An artificial intelligence-based groundwater pollution source identification method and system

Abstract

This invention relates to the field of machine learning technology, specifically to an artificial intelligence-based method and system for identifying groundwater pollution sources. The method includes the following steps: acquiring concentration data and hydrological parameters to calculate spatial connectivity vectors; performing differential processing to generate node representations; filtering concentration sequences based on retardation factors to generate time segments and performing reverse ordering; fusing transfer representations and reverse-order segments to generate spatiotemporal aggregated content; and generating aggregated coordinates based on the spatiotemporal aggregated content for distance filtering to obtain results. In this invention, by acquiring three-dimensional coordinates and multi-source hydrological data, graph structure construction and matrix differential processing are performed to deeply analyze the intrinsic correlation of the hydrodynamic network. Sequences are filtered by retardation factors and integrated in reverse order to generate time segments. Aggregated content of spatiotemporal decay trajectories is constructed by fusing diffusion features. Reverse mapping is performed according to the dissipation law to locate the source, avoiding local misinterpretations caused by numerical fitting and effectively improving the accuracy and efficiency of tracing the source of solute leakage in the groundwater layer.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method and system for identifying groundwater pollution sources based on artificial intelligence. Background Technology

[0002] The field of machine learning technology focuses on the autonomous evolution of computational logic through statistical laws and historical sample sets. It relies on physical computing nodes to construct multi-layer neural computing networks or multi-branch tree-like judgment structures. At the receiving end, a large-scale sample matrix containing feature values ​​is input. The numerical features of each layer are aggregated through forward propagation, and the deviation between the true value and the predicted value is calculated using a loss function. Then, the connection weights and bias values ​​between the nodes of each layer are finely adjusted in reverse along the gradient direction. Finally, a non-linear mathematical function relationship is fitted from the input parameters to the output parameters. Traditional AI-based groundwater pollution source identification methods refer to the technical aspects of retrieving the leakage points and release flux of chemical solutes within underground unconfined layers. These methods typically involve sampling at a fixed depth in physical monitoring wells with known coordinates and obtaining the concentration sequence of pollutants over time using chemical analysis instruments. This concentration sequence, along with the permeability, porosity, and hydraulic gradient of the aquifer, is used as input features and fed into a backpropagation multilayer perceptron for calculation. Latin hypercube sampling is then used to generate multiple sets of assumed pollution source latitude and longitude coordinates and release rate combinations. Subsequently, the root mean square error (RMSE) is calculated based on the simulated concentration values ​​from each monitoring well derived from the forward inference of the multilayer perceptron and the actual concentration values ​​obtained from on-site sampling. Through multiple iterations, the source latitude and longitude coordinates and mass release rate that minimize this RMS error are selected.

[0003] Existing groundwater pollution source identification technologies employ a sampling-based approach to generate assumed coordinate combinations and perform iterative forward extrapolation and root mean square error calculations. When faced with complex aquifers and variable hydraulic gradient conditions, these technologies incur enormous computational overhead and lead to lengthy iterative search processes. Relying solely on macroscopic error fitting of simulated concentration values ​​can easily cause the overall extrapolation to fall into local optima, making it difficult to accurately depict the actual diffusion and evolution paths and attenuation patterns of pollutants in three-dimensional space. Consequently, this severely hinders the accurate tracking and efficient location of groundwater solute leakage points. Summary of the Invention

[0004] To achieve the above objectives, the present invention adopts the following technical solution: a groundwater pollution source identification method based on artificial intelligence, comprising the following steps: S1: Obtain concentration information from monitoring wells and acquire three-dimensional spatial coordinate data, permeability data, and water flow gradient vector. Perform differential analysis based on the three-dimensional spatial coordinate data and use the permeability data to obtain an attenuation weight sequence. Calculate and generate a spatial connection vector from the attenuation weight sequence. S2: Obtain the spatial connection vector and input it into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix. Generate node representations based on the three-dimensional spatial coordinate data and the connection structure matrix. S3: Obtain pollutant concentration sequences and detect blocking factors; based on the node expression, filter pollutant concentration sequences to obtain filtered fragments; and based on the filtered fragments, determine the time span to generate time segments. S4: Perform time inversion calculation on the time segment according to the inverse convection dispersion equation, obtain dispersion field information and groundwater flow direction information, and calculate the migration expression. Then, merge the inverse segment with the migration expression to generate spatiotemporal aggregated content. S5: Utilize the spatiotemporal aggregated content and the node expression to perform spatial mapping to generate aggregated coordinates, obtain the coordinates of suspected leak points and distance screening thresholds, and perform distance screening based on the aggregated coordinates and the coordinates of suspected leak points to generate groundwater pollution source identification results.

