Intelligent environmental pollution traceability early warning system based on multi-source data fusion

Abstract

The present application relates to the technical field of environmental pollution monitoring and early warning and machine learning, and specifically discloses an intelligent environmental pollution tracing and early warning system based on multi-source data fusion. The technical scheme of the present application fills in the missing data and unifies the time-space coordinates of the original data of the multi-source environmental monitoring network, reaches an agreement on the time-space resolution and sampling frequency of different source data with the help of a time-space alignment model, dynamically deduces the time-space distribution prediction sequence of pollutants by using a machine learning agent model coupled with a pollutant physical diffusion equation, determines the pollution source location and release intensity parameters by combining a machine learning inversion algorithm, generates an early warning signal containing the pollution source location and diffusion trend when the pollutant concentration prediction value exceeds the early warning threshold, solves the problem that multi-source data cannot be deeply analyzed in collaboration and the tracing result is limited by the data island effect due to the dependence on a single monitoring network or a fixed threshold model, and improves the timeliness of environmental risk prevention and control and the scientific and refined level of environmental management.

Description

Technical Field

[0001] This invention relates to the fields of environmental pollution monitoring and early warning and machine learning technology, and in particular to an intelligent environmental pollution source tracing and early warning system that integrates multi-source data. Background Technology

[0002] Environmental pollution source tracing and early warning is a crucial aspect of ecological and environmental governance. With the widespread adoption of the Internet of Things (IoT) and sensor technologies, various monitoring stations for air quality, water quality, and soil have been extensively deployed, resulting in multi-source and heterogeneous monitoring data. However, existing source tracing methods often rely on single monitoring networks or fixed threshold models. Differences in spatiotemporal resolution, acquisition frequency, and format standards among data sources hinder in-depth collaborative analysis of multi-source data, limiting source tracing results due to data silos. Furthermore, traditional early warning mechanisms focus on passive responses after concentration exceedances, lacking the ability to dynamically predict pollution diffusion paths. This makes accurate source tracing and early warning difficult in the early stages of pollution events, restricting the timeliness of environmental risk prevention and control. Therefore, there is an urgent need for an environmental pollution early warning technology that can integrate multi-source monitoring data and possess intelligent source tracing and prediction capabilities to improve the scientific and precise level of environmental management. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention provides an intelligent environmental pollution source tracing and early warning system based on multi-source data fusion. The technical solution of this intelligent environmental pollution source tracing and early warning system based on multi-source data fusion is as follows: The multi-source data standardization module is used to fill in missing data and unify spatiotemporal coordinates of raw monitoring data from different types of environmental monitoring networks to generate standardized multi-source data. The spatiotemporal alignment module is used to input the standardized multi-source data into a pre-trained spatiotemporal alignment model to obtain aligned fused data with consistent spatiotemporal resolution and sampling frequency among data from different sources; The pollution diffusion simulation module is used to input the aligned and fused data into a machine learning proxy model coupled with the physical diffusion equation of pollutants, and dynamically simulate the spatiotemporal distribution prediction sequence of pollutants. The pollution source reverse tracing module is used to determine the location parameters and release intensity parameters of the pollution source based on the spatiotemporal distribution prediction sequence and real-time monitoring data using a machine learning inversion algorithm, and output the source tracing and location results. The early warning module is used to generate an early warning signal containing the location of the pollution source and its diffusion trend when the predicted value of the pollutant concentration in the spatiotemporal distribution prediction sequence exceeds the early warning threshold, in combination with the source tracing and location results.

[0004] The technical solution of this invention fills in missing data and unifies spatiotemporal coordinates in the raw data of a multi-source environmental monitoring network. By using a spatiotemporal alignment model, it achieves consistency in spatiotemporal resolution and sampling frequency of data from different sources. It uses a machine learning surrogate model coupled with the physical diffusion equation of pollutants to dynamically deduce the spatiotemporal distribution prediction sequence of pollutants. Combined with a machine learning inversion algorithm, it determines the location of pollution sources and release intensity parameters. When the predicted value of pollutant concentration exceeds the warning threshold, it generates a warning signal containing the location of pollution sources and diffusion trends. This solves the problems of difficulty in in-depth collaborative analysis of multi-source data and the limitation of source tracing results by the data silo effect caused by relying on a single monitoring network or a fixed threshold model. It improves the timeliness of environmental risk prevention and control and the scientific and refined level of environmental management.

