessed, and aligned with a unified timestamp to construct a meteorological spatiotemporal alignment set of diffraction spectra.

[0009] Extract the spectral energy distribution features and ring features into the deep inversion model, introduce humidity, wind speed and temperature from meteorological data, and combine preset thresholds to correct weights to invert the equivalent diameter distribution and concentration of particulate matter, and output the basic parameter set of particulate matter; based on the concentration gradient and meteorological wind field direction, co-infer the high probability source region and generate an initial source region estimation map.

[0010] Based on the initial source region estimation map and particulate matter basic parameter set, constraints are introduced to construct a constrained reverse trajectory tracing model, and the Lagrange reverse trajectory method is used to deduce the particulate matter propagation path.

[0011] A dual error feedback mechanism based on concentration and location is adopted to correct the propagation path and location distribution. A joint loss function is constructed to adjust the path weight and source area intensity distribution in reverse. Iterative optimization is used to generate online detection results and high-resolution spatial distribution location maps of particulate matter.

[0012] Specifically, the steps for retrieving the equivalent diameter distribution and concentration of particulate matter include:

[0013] The spectral energy distribution characteristics are obtained, and spectral feature vectors are generated using the local maximum method.

[0014] Extract the diffraction ring intensity curve, combine it with Hough circle transform to generate a ring-shaped feature vector, concatenate it with the spectral feature vector and combine it with linear transformation, and calculate the weighted enhanced image feature vector.

[0015] Meteorological data is normalized into meteorological feature vectors, which are used as query matrices and the enhanced image feature vectors to calculate calibration weights through a cross-attention mechanism.

[0016] The calibration weights are corrected by incorporating humidity, wind speed, and temperature from meteorological conditions and combining humidity thresholds, wind speed thresholds, and temperature thresholds.

[0017] The corrected calibration weights are multiplied by the enhanced image feature vector to obtain the meteorological corrected image features, which are then combined with the original meteorological features to generate a deep fusion feature vector.

[0018] The deep fusion feature vector is input into the deep inversion model, and through feature dimensionality reduction and nonlinear transformation, the preliminary inversion results of particulate matter equivalent diameter distribution (PSD) and particulate matter mass concentration (MC) are output.

[0019] Specifically, the steps for retrieving the equivalent diameter distribution and concentration of particulate matter also include:

[0020] Set the particle size threshold range and mass concentration threshold for the equivalent diameter distribution;

[0021] Construct particle size constraint rules. If the peak particle size of PSD exceeds the particle size threshold range, it is determined that the particle size constraint conditions are not met and the particle size is marked as abnormal; otherwise, it is determined that the particle size constraint conditions are met and the particle size is marked as normal.

[0022] Concentration constraint rules are constructed. If the concentration limit (MC) is greater than the mass concentration threshold, the concentration constraint rules are not met and the concentration is marked as abnormal; otherwise, the concentration constraint rules are met and the concentration is marked as normal.

[0023] If both the particle size constraint rule and the concentration constraint rule are satisfied, the preliminary inversion result is deemed reasonable; otherwise, it is deemed unreasonable.

[0024] All detection points are verified by time series and spatial grid inversion, and the effective inversion results are integrated to generate a set of basic particulate matter parameters.

[0025] Specifically, the steps for generating the initial source region estimation map include:

[0026] Based on the particulate matter basic parameter set, a concentration grid is generated by spatial interpolation of particulate matter concentration data, and a wind field grid is generated by spatial interpolation of wind field data, and the grids are aligned with the concentration grids in space.

[0027] Extract the proportion of fine and coarse particles in each grid cell, and set a first threshold and a second threshold;

[0028] For fine particulate matter, if the proportion of fine particulate matter exceeds the first threshold, it is determined that fine particulate matter has a dominant position; otherwise, it is determined that it does not have a dominant position.

[0029] For coarse particulate matter, if the proportion of coarse particles exceeds the second threshold, then coarse particulate matter is determined to be dominant; otherwise, it is determined not to be dominant.

[0030] If each grid cell has only one dominant type, then it is determined that there is a particle dominance type, including fine particle dominance and coarse particle dominance, and it is marked as the target grid;

[0031] Otherwise, it is determined that there is no particle-dominant type and it is marked as a non-target mesh;

[0032] Calculate the concentration difference between the target grid and each neighboring grid and the actual spatial distance. Divide the concentration difference by the actual spatial distance to obtain the concentration gradient magnitude. Determine the gradient direction based on the sign of the concentration difference.

[0033] Specifically, the steps for generating the initial source region estimation map also include:

[0034] Select the opposite direction of the gradient and calculate the included angle based on the wind field and wind direction. ;

[0035] like If the wind field and gradient direction are consistent, then the first collaborative weight is set.

[0036] like If the wind field and gradient part are aligned, then a second collaborative weight is set.

[0037] like If the wind field and gradient direction conflict, a third collaborative weight is set.

[0038] By combining the wind field direction weighted correction to obtain the cooperative pointing vector, candidate source units are divided, the number and intensity of cooperative pointing vectors are counted, and the convergence degree is obtained by calculating the product of the number ratio and the intensity ratio.

[0039] Set a convergence threshold and retain candidate source units that exceed the convergence threshold to form a high-probability source region in the initial screening.

[0040] Calculate the theoretical distance for each particulate matter from the high-probability source region of the initial screening to the high-concentration grid, and remove regions that exceed the theoretical distance;

[0041] When the high-probability source regions of the initial screening, dominated by coarse particles and dominated by fine particles, spatially overlap, the overlapping part is taken as the core source region and the comprehensive source probability is calculated. Combined with geographic information, an initial source region estimation map is generated.

[0042] Specifically, the steps for constructing a constrained reverse trajectory tracing model include:

[0043] By introducing constraints, a constrained reverse trajectory tracing model is constructed, including an input layer, a core computation layer, a constraint correction layer, and an output layer.

[0044] The input layer receives the initial source region estimation map and the particulate matter basic parameter set;

[0045] The core computing layer uses the detection point as the initial position and combines the reverse wind speed vector and time step to calculate the basic spatiotemporal coordinates of the trajectory.

[0046] Based on the trajectory-based spatiotemporal coordinates, the constraint correction layer converts the constraints into precipitation clearing functions, radiation attenuation functions, temperature and humidity settling functions, and wind speed steering functions, and then corrects them sequentially in conjunction with the constraint priority controller to obtain the corrected trajectory-based spatiotemporal coordinates.

[0047] The output layer transforms the basic spatiotemporal coordinates of the corrected trajectory into a multi-dimensional result;

[0048] By calling historical particulate matter source tracing cases, a training set is constructed. With the goal of minimizing the error between the multi-dimensional results and the actual source area coordinates and contribution, the key parameters of the constraint correction layer are calibrated to obtain an optimized constraint-type reverse trajectory source tracing model.

[0049] Specifically, the steps for deducing particulate matter propagation paths include:

[0050] Set up a reverse simulation period, trace back from the sampling time, take the detection point as the endpoint, set the initial height in combination with particle size and particle density, and set a first-level weight for each trajectory;

[0051] Particulate matter moving in the opposite direction of wind speed forms an initial trajectory cluster;

[0052] Based on the precipitation removal function, precipitation is statistically analyzed, and a moderate precipitation threshold and a heavy precipitation threshold are set to divide the area into a strong removal zone and a weak removal zone.

[0053] Based on the particulate matter basic parameter set, the particles are divided into highly hygroscopic particles, weakly hygroscopic particles, and biological particles. The removal rates in the strong removal zone and the weak removal zone are matched, and the trajectories affected by precipitation are removed from the initial trajectory cluster to form a first-level corrected trajectory cluster.

[0054] Based on the radiation attenuation function, the solar radiation intensity and duration are divided into a strong inactivation zone and a weak inactivation zone.

[0055] Radiation resistance types are classified based on the radiation resistance performance of particulate matter, radiation attenuation coefficients are matched from a preset lookup table, and secondary weights are calculated based on the primary weights.

[0056] If the secondary weight is lower than the preset weight threshold, the trajectory is removed; otherwise, the trajectory is retained to form a secondary corrected trajectory cluster.

[0057] Specifically, the steps for deducing particulate matter propagation paths also include:

[0058] Based on the aforementioned temperature and humidity sedimentation function, the sedimentation velocity is calculated in combination with particle size and particle density.

