A laser radar networking monitoring method based on intelligent collaboration and closed-loop optimization

Abstract

The application provides a laser radar networking monitoring method based on intelligent cooperation and closed-loop optimization, and relates to the technical field of data processing.The method comprises the following steps: constructing a deep neural network, taking a laser radar equation as a physical constraint, combining consistency of multiple radar overlapping areas, inverting extinction and backscattering coefficients and quantifying confidence; using uncertainty dynamic optimization scanning parameters to encrypt scanning in a high-uncertainty area and pre-scanning a high-information-gain area; embedding self-supervised calibration into inversion loss, based on physical residual to reverse repair abnormal points; fusing inversion, calibration and uncertainty confidence to construct an adaptive interpolation kernel and a space-time physical constraint, and reconstructing a four-dimensional data field; and fusing comprehensive confidence to map to transparency to construct an interactive four-dimensional visualization platform.The application realizes closed-loop self-optimization of inversion, calibration, scanning, interpolation, visualization and tracing, improves inversion precision and data reliability, and reduces operation and maintenance cost and energy consumption.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, specifically to a lidar network monitoring method based on intelligent collaboration and closed-loop optimization. Background Technology

[0002] Networked lidar monitoring technology has significant application value in atmospheric environmental monitoring, pollution source tracing and early warning. By coordinating observations with multiple lidar units, it is possible to obtain the three-dimensional distribution of aerosol extinction coefficient and backscattering coefficient over a large area with high spatiotemporal resolution, providing crucial data support for atmospheric pollution process analysis, pollution source location, and transport flux calculation.

[0003] Currently, lidar data inversion mainly employs traditional methods such as the Klett method and the Fernald method. These methods require pre-setting an empirical functional relationship (e.g., the aerosol wavelength index) between the extinction coefficient and the backscattering coefficient, and rely on boundary values ​​at a reference distance. However, the composition, particle size distribution, and mixing state of aerosols in the actual atmosphere vary drastically with space and time, making it difficult to accurately describe the real atmosphere with a fixed functional relationship, thus limiting the accuracy of the inversion. Furthermore, the selection of boundary values ​​depends on manually identifying the linear interval of the signal, which is highly subjective and prone to introducing integration errors. In multi-radar network scenarios, each radar typically performs inversion independently, followed by data fusion through simple interpolation or weighted averaging. Due to hardware differences and aging drift, different radars have system constant deviations, and the observation results in overlapping areas are often inconsistent. Traditional methods cannot utilize the physical consistency constraints of overlapping areas for self-calibration during the inversion process, resulting in low reliability of the fusion results. Moreover, existing inversion methods only output a definite value of the extinction coefficient, without providing any quantitative indicators of the reliability of the results, making it impossible for users to determine whether high-value areas are due to actual pollution or inversion uncertainties.

[0004] Regarding scanning strategies, existing lidar networks mostly employ fixed preset scanning modes (such as fixed angular resolution, period, and sector), which cannot be dynamically adjusted according to atmospheric conditions and uncertainties in inversion results. A few systems use event-driven scanning, adjusting the strategy only after detecting a pollution plume, essentially a passive response, and fail to utilize gradient information during the inversion process to predict information gain, making it difficult to conduct intensive observations of key areas in advance. In terms of data quality assurance, system calibration relies on external calibration equipment (such as standard scatterers carried by UAVs) for periodic execution, which is complex and cannot compensate for equipment drift in real time. Abnormal data is typically removed using statistical thresholds or repaired using autoencoders; the former may lose valid information, while the latter requires extensive training with historical normal data, and the repair results lack constraints from physical equations. Regarding 3D data reconstruction, traditional interpolation methods determine weights based solely on spatial distance, without considering prior information such as physical equation residuals and inversion confidence. The fixed shape of the interpolation kernel cannot adapt to spatial variations in data quality. The fusion weights of the digital twin background field are also fixed values, leading to large extrapolation errors in sparse data areas, while measured details are excessively smoothed in dense data areas. In terms of visualization, existing platforms primarily use two-dimensional static images, lacking four-dimensional spatiotemporal dynamic interactive functions, and fail to map confidence information to rendering parameters, making it impossible for users to intuitively distinguish between high-confidence areas and low-confidence blind spots. Regarding pollution source tracing, traditional methods rely on post-event offline analysis (such as forward diffusion simulation or reverse trajectory calculation), with response times reaching several hours, making it difficult to support emergency decision-making; the source tracing process does not integrate multi-source confidence information, resulting in significant errors in areas with unreliable wind fields or poor data quality. In terms of energy consumption management, each radar operates at full capacity 24 / 7, without dynamically adjusting its operating mode according to pollution risk and source tracing task status, leading to severe energy waste.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] To address the aforementioned problems in existing technologies, a LiDAR network monitoring method based on intelligent collaboration and closed-loop optimization is provided, aiming to solve at least one of the above problems.

[0007] The technical solution to achieve the purpose of this invention is as follows:

[0008] This invention provides a lidar network monitoring method based on intelligent collaboration and closed-loop optimization, which includes the following steps:

[0009] Step S1: By constructing a deep neural network, using the lidar equation as a physical constraint, and taking advantage of the observation consistency of multiple radars in the spatially overlapping area, the neural network parameters and the constants of each radar system are jointly optimized to achieve the inversion of extinction coefficient and backscattering coefficient without the need for preset functional relationships; and the Monte Carlo method is used to quantify the uncertainty of the inversion results and output three-dimensional grid data with inversion confidence.

[0010] Step S2: Using the inversion uncertainty field output from Step S1, calculate the comprehensive uncertainty index and uncertainty gradient of each spatial grid point; dynamically optimize the scanning parameters of each radar according to the uncertainty distribution, and use non-uniform scanning line density to perform intensified scanning of high uncertainty areas; continuously reduce global uncertainty through closed-loop iteration, and at the same time use neural network gradient information to predict information gain, pre-scan high information gain areas in advance, and feed the newly collected data back to Step S1 to reduce inversion uncertainty;

[0011] Step S3: Utilizing the observation consistency of the overlapping areas of multiple radars, the self-supervised calibration loss is embedded into the inversion loss function of step S1 to simultaneously optimize the radar system constants; at the same time, based on the physical residuals of the differential form of the lidar equation, the outliers are repaired through inverse optimization without pre-training, so that the repair results satisfy the physical equation constraints; a two-way feedback closed loop of calibration and repair is established, outputting highly consistent data with physical constraint confidence, and feeding it back to step S1 for dynamic adjustment of the inversion weights;

[0012] Step S4: Integrate the inversion confidence from Step S1, the physical constraint confidence from Step S3, and the uncertainty gradient information from Step S2 to construct interpolation weights negatively correlated with the residuals of the physical equations; design an anisotropic interpolation kernel that adaptively adjusts its bandwidth according to the uncertainty gradient; combine the digital twin background field and dynamically adjust the background field fusion weights based on the local observation data density and confidence; introduce physical transport constraints in the time dimension to achieve the reconstruction of a four-dimensional spatiotemporal data field that satisfies the laws of atmospheric evolution, and output continuous grid data with interpolation confidence.

[0013] Step S5: The inversion confidence from Step S1, the physical constraint confidence from Step S3, and the interpolation confidence from Step S4 are merged into a comprehensive confidence, which is then mapped to the transparency parameter of the 3D volume rendering. An interactive 4D visualization platform supporting time sliding windows, multi-section linkage, and confidence layer overlay is constructed. The pollution clump boundary, transmission path, and transmission flux features are automatically extracted and dynamically labeled. User attention to low-confidence areas is fed back to Steps S2 and S3 to drive encrypted scanning and calibration repair.

[0014] Step S6: Using the three-dimensional wind field and concentration field reconstructed in Step S4 and the comprehensive confidence level in Step S5, a confidence-weighted Lagrange inverse particle tracking method is used to generate the source region probability distribution field; the probability distribution is corrected by fusing the uncertainty high gradient region information from Step S2, the physical residual abnormal region information from Step S3, and the user attention information from Step S5; the radar is guided to perform a directional verification scan of the most likely source region, and the verification results are fed back to Steps S1 and S3 for inversion calibration and system constant update; at the same time, the radar operating mode is dynamically adjusted based on pollution risk prediction to form a closed-loop self-optimizing system for source tracing, scanning, and energy consumption.

[0015] Compared with the prior art, the beneficial effects of the present invention include:

[0016] (1) This invention uses the integral and differential forms of the lidar equation as dual physical constraints, and directly learns the extinction coefficient and backscattering coefficient as spatial functions through a deep neural network. It does not require any preset empirical function relationship, thus avoiding the inversion error caused by the mismatch of function assumptions in traditional methods. At the same time, it uses the consistency constraint of the overlapping area of ​​multiple lidars to realize the self-calibration of the system constant, so that the inversion results remain physically consistent between different observations.

[0017] (2) This invention utilizes multi-radar overlapping observation for self-supervised calibration, eliminating the need for manual calibration using external equipment such as standard scatterers carried by UAVs, which significantly reduces system operation and maintenance costs; at the same time, it can compensate for system deviations caused by equipment aging and drift in real time.

[0018] (3) The present invention uses the Monte Carlo method to quantify the uncertainty of the inversion results and outputs them in the form of confidence level in sync with the inversion results, so that users can clearly judge the reliability of the data; the confidence level information can be directly used for weight adjustment in subsequent modules such as three-dimensional interpolation and transmission throughput calculation, so as to avoid unreliable data from misleading decision-making.

[0019] (4) This invention uses the inversion uncertainty field as the driving signal for active scanning. By Bayesian optimization and non-uniform scan line density, the scanning resources are precisely configured in the region with the least information. The information gain is predicted by the neural network gradient to achieve forward-looking pre-scanning, so that the scanning strategy changes from passive response to active exploration.

