Non-ferrous metal industry cluster area soil and groundwater pollution intelligent early warning method
By constructing a deep learning model with topological constraints on groundwater flow field, the problems of static isolation and rigid thresholds in the early warning of soil and groundwater pollution in non-ferrous metal industrial clusters were solved. This model enables the simulation of pollutant migration paths and prediction of future trends, providing efficient and accurate adaptive early warning.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-03-16
- Publication Date
- 2026-06-16
AI Technical Summary
Existing technologies for early warning of soil and groundwater pollution in non-ferrous metal industrial clusters suffer from problems such as static and isolated indicator systems, insufficient integration of physical mechanisms, and rigid threshold classification. They cannot effectively predict future risks and ignore the migration paths and temporal evolution of pollutants.
A deep coupling model based on the topology and spatiotemporal characteristics of groundwater flow field is constructed. Through a deep learning model of "physical hydrological topological constraints + spatiotemporal sequence recursive prediction", combined with a time recursive layer, a spatial aggregation layer and a hierarchical judgment layer, spatiotemporal prediction and adaptive hierarchical early warning of pollutants are realized.
It achieves highly scientific and accurate early warning under complex hydrogeological conditions, can dynamically predict future trends, and lowers the barrier to entry through simple level output, making it suitable for front-line personnel.
Smart Images
Figure CN122222196A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of early warning technology for soil and groundwater pollution, specifically relating to the field of risk management technology for soil and groundwater pollution, and particularly to an intelligent pollution early warning method that integrates hydrogeological and physical mechanisms with deep learning algorithms and is applied to non-ferrous metal industrial clusters. Background Technology
[0002] Industrial clusters are typically spatially clustered around a leading industry, with closely connected or mutually supportive enterprises. Early warning models are constructed primarily based on the leading industry, supplemented by supporting enterprises. Non-ferrous metal industrial clusters are high-risk areas for heavy metal pollution in soil and groundwater. These areas are typically characterized by numerous pollution sources, complex hydrogeological conditions, and concealed pollutant migration pathways. Establishing an efficient and accurate intelligent early warning system is crucial for preventing environmental risks and ensuring human safety. Existing soil and groundwater pollution early warning methods have the following limitations: The indicator system is static and isolated: Traditional technologies rely on pollutant concentration data from a single point in time and a single monitoring well, ignoring the cumulative effect of pollutants and their evolution over time, thus failing to provide early warning of future risks.
[0003] Insufficient integration of physical mechanisms and fragmented spatial correlation: Existing evaluation models usually assume that each monitoring point is independent of each other, ignoring the mobility of groundwater as a pollutant carrier; in reality, upstream pollution plumes migrate downstream with the groundwater flow field, and simple statistical models cannot capture this spatial topological relationship based on physical flow direction.
[0004] Rigid threshold classification: Traditional technologies usually use national standards or background values as fixed thresholds (such as alarms when a certain value is exceeded); however, the geological background of mining areas varies greatly, and relying solely on concentration exceeding the standard is often lagging behind; there is a lack of an adaptive classification mechanism that can comprehensively consider multiple factors such as source, pathway, and sink.
[0005] Therefore, given the current situation of multi-source complex pollution with strong concealment and complex hydrogeological structure in non-ferrous metal industrial clusters, there is an urgent need to develop an intelligent early warning method that can deeply couple the physical mechanism of groundwater seepage with the spatiotemporal evolution characteristics of multi-dimensional monitoring data, and can output intuitive graded conclusions for front-line management and control needs. Summary of the Invention
[0006] The purpose of this invention is to provide an intelligent early warning method for soil and groundwater pollution in non-ferrous metal industrial clusters. This method is based on the synergy of groundwater flow field topology and spatiotemporal characteristics. By constructing a deep coupling model of "physical hydrological topological constraints + spatiotemporal sequence recursive prediction", it breaks through the limitations of spatiotemporal separation and threshold rigidity of traditional methods, and realizes adaptive early warning that conforms to hydrogeological laws and is easy to operate.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: A smart early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster includes the following steps: S1: Establish the environmental feature vector space: The monitoring points and key units within the study area are mapped as objects in a vector space. For each object, multidimensional data including soil pollutant concentration, pollutant generation, processes and equipment, resource and energy utilization, waste recycling, water quality vulnerability, artificial seepage prevention, human health assessment, and groundwater functional value are extracted. The standardized components are then concatenated in sequence to obtain the standardized feature vector of each point.