[0059] Extract the current trajectory node height, and calculate and correct the trajectory node height based on the settlement velocity and backtracking time;

[0060] The atmospheric boundary layer range is invoked, and the secondary weights are adjusted according to the height of the corrected trajectory nodes to obtain the tertiary weights; trajectories below the preset weight threshold are removed to generate a tertiary corrected trajectory cluster;

[0061] Based on the wind speed steering function, calculate the deviation of node direction and magnitude, dynamically set the deviation threshold, and count the proportion of abnormal nodes.

[0062] If the proportion of abnormal nodes is lower than the preset proportion threshold, the trajectory direction is corrected; otherwise, the corresponding trajectory is removed, resulting in a four-level corrected trajectory cluster.

[0063] Set an initial high-weight source region, calculate the angle with the trajectory direction and adjust the third-level weight to obtain a fourth-level weight, cluster the node density of the reverse trajectory, calculate the total weight of each type of node, filter the central source region in descending order and calculate the overlap rate with the estimated map of the initial source region.

[0064] When the overlap rate is lower than the preset overlap rate threshold, a high-resolution source region heatmap is obtained after correction, before screening. The node trajectory represents the propagation path.

[0065] Specifically, the steps for generating a high-resolution spatial distribution location map include:

[0066] Based on the boundary of the high-resolution source region heat map, the propagation path nodes are divided into source region nodes and non-source region nodes;

[0067] Statistical analysis of dwell time and acquisition of initial source region intensity; calculation of total weight of source region nodes and non-source region nodes; summation to obtain initial path weight.

[0068] Set the transmission coefficient, calculate the predicted concentration based on the initial weight of the path, and calculate the concentration error in combination with the actual concentration;

[0069] The source region node deviation is calculated based on the average Euclidean distance between the source region node and the center of the high-resolution source region heat map area, and the endpoint deviation is calculated based on the Euclidean distance between the theoretical propagation path endpoint and the actual propagation path endpoint. The position error is then calculated using a weighted average.

[0070] The error confidence level is dynamically adjusted by combining the actual concentration and wind speed fluctuation amplitude, and an error confidence level matrix is ​​generated.

[0071] Specifically, the steps for generating a high-resolution spatial distribution location map also include:

[0072] Based on the error confidence matrix, a joint loss function is constructed to calculate the joint loss, and the initial path weights are optimized using the gradient descent method to obtain the optimized path weights.

[0073] The source region intensity is adjusted according to the concentration error to obtain the optimized source region intensity. The optimized predicted concentration is calculated in combination with the transmission coefficient. The weighted calculation is then used to correct the detection value and generate the online particulate matter detection results.

[0074] Based on the optimized source region intensity, the optimized concentration value and spatial gradient distribution are calculated in conjunction with the diffusion law;

[0075] The optimized concentration value is introduced into the high-resolution source region heat map to generate a three-dimensional coordinate map. The propagation path corresponding to the optimized path weight and the spatial gradient distribution are added to generate a high-resolution spatial distribution location map.

[0076] The beneficial effects of this invention are:

[0077] 1. This invention simultaneously acquires aerosol diffraction images and meteorological data, constructs a meteorological spatiotemporal alignment set of diffraction spectra, extracts spectral energy distribution features and ring features, and combines humidity, wind speed, and temperature threshold correction weights to accurately deduce the basic parameter set of particulate matter. Based on the concentration gradient and wind field direction, it collaboratively infers high-probability source regions, generates an initial source region estimation map, and then uses a constrained reverse trajectory tracing model to infer the particulate matter propagation path, effectively improving the accuracy of source location and enhancing monitoring efficiency and precision.

[0078] 2. This invention constructs a constrained inverse trajectory tracing model with multiple constraints including precipitation removal, radiation attenuation, temperature and humidity settling, and wind speed guidance. It combines the Lagrange method to deduce the propagation path and matches and corrects rules for different types of particulate matter, significantly reducing interference from non-target factors. At the same time, it adopts a dual error feedback mechanism of concentration and location, combines the gradient descent method to optimize the path weight, and adjusts the grid intensity of the source area heat map according to the concentration error. Finally, it generates a three-dimensional high-resolution positioning map containing optimized concentration values ​​and spatial gradient distribution, effectively solving the problems of fuzzy positioning and easy deviation in traditional source tracing and adapting to complex meteorological and particulate matter types. Attached Figure Description

[0079] Figure 1 A flowchart of a lensless online detection and distribution localization method for airborne particulate matter;

[0080] Figure 2 This is a flowchart illustrating the inversion of the equivalent diameter distribution and concentration of particulate matter in this invention;

[0081] Figure 3 This is a flowchart illustrating the propagation path of particulate matter from the source region to the detection point in this invention;

[0082] Figure 4 This is a flowchart illustrating the generation of a high-resolution spatial distribution location map in this invention;

[0083] Figure 5 This is a schematic diagram of the principle of the initial source region estimation map in this invention. Detailed Implementation

[0084] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0085] refer to Figures 1 to 5 As shown in the figure, this embodiment introduces a lensless online detection and distribution localization method for airborne particulate matter, including the following steps:

[0086] Aerosol diffraction images are acquired synchronously using a multi-point lensless diffraction detection device and preprocessed, such as removing background noise. Meteorological data, such as wind direction and speed, temperature, humidity, and precipitation, are collected from multiple meteorological sensors and preprocessed, such as removing outliers. The preprocessed aerosol diffraction images and meteorological data are aligned with a unified timestamp to construct a meteorological spatiotemporal alignment set of diffraction spectra.

[0087] Fast Fourier Transform (FFT) is used to extract the spectral energy distribution features of aerosol diffraction images. A radial averaging algorithm is then used to generate ring-shaped features. These spectral energy distribution features and ring-shaped features are stitched together and input into a deep inversion model with meteorological corrections. Humidity, wind speed, and temperature from meteorological conditions are incorporated, and weights are adjusted using preset humidity, wind speed, and temperature thresholds. The equivalent diameter distribution and concentration of particulate matter are then inverted, outputting a set of basic particulate matter parameters. Based on the concentration gradient and meteorological wind field direction, high-probability source regions are co-inferred, generating an initial source region estimation map. The set of basic particulate matter parameters includes, but is not limited to, particle size and particle density.

[0088] Based on the initial source region estimation map and particulate matter basic parameter set, constraints are introduced to construct a constrained inverse trajectory source tracing model. The Lagrange inverse trajectory method is used to deduce the propagation path of particulate matter from the source region to the detection point. Among them, the constraints include wind speed vector guidance constraints, temperature and humidity settlement correction, precipitation clearing zone shielding and solar radiation inactivation zone attenuation.

[0089] A dual error feedback mechanism based on concentration and location is adopted to correct the propagation path and location distribution. A joint loss function is constructed to adjust the path weight and source area intensity distribution in reverse. Iterative optimization is used to generate online detection results and high-resolution spatial distribution location maps of particulate matter.

[0090] Specifically, the steps for retrieving the equivalent diameter distribution and concentration of particulate matter include:

[0091] Fast Fourier Transform is used to process aerosol diffraction images to obtain spectral energy distribution characteristics. Effective frequencies are detected and screened using the local maximum method to generate standardized spectral feature vectors. The spectral features include, but are not limited to, the proportion of spectral peak energy.

[0092] The radial averaging algorithm is used to extract the diffraction ring intensity curves of aerosol diffraction images. The ring morphology is analyzed by combining Hough circle transform to generate ring feature vectors. Among them, the ring features include, but are not limited to, ring intensity attenuation rate features and ring symmetry features.

[0093] The spectral feature vector and the annular feature vector are concatenated according to their dimensions to form a joint image feature vector. The joint image feature vector is then mapped to a query matrix, a key matrix, and a value matrix through a linear transformation. Attention weights are calculated and weighted summation is performed to obtain the enhanced image feature vector.

[0094] Meteorological data is normalized to obtain meteorological feature vectors and set as query matrices. Enhanced image feature vectors are set as key matrices and value matrices. The calibration weight distribution of meteorological data on image feature vectors is obtained through a cross-attention mechanism.