[0005] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and in order to make the above and other objects, features and advantages of the present invention more apparent and understandable, specific embodiments of the present invention are described below. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0007] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This is a schematic diagram of an embodiment of an intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to the present invention. Detailed Implementation

[0008] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0009] Figure 1 This diagram illustrates a structural schematic of an embodiment of an intelligent environmental pollution source tracing and early warning system based on multi-source data fusion provided by the present invention. Figure 1 As shown, this intelligent environmental pollution source tracing and early warning system based on multi-source data fusion includes: The multi-source data standardization module 110 is used to fill in missing data and unify spatiotemporal coordinates of the original monitoring data of different types of environmental monitoring networks to generate standardized multi-source data.

[0010] The raw monitoring data refers to the original, untreated pollutant concentration measurements directly collected by monitoring stations deployed in different environmental media. For example, air monitoring station A in an industrial park recorded a PM2.5 concentration of 80 μg / m³, water monitoring station B recorded a heavy metal concentration of 0.3 mg / L, and soil monitoring station C recorded an organic matter concentration of 5 mg / kg.

[0011] Missing data filling refers to the process of estimating and filling in missing data points in the original monitoring data due to equipment failure or communication interruption using valid measurements from nearby points. For example, if PM2.5 data for air monitoring station A at 10:00 AM on May 1, 2026 is missing, the PM2.5 concentration at 10:00 AM is estimated to be 82 μg / m³ using valid data from 9:00 AM and 11:00 AM, as well as data from nearby stations B and C, and then filled into the missing position. Spatiotemporal coordinate unification refers to the operation of converting and aligning data from different monitoring networks according to a unified time base and a unified spatial coordinate system. For example, the data recorded by water monitoring station B in UTC+8 time zone at 10:00 AM on May 1, 2026, and the GPS time data recorded by soil monitoring station C are uniformly converted to Beijing time and projected onto the same latitude and longitude grid.

[0012] Standardized multi-source data refers to a collection of environmental monitoring data of various types that has a complete time series and a consistent spatial reference system after missing data filling and spatiotemporal coordinate unification. For example, the PM2.5 concentration, heavy metal concentration and organic matter concentration data of monitoring stations A, B and C in the industrial park at every hour from 0:00 to 24:00 on May 1, 2026 are all filled in completely and unified to Beijing time and WGS84 coordinates.

[0013] The spatiotemporal alignment module 120 is used to input the standardized multi-source data into a pre-trained spatiotemporal alignment model to obtain aligned fusion data with consistent spatiotemporal resolution and sampling frequency among data from different sources.

[0014] Among them, the spatiotemporal alignment model refers to a pre-trained machine learning model used to eliminate the differences in spatiotemporal resolution and sampling frequency between data from different sources; for example, a deep residual alignment network spatiotemporally transforms the hourly data of air monitoring station A with the half-hourly data of water monitoring station B, making them comparable at the same time and spatial grid points.

[0015] Here, "source data" refers to raw or pre-processed data obtained from different environmental monitoring networks or different types of sensors, each possessing its own original spatiotemporal attributes; for example, hourly PM2.5 concentration data obtained from an air monitoring network and half-hourly heavy metal concentration data obtained from a water monitoring network. Spatiotemporal resolution refers to the minimum distinguishable distance in the spatial dimension and the minimum sampling interval in the temporal dimension for environmental monitoring data; for example, the spatial resolution of air monitoring station A is station location data, and the temporal resolution is 1 hour, while the spatial resolution of water monitoring station B is a cross-section every 500m along the river, and the temporal resolution is 30 minutes. Sampling frequency refers to the number of times the monitoring equipment collects data per unit time; for example, the sampling frequency of air monitoring station A is 1 time / hour, and the sampling frequency of water monitoring station B is 2 times / hour.