[0005] As a further embodiment of the present invention, the spatial connection vector includes spatial correlation, attenuation coefficient, and directional weight; the node representation includes node spatial features, node connection weight, and node structural factor; the time segment includes start time, end time, and duration; the spatiotemporal aggregation content includes reverse time information, migration feature information, and fusion expression parameters; and the groundwater pollution source identification result includes source location, distance threshold determination item, and risk assessment index.

[0006] As a further aspect of the present invention, the migration expression obtains the migration process and distribution characteristics of pollutants in groundwater by calculating dispersion field information and groundwater flow direction information.

[0007] As a further aspect of the present invention, the screening segment is determined by screening time periods from the pollutant concentration sequence based on the node expression, thereby judging the changes and temporal characteristics of the pollutants.

[0008] As a further aspect of the present invention, the specific steps of S1 are as follows: S101: Obtain three-dimensional spatial coordinate data and water flow gradient vector, construct a spatial grid based on the coordinate data, calculate the difference between adjacent points according to the grid and correct the water flow gradient direction, extract the direction change amount and form a direction consistency factor, and generate a coordinate difference matrix; S102: Call the coordinate difference matrix and permeability data, convert the difference components according to the permeability coefficient set and obtain the attenuation factor sequence, perform corresponding correction on the attenuation factor sequence and coordinate difference matrix and update the directional attenuation relationship to obtain the attenuation weight sequence; S103: Establish a weighted correlation value based on the attenuation weight sequence and the coordinate difference matrix, aggregate the directional components according to the correlation value to obtain the directional synthesis rate, and call the directional synthesis rate to perform concatenation and normalization processing on the directional sequence of the sampling points to generate a spatial connection vector.

[0009] As a further aspect of the present invention, the specific steps of S2 are as follows: S201: Obtain the spatial connection vector and input it into the graph structure construction process, perform inter-segment difference on the connection vector and retrieve the direction change of adjacent segments, filter data components that exceed the preset direction deviation threshold according to the fixed difference reference value and rearrange the remaining components to generate a connection structure matrix. S202: Call the connection structure matrix and three-dimensional spatial coordinate data, perform component splitting and linear superposition on the coordinates corresponding to the matrix index, form a coordinate association set based on the superposition result, and obtain the node coordinate mapping value; S203: Based on the node coordinate mapping value and the connection structure matrix, perform vector aggregation on the matrix node positions and filter out components with amplitude values ​​lower than a preset vector amplitude threshold. Perform directional merging and sequence recombination on the aggregation results to generate node representations.

[0010] As a further aspect of the present invention, the specific steps of S3 are as follows: S301: Collect pollutant concentration sequences and detect retardation factors, retrieve time components of concentration sequences and determine the range of numerical changes based on retardation factors, remove numerical components that are lower than the fixed change benchmark value based on a fixed change benchmark value, and generate a concentration screening matrix. S302: Call the node expression and the concentration screening matrix, perform numerical comparison on the node expression corresponding to the screening matrix index, extract the segment interval that meets the preset numerical difference threshold based on the comparison difference, and aggregate the continuous interval to obtain the screening segment; S303: Based on the selected segments, retrieve the segment time range and calculate the start and end time difference. Compare the time difference with a preset fixed time span threshold and retain the matching segments. Perform time series recombination on the retained segments to generate time segments.

[0011] As a further aspect of the present invention, the specific steps of S4 are as follows: S401: Obtain the time segment and retrieve the time tag, perform reverse order rearrangement of the time tag while maintaining the start and end correspondence, recombine the segment intervals according to the reverse order position and remove inconsistent intervals to generate reverse order segments; S402: Collect diffuse field information and groundwater flow direction information, detect the numerical gradient of diffuse field information and compare it with the directional component of groundwater flow direction information, distinguish the effective components according to the preset directional consistency threshold, and perform weighted superposition of the effective components to obtain the transfer expression. S403: Based on the reverse sequence fragment and the migration expression, match the corresponding time period value of the migration expression for the time segment of the reverse sequence fragment and perform numerical fusion, retrieve the component differences of the fused sequence and filter out abnormal components according to the fixed amplitude benchmark value, reconstruct the remaining sequence and establish the time and migration relationship structure to generate spatiotemporal aggregated content.