[0095] Incorporating meteorological conditions such as humidity, wind speed, and temperature, and considering the physical characteristics of aerosol particles—that high humidity easily causes particle aggregation leading to an increase in equivalent diameter, and that airflow disturbances affect the stability of diffraction ring morphology and that air refractive index changes with temperature, thus altering the light propagation path—a humidity threshold was set based on the typical characteristics of livestock farm sheds—closed and semi-closed spaces prone to high humidity, localized airflow disturbances, and high temperatures. 75%, wind speed threshold 0.8 meters per second and temperature threshold For a temperature of 32 degrees Celsius, a weighting adjustment is performed, including:

[0096] If the humidity exceeds the humidity threshold If the value is positive, it indicates that the equivalent diameter of the particles has increased, the ring strength attenuation has accelerated, and the calibration weight of the ring strength attenuation rate characteristic has been increased; otherwise, the baseline weight has been maintained.

[0097] If the wind speed exceeds the wind speed threshold If the deviation is positive, it indicates that the diffraction ring is deformed by the airflow, which reduces the reliability of the ring symmetry feature and lowers the calibration weight of the ring symmetry feature; otherwise, the baseline weight is maintained.

[0098] If the temperature exceeds the temperature threshold If the air refractive index deviates from the normal value, it indicates that the light propagation path has shifted due to the change in refractive index, and the weight of the peak energy ratio in the spectrum is reduced; otherwise, the baseline weight is maintained.

[0099] The meteorological corrected image features are obtained by calculating the product of the corrected calibration weights and the enhanced image feature vector; the original meteorological features and the meteorological corrected image features are fused together, and a meteorological corrected deep fusion feature vector is generated through two fully connected layers.

[0100] The meteorological-corrected deep fusion feature vector is input into the deep inversion model. Through feature dimensionality reduction and nonlinear transformation, cross-modal feature correlations are captured, and high-order abstract features are output. The deep inversion model includes an input layer, a feature mapping layer, a cross-modal fusion layer, and an output layer.

[0101] The input layer is used to receive the meteorologically corrected deep fusion feature vector;

[0102] The feature mapping layer is used to perform nonlinear transformation on the meteorologically corrected deep fusion feature vector through two fully connected neural networks combined with the ReLU activation function, to map low-dimensional features to a high-dimensional feature space and represent the nonlinear correlation between features; the embedded batch normalization layer stabilizes the feature distribution to adapt to the strong volatility of the environmental characteristics of livestock farms.

[0103] The cross-modal fusion layer is used to focus on the key correlation temperature between meteorological features and image features through a multi-head attention mechanism, such as the coupling features of humidity and ring intensity attenuation rate, and wind speed and ring symmetry. It uses weight allocation to strengthen effective cross-modal features and weaken noise features, and performs deep fusion of meteorological and image cross-modal features, outputting a fused high-order feature matrix.

[0104] The output layer is used to reduce the dimensionality of the high-order feature matrix to the target dimension through linear transformation, such as the inversion target dimension of the equivalent diameter and concentration of particulate matter. The output range is constrained by the Sigmoid activation function, and the output output is a high-order abstract feature characterizing the key physical parameters of aerosols. Among them, the key physical parameters include equivalent particle size and mass concentration.

[0105] The training set consists of a labeled dataset composed of aerosol diffraction images collected at different times in livestock farms, meteorological monitoring data, and laboratory-calibrated aerosol physical parameters. A deep inversion model is trained using a combination loss function of mean square error loss and cross-entropy loss, along with the Adam optimizer. After training, the meteorologically corrected deep fusion feature vector of the sample to be inverted is input, and high-order abstract features are output.

[0106] For example, the input layer is used to receive a 128-dimensional meteorologically corrected deep fusion feature vector;

[0107] The feature mapping layer is used to perform a nonlinear transformation on the 128-dimensional meteorological-corrected deep fusion feature vector through a two-layer fully connected neural network, with 256 neurons in the first hidden layer and 128 neurons in the second hidden layer, combined with the ReLU activation function, to complete the mapping from low-dimensional features to high-dimensional feature space and represent the nonlinear correlation between features; the embedded batch normalization layer stabilizes the feature distribution to adapt to the strong fluctuation of the environmental characteristics of livestock farms;

[0108] The cross-modal fusion layer is used to focus on the key correlation dimensions of meteorological features and image features through a 4-head attention mechanism with 32 dimensions each, such as the coupling features of humidity and ring intensity attenuation rate, wind speed and ring symmetry. It uses weight allocation to strengthen effective cross-modal features and weaken noise features, and performs deep fusion of meteorological and image cross-modal features, outputting a high-order feature matrix with 128×64 dimensions after fusion.

[0109] The output layer is used to reduce the dimensionality of the high-order feature matrix after the fusion of 128×64 dimensions to 2 dimensions through linear transformation, corresponding to the two inversion target dimensions of particulate matter equivalent particle size and mass concentration. The output range is constrained to 0~1 through the Sigmoid activation function, corresponding to the normalized range of equivalent particle size of 0-10 micrometers and mass concentration of 0-100 micrograms per cubic meter. The output is a 2-dimensional high-order abstract feature characterizing the key physical parameters of aerosols.

[0110] The training set consisted of 300 sets of aerosol diffraction images collected at different times in livestock farms, synchronous meteorological monitoring data, and laboratory-calibrated aerosol physical parameters. The meteorological monitoring data included relative humidity of 0-100%, wind speed of 0-5 m / s, and temperature of 15-35 degrees Celsius. The aerosol physical parameters included equivalent particle size of 0.1-10 μm and mass concentration of 10-100 μg / m³. A combined loss function was used, consisting of a mean squared error loss function with a weight of 0.7 and a cross-entropy loss function with a weight of 0.3, and an initial learning rate of 1×10⁻⁶. -4 The weight decay coefficient is 1×10 -5 The deep inversion model is trained using an Adam optimizer with a momentum of 0.9, with 500 training iterations and a batch size of 32. After training, the 128-dimensional meteorologically corrected deep fusion feature vector of the sample to be inverted is input, and the 2-dimensional high-order abstract features are output.

[0111] The high-order abstract features are processed by a multilayer perceptron decoder, and the equivalent diameter distribution of particles is output through a regression layer. With particulate matter mass concentration Preliminary inversion results;

[0112] Based on the particle size distribution characteristics of airborne aerosol particles in livestock farms, which are mainly inhalable fine particles, the particle size threshold range for equivalent diameter distribution is set to 0.3-10 micrometers. Based on the requirements for disease prevention and control in livestock farms, in order to reduce the risk of respiratory disease transmission and protect the respiratory health of livestock and poultry, the mass concentration threshold is set to 7.5 milligrams per cubic meter.

[0113] Construct particle size constraint rules, if If the peak particle size exceeds the particle size threshold range, it indicates that the detection range of the equipment is exceeded, and the particle size constraint rules are not met, and the particle size is marked as abnormal; otherwise, the particle size constraint rules are met, and the particle size is marked as normal.

[0114] Construct concentration constraint rules, if If the concentration is greater than the mass concentration threshold, it indicates that the concentration exceeds the normal aerosol concentration range, and the concentration is judged not to meet the concentration constraint rules and is marked as abnormal; otherwise, the concentration constraint rules are judged to meet the concentration constraint rules and the concentration is marked as normal.

[0115] If both particle size and concentration constraints are met, it indicates that the preliminary inversion results meet the equipment detection capability and aerosol concentration constraints, and the preliminary inversion results are deemed reasonable; otherwise, the preliminary inversion results are deemed unreasonable, and the specific anomaly type is recorded.

[0116] After completing the inversion and verification of a single detection point, all detection points are inverted and verified sequentially through time series and spatial grids. The effective inversion results of all detection points are integrated, and particulate matter from abnormal detection points with confidence levels below the confidence threshold is removed. The basic parameter set of particulate matter is then output. Among them, based on the accuracy requirements for aerosol monitoring in livestock farms, the relative error of aerosol particulate matter inversion results is less than 15%, and the data validity judgment standard is that the proportion of effective inversion data for a single detection point exceeds 80%. Therefore, the confidence threshold is determined to be 0.85.