[0016] Aligned and fused data refers to a dataset that is completely consistent in time frequency and spatial grid after standardized multi-source data is input into a spatiotemporal alignment model. For example, after alignment, joint data of PM2.5 concentration and heavy metal concentration are obtained for each 1km×1km grid in an industrial park every 30 minutes.

[0017] The pollution diffusion simulation module 130 is used to input the aligned and fused data into a machine learning proxy model that couples the pollutant physical diffusion equation to dynamically deduce the spatiotemporal distribution prediction sequence of pollutants.

[0018] Among them, coupling pollutant physical diffusion equations refers to embedding partial differential equations describing the migration and transformation of pollutants in environmental media into machine learning models as physical constraints; for example, incorporating convection-diffusion equations with source and attenuation terms as physical loss terms into the neural network training process.

[0019] Among them, machine learning proxy models refer to computational models trained using machine learning methods that replace high-precision physical models and are used to quickly simulate pollutant diffusion processes; for example, a physical information neural network takes the location of the pollution source and the release intensity as input and outputs the concentration distribution of the entire area over the next 24 hours.

[0020] Among them, the spatiotemporal distribution prediction sequence refers to the time series data formed by the predicted concentration values ​​of pollutants at various points in space at multiple consecutive future moments output by the machine learning surrogate model; for example, the model outputs the PM2.5 concentration values ​​at each 1km grid every 30 minutes in the next 24 hours of an industrial park, for a total of 48 time steps of concentration field.

[0021] The pollution source reverse tracing module 140 is used to determine the location parameters and release intensity parameters of the pollution source based on the spatiotemporal distribution prediction sequence and real-time monitoring data using a machine learning inversion algorithm, and output the source tracing and location results.

[0022] Real-time monitoring data refers to the pollutant concentration measurements obtained in real time from various types of environmental monitoring networks at the current or near real-time moment; for example, the PM2.5 concentration uploaded by air monitoring station A at 14:00 on May 1, 2026 was 85 μg / m³.

[0023] Among them, machine learning inversion algorithm refers to the calculation process of inferring pollution source parameters by using machine learning methods to reverse the difference between monitoring data and model predictions; for example, variational Bayesian inversion algorithm with sparse regularization constraint updates the probability distribution of pollution source location and release intensity based on the deviation between real-time monitoring concentration and concentration field predicted by diffusion model.

[0024] Here, "pollution source" refers to the origin or location of pollutants discharged into the environment; for example, an unapproved wastewater outlet in an industrial park continuously releases heavy metal pollutants into a river. "Location parameter" refers to the coordinate parameters describing the spatial location of the pollution source; for example, the longitude of the wastewater outlet is 116.123° and the latitude is 39.456°. "Release intensity parameter" refers to the parameter describing the amount of pollutants discharged into the environment by the pollution source per unit time; for example, the rate of heavy metal emission from the outlet is 2.5 kg / h.

[0025] The source tracing and location results refer to the final determined values ​​of the pollution source location parameters and release intensity parameters calculated by the inversion algorithm; for example, the inversion results show that the pollution source is located at 116.123° east longitude and 39.456° north latitude, and the heavy metal release intensity is 2.4 kg / h.

[0026] The early warning module 150 is used to generate an early warning signal containing the location of the pollution source and the diffusion trend when the predicted value of the pollutant concentration in the spatiotemporal distribution prediction sequence exceeds the early warning threshold, in combination with the source tracing and positioning results.

[0027] The predicted pollutant concentration refers to the estimated value of pollutant concentration at a specific time and location in the spatiotemporal distribution prediction sequence; for example, the model predicts that the PM2.5 concentration at a grid point in an industrial park will be 95 μg / m³ after 2 hours. The warning threshold refers to the pollutant concentration limit value used to determine whether an alarm needs to be issued; for example, an alarm is triggered when the predicted PM2.5 concentration exceeds 150 μg / m³.