[0008] For quantitative indicators related to soil pollutant concentration, water quality vulnerability, and human health assessment, a piecewise linear transformation method is used for standardization. The specific calculation logic is as follows: Based on the original classification boundary values in the relevant environmental quality standards, the original values are mapped to the 0-10 range; when the original values belong to the low-risk range, they are linearly mapped to the [0,4] range; when the original values belong to the medium-risk range, they are linearly mapped to the (4,7] range; when the original values belong to the high-risk range, they are linearly mapped to the (7,10] range. For indicators that only have qualitative classification descriptions, a discrete interval assignment method is used for standardization. The standard is described as low risk or level I and mapped to a representative value in the interval [0,4]; the standard is described as medium risk or level II and mapped to a representative value in the interval (4,7); the standard is described as high risk or level III and mapped to a representative value in the interval (7,10).
[0009] S2: Constructing a spatiotemporally coordinated intelligent pollution early warning model: A deep learning model comprising a temporal recursive layer, a spatial aggregation layer, and a hierarchical judgment layer is constructed to reconstruct features in the spatiotemporal dimensions of the initial feature vector. In the temporal dimension, the temporal recursive layer processes the time-series data of the object, extracts historical concentration change patterns through a gating mechanism, generates predicted concentration values for future moments, and dynamically updates the concentration components in the initial feature vector using these predicted values, forming a dynamic feature vector. In the spatial dimension, the spatial aggregation layer utilizes a topological structure constructed based on the groundwater flow field to perform calculations on the dynamic feature vector. The calculations on the dynamic feature vector are based on the groundwater flow direction and hydraulic connections, thereby calculating a comprehensive feature that integrates multiple locations. The comprehensive feature representation of each object is adaptively partitioned and distributed in the feature space. The warning level is determined based on the feature vector strength of the center point.
[0010] The specific construction process of the topology structure based on the groundwater flow field is as follows: The groundwater flow lines in the study area are calculated using a hydrogeological model. If the first object is located upstream of the groundwater flow of the second object, and there is a connected aquifer medium between the two, a directed connection edge is established from the first object to the second object to quantify the flux capacity of upstream pollutants migrating downstream with groundwater. The directed connection edge and the weight together constitute a physical constraint graph that restricts the computational path of the spatial aggregation layer.
[0011] The computational logic of the time recursion layer is as follows: A recurrent network structure with memory units is adopted to receive historical groundwater concentration data from multiple periods as input; redundant historical information is discarded through a forget gate, and current information is accepted through an input gate. The cell state is used to transmit long-term dependent features, and finally the concentration prediction value for the next time moment is output.
[0012] The operational logic of the spatial aggregation layer is as follows: For each target object, identify all its neighboring objects in the topology; multiply the dynamic feature vectors of all neighboring objects by their corresponding edge weights and perform a linear weighted summation; fuse the summation result with the target object's own dynamic feature vector and transform it through a nonlinear activation function to obtain the updated comprehensive feature representation of the target object, thereby simulating the physical diffusion and superposition process of pollutants along the groundwater flow field.
[0013] Compared with the prior art, the present invention has the following advantages: I. Deep integration of physical mechanism-driven and data-driven approaches: This invention differs from purely statistical data models by creatively utilizing "groundwater flow field topology" as a physical constraint. This enables the model's computation process to simulate the real physical process of pollutants migrating from upstream to downstream, significantly improving the scientific validity and accuracy of early warning under complex hydrogeological conditions.
[0014] II. Proactive Dynamic Early Warning: By introducing a time recursive network, traditional "static monitoring data" is transformed into "future trend prediction values". The graph convolutional neural network is combined with a hierarchical judgment mechanism to form a spatiotemporal pollution prediction model (LKGCN) that is not only based on the current concentration, but also on the future evolution trend, thus achieving true early warning.
[0015] Third, the results are simple and direct: Although a complex spatiotemporal collaborative algorithm is used internally, this invention directly outputs four clear levels of "no alarm, light alarm, medium alarm, and heavy alarm" through an adaptive mechanism. This end-to-end design greatly reduces the threshold for use, and front-line personnel can directly obtain the early warning conclusion without analyzing complex parameters.