[0117] Specifically, the steps for generating the initial source region estimation map include:

[0118] Based on a set of fundamental particulate matter parameters, spatial interpolation is performed on discrete particulate matter concentration data using co-kriging to generate... Continuous concentration grid, based on the equivalent diameter distribution of particles. The grids with a high proportion of fine and coarse particles in each concentration grid are marked to distinguish the types of pollution sources; among them, The number of rows and columns in the concentration grid indicates the number of elements generated. OK Concentration grid of columns;

[0119] The wind field data is spatially interpolated using the inverse distance weighting method to generate a wind field grid with the same resolution as the concentration grid. Each wind field grid includes the wind direction angle and wind speed.

[0120] Based on the spatial distribution rate of the concentration grid, the wind field grid is spatially aligned with the concentration grid to ensure that each concentration grid corresponds to a unique wind field data.

[0121] The proportion of fine and coarse particles in each grid cell was extracted. Based on the typical proportions of fine and coarse particles in livestock and poultry sheds (30%-40% fine particles, 60%-70% coarse particles), the pollution control grading requirements for different particle sizes were set as follows: areas with a fine particle proportion exceeding 45% were designated as key control areas, and areas exceeding 35% were designated as secondary key control areas. A first threshold was established. 35%, second threshold The percentage is 65%, and the particle dominance type is determined for each grid cell; fine particles are particles with an aerodynamic diameter of less than or equal to 2.5 micrometers, and coarse particles are particles with an aerodynamic diameter of greater than 2.5 micrometers and less than or equal to 10 micrometers.

[0122] For fine particulate matter, if the proportion of fine particles exceeds the first threshold If the condition is met, then fine particulate matter is determined to be dominant; otherwise, it is determined not to be dominant.

[0123] For coarse particulate matter, if the proportion of coarse particles exceeds the second threshold. If the coarse particles are dominant, then the coarse particles are determined to be dominant; otherwise, they are determined not to be dominant.

[0124] For each mesh cell, if only one dominant type exists, it is determined that there is a particle-dominant type, including fine-particle dominant and coarse-particle dominant, and it is marked as the target mesh;

[0125] If there is no dominant type or two dominant types exist, it is determined that there is no particle dominant type and it is marked as a non-target mesh;

[0126] For each target grid, neighboring grids in the top, bottom, left, right, and four diagonal directions are selected and their concentration values ​​are extracted. The concentration difference between the target grid and each neighboring grid is calculated, and the actual spatial distance is calculated using the grid's latitude and longitude. The ratio of the concentration difference to the actual spatial distance is calculated to obtain the concentration gradient amplitude, which reflects the rate of concentration change. Based on the sign of the concentration difference, the gradient direction is determined from the low-concentration grid to the high-concentration grid.

[0127] Since the source is usually located upwind of the pollution spread direction, the opposite direction of the gradient is selected, and the wind direction of the wind field grid is used for collaborative analysis to calculate the angle between the opposite direction of the gradient and the direction of the meteorological wind field. ,like If the wind field and gradient height are aligned, then the first collaborative weight is set to... ;like If the wind field and gradient part are aligned, then the second collaborative weight is set to... ;like If the wind field and gradient direction conflict, then a third collaborative weight is set. ;

[0128] Based on the gradient inverse direction, and combined with the meteorological wind field direction, a weighted correction is performed to obtain the cooperative pointing vector, where the magnitude of the cooperative pointing vector represents the pointing intensity. The weighted correction process includes: denoting the unit vector in the gradient inverse direction as... The unit vector of the meteorological wind field direction is denoted as Based on the angle between the gradient's opposite direction and the meteorological wind field direction. Select the corresponding collaborative weights ( , , ),pass Calculate the cooperative pointing vector ;

[0129] The area covered by the wind field grid is divided into candidate source cells with the same resolution as the concentration grid. For each candidate source cell, the number of vectors and the total intensity of all cooperative pointing vectors pointing to the candidate source cell are counted. The ratio of the number of vectors pointing to the candidate source cell to the total number of vectors is calculated, and the ratio of the total intensity of vectors pointing to the candidate source cell to the total intensity of all vectors is calculated. The convergence degree is obtained by multiplying the number ratio and the intensity ratio.

[0130] Based on the accuracy requirements for tracing aerosol pollution sources in livestock and poultry breeding houses, the source location error should be less than 1.5 meters and the actual distribution characteristics of the airflow field inside the house should be dominated by longitudinal ventilation and local vortex disturbance. The convergence threshold is set at 0.15. Candidate source units with convergence exceeding the convergence threshold are retained to form a high-probability source area in the initial screening, and the dominant pollution source type of each area is marked.

[0131] Combining the diffusion characteristics of particles of different sizes, such as coarse particles having a small diffusion coefficient and a short range of influence, the theoretical distance from the high-probability source area in the initial screening to the high-concentration grid is calculated based on the diffusion scale jointly formed by the product of wind speed and sampling time, and the product of diffusion coefficient and the square root of sampling time. If the actual distance from the high-probability source area in the initial screening to the high-concentration grid exceeds the theoretical distance, the corresponding high-probability source area in the initial screening is removed. Among them, the high-concentration grid is the grid where the particle concentration value exceeds the sum of the mean and standard deviation of the particle concentration in the high-probability source area in the initial screening.

[0132] When there is spatial overlap between the high-probability source regions dominated by coarse particles and those dominated by fine particles in the initial screening, the overlapping part is selected as the core source region. The coarse particle source probability is multiplied by the coarse particle source probability with the coarse particle proportion as the weight, and the fine particle source probability is multiplied by the fine particle source probability with the fine particle proportion as the weight. The two products are added together to obtain the comprehensive source probability. The coarse particle source probability and the fine particle source probability are obtained by normalizing the aggregation degree of particles of the corresponding particle size.

[0133] Based on the comprehensive source probability, the core source area is combined with geographic information to generate a visualized initial source area estimation map, where geographic information includes, but is not limited to, a basic map of road and building distribution.

[0134] Specifically, the steps for constructing a constrained reverse trajectory tracing model include:

[0135] Based on the experimental setting of constraints in the semi-enclosed environment of livestock and poultry breeding houses, a constrained reverse trajectory tracing model is constructed, including an input layer, a core calculation layer, a constraint correction layer and an output layer. The initial source area estimation map and the particulate matter basic parameter set are used as input layer data.

[0136] The core computing layer is equipped with a particulate matter tracking engine. It takes the detection point as the initial position, calculates the product of the inverse wind speed vector and the time step, and adds the product to the initial position to obtain the basic spatiotemporal coordinates of the trajectory.

[0137] The constraint correction layer adopts a multi-constraint priority processing mechanism. Taking the trajectory base spatiotemporal coordinates as the initial value, the constraint conditions are converted into precipitation removal function, radiation attenuation function, temperature and humidity settlement function, and wind speed guidance function respectively. Combined with the constraint priority controller, the correction processing is carried out in the order of precipitation removal, radiation attenuation, temperature and humidity settlement and wind speed guidance to obtain the corrected trajectory base spatiotemporal coordinates.

[0138] The output layer transforms the basic spatiotemporal coordinates of the corrected trajectory into multi-dimensional results, including but not limited to the source region probability distribution and the source region contribution.

[0139] By calling upon historical particulate matter source tracing cases, a training set is constructed, including input layer data of the constrained reverse trajectory source tracing model, real source region coordinates, and real source region contribution labels. With the goal of minimizing the errors between the multi-dimensional results and the real source region coordinates and real source region contribution, key parameters of the constraint correction layer, such as the radiation attenuation coefficient, are calibrated based on the training set. The optimization effect is evaluated by source region positioning error and relative contribution error, and finally, the optimized constrained reverse trajectory source tracing model is obtained.

[0140] For example, the initial source area estimation map covering the entire area of ​​the breeding house and a 10-meter radius around it, along with a set of basic particulate matter parameters with a particle size of 2.5 micrometers, an initial concentration of 0.35 milligrams per cubic meter, and a diffusion coefficient of 0.002 square meters per second, are used as input layer data.

[0141] The core computing layer is equipped with a particulate matter tracking engine. Three detection points within the livestock shed are used as initial positions, with coordinates (10m, 10m, 1.5m), (20m, 10m, 1.5m), and (40m, 10m, 1.5m). The reverse wind speed vector is the measured average reverse wind speed of 0.6 m / s, pointing from the detection points towards the source within the livestock shed. A time step of 10 seconds is selected. The product of the reverse wind speed vector and the time step is calculated to be 0.6 × 10 = 6 meters. This product is then added to the initial position to obtain the basic spatiotemporal coordinates of the trajectory. For example, with (20m, 10m, 1.5m) as the initial point, the basic spatiotemporal coordinates of the trajectory at the first time step are (14m, 10m, 1.5m).