[0020] (5) This invention embeds the self-supervised calibration loss into the inversion loss function, and uses the generation and repair method guided by the physical equation gradient to repair outliers without pre-training, and the repair results are forced to meet the constraints of the lidar equation; calibration and repair form a positive feedback closed loop;

[0021] (6) This invention uses confidence-weighted inverse particle tracking to generate the probability distribution field of the source region, integrates multi-source information for correction, and dynamically adjusts the radar working mode (energy saving / conventional / enhanced) based on pollution risk prediction. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a LiDAR network monitoring method based on intelligent collaboration and closed-loop optimization. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0025] Therefore, the following detailed description of embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely illustrates some embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that, unless otherwise specified, the embodiments and features and technical solutions in the present invention can be combined with each other.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] The present invention will be further described in detail below with reference to embodiments.

[0029] like Figure 1 As shown, this invention provides a lidar network monitoring method based on intelligent collaboration and closed-loop optimization, which includes the following steps:

[0030] Step S1: By constructing a deep neural network, using the lidar equation as a physical constraint, and taking advantage of the observation consistency of multiple radars in the spatially overlapping area, the neural network parameters and the constants of each radar system are jointly optimized to achieve the inversion of extinction coefficient and backscattering coefficient without the need for preset functional relationships; and the Monte Carlo method is used to quantify the uncertainty of the inversion results and output three-dimensional grid data with inversion confidence.

[0031] Step S2: Using the inversion uncertainty field output in Step S1, calculate the comprehensive uncertainty index and uncertainty gradient of each spatial grid point; dynamically optimize the scanning parameters (angular resolution, scanning period, sector width) of each radar according to the uncertainty distribution, and use non-uniform scanning line density to perform densified scanning of high uncertainty areas; continuously reduce global uncertainty through closed-loop iteration, and at the same time use neural network gradient information to predict information gain and pre-scan high information gain areas in advance.

[0032] Step S3: Utilize the observation consistency of the overlapping areas of multiple radars to embed the self-supervised calibration loss into the inversion loss function of step S1, and simultaneously optimize the radar system constants; at the same time, based on the physical residuals of the differential form of the lidar equation, perform inverse optimization to generate repairs for outliers without pre-training, so that the repair results satisfy the constraints of the physical equation; and establish a two-way feedback closed loop of calibration and repair to output highly consistent data with physical constraint confidence.

[0033] Step S4: Integrate the inversion confidence from Step S1, the physical constraint confidence from Step S3, and the uncertainty gradient information from Step S2 to construct interpolation weights negatively correlated with the residuals of the physical equations; design an anisotropic interpolation kernel with bandwidth adaptively adjusted according to the uncertainty gradient; combine the digital twin background field and dynamically adjust the background field fusion weights based on the local observation data density and confidence; and introduce physical transport constraints in the time dimension to achieve the reconstruction of a four-dimensional spatiotemporal data field that satisfies the laws of atmospheric evolution, outputting continuous grid data with interpolation confidence.

[0034] Step S5: The inversion confidence from Step S1, the physical constraint confidence from Step S3, and the interpolation confidence from Step S4 are merged into a comprehensive confidence, which is then mapped to the transparency parameter of the 3D volume rendering. An interactive 4D visualization platform supporting time sliding windows, multi-section linkage, and confidence layer overlay is constructed. Features such as contamination clump boundaries, transmission paths, and transmission flux are automatically extracted and dynamically labeled. User attention to low-confidence areas is fed back to Steps S2 and S3 to drive encrypted scanning and calibration repair.

[0035] Step S6: Using the three-dimensional wind field and concentration field reconstructed in Step S4 and the comprehensive confidence level in Step S5, a confidence-weighted Lagrange inverse particle tracking method is used to generate the source region probability distribution field; the probability distribution is corrected by fusing the uncertainty high gradient region information from Step S2, the physical residual abnormal region information from Step S3, and the user attention information from Step S5; the radar is guided to perform a directional verification scan of the most likely source region, and the verification results are fed back to Steps S1 and S3 for inversion calibration; at the same time, the radar operating mode (energy-saving mode, normal mode, enhanced mode) is dynamically adjusted based on pollution risk prediction to form a closed-loop self-optimizing system for source tracing, scanning, and energy consumption.

[0036] It should be noted that this invention constructs an intelligent collaborative monitoring ecosystem with "multi-dimensional confidence" as the pervasive information carrier, "LiDAR physical equations" as the underlying constraint, and "closed-loop feedback" as the driving mechanism, coupling six major functional modules—inversion, scanning, calibration, interpolation, visualization, traceability, and energy consumption management—into an organic whole.

[0037] Physically constrained deep inversion abandons the assumption of pre-defined functional relationships, constructs a deep neural network with the lidar equation (integral and differential forms) as the physical constraint, learns the extinction coefficient and backscattering coefficient as spatial functions, jointly optimizes the neural network parameters and constants of each lidar system, and uses the Monte Carlo method to quantify the inversion uncertainty, outputting three-dimensional grid data with inversion confidence.

[0038] Uncertainty-guided active adaptive scanning uses the inverted uncertainty field as the driving signal, dynamically adjusts each radar scanning parameter through Bayesian optimization and non-uniform scan line density, uses neural network gradient prediction information gain to achieve forward-looking pre-scanning, and feeds the new data back to the inversion module to form a closed-loop iteration of "inversion → scanning → inversion", which transforms the scanning strategy from passive response to active exploration.

[0039] The two-way closed loop of self-supervised calibration and physical-guided repair embeds the self-supervised calibration loss into the inversion loss function and uses the consistency of the overlapping areas of multiple radars to synchronously optimize the radar system constants; based on the physical residuals of the differential form of the lidar equation, anomaly point repair values ​​that satisfy physical constraints are generated through inverse optimization; a positive feedback loop of calibration and repair is established to output highly consistent data with physical constraint confidence.

[0040] The adaptive spatiotemporal interpolation of confidence propagation integrates inversion confidence, physical constraint confidence, and uncertainty gradient information, constructs interpolation weights negatively correlated with the physical equation residuals, designs an anisotropic interpolation kernel with bandwidth adaptively adjusted according to uncertainty, introduces confidence-driven fusion and time-dimensional physical transport constraints from the digital twin background field, reconstructs a four-dimensional spatiotemporal data field, and outputs the interpolation confidence.

[0041] Confidence-driven transparent visualization and user feedback integrates inversion confidence, physical constraint confidence, and interpolation confidence into a comprehensive confidence, which is then mapped to the transparency parameters of 3D volume rendering. This constructs an interactive 4D visualization platform that supports time-sliding, multi-section linkage, and confidence layer overlay. User attention to low-confidence areas is fed back to the scanning and calibration modules, enabling human-machine collaboration.

[0042] The source tracing and energy consumption closed loop of multi-source confidence fusion uses confidence-weighted inverse particle tracking to generate the source region probability distribution field, and integrates information from uncertain high gradient regions, physical residual abnormal regions, and user-concerned information for correction, guiding the radar to perform directional verification scanning and feeding back to the inversion and calibration module; based on pollution risk prediction, the radar operating mode (energy saving / normal / enhanced) is dynamically adjusted to form a closed-loop self-optimizing system for source tracing, scanning, and energy consumption.

[0043] Furthermore, in some embodiments of the present invention, in step S1, lidar inversion is performed without pre-setting the functional relationship between the extinction coefficient and the backscattering coefficient. Simultaneously, multi-radar collaborative inversion and confidence quantification of the inversion results are achieved, which includes the following steps:

[0044] Step S11: Construct a deep neural network as the inversion model. The input to the neural network is spatial coordinates, and the output is the extinction coefficient and backscattering coefficient at those spatial coordinates. Let the spatial point coordinates be... Let (x, y, z) be a three-dimensional vector, where x, y, and z are the coordinates of the point in the east-west, north-south, and vertical directions, respectively, all in meters (m). Neural Network The output is defined as:

[0045]

[0046] in, For spatial points The extinction coefficient of the atmosphere, expressed in meters (m²). -1 ), which represents the proportion of laser light that attenuates due to absorption and scattering per unit distance traveled in the atmosphere;

[0047] For spatial points The backscattering coefficient of the atmosphere, expressed in meters per steradian (m). -1 ·sr -1 This indicates the ability of the atmospheric medium to reflect incident laser energy back to the radar receiver.

[0048] θ represents the parameters to be trained in the neural network, including the connection weights and bias terms between neurons in each layer, which are automatically determined by an optimization algorithm.

[0049] The neural network employs a fully connected structure, comprising an input layer, multiple hidden layers, and an output layer. The input layer contains three neurons, corresponding to the three components of spatial coordinates; the hidden layers use the hyperbolic tangent function (tanh) as the activation function, enabling the network to fit nonlinear mappings; the output layer contains two neurons, outputting the extinction coefficient and backscattering coefficient, respectively. No activation function is set for the output layer to ensure that the output value can be positive or negative, but its physical validity is constrained by a loss function.

[0050] It should be noted that spatial coordinates refer to the position coordinates in a three-dimensional Cartesian coordinate system established with the center of the monitoring area as the origin. This coordinate system is converted to the geographic coordinate system using BeiDou positioning information to ensure that the observation data of each radar is mapped to the same spatial coordinate system. Neural network parameters refer to trainable variables optimized through the backpropagation algorithm. Their initial values ​​are initialized using the Xavier initialization method to keep the variance of the gradients at each layer stable.

[0051] Step S12 involves using the lidar equation as a physical constraint to construct data fitting loss terms and physical equation loss terms, ensuring that the neural network's inversion results conform to the physical laws governing laser propagation in the atmosphere; this includes the following steps:

[0052] Step S121, the power of the echo signal received by the lidar The parameters to be inverted satisfy the following lidar equation:

[0053]

[0054] Where the subscript i represents the i-th radar, i=1,2,…,N radar N radar This represents the total number of radars in the network.