[0028] The location of the pollution source refers to its specific spatial location; for example, the pollution source is a rainwater well located 20m west of the sewage treatment pond in the northeast corner of the industrial park. The diffusion trend refers to the direction, rate, and potential impact range of the pollutant's concentration changes over time and space; for example, a forecast shows PM2.5 spreading southeast at a speed of 2km / h, covering two adjacent residential areas within the next 3 hours.

[0029] Among them, the warning signal refers to the alarm notification generated by the system when the predicted concentration exceeds the warning threshold, which includes information on the pollution source and diffusion. For example, a red warning signal is generated, which reads: "The pollution source in the northeast corner of the industrial park is releasing PM2.5 at an intensity of 3 kg / h. It is expected to affect Community A in 2 hours. Residents are advised to reduce outdoor activities."

[0030] The technical solution of this embodiment fills in missing data and unifies spatiotemporal coordinates of the original data from a multi-source environmental monitoring network. By using a spatiotemporal alignment model, it achieves consistency in spatiotemporal resolution and sampling frequency of data from different sources. It uses a machine learning surrogate model coupled with the physical diffusion equation of pollutants to dynamically deduce the spatiotemporal distribution prediction sequence of pollutants. Combined with a machine learning inversion algorithm, it determines the location of pollution sources and release intensity parameters. When the predicted value of pollutant concentration exceeds the warning threshold, it generates a warning signal containing the location of pollution sources and diffusion trends. This solves the problems of difficulty in in-depth collaborative analysis of multi-source data and the limitation of source tracing results by the data silo effect caused by relying on a single monitoring network or a fixed threshold model. It improves the timeliness of environmental risk prevention and control and the scientific and refined level of environmental management.

[0031] In one alternative approach, the different types of environmental monitoring networks include air quality monitoring networks, water quality monitoring networks, and soil monitoring networks.

[0032] The air quality monitoring network refers to a multi-site monitoring system deployed in the atmospheric environment for real-time collection of gaseous pollutant and particulate matter concentrations; for example, an industrial park has five air monitoring stations, each uploading PM2.5, PM10, and sulfur dioxide concentration data hourly. The water quality monitoring network refers to a multi-section monitoring system deployed in rivers, lakes, groundwater, and other water environments for continuous monitoring of water quality parameters; for example, three sections (upstream, midstream, and downstream) of a river within the industrial park collect heavy metal and ammonia nitrogen concentration data every 30 minutes. The soil monitoring network refers to a multi-point monitoring system deployed in the soil environment for periodic testing of soil pollutant content; for example, an industrial park has 10 soil monitoring points arranged in a grid pattern, collecting organic matter and heavy metal content data weekly.

[0033] Among the above-mentioned optional methods, further integration of data from three types of monitoring networks—air quality, water quality, and soil—broadens the dimensions of pollution source information acquisition, making source tracing analysis no longer limited to a single medium, and improving the coverage and data integrity of multi-media collaborative monitoring.

[0034] In one alternative embodiment, the multi-source data standardization module 110 is specifically used for: The missing value in the original monitoring data is obtained by performing spatiotemporal inverse distance weighted interpolation on the nearest valid data points around the spatiotemporal location of the missing value.

[0035] Missing values ​​refer to gaps in the original monitoring data due to equipment malfunction, communication interruption, or sampling failure. For example, the PM2.5 data record for air monitoring station A at 10:00 AM on May 1, 2026, is blank or NULL. Neighboring valid data points refer to data points that are geographically close to the missing value and have valid measurements. For example, the PM2.5 data at 9:00 AM one hour before the missing value is 78 μg / m³, and the data at 11:00 AM one hour after is 84 μg / m³; monitoring station B, within a 1km radius of the missing value, has a data of 80 μg / m³. Imputed values ​​are values ​​calculated using interpolation or estimation methods to replace missing values. For example, the imputed PM2.5 value for 10:00 AM is 82 μg / m³, calculated using inverse distance weighted interpolation between the preceding and following time points and neighboring spatial points.