[0142] The constraint correction layer employs a multi-constraint priority processing mechanism. Using the trajectory's basic spatiotemporal coordinates as initial values, the constraints are converted into: a precipitation removal function (0.2 mm of measured precipitation per day with a removal coefficient of 0.05); a radiation attenuation function (550 watts per square meter of measured radiation intensity with an initial radiation attenuation coefficient of 0.00012 watts per square meter of initial radiation attenuation); a temperature and humidity settling function (22 degrees Celsius and 65% humidity in the livestock shed with a settling coefficient of 0.003 per degree Celsius per percentage of relative humidity); and a wind speed guidance function (0.7 meters per second of measured real-time wind speed with a guidance coefficient of 0.8). Combined with the constraint priority controller, corrections are performed sequentially according to the order of precipitation removal, radiation attenuation, temperature and humidity settling, and wind speed guidance. In this embodiment... The coordinates represent the lateral position of the livestock shed. Pollutants diffuse radially in a straight line from the detection point to the source, with no significant lateral offset. Therefore, only [the following is observed]... coordinates and Coordinate constraints are corrected. The coordinates remain unchanged; for example, the basic spatiotemporal coordinates of the trajectory at the first time step (14 meters, 10 meters, 1.5 meters) are corrected to obtain the corrected basic spatiotemporal coordinates of the trajectory (9.9 meters, 10 meters, 1.44 meters); among them, for The coordinates were corrected as follows: after precipitation removal, the value is 14 × (1 - 0.05) = 13.3 meters; after radiation attenuation, the value is 13.3 × (1 - 0.00012 × 550) ≈ 12.42 meters; after wind speed guidance, the value is 12.42 × 0.8 ≈ 9.9 meters; and after temperature and humidity settling... After coordinate correction, we get 1.5 × (1 - 0.003 × 22 × 65%) ≈ 1.44 meters;

[0143] The output layer transforms the basic spatiotemporal coordinates of the corrected trajectory into multi-dimensional results, including the probability distribution of source regions with a core source region probability of over 85% and a secondary source region probability of 30%-85%, and the source region contribution of the core source region of 72% and the secondary source region contribution of 28%.

[0144] A training set was constructed using 100 historical particulate matter source tracing cases, including input layer data of the constrained reverse trajectory source tracing model, real source region coordinates, and real source region contribution labels. With the goal of minimizing the errors between the multi-dimensional results and the real source region coordinates and contributions, key parameters of the constraint correction layer were calibrated based on the training set. For example, the radiation attenuation coefficient was calibrated to 0.00015 per watt-square meter. The optimization effect was evaluated using the source region positioning error (average error before calibration: 1.2 meters; average error after calibration: 0.35 meters), and the contribution relative error (average error before calibration: 8.3%; average error after calibration: 2.1%). The optimized constrained reverse trajectory source tracing model was finally obtained.

[0145] Specifically, the steps for deducing particulate matter propagation paths include:

[0146] Based on the optimized constrained reverse trajectory tracing model, the reverse extrapolation period is set according to the atmospheric residence time of particulate matter, and the model traces back from the sampling time, with the detection point as the trajectory endpoint, and the initial height is set in combination with particle size and particulate matter density.

[0147] The reverse deduction period is divided into fixed intervals. Within a time period, a trajectory cluster is formed within the adjacent range of the initial altitude, and a first-level weight is assigned to each trajectory; wherein, based on the particulate matter atmospheric diffusion timescale of 1 hour and the sampling resolution of 10 minutes, a fixed interval is determined. It lasts for 10 minutes;

[0148] The step size is dynamically adjusted according to wind speed. Based on the measured wind speed range of 0.3-1.2 m / s in the semi-enclosed environment of livestock and poultry sheds, the wind speed threshold is set at 0.7 m / s. When the wind speed is lower than the wind speed threshold, the step size is increased to avoid calculation redundancy. When the wind speed exceeds the wind speed threshold, the step size is decreased to ensure the accuracy of the trajectory nodes. The particles are moved in the opposite direction of the wind speed, and the inverse spatiotemporal coordinates of each trajectory are recorded to form an initial trajectory cluster.

[0149] Based on the precipitation removal function, statistics For each time period, the precipitation threshold is set at 0.2 mm based on the precipitation characteristics of daily precipitation below 0.5 mm, and the heavy precipitation threshold is set at 0.3 mm based on the regional short-term precipitation extreme value. If the precipitation exceeds the moderate precipitation threshold but does not exceed the heavy precipitation threshold, it is classified as a weak clearing area; if the precipitation exceeds the heavy precipitation threshold, it is classified as a strong clearing area.

[0150] Based on the set of basic particulate matter parameters, particulate matter is classified into different types: particulate matter containing sulfate-soluble components is classified as strongly hygroscopic particles; particulate matter that is not easily soluble in water is classified as weakly hygroscopic particles; and particulate matter containing pollen, bacteria, or fungal spores that has biological activity is classified as biological particles.

[0151] Strongly hygroscopic particles, due to their tendency to adsorb water vapor and form large particles, have a moderate removal rate in weak removal areas and a high removal rate in strong removal areas; weakly hygroscopic particles, due to their difficulty in binding with water vapor, have a low removal rate in weak removal areas and a moderate removal rate in strong removal areas; biological particles, due to their cell walls being easily destroyed and deactivated by rainwater, have a moderate removal rate in weak removal areas and a near-complete removal rate in strong removal areas; after removing trajectories affected by precipitation from the initial trajectory cluster, the remaining trajectories constitute the first-level corrected trajectory cluster;

[0152] Based on the first-level corrected trajectory cluster, and combined with the preset division rules for solar radiation intensity and duration, strong inactivation zones and weak inactivation zones are divided. The preset division rules are set based on the light monitoring data of livestock and poultry breeding houses. Areas with solar radiation intensity exceeding 700 lux and duration exceeding 30 minutes are classified as strong inactivation zones; areas with solar radiation intensity below 700 lux or duration below 30 minutes are classified as weak inactivation zones.

[0153] Based on the radiation resistance performance of particulate matter, the radiation resistance types are classified as follows: particulate matter with stable inorganic mineral particle structure and resistance to radiation damage is classified as high radiation resistance particles; particulate matter with secondary organic aerosol components that are easily decomposed by radiation is classified as medium radiation resistance particles; and particulate matter with structures such as bacteria, viruses, and fungal spores that are easily destroyed and inactivated by radiation is classified as low radiation resistance particles.

[0154] Based on the radiation attenuation function, and combining the radiation resistance type of each trajectory with the radiation intensity and duration characteristics of the strong and weak inactivation zones it passes through, the radiation attenuation coefficient is matched from a preset lookup table, and the total radiation attenuation coefficient of the trajectory is accumulated and calculated. The total radiation attenuation coefficient is multiplied by the first-level weight to obtain the second-level weight. If the second-level weight is lower than the preset weight threshold, the corresponding trajectory is removed; otherwise, the trajectory is retained and the total radiation attenuation coefficient is recorded to obtain the second-level corrected trajectory cluster. The preset lookup table is constructed using radiation attenuation test data of different radiation resistance types of particulate matter in livestock and poultry breeding houses. The test uses an actual light intensity gradient of 300-1000 lux in the house and different radiation durations of 10-60 minutes to measure the radiation attenuation coefficient of each type of particulate matter. The data are then compiled and summarized to form the preset lookup table.

[0155] Specifically, the steps for deriving the propagation path of particulate matter from the source area to the detection point also include:

[0156] Based on the secondary corrected trajectory cluster, and using the temperature and humidity settling function, combined with the physical law that high temperature and high humidity result in low air density and real-time temperature and humidity, the air density of each node in each trajectory is calculated using a preset correlation formula; where the preset correlation formula is the air density... , Atmospheric pressure, The molar mass of dry air, This is the universal gas constant. Absolute temperature It is the water vapor pressure;

[0157] For weakly hygroscopic particles, the physical settling velocity is calculated using Stokes' formula by combining air density, particle size, and particle density; for strongly hygroscopic particles, the particle size expansion ratio is determined based on the correlation between humidity and particle size expansion, and the overall settling velocity is calculated by combining the physical settling velocity.