[0055] The detection range is expressed in meters (m), representing the slant distance from the radar transmitter to a scattering object in the atmosphere.

[0056] For radar i at range The power of the backscattered signal received at point C, measured in watts (W), was obtained from actual measurements by the radar receiver; i is the system constant of radar i, with the unit being watts multiplied by meters cubed (W·m³), which includes hardware parameters such as transmit power, receive area, and optical efficiency. This constant can be determined through laboratory calibration or the self-calibration mechanism of this invention.

[0057] Indicates the distance from radar i along the detection direction. The spatial coordinates of the location, in meters (m), are calculated using the radar's BeiDou position coordinates, scanning azimuth angle, and elevation angle.

[0058] For spatial points The backscattering coefficient at a given location is expressed in units of steradian per meter (m). -1 ·sr -1 ), output by a neural network;

[0059] To the distance from radar The optical thickness at a given point is dimensionless and represents the total degree of laser attenuation due to extinction along the path, where the integral variable is... From 0 to The path length.

[0060] The data fitting loss is defined as the mean square error of all radar detection points:

[0061]

[0062] in, For data fitting loss; For radar i at a distance r j The measured echo signal power at the location is expressed in watts (W), which is the raw observation data of the lidar. This refers to the number of detection points for a single radar. This represents the square norm, which is the sum of squared errors.

[0063] Step S122: After performing range square correction on the lidar equation, take the logarithm to obtain the logarithmic range correction signal:

[0064]

[0065] Where ln represents the natural logarithm, Dimensionless, it is the logarithmic power after eliminating the geometric diffusion effect.

[0066] right Differentiating with respect to distance r, we get:

[0067]

[0068] in, The logarithmic correction signal is a dimensionless rate of change with distance and can be calculated by numerical difference of measured data or automatic differentiation by neural network.

[0069] This represents the rate of change of the backscattering coefficient with distance, expressed in meters per steradian (m). -2 ·sr -1 ), obtained through automatic differentiation via neural network.

[0070] This equation is the differential form of the lidar equation, and is equivalent to the original integral form. The physical equation loss forces the neural network to satisfy its equivalence relation:

[0071]

[0072] in, This represents the loss in the physical equations.

[0073] It should be noted that automatic differentiation refers to the derivative calculation method built into neural network frameworks (such as TensorFlow and PyTorch). It accurately calculates the derivative of the output with respect to the input through backpropagation of the computation graph, avoiding the discretization errors of traditional numerical differential algorithms. Integration is implemented using numerical integration methods. During neural network training, the trapezoidal rule or Simpson's rule is used to perform integration calculations along discrete points on the detection path, with the integration step size consistent with the range resolution of the lidar.

[0074] Furthermore, in some embodiments of the present invention, in step S13, the observation consistency of multiple lidars in the spatially overlapping area is used to construct a multi-radar consistency loss, so that the inversion results of different radars for the same spatial point are consistent.

[0075] For any two radars i and j, define their spatial overlap region Ω. ij :

[0076]

[0077] in, Represents a three-dimensional real space, representing the entire monitoring area; This indicates that the detection rays of radar i and radar j pass through the spatial point simultaneously. There is a detection range R i and R j All are real numbers, and the unit is meters; R represents the distance traveled from radar i along the detection direction. i The coordinates of the next reached point in space; R represents the distance traveled from radar j along the detection direction. j The coordinates of the space point reached later; j represents the j-th radar, which is a natural number.

[0078] In the overlapping region Ω ij Within the same physical point r, the inversion results of radar i and radar j should satisfy:

[0079] ,

[0080] in, This indicates that the spatial points are obtained by inverting the observation data of radar i. Atmospheric extinction coefficient, This indicates that the spatial point is obtained by inverting the observation data of radar j. Atmospheric extinction coefficient, This indicates that the spatial points are obtained by inverting the observation data of radar i. Atmospheric backscattering coefficient, This indicates that the spatial point is obtained by inverting the observation data of radar j. Atmospheric backscattering coefficient. Multi-radar consistency loss is defined as:

[0081]

[0082] in, Let ε be the set of all radar pairs with overlapping regions, where ε = {(i,j) | i < j, Ω}. ij ≠∅}; This indicates that all radar pairs with overlapping regions are traversed. This represents the summation of all spatially discrete points within the overlapping region. In actual calculations, a grid sampling method is used, generating sampling points within the overlapping region at a fixed resolution (e.g., 50m × 50m × 50m). When the inversion results for the same spatial point from different radars are inconsistent, this loss function will produce a positive value; when the inversion results tend to be consistent, the loss function tends to zero. By minimizing this loss, the inversion results of each radar are forced to remain consistent within the overlapping region.

[0083] It should be noted that the overlapping area is determined through geometric calculations. Based on the BeiDou system coordinates, scanning range, and detection range of each radar, the intersection or near-intersection of the detection rays from different radars in space is calculated. When the shortest distance between two rays is less than a preset threshold (e.g., 10 meters), the spatial point is considered to belong to the overlapping area. A physical point refers to a fixed location in space whose coordinates do not change with the change of the observing radar; it is a physical property of the atmosphere itself.

[0084] Step S14: The data fitting loss, physical equation loss, and multi-radar consistency constraint loss function values ​​are weighted and summed to obtain the total loss function:

[0085]

[0086] in, λ1 is the total loss function; λ2 is the weighting coefficient of the physical equation loss term, dimensionless, ranging from 0.1 to 10.0, used to balance data fitting and physical consistency; λ3 is the weighting coefficient of the multi-radar consistency loss term, dimensionless, ranging from 0.1 to 10.0, used to balance single-radar inversion accuracy and multi-radar consistency; λ1 and λ2 are determined by cross-validation, selecting the parameter combination that minimizes the inversion error of the validation set.

[0087] The backpropagation algorithm and Adam optimizer are used to iteratively optimize the neural network parameters θ to achieve the desired total loss function. Minimize. During the optimization process, the constants C of each radar system are simultaneously minimized. i It can be used as an optimizable variable in training to achieve self-calibration of system constants.

[0088] Step S15: After the neural network training is complete, it can be used for inversion calculation. However, the inversion result obtained at this time is a single numerical value, which cannot reflect its reliability. The Monte Carlo method is used to quantify the uncertainty of the inversion result. The principle of this method is: during the neural network testing phase, the random deactivation mechanism used during training is kept active, that is, some neurons are randomly ignored during each forward propagation. Due to the randomness of random deactivation, multiple forward propagations will yield multiple slightly different inversion results.

[0089] Statistical analysis is performed on the inversion results, calculating their mean and standard deviation. The mean is used as the final inversion result output, while the standard deviation reflects the uncertainty of the inversion result. When the standard deviation is less than a threshold, it indicates that multiple inversion results are consistent, and the neural network has high confidence in the inversion of that point; when the standard deviation is greater than the threshold, it indicates that the inversion result is sensitive to the network structure and has high uncertainty.

[0090] The confidence level is defined based on the ratio of the standard deviation to the mean. This confidence level ranges from 0 to 1. A ratio close to 0 indicates a confidence level close to 1, signifying highly reliable inversion results; a large ratio indicates a confidence level close to 0, signifying unreliable inversion results. The confidence level directly reflects the reliability of the inversion results.

[0091] In step S16, the average values ​​of the calculated extinction coefficient and backscattering coefficient are used as the final inversion result and saved as 3D grid data along with the confidence level. The 3D grid covers the entire monitoring area, and each grid point stores the inversion result and corresponding confidence level at that location. This data is synchronously stored in a database for subsequent processing modules to access. Subsequent modules such as 3D interpolation, transmission flux calculation, and visualization analysis can assign different weights to data in different areas based on the confidence level information. This allows high-confidence data to play a greater role in subsequent processing, while automatically reducing the weight of low-confidence data to avoid unreliable data misleading decision-making.

[0092] It should be noted that the constraints of the lidar equations are directly derived from the fundamental physical laws governing laser propagation in the atmosphere. This anchors the neural network's inversion results to these physical laws, avoiding non-physical results that might arise from purely data-driven methods. Without this foundation, subsequent multi-radar coordination and uncertainty quantification would lack a physical basis.

[0093] It should be noted that the multi-radar consistency constraint, based on physical constraints, further introduces redundant information from space observations. Observations of the same space point by different radars constitute an overdetermined system. This overdeterminism reduces the ill-conditioned nature of the inversion problem and provides the possibility for self-calibration of system constants. More importantly, its constraints depend on the established physical framework—only when the physical equations are used as a common foundation can the inversion results of different radars be comparable.

[0094] It should be noted that the neural network parameters are optimized synchronously with the radar system constants, allowing physical constraints and multi-radar constraints to work synergistically within the same framework. This design achieves two key functions: first, automatically calibrating system biases through consistency constraints; and second, ensuring the physical rationality of the inversion results through physical constraints. These two constraints support each other during the optimization process, forming a positive feedback loop—physical constraints ensure the inversion results conform to established patterns, while multi-radar constraints ensure consistency across different observations. Together, they ensure that the inversion results satisfy both physical properties and maintain consistency.

[0095] It should be noted that uncertainty quantification relies on the neural network model established in the preceding steps. The Monte Carlo method requires the neural network to have a random deactivation mechanism, and the definition of confidence depends on commonly determined physical parameters. Without the physical framework established in the preceding steps, uncertainty quantification will lose its physical meaning.

[0096] It should be noted that outputting the confidence level and inversion results simultaneously allows uncertainty information to be passed to subsequent processing modules. This design enables high-confidence data to play a greater role in subsequent applications, while low-confidence data automatically has its weight reduced, forming a complete closed loop from physical inversion to trusted applications. The realization of this closed loop depends on the accuracy and reliability of the confidence level information provided in the preceding steps.

[0097] It should be further explained that physical constraints provide the foundation, multi-radar collaboration provides enhancement, joint optimization achieves synergy, uncertainty quantification provides extension, and confidence propagation realizes value.