[0016] IV. Integration of Multi-Dimensional Features Across the Entire Chain: A standardized indicator system covering the characteristics of the non-ferrous metal industry cluster area was constructed, and the problem of heterogeneous data integration was solved through normalization mapping, which comprehensively reflects the comprehensive environmental characteristics of the cluster area. Attached Figure Description
[0017] Figure 1 This is a flowchart of the present invention; Figure 2 This is a feature point map extracted from the non-ferrous metal industrial cluster area according to an embodiment of the present invention; Figure 3 The training and convergence process of the LSTM model is demonstrated. Figure 4 The comparison between the true and predicted values of the LSTM model is shown. Figure 5 This is a schematic diagram of the distribution of aquifer types in the non-ferrous metal industrial cluster area according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the groundwater flow direction in the non-ferrous metal industrial cluster area according to an embodiment of the present invention; Figure 7 It is a graph structure constructed based on the characteristics of the study area; Figure 8 The results are shown after prediction by the Space-Time Prediction Model (LKGCN). Figure 9 The diagram illustrates the warning results for the example. Detailed Implementation
[0018] A smart early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster includes the following steps: Figure 1 As shown; S1: Establish the environmental feature vector space: The monitoring points and key units within the study area are mapped as objects in a vector space. For each object, multidimensional data is extracted, including soil pollutant concentration, pollutant generation, processes and equipment, resource and energy utilization, waste recycling, water quality vulnerability, artificial seepage prevention, human health assessment, and groundwater functional value. The standardized mapped components are then sequentially concatenated to obtain the standardized feature vector for each point. The feature vectors of all points together constitute the feature matrix. H （0） ; For quantitative indicators related to soil pollutant concentration, water quality vulnerability, and human health assessment, a piecewise linear transformation method is used for standardization. The specific calculation logic is as follows: Based on the original classification thresholds in the relevant environmental quality standards, the original values are mapped to the 0-10 range. When the original value belongs to the low-risk range, it is linearly mapped to the [0,4] range; when the original value belongs to the medium-risk range, it is linearly mapped to the (4,7] range; when the original value belongs to the high-risk range, it is linearly mapped to the (7,10] range. For indicators with only qualitative classification descriptions, a discrete interval assignment method is used for standardization. Standards described as low-risk or Level I are mapped to representative values in the [0,4] range; standards described as medium-risk or Level II are mapped to representative values in the (4,7] range; and standards described as high-risk or Level III are mapped to representative values in the (7,10] range.
[0019] S2: Constructing a spatiotemporally coordinated intelligent pollution early warning model: A deep learning model comprising a temporal recursive layer, a spatial aggregation layer, and a hierarchical judgment layer is constructed to reconstruct features in the spatiotemporal dimensions of the initial feature vector. In the temporal dimension, the temporal recursive layer processes the time-series data of the object, extracts historical concentration change patterns through a gating mechanism, generates predicted concentration values for future moments, and dynamically updates the concentration components in the initial feature vector using these predicted values, forming a dynamic feature vector. In the spatial dimension, the spatial aggregation layer utilizes a topological structure constructed based on the groundwater flow field to perform calculations on the dynamic feature vector. The calculations on the dynamic feature vector are based on the groundwater flow direction and hydraulic connections, thereby calculating a comprehensive feature that integrates multiple locations. The comprehensive feature representation of each object is adaptively partitioned and distributed in the feature space. The warning level is determined based on the feature vector strength of the center point.
[0020] The specific construction process of the topology based on the groundwater flow field is as follows: Based on the hydrogeological conditions, the groundwater flow direction in the study area was determined, and two monitoring points were established within the study area. i and j If the location i Located at the point j If the groundwater flows upstream and the two points are adjacent or located in the same hydrogeological unit, then the point is determined. i right j There is a direct impact; if the location i Located at the point j If the downstream of the point or if there is no hydraulic connection between the two, it is determined that there is no direct influence. The directed connecting edges and points constitute the physical constraint graph of the spatial aggregation layer operation path, which is expressed as adjacency matrix elements. A ij (A value of 1 is assigned if there is a direct impact, otherwise 0).
[0021] The computational logic of the time recursion layer is as follows: A recurrent network structure including memory units is used to receive the groundwater pollutant concentration sequence of a specified point over the past T time steps. As input, redundant historical information is discarded through a forget gate, and current information is accepted through an input gate. Utilizing the long-term dependency characteristics of cell state propagation, the predicted groundwater pollutant concentration at that location at time T+1 is output. P pred The eigenvectors are reconstructed using the eigenvalues to build a dynamic eigenvector matrix. H （0） .