[0036] Among the above-mentioned optional methods, spatiotemporal inverse distance weighted interpolation is further used to fill in the missing values ​​in the original monitoring data, which improves the data loss problem caused by equipment failure or transmission interruption, and makes subsequent spatiotemporal alignment and diffusion inference based on a more complete data foundation.

[0037] In one alternative approach, the spatiotemporal alignment model is a deep residual alignment network, which maps the standardized multi-source data to a unified spatiotemporal grid through a cross-scale feature mapping module.

[0038] Among them, deep residual alignment network refers to a deep neural network that uses a residual connection structure to learn the nonlinear mapping relationship between data of different resolutions; for example, a convolutional neural network containing 3 residual blocks is input to a sparse time series of an air monitoring station and outputs an encrypted time series every half hour.

[0039] Among the above-mentioned optional approaches, a deep residual alignment network is further introduced as a spatiotemporal alignment model. Through cross-scale feature mapping, heterogeneous data is mapped to a unified spatiotemporal grid, which alleviates the differences in spatial resolution and data format between different monitoring networks and enables multi-source data to be fused.

[0040] In an alternative embodiment, the spatiotemporal alignment module 120 is specifically used for: The monitoring data with the lowest time resolution in the standardized multi-source data is upsampled to the time scale of the monitoring data with the highest time resolution through the time attention mechanism of the deep residual alignment network.

[0041] The cross-scale feature mapping module refers to a sub-network component in the deep residual alignment network responsible for transforming feature maps of different scales to the same spatial and temporal scale; for example, this component transforms hour-level temporal features and kilometer-level spatial features into half-hour-level and hundred-meter-level features through transpose convolution operations.

[0042] The lowest temporal resolution refers to the time-discriminating capability of the data with the longest sampling interval among the multi-source data participating in the alignment; for example, the temporal resolution of air monitoring station A is 1 hour, which is lower than the 30-minute resolution of water monitoring station B, therefore 1 hour is the lowest temporal resolution. Monitoring data refers to the raw or pre-processed pollutant concentration measurements obtained from various types of environmental monitoring networks; for example, the PM2.5 value of 80 μg / m³ recorded by air monitoring station A.

[0043] The time attention mechanism refers to the component in the deep residual alignment network that automatically learns the weights of the contributions of data from different historical moments to the alignment result at the current moment. For example, this mechanism assigns a weight of 0.6 to the data from the previous hour, a weight of 0.3 to the data from the previous two hours, and a weight of 0.1 to the data from the previous three hours, and then sums the weights to obtain the alignment value at the current moment.

[0044] Among the above-mentioned optional methods, the time attention mechanism is further utilized to upsample the low time resolution monitoring data to a high time resolution scale, which solves the problem of time sequence misalignment caused by inconsistent sampling frequencies of different monitoring networks, and keeps the fused data synchronized in the time dimension.

[0045] In one alternative approach, the machine learning proxy model is a physical information neural network, the pollutant physical diffusion equation is a convection-diffusion equation with a nonlinear reaction term, and the physical information neural network is trained using the residuals of the convection-diffusion equation and the residuals of the boundary conditions as a loss function.

[0046] Among them, physical information neural network refers to a type of machine learning model that incorporates the residuals of physical equations as a loss function term into the training of the neural network; for example, a neural network that simultaneously fits monitoring data and convection-diffusion equation residuals has a total loss function that includes data error terms and physical residual terms.

[0047] The convection-diffusion equation refers to a partial differential equation describing the movement of pollutants in a fluid medium and their dispersion process caused by concentration gradients; for example, the equation has the following form: Where C is the concentration, Where is the flow velocity and D is the diffusion coefficient. Let S be the attenuation coefficient, and S be the source term.