[0098] Regarding functional assumptions, existing methods like Klett and Fernald forcefully presuppose an empirical functional relationship between the extinction coefficient and the backscattering coefficient. These assumptions stem from the statistical characteristics of atmospheric aerosols, but are not universal physical laws. When aerosol types vary spatially, a fixed functional relationship cannot accurately describe the actual atmosphere. This invention outputs the extinction coefficient and backscattering coefficient as two independent spatial functions simultaneously, without presupposing any functional relationship. The relationship is entirely determined by the physical laws of the lidar equations and observational data, achieving a fundamental shift from "assumption-driven" to "physics-driven" approaches.

[0099] Regarding boundary value determination, existing methods require the known extinction coefficient at a reference distance as the boundary value. Traditional methods determine the reference point by manually identifying the linear interval of the signal curve. This process is highly subjective and difficult to accurately select when atmospheric stratification is complex, and the boundary value error can propagate and amplify along the integration path. This invention uses neural networks for automatic learning, eliminating the subjectivity of boundary value selection without human intervention.

[0100] In terms of multi-radar utilization, existing technologies involve independent inversion by each radar, with data fusion only performed post-processing, neglecting the physical constraints of overlapping observation areas. This invention utilizes consistency constraints in overlapping areas during the inversion process, achieving a leap from "post-processing" to "forward coupling," enabling mutual calibration of multi-radar data during the inversion stage. Existing technologies only use multi-radar networking to expand monitoring range or for post-processing data fusion, failing to leverage the physical constraints of overlapping observation areas. This invention constructs multi-radar consistency constraints, forcing different radars to maintain consistent inversion results for the same spatial point. Its design allows for mutual calibration of multi-radar data during the inversion process, eliminating the need for manual calibration using UAVs equipped with standard scatterers, significantly reducing system operation and maintenance costs, and providing real-time compensation for system deviations caused by equipment aging and drift.

[0101] In terms of system calibration, existing technologies rely on external calibration equipment for periodic calibration, which is complex to operate and cannot compensate for equipment drift in real time. This invention utilizes overlapping observation self-calibration, achieving an upgrade from "manual calibration" to "automatic calibration," significantly reducing operation and maintenance costs.

[0102] Regarding uncertainty, existing technologies do not provide any reliability information, and users cannot determine the credibility of the data. This invention quantifies and outputs confidence levels, achieving a progress from "deterministic output" to "credible output".

[0103] Regarding physical consistency, existing technologies only satisfy integral form equations, while this invention satisfies both integral and differential form equations, forming dual physical constraints and enhancing the physical self-consistency of the inversion results.

[0104] It should be further explained that existing technologies only output a single value of the extinction coefficient, which cannot characterize the uncertainty of the inversion results, and users cannot judge the reliability of the data. This invention quantifies the uncertainty of the inversion results using the Monte Carlo method and outputs it simultaneously with the inversion results in the form of a confidence level. The confidence level information can be directly used for weight adjustment in subsequent modules such as 3D interpolation and transmission flux calculation, allowing high-confidence data to play a greater role in subsequent processing, while automatically reducing the weight of low-confidence data, thus avoiding misleading decisions based on unreliable data.

[0105] This invention simultaneously constructs data fitting constraints and physical equation constraints, ensuring that the inversion results not only fit the observed data but also satisfy the differential form of the lidar equations. This dual constraint reduces the ill-conditioned nature of the inversion problem, making the results more physically consistent. Even under low signal-to-noise ratio conditions, the inversion results of this invention maintain stable spatial continuity, while traditional methods exhibit significant oscillations under these conditions.

[0106] This invention uses the system constants of each radar as optimizable variables in neural network training, automatically determining the relative system constants by leveraging the observation consistency of overlapping regions. When the system constant of a particular radar is known, the absolute system constants of all radars can be obtained. Its self-calibration mechanism enables the network system to have long-term automatic maintenance capabilities, eliminating the need for frequent manual calibration.

[0107] Furthermore, in some embodiments of the present invention, in step S2, intelligent linkage of radar scanning strategies is realized based on event-driven adaptive collaborative scanning; this includes the following steps:

[0108] Step S21: The central processing server receives the mean extinction coefficient, mean backscattering coefficient, and their corresponding variances for each spatial grid point output in step S1. The system defines a comprehensive uncertainty index, which is the weighted sum of the relative standard deviations of the extinction coefficient and the backscattering coefficient, with weighting coefficients of 0.6 and 0.4, respectively. This index ranges from 0 to 1; a larger value indicates a less reliable inversion result for that point.

[0109] The system divides the monitoring area into a three-dimensional grid (100 meters by 100 meters horizontally and 50 meters vertically) and calculates the average uncertainty of each grid cell. Simultaneously, it uses the finite difference method to calculate the spatial gradient of the uncertainty field, which is used to identify abrupt uncertainty boundaries, such as the edges of blind zones or strong attenuation regions. When the magnitude of the spatial gradient exceeds a preset threshold, the location is determined to be a high-gradient region of uncertainty.

[0110] Step S22: For each radar, the system assesses the expected reduction in global uncertainty under its current scanning parameters (angular resolution, scan period, sector width). The system establishes response relationships based on historical data statistics: the average reduction in inversion uncertainty within the region when the angular resolution is increased; and the reduction in uncertainty when the scan period is shortened.

[0111] A Bayesian optimization method was employed, with the objective function being the reduction in global uncertainty per unit time, to solve for the optimal combination of scanning parameters for each radar. This objective function comprehensively considers the reduction in uncertainty, scanning time, and radar energy consumption. The search space for Bayesian optimization is: angular resolution from 0.5 degrees to 2.0 degrees, scanning period from 1 minute to 10 minutes, and sector width from 30 degrees to 120 degrees.

[0112] For regions with high gradient uncertainty, the system employs a non-uniform scan line density: the density of scan lines along the direction of the uncertainty gradient is proportional to the projected uncertainty value in that direction. This means that denser scan lines are placed in the direction where uncertainty changes most dramatically, while other directions maintain a normal density. This design allows scanning resources to be precisely focused on the areas with the least information.

[0113] Step S23: After the radar collects new data according to the optimized scanning parameters, step S1 uses the new data to perform neural network inversion again and update the uncertainty field. The system compares the change in global average uncertainty before and after the update. If the change is greater than a preset threshold (e.g., 0.05), it indicates that the scan adjustment is effective, the system maintains the current optimization direction, and continues to execute the Bayesian optimization in step S22; if the change is less than the threshold, it is determined that the current scanning parameters are close to the hardware performance limit, the system reallocates scanning resources to other grid cells with higher average uncertainty, and reduces the scanning intensity of the current area (e.g., relax the angular resolution, extend the scanning cycle).

[0114] The closed-loop iteration frequency is synchronized with the radar scanning cycle, typically performed every 2 to 5 minutes. After 3 to 5 iterations, the average uncertainty of the network coverage area can converge to below 0.15.

[0115] Step S24: Based on the gradient information from the neural network in Step S1, the system can predict the expected reduction in inversion uncertainty, i.e., the information gain, after probing a region that has not yet been scanned or has been sparsely scanned. Specifically, the system calculates the absolute value of the gradient of the total loss function with respect to the observation at the spatial coordinates. This gradient value reflects the degree of influence of the observation data at that location on the inversion result. The predicted information gain is defined as the product of this absolute gradient value and the current uncertainty.

[0116] When the predicted information gain exceeds a preset threshold (e.g., 0.7) and no radar is currently scanning the area, the system marks the area as a high information gain region and sends a pre-scan command to the radar covering the area 10 to 20 minutes in advance, requiring it to conduct at least one encrypted detection of the area in the next scan cycle (angular resolution encrypted to 0.5 degrees, scan cycle shortened to 1 minute). This pre-scan mechanism does not rely on pollution event detection and is an active exploration behavior, enabling the system to obtain key area data in advance, thereby improving its ability to predict future pollution evolution.

[0117] Step S25: When the difference in inversion uncertainty between the two radars in the spatially overlapping area exceeds 0.3, it is determined that the radar with higher uncertainty has an observation defect in that area. The central server coordinates the two radars to conduct a joint encrypted scan of the overlapping area simultaneously, requiring the two radars to use exactly the same scanning geometry parameters (azimuth range, elevation angle, angular resolution, and scanning period), and the scanning time to be strictly synchronized (time difference less than 0.1 seconds).

[0118] The joint scan data is used simultaneously for online calibration (correcting radar system constants) in step S1 and for fusion updating of the uncertainty field. If the uncertainty difference is still greater than 0.2 after the joint scan, absolute calibration (UAV calibration) in step S3 is triggered. This mechanism ensures that the inversion results of multiple radars in the overlapping area are both physically consistent and have similar reliability.

[0119] Step S26: The system accumulates uncertainty change data after each scan strategy adjustment, constructing a "strategy, effect" database. Each record includes the scan parameters before adjustment, the statistical characteristics of the uncertainty field, and the amount of uncertainty reduction after adjustment. Using this database, a meta-learning model is trained offline. This model is a three-layer fully connected neural network, with the input being the statistical characteristics of the uncertainty field (mean, variance, gradient distribution, etc., a 20-dimensional vector), and the output being the recommended combination of scan parameters.

[0120] After deployment, the meta-model is used to recommend scanning parameters in real time, reducing the decision time for scanning strategies from minutes to seconds in Bayesian optimization. When deploying a new network system, transfer learning can be performed using the existing meta-model: the parameters of the first two network layers are frozen, only the output layer is fine-tuned, and the system is trained for 10 epochs using the initial 24 hours of data from the new system's operation. Transfer learning allows the new system to achieve the scanning performance of a mature system within 24 hours, without needing to optimize from scratch.