[0158] Because particulate matter is reduced in height due to gravity settling as it travels from the source area to the detection point, the height of the trajectory nodes needs to be corrected upwards. The backtracking time of each trajectory is calculated, and the current trajectory node height is extracted from the secondary corrected trajectory cluster. For weakly hygroscopic particles, the product of physical settling velocity and backtracking time is calculated, and the product is added to the current trajectory node height to obtain the corrected trajectory node height. For strongly hygroscopic particles, the product of comprehensive settling velocity and backtracking time is calculated, and the product is added to the current trajectory node height to obtain the corrected trajectory node height.

[0159] The atmospheric boundary layer range obtained by lidar is used, and the secondary weights are adjusted according to the height of the corrected trajectory nodes. If the height of the corrected trajectory nodes is within the atmospheric boundary layer range, it indicates that the trajectory conforms to the actual atmospheric laws of particulate matter, and the secondary weight of the trajectory remains unchanged; otherwise, it indicates that the trajectory deviates from the actual atmospheric laws of particulate matter, the secondary weight of the trajectory is reduced to obtain the tertiary weight, and trajectories below the preset weight threshold are removed to generate a tertiary corrected trajectory cluster. Among them, by statistically analyzing the secondary weight distribution of all trajectories in the secondary corrected trajectory cluster, the 25th percentile of the weight distribution is calculated to determine the preset weight threshold as 0.2.

[0160] Based on the wind speed steering function, for each node of each trajectory, the actual wind speed of the modified trajectory's basic spatiotemporal coordinates is extracted. The theoretical wind speed is obtained by calculating the ratio of the node's movement distance to time. The directional deviation is calculated using the absolute difference between the actual wind direction and the theoretical wind direction. The magnitude deviation is obtained by dividing the absolute difference between the actual wind speed and the theoretical wind speed by the actual wind speed.

[0161] A dynamic threshold is set based on the height of the corrected trajectory nodes in the three-level corrected trajectory cluster. If the height of a corrected trajectory node is lower than the height threshold, it is determined to be in the near-ground layer, and the first directional deviation threshold is set to... The first size deviation threshold is Otherwise, it is determined to be an upper-level layer, and the second directional deviation threshold is set to... The second size deviation threshold is ; mark abnormal nodes whose directional deviation exceeds the corresponding first directional deviation threshold or second directional deviation threshold, and mark abnormal nodes whose magnitude deviation exceeds the corresponding first magnitude deviation threshold or second magnitude deviation threshold, and count the percentage of abnormal nodes for each trajectory; wherein, the height threshold is determined to be 20 meters based on 20% of the atmospheric boundary layer height obtained by the lidar;

[0162] Based on the statistical distribution of the proportion of abnormal nodes in all trajectories in the three-level corrected trajectory cluster, the 75th percentile of the statistical distribution is calculated, and the proportion threshold is set to 0.2. If the proportion of abnormal nodes is lower than the proportion threshold, the trajectory direction is corrected; otherwise, the trajectory corresponding to the abnormal node is removed. Finally, the four-level corrected trajectory cluster is obtained.

[0163] Based on the trajectory direction of the four-level corrected trajectory cluster, an initial high-weight source region is set according to the spatial distribution characteristics of the initial source region estimation map. The angle between the trajectory direction and the initial high-weight source region is calculated. If the angle exceeds a preset angle threshold, the trajectory weight is increased, for example, by setting an increase coefficient based on expert experience and calculating the product of the trajectory weight and the increase coefficient to increase the trajectory weight. Otherwise, the trajectory weight is decreased, for example, by setting a decrease coefficient based on expert experience and calculating the product of the trajectory weight and the decrease coefficient to decrease the trajectory weight. This yields the four-level weight.

[0164] Four levels of weights are defined as weighting factors. Density clustering is performed on the inverse trajectory nodes. Based on the average distance between the detection point and the potential source region being 5-50 kilometers and the spatial resolution of the trajectory nodes being 200-1000 meters, a baseline distance of 500-2000 meters is set. Nodes with a distance less than the baseline distance are classified into the same category. The total weight of each category is calculated, and the nodes are filtered in descending order before... The class node is defined as the central source region; in this embodiment, the propagation of particulate matter from the detection point to the source region is defined as the reverse trajectory;

[0165] The estimated maps of the central source region and the initial source region are superimposed. The spatial overlap area is calculated by spatial intersection operation, and the total area is calculated by spatial union operation. The overlap rate is obtained by dividing the spatial overlap area by the total area. Based on the source region identification accuracy requirement of sub-kilometer level spatial accuracy and the statistical value of the overlap rate of historical source data of 0.5~0.7, the overlap rate threshold is set to 0.6. When the overlap rate is lower than the overlap rate threshold, the estimated map of the initial source region is corrected using the central source region to obtain a high-resolution source region heat map.

[0166] Based on high-resolution source region heatmaps, the total weight of each type of node in each trajectory is traversed to filter the pre-selection parameters. The node trajectory represents the propagation path.

[0167] Specifically, the steps for generating a high-resolution spatial distribution location map include:

[0168] Based on the region boundary of the high-resolution source region heat map, the nodes of each propagation path are divided into source region nodes and non-source region nodes. Specifically, through spatial location matching, if the node coordinates of the propagation path are within the region boundary, it is determined to be a source region node; otherwise, it is determined to be a non-source region node.

[0169] The residence time of source region nodes and non-source region nodes along each propagation path is statistically analyzed. The initial source region intensity is obtained from a high-resolution source region heatmap. The product of the residence time and initial source region intensity for each source region node along each propagation path is calculated, and the sum of these products for all source region nodes is obtained to get the total weight of the source region nodes. Considering the distance attenuation characteristic of propagation paths—that particulate matter diffusion decreases with increasing propagation distance—a propagation attenuation coefficient of 0.85–0.95 is set. The product of the residence time, initial source region intensity, and propagation attenuation coefficient for each non-source region node along each propagation path is calculated, and the sum of these products for all non-source region nodes is obtained to get the total weight of the non-source region nodes. The total weight of the source region nodes is then added to the total weight of the non-source region nodes to obtain the initial weight of the path.

[0170] Based on the diffusion characteristics of particulate matter, which are diffused with airflow, the transmission coefficient is set to 0.78~0.92. The product of the initial path weight, the initial source region intensity, and the transmission coefficient is calculated to obtain the predicted concentration of the propagation path. The actual concentration of the propagation path is measured by an online sensor, and the absolute difference between the predicted concentration and the actual concentration is calculated to obtain the concentration error.

[0171] The source node deviation is obtained by calculating the average Euclidean distance between the source node and the center of the high-resolution source heat map region for each propagation path; the theoretical propagation path endpoint is derived using a constrained reverse trajectory tracing model, and the actual propagation path endpoint is measured using GNSS positioning. The Euclidean distance between the theoretical and actual propagation path endpoints is calculated to obtain the endpoint deviation; based on the particulate matter tracing accuracy requirement of ±5 meters and the influence weight ratio of source node deviation and endpoint deviation of 45:55, the source node deviation coefficient is set to 0.45 and the endpoint deviation coefficient is set to 0.55. The position error is calculated by multiplying the source node deviation by the source node deviation coefficient and the endpoint deviation by the endpoint deviation coefficient.

[0172] Based on the requirement of ±0.02 μg / m³ for online particulate matter detection and the actual background concentration of 0.03 μg / m³, a lower limit of concentration of 0.05 μg / m³ is set. For propagation paths where the actual concentration exceeds the lower limit, the reliability of concentration error is improved. Based on the requirement of stable meteorological data, where the coefficient of variation of wind speed fluctuation is less than 0.2, an amplitude threshold of 0.8 m / s is set. Based on meteorological data, the standard deviation is calculated through a sliding time window to obtain the wind speed fluctuation amplitude. For propagation paths where the wind speed fluctuation amplitude is less than the amplitude threshold, the reliability of location error is improved. An error reliability matrix is ​​generated by combining the reliability of concentration error, and the concentration error weight coefficient and location error weight coefficient of each propagation path are recorded.