[0048] Boundary conditions refer to the constraints on pollutant concentration or flux imposed on the boundary of the simulated region; for example, the upstream boundary is set as a known concentration time series, and the downstream boundary is set as a free outflow condition. The loss function is a mathematical function used to measure the difference between the model's predicted value and the target value when training a machine learning model; for example, loss function = mean squared error of data + mean squared error of physical equation residuals + mean squared error of boundary conditions.

[0049] Among the above-mentioned optional methods, a physical information neural network is further adopted as a machine learning proxy model, which incorporates the physical residuals and boundary condition residuals of the convection-diffusion equation into the loss function, so that the pollution diffusion simulation process follows the physical laws of pollutant transport, thereby improving the physical consistency and interpretability of the predicted sequence.

[0050] In an alternative embodiment, the pollution diffusion simulation module 130 is specifically used for: The spatial discrete grid resolution for the dynamic extrapolation of the spatiotemporal distribution prediction sequence is adaptively determined based on the spatial distribution density of all monitoring stations in the aligned and fused data.

[0051] Here, a monitoring station refers to the specific location of equipment fixedly installed in the environment for continuously or periodically collecting pollutant data; for example, air monitoring station A in an industrial park is located on the rooftop of a building at 116.123°E, 39.456°N. Spatial distribution density refers to the number of monitoring stations per unit area or the average distance between stations; for example, there are 2 air monitoring stations per square kilometer in the industrial park, with an average station spacing of 700m.

[0052] Spatial discrete grid resolution refers to the actual spatial size represented by each grid cell when a continuous space is divided into discrete cells; for example, if an industrial park is divided into a 100m×100m grid, the spatial discrete grid resolution is 100m.

[0053] Among the above-mentioned optional methods, the spatial discrete grid resolution is further adaptively determined based on the spatial distribution density of the monitoring stations, which avoids the problem of insufficient resolution of the fixed grid in sparse areas of the stations or computational redundancy in dense areas of the stations, making the simulation calculation more reasonable in terms of spatial resource allocation.

[0054] In one alternative approach, the machine learning inversion algorithm is a variational Bayesian inversion algorithm with sparse regularization constraints, which iteratively updates the distribution of pollution source parameters by minimizing the expected posterior divergence between the spatiotemporal distribution prediction sequence and the real-time monitoring data.

[0055] Sparse regularization constraints refer to adding an L1 norm term to the loss function of the inversion algorithm to ensure the solution exhibits sparse characteristics. For example, adding the L1 norm of the pollution source parameters to the objective function of variational Bayesian inversion ensures that only a few grid points in the final solution have non-zero source strength. Variational Bayesian inversion algorithms are Bayesian inversion methods that use variational inference to approximate posterior distributions to estimate the probability distribution of unknown parameters from observed data. For example, assuming the pollution source parameters follow a Gaussian distribution, the mean and variance of the distribution are iteratively updated by minimizing the KL divergence between the approximate posterior and the true posterior.

[0056] Here, expected posterior divergence refers to the expected value of the KL divergence used in variational Bayesian inversion to measure the difference between the approximate distribution and the true posterior distribution; for example, calculating the approximate posterior... With the true posterior The KL divergence between the two is calculated, and this divergence is minimized to approximate the true posterior. Pollution source parameters refer to adjustable parameters describing the characteristics of a pollution source, including location parameters and release intensity parameters; for example, a pollution source located at longitude 116.123°, latitude 39.456°, with a release intensity of 2.5 kg / h.

[0057] Among the above-mentioned alternative methods, a variational Bayesian inversion algorithm with sparse regularization constraints is further adopted. By minimizing the expected posterior divergence, the pollution source parameter distribution is updated iteratively. It can still converge stably under the conditions of noise interference and limited monitoring data, thus improving the robustness of source tracing and localization.