[0121] Existing lidar network scanning technologies mainly fall into two categories: one is fixed preset scanning modes, which cannot adapt to environmental changes; the other is event-driven scanning, which only adjusts the strategy after detecting a contamination clump, essentially a passive response, and does not utilize the uncertainty information in the inversion results. By using the inversion uncertainty as the driving signal for active scanning, the approach transforms from "passive response" to "active exploration," and through closed-loop iteration, the scanning strategy and inversion quality mutually reinforce each other.

[0122] Furthermore, in some embodiments of the present invention, in step S3, after completing the physical constraint-based neural network inversion and obtaining the extinction coefficient, backscattering coefficient, and uncertainty of each spatial point in step S1, step S3 utilizes the uncertainty field and physical equation gradient information output in step S1 to construct a self-supervised mutual calibration and data generation and repair method. This upgrades traditional post-calibration and anomaly removal to online self-supervised learning and physically guided generation and repair, deeply integrating the data quality assurance process with the inversion process, forming a positive feedback loop of inversion-calibration-repair. It includes the following steps:

[0123] Step S31 utilizes the observation consistency of multiple radars in spatially overlapping areas to construct a self-supervised calibration mechanism that requires no external calibration equipment. Unlike existing technologies, this step directly embeds the calibration loss into the loss function of step S1, enabling the calibration process to proceed synchronously with neural network inversion, rather than as a post-processing step.

[0124] For any two radars i and j, in the overlapping region Ω ij Within the system, it is mandatory that the calibrated observations conform to the actual physical quantities. The self-supervised calibration loss L is defined. cal This loss measures the difference between the extinction coefficient and backscattering coefficient of different radars after calibration for the same spatial point. cal Further increase the radar system constant C i As a learnable parameter, confidence weighting is introduced into the loss function.

[0125] The system accumulates matching point pairs in overlapping areas using a sliding time window (window length 10 to 30 minutes), and dynamically updates the system constants of each radar using weighted least squares. The weights consist of three parts: the distance of the point to the radar (closer distance, higher weight), the uncertainty of the output in step S1 (lower uncertainty, higher weight), and the observation frequency (more observations, higher weight). The calibrated data is fed back in real-time to the inversion module in step S1 for the next round of neural network training, forming a real-time closed loop of calibration and inversion.

[0126] Step S32: Utilize the physical equation loss gradient of the neural network from step S1 to achieve anomaly detection and physically guided generative repair without pre-training. Traditional methods use autoencoders for anomaly detection, requiring extensive training on historical normal data, and the repair results lack physical constraints.

[0127] The system calculates the differential form residual of the lidar equation at each observation point in real time. This residual reflects the degree of deviation between the observed data and physical laws. When the residual exceeds a dynamic threshold, it is identified as an outlier. The dynamic threshold is adaptively determined based on the statistical distribution of the residuals over the past 24 hours.

[0128] For outliers, the system does not directly remove them, but generates repair values ​​by solving an inverse optimization problem: taking the extinction coefficient and backscattering coefficient output by the neural network in step S1 as initial values, and using the lidar equation and spatial smoothness as constraints, the system minimizes the physical equation residuals and spatial discontinuities.

[0129] The repaired data carries generation labels (including the physical residuals upon which the repair was based, the number of iterations, and the changes before and after the repair), and participates in subsequent processing together with normal data. This generative repair method does not require offline training, can adapt to any unknown atmospheric conditions, and the repair results naturally satisfy the constraints of physical equations.

[0130] Step S33: The system constant output by the self-supervised calibration module is used to correct the original observation data. The corrected data is then input into the repair module. The high-confidence data output by the repair module is fed back to the self-supervised calibration module, increasing the number of effective matching point pairs in the overlapping area. This creates a positive feedback loop: calibration improves data consistency, making physical residual calculations more accurate, thus making the repair more reliable; repair fills data gaps, increasing the available data in the overlapping area, thus making the calibration more stable.

[0131] Step S34: Based on the comprehensive data quality assessment results, generate the physical constraint confidence score for each data point. This confidence score is obtained by a weighted product of three parts: calibration confidence score (determined by the estimated variance of the system constant), physical residual confidence score (obtained by normalizing the inverse of the residual of the differential form of the lidar equation at the observation point), and generation repair confidence score (the highest value is taken for the original data, and a lower value is set for the generated repair data based on the optimization convergence residual).

[0132] The physical constraint confidence level is stored along with the data and is used by the 3D interpolation module in step S4 and the visualization module in step S5. Simultaneously, the confidence level information is fed back to the loss function in step S1 as a weight for each observation point, allowing high-confidence data to play a greater role in the inversion process, while automatically reducing the impact of low-confidence data.

[0133] Step S35: The system monitors the physical residual distribution, system constant drift trend, and repair frequency of each radar in real time. When the physical residual of a radar exceeds twice the historical average for one consecutive hour, the online calibration acceleration update in step S31 is automatically triggered; when the repair frequency exceeds 20%, it is determined that there may be continuous interference or equipment failure in the area, an alarm is generated, and maintenance personnel are advised to check.

[0134] Existing lidar network monitoring technologies suffer from the following shortcomings in data quality assurance: First, dynamic mutual calibration and inversion processes are separated, resulting in calibration lag and an inability to utilize inversion uncertainty information; second, anomaly detection relies on offline-trained autoencoders, requiring a large amount of historical normal data and failing to adapt to unknown atmospheric conditions; third, data repair employs simple interpolation or autoencoder reconstruction, lacking physical equation constraints, and the repair results may violate lidar equations; fourth, calibration and repair are independent of each other, failing to form a collaborative closed loop. This invention embeds self-supervised calibration into the inversion loss function, synchronizing calibration and inversion; utilizes gradient-guided generative repair based on physical equations, eliminating the need for pre-training and ensuring the repair results satisfy physical constraints; and establishes a joint optimization positive feedback loop for calibration and repair, enabling the system to continuously self-enhance during operation.

[0135] It should be noted that this step is deeply coupled with step S1. Step S1 outputs an uncertainty field and the gradient of the physical equation, driving the anomaly detection and repair generation in step S3. The high-confidence data and calibrated system constants output by step S3 are fed back to step S1, reducing inversion uncertainty. The two are mutually causal, jointly constituting a data quality self-healing system under physical constraints. This invention jointly optimizes calibration loss and inversion loss, improving the convergence speed of system constants by approximately 60% compared to traditional post-calibration, and can track equipment aging drift in real time. Repair values ​​are directly generated using the gradient of the lidar equation, eliminating dependence on historical datasets, allowing for immediate operation after deployment, and adapting to any sudden atmospheric conditions. The generated repair values ​​are forced to satisfy the differential form of the lidar equation. Confidence information is fed back to the inversion weights of step S1 and the interpolation weights of step S4, forming a complete and reliable chain from data acquisition to inversion to application.

[0136] Furthermore, in some embodiments of the present invention, in step S4, after extracting the extinction coefficient field and backscattering coefficient field with physical constraint confidence obtained in step S1, the uncertainty-driven adaptive scanning data obtained in step S2, and the consistency data after self-supervised calibration and generation repair obtained in step S3, step S4 utilizes the above multi-source information to construct an adaptive spatiotemporal interpolation method based on the propagation of physical constraint confidence, organically fusing the lidar equation residual, uncertainty field, and digital twin background field to achieve high-fidelity reconstruction from sparse observations to a continuous four-dimensional data field. This includes the following steps:

[0137] Step S41: The system obtains the extinction coefficient, backscattering coefficient, and corresponding physical equation residuals (i.e., the degree of deviation of the differential form of the lidar equation) for each observation point from step S1. From step S3, the system obtains the comprehensive physical constraint confidence score for each point, which integrates the calibration confidence score, physical residual confidence score, and generation and repair confidence score.

[0138] The system defines interpolation weights that are positively correlated with the confidence level of physical constraints and negatively correlated with the residuals of the physical equations. Points with smaller physical equation residuals (i.e., points that better conform to the lidar equations) receive higher interpolation weights, while points with larger physical equation residuals receive lower weights. Simultaneously, the inversion uncertainty output from step S1 is also used as an auxiliary factor: when the uncertainty exceeds a preset threshold, the interpolation weight for that point is further reduced. Unlike traditional interpolation that relies solely on spatial distance, the weights in this step consider both physical consistency and inversion reliability.

[0139] Step S42: Based on the global uncertainty field output in step S2 and the physical constraint confidence level output in step S3, dynamically design the anisotropic interpolation kernel. Traditional interpolation methods use fixed distance decay functions (such as inverse distance weighting or Gaussian kernels), which cannot adapt to spatial variations in data quality.

[0140] For a target grid point to be interpolated, the system searches for a set of observation points in its neighborhood. For each observation point, the spatial distance to the target point is calculated, and an adaptive kernel function is defined. This kernel function includes a distance decay term and a physical confidence weighting term, where the bandwidth of the kernel function is adaptively adjusted according to the local uncertainty gradient at the observation point: in regions with large uncertainty gradients, the bandwidth is reduced to preserve details; in regions with uniform uncertainty, the bandwidth is increased to smooth noise. This adaptive kernel enables the interpolation process to automatically adjust the smoothness according to spatial variations in data quality, avoiding over-extrapolation in high-uncertainty regions.

[0141] Step S43: Introduce the background field (from numerical weather prediction and topographic data) provided by the digital twin model. However, unlike traditional methods, the fusion weights of the background field are not fixed, but are determined by the confidence of local physical constraints and the density of observation data.

[0142] When the observed data density is high and the confidence level is high, the background field weight approaches zero, and the interpolation result is dominated by the measured data. When the observed data is sparse or the confidence level is low, the background field weight increases, and it is filled by the digital twin background field. The final interpolation result is a weighted fusion of the measured data-dominated Kriging interpolation and the background field. This fusion strategy allows the interpolation to retain measured characteristics in the reliable data region and maintain physical plausibility in the blind zone.