[0173] The concentration loss is obtained by calculating the product of the concentration error of each propagation path, the initial weight of the path, and the concentration error weight coefficient; the position error loss is obtained by calculating the difference between the base weight coefficient and the initial weight of the path, multiplying the difference by the position error and the position error weight coefficient; in this embodiment, the base weight coefficient is set to 1.

[0174] A dynamic balance rule is constructed: if the average concentration error along the propagation path exceeds the average location error, the concentration balance coefficient is set to 0.6 and the location balance coefficient is set to 0.4; otherwise, the concentration balance coefficient is set to 0.4 and the location balance coefficient is set to 0.6.

[0175] Construct a joint loss function, multiply the concentration loss by the concentration balance coefficient to obtain the concentration loss term, calculate the product of the position error loss and the position balance coefficient to obtain the position loss term, and add the concentration loss term and the position loss term to obtain the joint loss.

[0176] The gradient descent method is used to calculate the partial derivative of the joint loss function with respect to the initial weights of the path. If the partial derivative is positive, it is determined that the current weights cause an increase in the joint loss, and the path weights are reduced. If the partial derivative is negative, it is determined that the current weights cause a decrease in the joint loss, and the path weights are increased to ensure that the path weights are non-negative. Finally, the optimized path weights are obtained.

[0177] A correlation table between propagation paths and source regions is constructed. For high-resolution source region heat maps, based on the particulate matter source tracing accuracy requirement of ±5 meters and the normal distribution of concentration error characteristics, a maximum threshold of 0.3 micrograms per cubic meter and a minimum threshold of 0.1 micrograms per cubic meter are set. If the concentration error of the propagation path corresponding to the grid in the correlation table exceeds the maximum threshold, the grid is considered to have a low matching degree with the propagation path, and the grid intensity is reduced. If the concentration error of the propagation path corresponding to the grid in the correlation table is lower than the minimum threshold, the grid is considered to have a high matching degree with the propagation path, and the grid intensity is increased. Finally, the optimized source region intensity is obtained.

[0178] By combining the transmission coefficient, the optimized path weight, optimized source area intensity, and the product of the transmission coefficient are calculated to obtain the optimized predicted concentration. Based on the online sensor measurement accuracy of ±0.02 micrograms per cubic meter and the actual concentration confidence level of 95%, the actual concentration coefficient is set to 0.55. Based on the goodness of fit of the optimized predicted concentration of 0.93 and the accuracy of the constrained reverse trajectory tracing model of ±4.8 meters, the optimized predicted concentration coefficient is set to 0.45. The corrected detection value is calculated by combining the actual concentration × actual concentration coefficient and the optimized predicted concentration × optimized predicted concentration coefficient, thus generating the online particulate matter detection result.

[0179] Based on optimizing the source region intensity and considering the diffusion law that concentration decreases with increasing source region distance, the following calculations are performed. The optimized concentration values ​​of the grid are calculated, and the rate of change of source region intensity between adjacent grids is obtained to obtain the spatial gradient distribution of the optimized source region intensity. The optimized concentration values ​​are then introduced into the high-resolution source region heatmap to generate a 3D coordinate map. The propagation paths corresponding to the optimized path weights and the spatial gradient distribution of the optimized source region intensity are added to generate a high-resolution spatial distribution location map. To optimize the number of rows and columns in the concentration value grid, representing OK Optimized concentration value grid for columns.

[0180] In summary, this invention aims to address the problem that traditional methods, lacking the integration of meteorological data, struggle to accurately trace the source of airborne particulate matter in complex environments. It utilizes a multi-point lensless diffraction device to acquire aerosol diffraction images, combining them with multi-source meteorological data to construct a meteorological spatiotemporal alignment set of diffraction spectra. The energy distribution features of the image spectrum are extracted using Fast Fourier Transform, and a radial averaging algorithm is used to generate diffraction ring features, which are then stitched together and input into a depth inversion model. Humidity, wind speed, and temperature from the meteorological data are incorporated, and calibration weights are corrected using preset thresholds. By constructing particle size and concentration constraint rules, the equivalent diameter distribution and mass concentration of particulate matter are inverted, generating a basic set of particulate matter parameters. Based on the collaborative analysis of concentration gradient and meteorological wind field direction, co-kriging is used for spatial interpolation, combined with wind field inverse distance weighted interpolation. The consistency between the gradient inverse direction and wind direction is used to generate an initial source region estimation map. A constrained inverse trajectory source tracing model is further constructed, integrating the Lagrange inverse trajectory method with a multi-constraint correction mechanism. This includes a precipitation clearing function to shield areas of heavy precipitation, a radiation attenuation function to correct trajectory weights based on particulate matter radiation resistance type, a temperature and humidity settling function to calculate settling velocity based on Stokes' theorem and correct trajectory height upwards, and a wind speed steering function to dynamically adjust the trajectory through directional and magnitude deviations. A dual error feedback mechanism based on concentration and location is employed. The deviation between predicted concentration and endpoint location is calculated based on actual concentration and GNSS positioning data. A joint loss function is constructed, and the initial path weights and source area intensity are optimized inversely using gradient descent. A high-resolution spatial distribution positioning map is iteratively generated, achieving the integration of lensless optical detection, meteorological coupling inversion, and dynamic source tracing positioning, thus improving the online monitoring and accurate positioning capabilities of airborne particulate matter in complex environments.

[0181] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A lensless online detection and distribution localization method for airborne particulate matter, characterized in that, include: Aerosol diffraction images and meteorological data were collected, preprocessed, and aligned with a unified timestamp to construct a meteorological spatiotemporal alignment set of diffraction spectra. Extract the spectral energy distribution features and ring features into the deep inversion model, introduce humidity, wind speed and temperature from meteorological data, and combine preset thresholds to correct weights to invert the equivalent diameter distribution and concentration of particulate matter, and output the basic parameter set of particulate matter; based on the concentration gradient and meteorological wind field direction, co-infer the high probability source region and generate an initial source region estimation map. Based on the initial source region estimation map and particulate matter basic parameter set, constraints are introduced to construct a constrained reverse trajectory tracing model, and the Lagrange reverse trajectory method is used to deduce the particulate matter propagation path. A dual error feedback mechanism of concentration and location is adopted to correct the propagation path and location distribution. A joint loss function is constructed to adjust the path weight and source area intensity distribution in reverse. Iterative optimization is used to generate online detection results and high-resolution spatial distribution location map of particulate matter. By introducing constraints, a constrained reverse trajectory tracing model is constructed, including an input layer, a core computation layer, a constraint correction layer, and an output layer. The input layer receives the initial source region estimation map and the particulate matter basic parameter set; The core computing layer uses the detection point as the initial position and combines the reverse wind speed vector and time step to calculate the basic spatiotemporal coordinates of the trajectory. Based on the trajectory-based spatiotemporal coordinates, the constraint correction layer converts the constraints into precipitation clearing functions, radiation attenuation functions, temperature and humidity settling functions, and wind speed steering functions, and then corrects them sequentially in conjunction with the constraint priority controller to obtain the corrected trajectory-based spatiotemporal coordinates. The output layer transforms the basic spatiotemporal coordinates of the corrected trajectory into a multi-dimensional result; By calling historical particulate matter source tracing cases, a training set is constructed. With the goal of minimizing the error between the multi-dimensional results and the actual source area coordinates and contribution, the key parameters of the constraint correction layer are calibrated to obtain an optimized constraint-type reverse trajectory source tracing model. The steps for deriving particulate matter propagation paths include: Set up a reverse simulation period, trace back from the sampling time, take the detection point as the endpoint, set the initial height in combination with particle size and particle density, and set a first-level weight for each trajectory; Particulate matter moving in the opposite direction of wind speed forms an initial trajectory cluster; Based on the precipitation removal function, precipitation is statistically analyzed, and a moderate precipitation threshold and a heavy precipitation threshold are set to divide the area into a strong removal zone and a weak removal zone. Based on the particulate matter basic parameter set, the particles are divided into highly hygroscopic particles, weakly hygroscopic particles, and biological particles. The removal rates in the strong removal zone and the weak removal zone are matched, and the trajectories affected by precipitation are removed from the initial trajectory cluster to form a first-level corrected trajectory cluster. Based on the radiation attenuation function, the solar radiation intensity and duration are divided into a strong inactivation zone and a weak inactivation zone. Radiation resistance types are classified based on the radiation resistance performance of particulate matter, radiation attenuation coefficients are matched from a preset lookup table, and secondary weights are calculated based on the primary weights. If the secondary weight is lower than the preset weight threshold, the trajectory is removed; otherwise, the trajectory is retained to form a secondary corrected trajectory cluster. Based on the aforementioned temperature and humidity sedimentation function, the sedimentation velocity is calculated in combination with particle size and particle density. Extract the current trajectory node height, and calculate and correct the trajectory node height based on the settlement velocity and backtracking time; The atmospheric boundary layer range is invoked, and the secondary weights are adjusted according to the height of the corrected trajectory nodes to obtain the tertiary weights; trajectories below the preset weight threshold are removed to generate a tertiary corrected trajectory cluster; Based on the wind speed steering function, calculate the deviation of node direction and magnitude, dynamically set the deviation threshold, and count the proportion of abnormal nodes. If the proportion of abnormal nodes is lower than the preset proportion threshold, the trajectory direction is corrected; otherwise, the corresponding trajectory is removed, resulting in a four-level corrected trajectory cluster. Set an initial high-weight source region, calculate the angle with the trajectory direction and adjust the third-level weight to obtain a fourth-level weight, cluster the node density of the reverse trajectory, calculate the total weight of each type of node, filter the central source region in descending order and calculate the overlap rate with the estimated map of the initial source region. When the overlap rate is lower than the preset overlap rate threshold, a high-resolution source region heatmap is obtained after correction, before screening. The node trajectory represents the propagation path.