[0022] The operational logic of the spatial aggregation layer is as follows: For each target object, identify all its neighboring objects in the topology; multiply the dynamic feature vectors of all neighboring objects by their corresponding edge weights and perform a linear weighted sum; fuse the sum with the target object's own dynamic feature vector and transform it using a nonlinear activation function to obtain the updated comprehensive feature representation of the target object, expressed by the formula: ;in, It is an adjacency matrix. H （i） For the first i The feature matrix of the initial layer (the dynamic feature vector matrix obtained in step S2) H （0） ), W （i） σ is the weight matrix, and σ is the activation function, which is used to simulate the physical diffusion and superposition process of pollutants along the groundwater flow field.
[0023] The operational logic of the hierarchical determination layer is as follows: The spatiotemporal collaborative feature vectors output in step S2 are mapped to a high-dimensional feature space. The distribution and topological distance of each object in the space are calculated. Using the minimum distance iterative optimization algorithm, based on the intrinsic affinity of the feature vectors, the points in the entire domain are automatically converged and divided into four state sets with high internal consistency. The feature magnitude is analyzed, and this index directly quantifies the comprehensive intensity of the domain. The four domains are sorted according to the magnitude of the intensity. According to the physical meaning of the intensity from low to high, they are successively classified into four response levels: no alarm, light alarm, medium alarm, and heavy alarm.
[0024] Specific examples: Step 1: Data collection and organization; Data on a specific non-ferrous metal industrial cluster was collected, organized, and analyzed according to the indicator system. This included: the boundary of the cluster and surrounding sensitive targets; mining and beneficiation processes, pollutant emissions, ore recovery, energy consumption, and seepage prevention measures within the cluster; hydrogeological data on aquifer types, thickness, permeability coefficient, slope, and precipitation in the area; and existing monitoring well data within the cluster, including groundwater flow direction and pollutant concentration data.
[0025] Step 2: Determine the indicator scores based on the indicator system and construct the environmental feature vector; Based on the established indicators and indicator scoring rules, all data are organized into a matrix as input for subsequent models. The matrix is then standardized to ensure that the dimensions of each indicator are consistent.
[0026] The table below shows the original scores of indicators for a certain non-ferrous metal industrial cluster: serial number Pollutant generation Process and Equipment Resource and energy utilization Waste recycling Human health assessment Functional value of groundwater Artificial seepage prevention Water quality vulnerability Soil pollutant concentration (mg / kg) 1 Level 3 Secondary mineral processing Secondary mineral processing Mineral processing level 1 0.000054 Low / 76 15.77 2 Level 3 Secondary mineral processing Secondary mineral processing Mineral processing level 1 0.002238 Low / 7.84 89.26 3 / / / / 0.004652 Low / 128 157.40 4 Level 3 Secondary mineral processing Secondary mineral processing Mineral processing level 1 0.004857 Low Key seepage prevention 74 174.88 5 Level 3 Secondary mineral processing Secondary mineral processing Mineral processing level 1 0.005030 Low Simple waterproofing 136 171.39 6 / / / / 0.003120 Low / 88 116.13 7 Level 3 Secondary mineral processing Secondary mineral processing Mineral processing level 1 0.007750 Low General seepage prevention 90 203.64 8 / / / / 0.002612 Low / 76 110.99 9 / Mining Level 2 Mining Level 2 Mining Level 1 0.002247 Low / 85 104.87 10 / Mining Level 2 Mining Level 2 Mining Level 1 0.001719 Low / 108 90.17 11 / Mining Level 2 Mining Level 2 Mining Level 1 0.003717 Low / 69 157.20 12 / Mining Level 2 Mining Level 2 Mining Level 1 0.004825 Low / 72 166.21 13 / / / / 0.004600 high / 6.4 156.73 14 / / / / 0.003996 high / 5.59 140.20 15 Level 3 Secondary mineral processing Secondary mineral processing Secondary mineral processing 0.006102 high / 58 179.89 16 / Mining Level 2 Mining Level 2 Mining Level 2 0.001504 Low / 82 65.35 17 / Mining Level 2 Mining Level 2 Mining Level 2 0.001169 Low / 38 83.12 18 / / / / 0.005977 high / 62 174.13 19 Level 3 Secondary mineral processing Secondary mineral processing Secondary mineral processing 0.004663 Low / 73 161.70 Note: Arsenic (As), the most polluted pollutant in the study area, was selected as the characteristic pollutant to represent the soil pollutant concentration.