[0143] Step S44: The physical equation constraints of Step S1 are extended to the time dimension, introducing a physical constraint for time smoothing. Traditional spatiotemporal interpolation handles the spatial and temporal dimensions independently, which may cause the interpolation results to violate the laws of atmospheric evolution (such as non-physical abrupt changes in the concentration field within a short period of time).

[0144] For adjacent moments in the time series, the system mandates that the rate of change of the interpolation results satisfy the approximate form of the atmospheric transport equation, meaning that concentration changes should be consistent with wind field divergence and source-sink terms. During spatiotemporal interpolation, the system uses this temporal consistency loss as a regularization term, jointly optimizing it with spatial interpolation errors to ensure that the interpolation results satisfy physical transport laws in the time dimension. Specifically, a scheme combining spatiotemporal kriging and physical constraints is adopted. The interpolation results are obtained by solving a system of physically constrained linear equations, ensuring that the interpolation results remain physically self-consistent in both the spatiotemporal dimensions.

[0145] Step S45: After interpolation is completed, the system uses the confidence propagation mechanism from step S1 to calculate the interpolation uncertainty for each interpolation point. This uncertainty consists of three parts: uncertainty propagation of the observation data (weighted propagation based on the uncertainty of the observation points and the interpolation weights), uncertainty of the interpolation kernel (introduced by the estimated variance of the adaptive bandwidth), and uncertainty of the background field (provided by the uncertainty of the digital twin model itself). This uncertainty is saved along with the interpolation results and used as a confidence layer overlay during visualization in step S5, allowing users to clearly understand the reliable and unreliable regions of the interpolation results.

[0146] Step S46: The system compares the interpolation result with the original inversion result from Step S1 and calculates the interpolation residual. When the interpolation residual continuously exceeds a preset threshold in a certain region, it is determined that there is a problem with the data quality or interpolation parameters in that region. The system automatically performs the following feedback operations: feeds back the interpolation residual information of that region to Step S2, triggering a denser scan of that region (increasing the observation density); feeds back the interpolation residual information of that region to Step S3, indicating that there may be undetected abnormal data or calibration bias; adjusts the adaptive bandwidth parameter in Step S42, optimizing it through a sliding window to better match the spatial variability of the current atmospheric state. This feedback mechanism enables the interpolation strategy to continuously self-optimize as the system runs, and the interpolation accuracy continuously improves over long-term operation.

[0147] Existing lidar network monitoring technologies suffer from the following drawbacks in 3D interpolation: First, interpolation weights are based solely on spatial distance, ignoring physical equation residuals and inversion confidence levels; second, the interpolation kernel is fixed, failing to adapt to spatial variations in data quality; third, the fusion weights for the digital twin background field are fixed, unable to be dynamically adjusted based on measured data density; fourth, spatiotemporal interpolation lacks physical constraints, easily leading to non-physical mutations; and fifth, there is no quantification of interpolation uncertainty, making it impossible for users to assess the reliability of the interpolation results. This paper addresses this issue by using physical constraint confidence levels as the core of the interpolation weights, designing an adaptive anisotropic interpolation kernel to achieve confidence-driven fusion of the background field and measured data. It introduces physical transport constraints in the time dimension and establishes a closed-loop optimization mechanism for the propagation and feedback of interpolation uncertainty.

[0148] It should be noted that this step is in deep collaboration with steps S1, S2 and S3. Step S1 provides the physical constraint confidence and uncertainty field, step S2 provides the uncertainty-driven encrypted observation data, step S3 provides consistent calibration data, and step S4 uses the above information to achieve high-fidelity interpolation and feeds back the interpolation residual to the previous step, forming a complete intelligent closed loop from inversion to scanning to interpolation.

[0149] Furthermore, in some embodiments of the present invention, in step S5, after obtaining the extinction coefficient field and backscattering coefficient field with physical constraint confidence in step S1, obtaining the uncertainty-driven adaptive scanning data in step S2, obtaining the consistency data after self-supervised calibration and generation repair in step S3, and obtaining the adaptive spatiotemporal interpolation result based on the propagation of physical constraint confidence in step S4, the multi-dimensional reliability indicators such as inversion confidence, calibration confidence, interpolation confidence, and radar observation conditions are organically integrated to achieve a quantitative expression of the credibility of the monitoring results and an intuitive dynamic presentation of the pollution process, solving the problems of traditional visualization methods ignoring data reliability, lacking continuous spatiotemporal dynamic display, and making it difficult for users to identify credible areas. It includes the following steps:

[0150] Step S51: The system obtains the inversion confidence of each observation point from step S1 (inversion variance obtained based on the Monte Carlo method), the physical constraint confidence of each data point from step S3 (integrating calibration confidence, physical residual confidence and generation repair confidence), and the interpolation confidence of each interpolation grid point from step S4 (weighted by observation data uncertainty, interpolation kernel uncertainty and background field uncertainty).

[0151] The overall confidence level for each spatial grid point is defined as the weighted product of the three confidence levels mentioned above. The weights are dynamically adjusted based on the data source: for original observation points, the inversion confidence level and physical constraint confidence level dominate; for interpolation points, the weight of the interpolation confidence level increases. The overall confidence level ranges from 0 to 1, with a higher value indicating more reliable monitoring results for that point.

[0152] The system propagates the comprehensive confidence score within a four-dimensional spatiotemporal data field: for adjacent frames in a continuous time series, spatiotemporal smoothing filtering is used to maintain the continuity of the confidence score in the time dimension, avoiding drastic jumps in confidence score caused by anomalies in a single frame. The propagated confidence score field is stored together with the extinction coefficient field and wind field for use in visualization rendering.

[0153] Step S52: Directly map the overall confidence level to the transparency parameter of the 3D volume rendering. Traditional visualization methods use the same rendering intensity for all regions, making it impossible for users to intuitively distinguish between reliable and unreliable regions.

[0154] In the 3D visualization scene, the system sets the opacity of each voxel to be positively correlated with the overall confidence level: areas with a confidence level higher than 0.8 are completely opaque and have saturated colors; areas with a confidence level between 0.4 and 0.8 are semi-transparent; and areas with a confidence level lower than 0.4 are highly transparent and have a gray mask overlaid to indicate to the user that the data in that area is unreliable.

[0155] Meanwhile, the system adaptively adjusts rendering details based on the confidence spatial gradient: at boundaries where confidence changes drastically (such as the edge of a detection blind zone), it automatically enhances the edge contour lines to highlight the boundary between credible and untrustworthy areas; in low-value areas with uniform confidence, it uses blurring to prevent users from over-interpreting untrustworthy data.

[0156] Step S53: The system organizes the four-dimensional data field generated in step S4 (including the three-dimensional extinction coefficient field, wind field, and comprehensive confidence field in the time dimension) into an interactive spatiotemporal data cube. The visualization interface provides the following core functions:

[0157] Users can continuously play the entire process of pollution cloud formation, movement, and dissipation using the timeline slider. The playback speed is adjustable (0.5x to 4x), and forward playback, reverse playback, and loop playback are supported. During playback, the comprehensive confidence field is updated synchronously, allowing users to clearly understand the changes in data reliability at different time periods.

[0158] Employing ray projection rendering technology, the extinction coefficient field is rendered as a semi-transparent colored volume, with colors grading from blue (low concentration) to red (high concentration). Simultaneously, three-dimensional streamlines are overlaid; streamline density and color represent wind speed, and streamline direction represents wind direction. Users can freely rotate and zoom the viewpoint to observe the spatial distribution and transport path of pollutants from any angle.

[0159] The system supports arbitrary vertical and horizontal cross-sections within a 3D scene. When a user clicks on a location, it automatically generates a vertical cross-section and a horizontal slice at that point, and plots the concentration change curve for that point in the time-series view. All views are synchronized: as the user drags the time slider, the 3D scene, cross-sections, and graphs update synchronously.

[0160] Users can enable the confidence level layer separately to display the spatial distribution of the overall confidence level in the form of a heatmap, with colors gradation from red (high confidence) to blue (low confidence). The confidence level layer can be semi-transparently overlaid with the pollution concentration layer to visually display high-confidence pollution areas and low-confidence blind spots.

[0161] Step S54: The system calculates key features of pollution transmission in real time based on a four-dimensional data field and dynamically annotates them in a visualization interface.

[0162] An isosurface extraction algorithm is used to automatically identify the 3D boundary contour of contaminant blobs based on the extinction coefficient threshold, and the blobs are dynamically wrapped in a semi-transparent curved surface. The color of the boundary contour is dynamically adjusted according to the internal average confidence level: the boundary of high-confidence regions is a solid green line, and the boundary of low-confidence regions is a dashed red line.

[0163] Based on the wind field and concentration gradient, the trajectory of the pollutant plume's centroid is calculated, and the transmission path is dynamically displayed in a 3D scene using a gold curve with arrows. The width of the trajectory segment is proportional to the average confidence level along the path, allowing users to intuitively judge the reliability of the transmission path.

[0164] At user-specified key cross-sections (such as city boundaries or valley mouths), particulate matter transport flux (the product of concentration and the normal component of wind speed) is calculated in real time and displayed as a superimposed heatmap. The magnitude and direction of the flux value are represented by color and arrows, and a confidence profile is superimposed on the cross-section to indicate the confidence interval of the flux calculation.

[0165] Step S55: The system monitors the response performance of the visual interaction and establishes a feedback optimization mechanism. When the frame rate drops due to the user performing complex operations (such as large-scale view rotation or high-resolution volume rendering), the system automatically adopts the following optimization strategies: dynamically reducing the spatial resolution of distant areas while maintaining details in near areas; preloading data from subsequent time frames into memory to reduce waiting time; and intelligently predicting the areas that the user may focus on in the next frame based on the user's historical operation habits and rendering them in advance.