[0058] In an alternative approach, the pollution source reverse tracing module 140 calculates the release intensity parameter of the pollution source using the following formula:

[0059] in, This represents an estimated value of the release strength parameter. This represents the total number of monitoring points involved in the calculation. This represents the total number of historical time steps involved in the calculation. Indicates the first The monitoring point at the 1st The difference between the measured concentration value and the background concentration value at each time step Indicates the first The spatial distance between each monitoring point and the location of the candidate pollution source Indicates spatially related length parameters, Represents the kernel function. Represents the attenuation coefficient function. For integration variables, and They represent the first The and the first At that moment in time step This indicates the first prediction sequence obtained based on the spatiotemporal distribution. The monitoring point at the 1st The unit source strength response coefficient at each time step This represents the adjustment coefficient. This represents the spatial partial derivative of the difference along the prevailing wind direction.

[0060] It should be noted that the above formula is based on the difference between the measured concentration and the background concentration at multiple monitoring points at multiple historical time steps. It introduces a spatial kernel function to weight the attenuation by distance, and considers the attenuation effect of pollutants during the propagation process (by integrating the attenuation coefficient function over time). The denominator introduces a unit source strength response coefficient and a spatial partial derivative correction term for the prevailing wind direction, thereby constructing an inversion estimate that integrates spatial correlation, temporal attenuation effect and wind direction sensitivity. The above formula calculates the ratio of the spatiotemporal weighted accumulation of the measured differences in the numerator to the spatiotemporal weighted accumulation of the response coefficient in the denominator, and estimates the release intensity parameters of the pollution source from multi-source spatiotemporal monitoring data, overcoming the shortcomings of traditional single-point inversion methods that are insensitive to wind direction fluctuations and diffusion attenuation.

[0061] Among the above-mentioned optional methods, a formula that includes kernel function weighting, attenuation integral and spatial partial derivative correction is further used to calculate the release intensity parameter. The monitoring concentration difference, spatial distance and prevailing wind direction changes are comprehensively incorporated into the estimation process, so that the quantitative result of the pollution source release intensity is closer to the actual emission situation.

[0062] In an alternative approach, the warning threshold in the warning module 150 is determined using the following dynamic calculation formula:

[0063] in, Indicates spatial location and time The warning threshold Indicated at the location At any time The average of historical pollutant concentration data within the previously preset time window. Indicates at time Historical data standard deviation , , All are positive adjustment coefficients. For integration variables, Indicates the location in the spatiotemporal distribution prediction sequence. and time Predicted pollutant concentrations The spatial gradient vector, Describes the Euclidean norm. This represents the maximum value of the spatial gradient norm at all points in time and space.

[0064] It should be noted that the above formula decomposes the warning threshold into the sum of the historical concentration statistical mean and the dynamic adjustment term. The dynamic adjustment term is in the form of an integral from the current moment to infinite past time. The integral kernel includes the historical standard deviation, the exponential decay factor, and the normalization correction of the spatial gradient norm. This makes the influence of historical data on the current threshold decay exponentially over time. At the same time, the spatial gradient information amplifies the risk weight of areas with drastic concentration changes. The above formula realizes the adaptive adjustment of the warning threshold with time and space. When the historical fluctuation of pollutant concentration is large or the spatial gradient is steep, the threshold is automatically raised to avoid false alarms. During the rapid rise of concentration, the gradient term suppresses the excessive rise of the threshold to ensure the sensitivity of the warning. This solves the problem that fixed threshold models cannot balance sensitivity and robustness.

[0065] Among the above-mentioned optional methods, a dynamic formula that integrates historical average concentration, standard deviation decay integral and spatial gradient normalization information is further adopted to determine the warning threshold, so that the warning threshold can be adaptively adjusted with the pollution diffusion trend and spatiotemporal distribution characteristics, avoiding false alarms or missed alarms in complex pollution scenarios with fixed thresholds.