[0027] The table below shows the standardized scores of indicators for a certain non-ferrous metal mining industrial cluster: serial number Pollutant generation Process and Equipment Resource and energy utilization Waste recycling Human health assessment Functional value of groundwater Artificial seepage prevention Water quality vulnerability Soil pollutant concentration 1 8 6 6 3 6 2 10 0.6 2.3 2 8 6 6 3 10 4 10 7.8 5.1 3 1 1 1 1 10 3 10 8.2 7.7 4 9 6 6 3 10 2 2 0.1 8.3 5 9 6 6 3 10 2 5 9.4 8.2 6 1 1 1 1 10 3 10 2.2 6.1 7 9 6 6 3 10 3 4 2.5 9.4 8 1 1 1 1 10 2 10 0.6 5.9 9 1 5 5 2 10 3 10 1.8 5.7 10 1 5 5 2 10 3 10 5.2 5.1 11 1 5 5 2 10 3 10 0.1 8.6 12 1 5 5 2 10 2 10 0.1 8 13 1 1 1 1 10 10 10 6.4 7.6 14 1 1 1 1 10 8 10 5.6 7 15 8 5 5 2 10 8 10 0.1 8.5 16 1 5 5 2 10 3 10 1.3 4.2 17 1 5 5 2 10 3 10 0.1 4.9 18 1 1 1 1 10 8 10 0.1 8.3 19 8 5 5 2 10 2 10 0.1 7.8 Note: Pollutant generation, processes and equipment, resource and energy utilization, waste recycling, and artificial seepage prevention are qualitatively graded indicators, which are standardized using a discrete interval assignment method based on the specific conditions of the study area; Human health assessment, water quality vulnerability, and soil pollutant concentration are quantitative indicators, which are scored first and then linearly mapped; Groundwater functional value is a semi-quantitative indicator, and its score is obtained by matrix superposition after water abundance and current water quality rating, as detailed in step S1.
[0028] Step 3: Construct a spatiotemporal collaborative early warning model (LKGCN); The collected groundwater pollutant concentration data (Zn, Cd, Pb, Cu, As, locations as follows) over 6 years were compiled. Figure 2 As shown), the historical monitoring data of each location is processed using a Long Short-Term Memory (LSTM) network (the model training and convergence process is as follows). Figure 3 (As shown); the LSTM model captures long-term dependencies in time series data through its unique gating mechanism, predicts the changing trends of groundwater pollutant concentrations, and outputs predicted concentration values for future times. P pred ,like Figure 4 As shown; the initial environmental feature matrix is reconstructed into a dynamic feature matrix by selecting the year data for which the early warning is expected. In this example, the data for 2026 is selected.
[0029] Based on the hydrogeological data of the study area, the distribution of aquifers in the study area is delineated. For example... Figure 5 As shown and the direction of groundwater flow, as Figure 6 As shown; based on data from the actual study area, the hydraulic connections between each monitoring unit are analyzed. Using this as a physical constraint basis, a spatial topology is constructed, i.e., a graph structure as shown. Figure 7 As shown, nodes represent feature points, and directed edges represent the actual flow path of groundwater (from upstream to downstream), thus transforming discrete monitoring points into a holistic network that conforms to physical laws.
[0030] The aforementioned dynamic feature vectors are used as attribute inputs for graph nodes, and the "spatial topology" is used as the adjacency matrix input. A Graph Convolutional Network (GCN) is employed for computation. The GCN combines the topology with spatial aggregation of the feature vectors; specifically, based on the groundwater flow direction, it weights and transfers pollution features from upstream nodes to downstream nodes, simulating the physical migration and diffusion of pollutants in groundwater, generating deep, comprehensive features that integrate spatiotemporal multidimensional information. Adaptive state partitioning is performed within the high-dimensional deep feature manifold space constructed by the GCN. This strategy utilizes the spatial distribution of feature vectors and Euclidean distance to automatically identify and lock "risk homogeneous clusters" with high internal consistency, such as... Figure 8 As shown; by calculating the magnitude of the feature vector of each cluster center point (based on the evaluation system's actual physical meaning, a higher standardized score indicates a greater potential risk), each cluster is mapped to a preset early warning level system; the system directly outputs the following four warning levels, and the early warning results for the study area are obtained through visualization analysis, such as... Figure 9 As shown: No alarm: The cluster with the lowest corresponding feature intensity represents a safe regional environment or is at the background value level; Light alert: Clusters with lower characteristic intensity indicate the presence of minor pollution disturbances but no risk. Medium alert: Clusters with moderate intensity of pollution characteristics indicate an upward trend in pollution or localized exceedances, requiring close monitoring; Severe Alert: The cluster with the highest corresponding feature intensity indicates serious pollution and a risk of spread, requiring immediate control measures.