[0166] Meanwhile, the system records the duration of user attention and interaction frequency for areas with different confidence levels, and feeds this data back to steps S2 and S3: if the user frequently zooms in to view a low-confidence area, it indicates that there is a problem with the actual pollution risk or data quality of the area. The system automatically triggers step S2 to perform an encrypted scan of the area and prompts step S3 to strengthen calibration and repair.

[0167] Existing lidar network monitoring technologies suffer from the following shortcomings in result evaluation and visualization: First, they lack a multi-source confidence fusion mechanism, preventing users from ascertaining the reliability of each stage of inversion, calibration, and interpolation. Second, visualization rendering does not consider confidence levels, resulting in identical visual representations of high-noise and high-confidence areas, potentially misleading decision-making. Third, visualization dimensions are limited, primarily using two-dimensional static images, failing to dynamically display the pollution process. Fourth, they lack transmission feature extraction and dynamic annotation functions, making it difficult for users to quickly understand the source and destination of pollution. Fifth, they lack closed-loop feedback for visualization effectiveness, making it impossible to optimize scanning strategies based on user concerns. This invention integrates the confidence levels from inversion, calibration, and interpolation into a comprehensive confidence level, driving transparency mapping and adaptive rendering. It constructs a four-dimensional spatiotemporal dynamic interactive visualization platform, supporting time sliding, multi-profile linkage, and confidence layer overlay. It automatically extracts pollution boundaries, transmission paths, and flux heatmaps, achieving closed-loop feedback between visualization effectiveness and scanning strategies.

[0168] It should be noted that the integrated inversion, calibration, and interpolation three-source confidence level of this invention allows users to intuitively distinguish between high-confidence and low-confidence regions, reducing the decision-making error rate compared to traditional no-confidence visualization. The semi-transparent or blurred processing of low-confidence regions prevents users from over-interpreting unreliable data, improving their accurate understanding of pollution distribution. The time-sliding window and multi-profile linkage function allow users to continuously observe the evolution of pollution plumes. Functions such as pollution plume boundary tracking, transport path animation, and flux heatmaps enable users to quickly locate pollution sources and transport channels without manual calculations, shortening source tracing time. User attention to low-confidence regions is fed back to steps S2 and S3, driving encrypted scanning and calibration repair, achieving synergistic optimization of visualization and monitoring. Step S1 provides the inversion confidence score, step S3 provides the physical constraint confidence score, step S4 provides the interpolation confidence score, and step S5 integrates the three to generate a comprehensive confidence score. At the same time, the user attention data in step S5 is fed back to steps S2 and S3, forming a complete intelligent closed loop of inversion, calibration, interpolation, visualization, and re-observation.

[0169] Furthermore, in some embodiments of the present invention, in step S6, the real-time reconstructed three-dimensional wind field, concentration field, and confidence information of each stage are organically integrated to achieve rapid and accurate location of pollution sources and dynamic intelligent adjustment of radar energy consumption, and to establish a complete closed-loop feedback control system from source tracing to scanning to energy consumption management. This includes the following steps:

[0170] Step S61: The system obtains the real-time reconstructed three-dimensional wind field and extinction coefficient field from step S4, and obtains the comprehensive confidence level of each spatial grid point from step S5. Starting from the location of the pollutant plume observed at the current moment, the system traces the trajectory of particles in reverse in the three-dimensional wind field to simulate the reverse transport process of pollutants from the observation point to the source region.

[0171] Unlike traditional reverse tracing, this step introduces confidence weighting in particle trajectory calculation: for regions with a comprehensive confidence level higher than 0.7, particle velocity is entirely determined by the reconstructed wind field; for regions with a confidence level between 0.4 and 0.7, a small random perturbation is added to the particle velocity to reflect the uncertainty of the wind field; for regions with a confidence level lower than 0.4, the weight of the random perturbation is increased in the particle velocity to avoid source tracing bias caused by unreliable wind fields. Simultaneously, each time a particle passes through a grid point during reverse tracing, the comprehensive confidence level of that point is accumulated into the confidence integral of the particle trajectory, ultimately resulting in each reverse trajectory carrying a comprehensive confidence score.

[0172] The system uses the Monte Carlo method to generate a large number of inverse particle trajectories (no fewer than 10,000). The initial position and orientation of each trajectory are randomly sampled based on the spatial distribution probability of the contamination plume. The trajectory integration time is dynamically set according to the possible upstream distance of the contamination plume, typically ranging from 30 minutes to 2 hours.

[0173] Step S62: The system performs statistical analysis on the particle landing points reached on the ground through reverse tracking, generating a probability distribution field for the pollution source area. Unlike traditional methods that only count the number of particle landing points, this step uses the comprehensive confidence score of each particle as a weight to perform weighted statistics on the landing points. The weighted probability distribution field is defined as follows: the probability density of each grid cell is equal to the sum of the confidence scores of the particles arriving at that cell divided by the sum of the confidence scores of all particles.

[0174] Simultaneously, the system integrates the information on high-gradient uncertainty regions output in step S2, the information on abnormal physical constraint confidence regions output in step S3, and the information on low-confidence regions of key concern to the user in step S5 to correct the initial probability distribution field. The correction rules are as follows: for the high-gradient uncertainty regions marked in step S2, the probability density is multiplied by an enhancement coefficient greater than 1 (e.g., 1.5), because these regions may contain insufficiently observed sources of contamination; for the regions marked in step S3 with persistently high physical residuals, the probability density is multiplied by a decay coefficient (e.g., 0.7), because the data quality in these regions is poor and the reliability of the source tracing results is low; for the low-confidence regions frequently magnified by the user in step S5, the probability density is multiplied by an enhancement coefficient, reflecting the user's prior concerns.

[0175] The corrected probability distribution field outputs the following characteristic parameters: the most likely source region location (the grid cell corresponding to the maximum probability density), the source region confidence interval (the spatial region containing 80% probability quality), and the multi-source separation indicator (when the probability distribution shows multiple clearly separated peaks, it is determined that there are multiple independent pollution sources).

[0176] Step S63: The system sends the most likely source region location and source region confidence interval to surrounding radars, guiding them to perform directional high-frequency verification scans. The verification scan strategy is dynamically adjusted based on the overall confidence distribution within the source region confidence interval: in sub-regions with an overall confidence level below 0.4, the highest resolution encrypted scan (angular resolution 0.5 degrees, scan cycle 1 minute) is used; in sub-regions with an overall confidence level above 0.7, conventional scanning is sufficient.

[0177] During the verification scan, the system compares the extinction coefficient change trend at the source region location with the inversion results of step S1 in real time. If the verification scan finds that the extinction coefficient of the source region continues to increase or exhibits pulse emission characteristics, the location is confirmed as an actual pollution source, and the verification result is fed back as a high-confidence sample to the neural network training set in step S1 to improve the accuracy of subsequent inversions. If the verification scan does not find any anomalies, the system backtracks to the second peak position of the probability distribution field and repeats the verification process. The data obtained from the verification scan is also used for online calibration in step S3 to update the system constants of the relevant radars.

[0178] Once a pollution source is verified and confirmed, the system outputs its geographical coordinates, emission intensity estimate (calculated based on peak extinction coefficient and wind field divergence), emission time window, and the overall confidence level of the source tracing result, which are then used for visualization overlay and emergency decision-making in step S5.

[0179] Step S64: The system constructs a pollution risk prediction model. The model's inputs include: the current uncertainty field distribution output from Step S2, the physical residual trend output from Step S3, the four-dimensional concentration field evolution trend output from Step S4, upstream regional pollution monitoring information (acquired through network collaboration), and real-time meteorological data (wind speed, wind direction, atmospheric stability, and boundary layer height). The model outputs the pollution risk level of the monitoring area for the next 1 to 6 hours, categorized into low, medium, and high risk levels. The risk level is updated every 30 minutes.

[0180] The system dynamically adjusts the operating mode of each radar based on the current pollution risk level and the ongoing source tracing verification status:

[0181] Low-risk periods: When the risk level is low and there are no ongoing traceability verification tasks, the system switches the radar to energy-saving mode. In energy-saving mode, the scanning cycle is extended to 2 to 3 times that of the regular mode (e.g., from 5 minutes to 12 minutes), the angular resolution is widened to 2 to 3 degrees, the number of vertical scanning layers is reduced, and pulse emission is used to reduce the average power of the laser.

[0182] Medium-risk period: When the risk level is medium, the system maintains the normal scanning mode, maintains normal spatiotemporal resolution, and ensures timely capture of potential contamination events.

[0183] High-risk periods: When the risk level is high or traceability verification scanning is in progress, the system automatically switches to enhanced scanning mode. In enhanced mode, the scanning cycle is shortened to one-half to one-third of that in regular mode (e.g., from 5 minutes to 2 minutes), the angular resolution is increased to 0.5 degrees, the number of vertical scanning layers is increased, and continuous wave transmission is used to improve the signal-to-noise ratio.

[0184] Mode switching is performed gradually over 3 to 5 scan cycles to avoid data gaps. For edge radars that have been in a low-risk state for a long time, the system supports a deep sleep mode: the radar stops scanning and only maintains communication and status monitoring. When the risk level increases or a traceability verification command is received, it can be woken up within 30 seconds.

[0185] Step S65: The system incorporates the source tracing verification results, energy consumption adjustment effects, and user adoption of the source tracing results (from user interaction records in step S5) into a unified closed-loop feedback framework:

[0186] The source tracing feedback is fed back to the scanning strategy. The confirmed pollution source location and emission characteristics are fed back to step S2 to optimize subsequent scanning strategies. For stationary pollution sources, the system includes them in key monitoring areas and increases sampling density in the regular scanning mode. For intermittent emission sources, the system automatically increases the monitoring frequency during periods of high emission based on historical emission time patterns. For mobile pollution sources, the system uses the source tracing results to train a mobile source trajectory prediction model to improve tracking accuracy.