2. The method for online detection and distribution localization of airborne particulate matter without lenses based on diffraction according to claim 1, characterized in that, The specific steps for retrieving the equivalent diameter distribution and concentration of particulate matter include: The spectral energy distribution characteristics are obtained, and spectral feature vectors are generated using the local maximum method. Extract the diffraction ring intensity curve, combine it with Hough circle transform to generate a ring-shaped feature vector, concatenate it with the spectral feature vector and combine it with linear transformation, and calculate the weighted enhanced image feature vector. Meteorological data is normalized into meteorological feature vectors, which are used as query matrices and the enhanced image feature vectors to calculate calibration weights through a cross-attention mechanism. The calibration weights are corrected by incorporating humidity, wind speed, and temperature from meteorological conditions and combining humidity thresholds, wind speed thresholds, and temperature thresholds. The corrected calibration weights are multiplied by the enhanced image feature vector to obtain the meteorological corrected image features, which are then combined with the original meteorological features to generate a deep fusion feature vector. The deep fusion feature vector is input into the deep inversion model, and through feature dimensionality reduction and nonlinear transformation, the preliminary inversion results of particulate matter equivalent diameter distribution (PSD) and particulate matter mass concentration (MC) are output.

3. The method for online detection and distribution localization of airborne particulate matter without lenses based on diffraction according to claim 2, characterized in that, The specific steps for retrieving the equivalent diameter distribution and concentration of particulate matter also include: Set the particle size threshold range and mass concentration threshold for the equivalent diameter distribution; Construct particle size constraint rules. If the peak particle size of PSD exceeds the particle size threshold range, it is determined that the particle size constraint conditions are not met and the particle size is marked as abnormal; otherwise, it is determined that the particle size constraint conditions are met and the particle size is marked as normal. Concentration constraint rules are constructed. If the concentration limit (MC) is greater than the mass concentration threshold, the concentration constraint rules are not met and the concentration is marked as abnormal; otherwise, the concentration constraint rules are met and the concentration is marked as normal. If both the particle size constraint rule and the concentration constraint rule are satisfied, the preliminary inversion result is deemed reasonable; otherwise, it is deemed unreasonable. All detection points are verified by time series and spatial grid inversion, and the effective inversion results are integrated to generate a set of basic particulate matter parameters.

4. The method for online detection and distribution localization of airborne particulate matter without lenses based on diffraction according to claim 3, characterized in that, The specific steps for generating the initial source region estimation map include: Based on the particulate matter basic parameter set, a concentration grid is generated by spatial interpolation of particulate matter concentration data, and a wind field grid is generated by spatial interpolation of wind field data, and the grids are aligned with the concentration grids in space. Extract the proportion of fine and coarse particles in each grid cell, and set a first threshold and a second threshold; For fine particulate matter, if the proportion of fine particulate matter exceeds the first threshold, it is determined that fine particulate matter has a dominant position; otherwise, it is determined that it does not have a dominant position. For coarse particulate matter, if the proportion of coarse particles exceeds the second threshold, then coarse particulate matter is determined to be dominant; otherwise, it is determined not to be dominant. If each grid cell has only one dominant type, then it is determined that there is a particle dominance type, including fine particle dominance and coarse particle dominance, and it is marked as the target grid; Otherwise, it is determined that there is no particle-dominant type and it is marked as a non-target mesh; Calculate the concentration difference between the target grid and each neighboring grid and the actual spatial distance. Divide the concentration difference by the actual spatial distance to obtain the concentration gradient magnitude. Determine the gradient direction based on the sign of the concentration difference.

5. The method for online detection and distribution localization of airborne particulate matter without lenses based on diffraction according to claim 4, characterized in that, The specific steps for generating the initial source region estimation map also include: Select the opposite direction of the gradient and calculate the included angle based on the wind field and wind direction. ; like If the wind field and gradient direction are consistent, then a cooperative weight is set. ; like If the wind field and gradient part are aligned, then a cooperative weight is set. ; like If the wind field and gradient direction conflict, then a cooperative weight is set. ; By combining the wind field direction weighted correction to obtain the cooperative pointing vector, candidate source units are divided, the number and intensity of cooperative pointing vectors are counted, and the convergence degree is obtained by calculating the product of the number ratio and the intensity ratio. Set a convergence threshold and retain candidate source units that exceed the convergence threshold to form a high-probability source region in the initial screening. Calculate the theoretical distance for each particulate matter from the high-probability source region of the initial screening to the high-concentration grid, and remove regions that exceed the theoretical distance; When the high-probability source regions of the initial screening, dominated by coarse particles and dominated by fine particles, spatially overlap, the overlapping part is taken as the core source region and the comprehensive source probability is calculated. Combined with geographic information, an initial source region estimation map is generated.

6. The method for online detection and distribution localization of airborne particulate matter using lensless diffraction according to claim 5, characterized in that, The specific steps for generating a high-resolution spatial distribution map include: Based on the boundary of the high-resolution source region heat map, the propagation path nodes are divided into source region nodes and non-source region nodes; Statistical analysis of dwell time and acquisition of initial source region intensity; calculation of total weight of source region nodes and non-source region nodes; summation to obtain initial path weight. Set the transmission coefficient, calculate the predicted concentration based on the initial weight of the path, and calculate the concentration error in combination with the actual concentration; The source region node deviation is calculated based on the average Euclidean distance between the source region node and the center of the high-resolution source region heat map area, and the endpoint deviation is calculated based on the Euclidean distance between the theoretical propagation path endpoint and the actual propagation path endpoint. The position error is then calculated using a weighted average. The error confidence level is dynamically adjusted by combining the actual concentration and wind speed fluctuation amplitude, and an error confidence level matrix is ​​generated.

7. The method for online detection and distribution localization of airborne particulate matter without lenses based on diffraction according to claim 6, characterized in that, The specific steps for generating a high-resolution spatial distribution map also include: Based on the error confidence matrix, a joint loss function is constructed to calculate the joint loss, and the initial path weights are optimized using the gradient descent method to obtain the optimized path weights. The source region intensity is adjusted according to the concentration error to obtain the optimized source region intensity. The optimized predicted concentration is calculated in combination with the transmission coefficient. The weighted calculation is then used to correct the detection value and generate the online particulate matter detection results. Based on the optimized source region intensity, the optimized concentration value and spatial gradient distribution are calculated in conjunction with the diffusion law; The optimized concentration value is introduced into the high-resolution source region heat map to generate a three-dimensional coordinate map. The propagation path corresponding to the optimized path weight and the spatial gradient distribution are added to generate a high-resolution spatial distribution location map.

Images