[0066] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and do not imply a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0067] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-source data fusion intelligent environmental pollution source tracing and early warning system, characterized in that, include: The multi-source data standardization module is used to fill in missing data and unify spatiotemporal coordinates of raw monitoring data from different types of environmental monitoring networks to generate standardized multi-source data. The spatiotemporal alignment module is used to input the standardized multi-source data into a pre-trained spatiotemporal alignment model to obtain aligned fused data with consistent spatiotemporal resolution and sampling frequency among data from different sources; The pollution diffusion simulation module is used to input the aligned and fused data into a machine learning proxy model coupled with the physical diffusion equation of pollutants, and dynamically simulate the spatiotemporal distribution prediction sequence of pollutants. The pollution source reverse tracing module is used to determine the location parameters and release intensity parameters of the pollution source based on the spatiotemporal distribution prediction sequence and real-time monitoring data using a machine learning inversion algorithm, and output the source tracing and location results. The early warning module is used to generate an early warning signal containing the location of the pollution source and its diffusion trend when the predicted value of the pollutant concentration in the spatiotemporal distribution prediction sequence exceeds the early warning threshold, in combination with the source tracing and location results.

2. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 1, characterized in that, The different types of environmental monitoring networks include air quality monitoring networks, water quality monitoring networks, and soil monitoring networks.

3. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 2, characterized in that, The multi-source data standardization module is specifically used for: The missing value in the original monitoring data is obtained by performing spatiotemporal inverse distance weighted interpolation on the nearest valid data points around the spatiotemporal location of the missing value.

4. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 1, characterized in that, The spatiotemporal alignment model is a deep residual alignment network, which maps the standardized multi-source data to a unified spatiotemporal grid through a cross-scale feature mapping module.

5. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 4, characterized in that, The spatiotemporal alignment module is specifically used for: The monitoring data with the lowest time resolution in the standardized multi-source data is upsampled to the time scale of the monitoring data with the highest time resolution through the time attention mechanism of the deep residual alignment network.

6. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 1, characterized in that, The machine learning proxy model is a physical information neural network, and the pollutant physical diffusion equation is a convection-diffusion equation with a nonlinear reaction term. The physical information neural network is trained by using the residuals of the convection-diffusion equation and the residuals of the boundary conditions together as a loss function.

7. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 6, characterized in that, The pollution diffusion simulation module is specifically used for: The spatial discrete grid resolution for the dynamic extrapolation of the spatiotemporal distribution prediction sequence is adaptively determined based on the spatial distribution density of all monitoring stations in the aligned and fused data.

8. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 1, characterized in that, The machine learning inversion algorithm is a variational Bayesian inversion algorithm with sparse regularization constraints. The variational Bayesian inversion algorithm iteratively updates the distribution of pollution source parameters by minimizing the expected posterior divergence between the spatiotemporal distribution prediction sequence and the real-time monitoring data.

9. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 8, characterized in that, The pollution source reverse tracing module uses the following formula to calculate the release intensity parameter of the pollution source: ； in, This represents an estimated value of the release strength parameter. This represents the total number of monitoring points involved in the calculation. This represents the total number of historical time steps involved in the calculation. Indicates the first The monitoring point at the 1st The difference between the measured concentration value and the background concentration value at each time step Indicates the first The spatial distance between each monitoring point and the location of the candidate pollution source Indicates spatially related length parameters, Represents the kernel function. Represents the attenuation coefficient function. For integration variables, and They represent the first The and the first At that moment in time step This indicates the first prediction sequence obtained based on the spatiotemporal distribution. The monitoring point at the 1st The unit source strength response coefficient at each time step. This represents the adjustment coefficient. This represents the spatial partial derivative of the difference along the prevailing wind direction.

10. The intelligent environmental pollution source tracing and early warning system based on multi-source data fusion according to claim 1, characterized in that, The warning threshold in the warning module is determined using the following dynamic calculation formula: ； in, Indicates spatial location and time The warning threshold Indicated at the location At any time The average of historical pollutant concentration data within the previously preset time window. Indicates at time Historical data standard deviation , , All are positive adjustment coefficients. For integration variables, Indicates the location in the spatiotemporal distribution prediction sequence. and time Predicted pollutant concentrations The spatial gradient vector, Denotes the Euclidean norm. This represents the maximum value of the spatial gradient norm at all points in time and space.

Images