[0187] The risk prediction model is adaptively updated, comparing actual pollution events with the risk prediction results to calculate the prediction accuracy. When the accuracy consistently falls below a threshold, the system triggers model parameter retraining, adjusting the weight coefficients of each influencing factor based on the characteristics of recent pollution events, enabling risk prediction to adapt to seasonal changes in meteorological conditions or emission source distribution.

[0188] Energy consumption strategy effectiveness assessment: The system regularly evaluates the balance between energy savings and pollution incident underreporting rate in energy-saving mode, and between additional energy consumption and capture rate improvement in enhanced mode.

[0189] The system self-verifies the credibility of the tracing results. It cross-compares multiple independent tracing results. When multiple consecutive tracing results point to the same source area and the verification scan confirms it, the confidence level of the source area is automatically increased. When the tracing results at different times and under different weather conditions are contradictory, the system triggers step S3 to review the calibration data of the area and prompts the maintenance personnel to check the equipment status.

[0190] Step S66: The system establishes an operational status monitoring module to monitor the following indicators in real time and generate corresponding alarms:

[0191] The source tracing confidence trend indicates that when the overall confidence level is below the threshold for more than three consecutive source tracing results, it suggests that the source tracing may be difficult due to inaccurate wind field models or insufficient observation data. It is recommended to supplement radar sites or adjust the scanning strategy.

[0192] Energy consumption anomaly monitoring: When the actual energy consumption of a radar deviates from the theoretical energy consumption by more than a threshold, it indicates that there may be equipment failure or parameter configuration error, and generates a maintenance alarm.

[0193] The system monitors the health status of each radar, including key parameters such as transmit power, receive sensitivity, and signal-to-noise ratio baseline. When these parameters deviate from the normal range and exceed the threshold, an equipment maintenance alarm is generated.

[0194] Closed-loop response delay: Monitor the end-to-end delay from the occurrence of a pollution event to the output of the source tracing results. When the delay exceeds a threshold, trigger system performance diagnosis to check data links and computing resources.

[0195] Existing lidar network monitoring technologies have the following shortcomings in pollution source tracing and energy consumption management: First, pollution source tracing relies on post-event offline analysis, with response times as long as several hours, making it unsuitable for emergency decision-making; second, the source tracing methods do not utilize multi-source confidence information, resulting in significant errors in areas with unreliable wind fields or poor data quality; third, energy consumption management is inefficient, with the radar operating at full capacity 24 / 7, leading to serious energy waste; fourth, there is a lack of closed-loop feedback between source tracing, scanning, and energy consumption, preventing monitoring strategies from being dynamically optimized based on source tracing results and user concerns; and fifth, risk prediction is disconnected from energy consumption management, failing to achieve on-demand energy saving. This invention integrates the confidence levels of each stage—inversion, calibration, interpolation, and visualization—into reverse tracing and source area probability calculation; establishes a closed loop for source area verification scanning and inversion calibration to improve source tracing reliability; constructs adaptive energy consumption management based on pollution risk prediction to achieve on-demand energy saving; and forms a closed-loop self-optimizing system for source tracing, scanning, and energy consumption.

[0196] It should be noted that step S1 provides inversion confidence, step S2 provides uncertainty-driven data, step S3 provides physical constraint confidence, step S4 provides interpolation confidence, step S5 provides user attention feedback, and step S6 integrates all the above information to complete source tracing and energy consumption optimization, and feeds back the source tracing results and energy consumption assessment to the preceding steps, forming a complete intelligent closed loop from inversion, scanning, calibration, interpolation, visualization to source tracing and energy consumption management.

[0197] It should be noted that technical features that are not fully explained will be addressed using conventional technical methods.

[0198] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described herein. Although the present invention has been described in detail with reference to the above embodiments, the present invention is not limited to the specific embodiments described above. Therefore, any modifications or equivalent substitutions to the present invention, as well as all technical solutions and improvements that do not depart from the spirit and scope of the invention, are covered within the scope of the claims of the present invention.

Claims

1. A lidar network monitoring method based on intelligent collaboration and closed-loop optimization, characterized in that, It includes the following steps: Step S1: Construct a deep neural network, using the lidar equation as a physical constraint, and take advantage of the observation consistency of multiple radars in the overlapping area to jointly optimize the neural network parameters and the constants of each radar system, so as to realize the inversion of the extinction coefficient and backscattering coefficient, while quantifying the inversion uncertainty and outputting three-dimensional grid data with inversion confidence. Step S2: Utilize the inversion uncertainty to dynamically optimize radar scanning parameters, perform intensified scanning of high uncertainty areas, use neural network gradient prediction information gain for pre-scanning, and feed the new data back to step S1; Step S3: Embed the self-supervised calibration loss into the inversion loss function of step S1 to calibrate the radar system constants, and perform reverse repair on outliers based on physical residuals to establish a two-way feedback between calibration and repair, output the physical constraint confidence and feed it back to step S1. Step S4: Integrate the inversion confidence, physical constraint confidence, and uncertainty to construct interpolation weights and adaptive interpolation kernels related to the physical residuals. Combine the digital twin background field and time-dimensional physical transport constraints to reconstruct the four-dimensional spatiotemporal data field and output the interpolation confidence. Step S5: Integrate the inversion confidence, physical constraint confidence, and interpolation confidence into a comprehensive confidence, map it to the transparency of the 3D volume drawing, construct an interactive 4D visualization platform, and feed back the user's attention to the low confidence area to steps S2 and S3; Step S6: Using the reconstructed three-dimensional wind field, concentration field and the comprehensive confidence level, the source region probability distribution is generated by confidence-weighted inverse particle tracking, which guides the radar verification scan and feeds back to steps S1 and S3. At the same time, the radar working mode is dynamically adjusted based on pollution risk prediction to form a closed-loop self-optimization.

2. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S1, using the lidar equation as a physical constraint includes: Construct a data fitting loss term so that the extinction coefficient and backscattering coefficient output by the neural network satisfy the integral form of the lidar equation; Construct a loss term for the physical equation so that the extinction coefficient and backscattering coefficient output by the neural network satisfy the differential form of the lidar equation; When jointly optimizing the neural network parameters and the constants of each radar system, the data fitting loss term and the physical equation loss term are weighted and summed as part of the total loss function.

3. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S1, utilizing the observation consistency of multiple radars in a spatially overlapping region includes: For any two radars, define the spatial overlap region of their detection rays; A multi-radar consistency loss is constructed, which is used to measure the difference between the extinction coefficient and backscattering coefficient of the same spatial point obtained by different radars in the overlapping region; The multi-radar consistency loss is added to the total loss function, and together with the physical constraints of the lidar equation, the neural network parameters and the constants of each radar system are optimized.

4. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S1, the quantification of the uncertainty of the inversion result using the Monte Carlo method includes: During the neural network testing phase, the random deactivation mechanism used during training is kept active, and multiple forward propagations are performed to obtain multiple inversion results. Statistical analysis was performed on the multiple inversion results to calculate their mean and standard deviation; The mean is output as the final inversion result, the standard deviation is used as the uncertainty of the inversion result, and the inversion confidence is defined based on the ratio of the standard deviation to the mean.

5. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S1, the joint optimization of neural network parameters and radar system constants includes: The system constants of each radar are treated as optimizable variables and iteratively optimized in sync with the parameters of the neural network. By leveraging the observation consistency of multiple radars in spatially overlapping areas, the system constant is automatically calibrated by minimizing the consistency loss of multiple radars, thus achieving self-calibration without the need for external calibration equipment.

6. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S2, the dynamic optimization of radar scanning parameters includes: Calculate the comprehensive uncertainty index and its spatial gradient based on the aforementioned uncertainty field; Using the Bayesian optimization method, with the reduction in global uncertainty per unit time as the objective function, the radar's angular resolution, scanning period, and sector width are solved. For regions with high gradient uncertainty, a non-uniform scan line density is used, with denser scan lines set in the direction of the uncertainty gradient.

7. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S3, the reverse repair of outliers based on physical residuals includes: The differential form residual of the lidar equation at each observation point is calculated in real time, and when the residual exceeds the dynamic threshold, it is determined to be an anomaly. For the aforementioned anomaly points, the extinction coefficient and backscattering coefficient output by the neural network are used as initial values. With the lidar equation and spatial smoothness as constraints, the repair value is generated through inverse optimization so that the repair result satisfies the physical equation constraints.

8. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S4, the construction of interpolation weights and adaptive interpolation kernels related to the physical residuals includes: Define interpolation weights, which are positively correlated with the confidence level of physical constraints and negatively correlated with the residuals of physical equations; An anisotropic interpolation kernel is designed, the bandwidth of which is adaptively adjusted according to the local uncertainty gradient at the observation point. The bandwidth is reduced in regions with large uncertainty gradients and increased in regions with uniform uncertainty.

9. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S5, the construction of the interactive four-dimensional visualization platform includes: The overall confidence level is mapped to the transparency parameter of the 3D volume rendering. Areas with high confidence are completely opaque, while areas with low confidence are semi-transparent or blurred. It offers features such as time-sliding window playback, 3D stereo rendering, multi-section linkage analysis, and confidence layer overlay. Automatically extract and dynamically label the boundaries, transport paths, and transport flux characteristics of contamination blobs.

10. The lidar network monitoring method based on intelligent collaboration and closed-loop optimization according to claim 1, characterized in that, In step S6, the dynamic adjustment of the radar operating mode based on pollution risk prediction includes: A pollution risk prediction model is constructed, and the pollution risk level for future periods is output based on the uncertainty field, physical residual trend, concentration field evolution trend and meteorological data. Based on the pollution risk level and source tracing verification status, the radar operating mode will be dynamically switched between energy-saving mode, normal mode and enhanced mode. During low-risk periods, switch to energy-saving mode to extend the scanning cycle and widen the angular resolution to reduce energy consumption; during high-risk periods or when conducting traceability verification, switch to enhanced mode to shorten the scanning cycle and increase the angular resolution to improve data capture capabilities.

Images