Claims
1. A method for intelligent early warning of soil and groundwater pollution in a non-ferrous metal industrial cluster, characterized in that, Includes the following steps: S1: Establish the environmental feature vector space: The monitoring points and key units within the study area are mapped to objects in a vector space. For each object, multidimensional data including soil and groundwater pollutant concentrations, pollutant generation, processes and equipment, resource and energy utilization, waste recycling, water quality vulnerability, artificial seepage prevention, human health assessment, and groundwater functional value are extracted and combined into an initial feature vector after standardization. S2: Constructing a spatiotemporally coordinated intelligent pollution early warning model: A deep learning model comprising a temporal recursive layer, a spatial aggregation layer, and a hierarchical decision layer is constructed to reconstruct features in the spatiotemporal dimensions of the initial feature vector. In the temporal dimension, the temporal recursive layer processes the time-series data of the object, extracts historical concentration change patterns through a gating mechanism, generates predicted concentration values for future moments, and dynamically updates the concentration components in the initial feature vector using these predicted values, forming a dynamic feature vector. In the spatial dimension, the spatial aggregation layer utilizes a topological structure constructed based on the groundwater flow field to perform calculations on the dynamic feature vector. This calculation, based on the groundwater flow direction and hydraulic connections, calculates a comprehensive feature integrating multiple locations. The comprehensive feature representation of each object is adaptively divided into its distribution in the feature space; the warning level is determined based on the feature vector strength of the center point.
2. The intelligent early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster area according to claim 1, characterized in that: In step S1, for quantitative indicators of soil pollutant concentration, water quality vulnerability, and human health assessment, a piecewise linear transformation method is used for standardization. The specific calculation logic is as follows: based on the original classification boundary values in the relevant environmental quality standards, the original values are mapped to the 0-10 interval; when the original values belong to the low-risk interval, they are linearly mapped to the [0,4] interval; when the original values belong to the medium-risk interval, they are linearly mapped to the (4,7] interval; when the original values belong to the high-risk interval, they are linearly mapped to the (7,10] interval. For indicators with only qualitative classification descriptions, a discrete interval assignment method is used for standardization. Standards described as low-risk or Level I are mapped to representative values in the [0,4] interval; standards described as medium-risk or Level II are mapped to representative values in the (4,7] interval; and standards described as high-risk or Level III are mapped to representative values in the (7,10] interval.
3. The intelligent early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster area according to claim 1, characterized in that: In step S2, the specific construction process of the topology based on the groundwater flow field is as follows: The groundwater flow lines in the study area are calculated using a hydrogeological model. If the first object is located upstream of the groundwater flow of the second object, and there is a connected aquifer medium between the two, a directed connection edge is established from the first object to the second object to quantify the flux capacity of upstream pollutants migrating downstream with groundwater. The directed connection edge and the weight together constitute a physical constraint graph that restricts the computational path of the spatial aggregation layer.
4. The intelligent early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster area according to claim 1, characterized in that: In step S2, the operation logic of the time recursion layer is as follows: A recurrent network structure with memory units is adopted to receive historical groundwater concentration data from multiple periods as input; redundant historical information is discarded through a forget gate, and current information is accepted through an input gate. The cell state is used to transmit long-term dependent features, and finally the concentration prediction value for the next time moment is output.
5. The intelligent early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster area according to claim 1, characterized in that: In step S2, the operational logic of the spatial aggregation layer is as follows: For each target object, identify all its neighboring objects in the topology; multiply the dynamic feature vectors of all neighboring objects by their corresponding edge weights and perform a linear weighted summation; fuse the summation result with the target object's own dynamic feature vector and transform it through a nonlinear activation function to obtain the updated comprehensive feature representation of the target object, thereby simulating the physical diffusion and superposition process of pollutants along the groundwater flow field.
6. The intelligent early warning method for soil and groundwater pollution in a non-ferrous metal industrial cluster area according to claim 1, characterized in that: In step S2, the operation logic of the hierarchical determination layer is as follows: The spatiotemporal collaborative feature vectors output in step S2 are mapped to a high-dimensional feature space. The distribution and topological distance of each object in the space are calculated. Using the minimum distance iterative optimization algorithm, based on the intrinsic affinity of the feature vectors, the points in the entire domain are automatically converged and divided into four state sets with high internal consistency. The feature magnitude is analyzed, and this index directly quantifies the comprehensive intensity of the domain. The four domains are sorted according to the magnitude of the intensity. According to the physical meaning of the intensity from low to high, they are successively classified into four response levels: no alarm, light alarm, medium alarm, and heavy alarm.