[0012] As a further aspect of the present invention, the specific steps of S5 are as follows: S501: Call the spatiotemporal aggregated content and the node expression, retrieve the positional component in the spatiotemporal aggregated content and perform mapping according to the spatial index of the node expression, perform linear conversion on the mapping result according to a fixed weight factor and perform directional merging on the conversion sequence to generate aggregated coordinates; S502: Based on the aggregated coordinates, collect the coordinates of suspected leakage points from the hydrodynamic parameter group of the aquifer, detect the spatial components of the coordinates of the suspected leakage points, define the distance reference interval according to the preset threshold, compare the distance reference interval with the numerical range of the aggregated coordinates and form a distance judgment sequence to obtain the leakage distance screening value. S503: Based on the leakage distance screening value and the coordinates of the suspected leakage points, perform coordinate aggregation on the points that meet the benchmark interval in the distance screening value and remove those that do not meet the benchmark interval, reorganize the remaining coordinate sequence and establish a point association set to generate groundwater pollution source identification results.

[0013] An artificial intelligence-based groundwater pollution source identification system includes: The attenuation weight and spatial connection module is used to implement S1: acquiring monitoring well concentration information and acquiring three-dimensional spatial coordinate data, permeability data and water flow gradient vector; performing differential analysis based on three-dimensional spatial coordinate data and using permeability data to obtain an attenuation weight sequence; and calculating and generating a spatial connection vector from the attenuation weight sequence. The node representation and structure matrix module is used to implement S2: obtain the spatial connection vector and input it into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix, and generate node representation based on the three-dimensional spatial coordinate data and the connection structure matrix; The pollutant screening and time segmentation module is used to implement S3: acquire pollutant concentration sequences and detect retardation factors, screen pollutant concentration sequences based on the node expression to obtain screening segments, and determine the time span based on the screening segments to generate time segments. The spatiotemporal aggregation and migration fusion module is used to implement S4: performing time inversion calculation on the time segment according to the inverse convection dispersion equation, obtaining dispersion field information and groundwater flow direction information, and calculating the migration expression, and fusing the reverse segment with the migration expression to generate spatiotemporal aggregated content; The pollution source identification and distance screening module is used to implement S5: using the spatiotemporal aggregated content and the node expression to perform spatial mapping to generate aggregated coordinates, obtain the coordinates of suspected leakage points and distance screening thresholds, and perform distance screening based on the aggregated coordinates and the coordinates of suspected leakage points to generate groundwater pollution source identification results.

[0014] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, by acquiring three-dimensional spatial coordinates and multi-source hydrological data, and performing graph structure construction and matrix difference processing based on permeability characteristics, the characteristics of the medium are deeply analyzed, and the intrinsic correlation of the hydrodynamic network is revealed at the node expression level. Concentration sequences are screened by combining resistance factors and inversely integrated to generate time segments. Aggregated content containing spatiotemporal decay trajectories is constructed by fusing dispersion field and flow direction characteristics. The inefficient framework of blind sampling iteration is abandoned, and reverse mapping positioning is performed according to the physical dissipation law of pollutants. This avoids local misleading and delay caused by numerical fitting. Accurate distance verification is completed under the constraint of suspected points by relying on multi-dimensional features, which effectively improves the analytical accuracy and execution efficiency of tracing the source of solute leakage in the groundwater layer. Attached Figure Description

[0015] 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 these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the steps of the present invention; Figure 2 This is a detailed schematic diagram of S1 of the present invention; Figure 3 This is a detailed schematic diagram of S2 of the present invention; Figure 4 This is a detailed schematic diagram of S3 of the present invention; Figure 5 This is a detailed schematic diagram of S4 of the present invention; Figure 6 This is a detailed schematic diagram of S5 of the present invention; Figure 7 This is a system module diagram of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0018] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0019] Please see Figure 1 This invention provides an artificial intelligence-based method for identifying groundwater pollution sources, comprising the following steps: S1: By collecting concentration data from monitoring wells deployed in the aquifer area, and obtaining three-dimensional spatial coordinate data, permeability data and water flow gradient vector, differential analysis is performed based on the three-dimensional spatial coordinate data, the attenuation weight sequence is obtained by calling the permeability data, and the spatial connection vector is generated by performing calculation on the attenuation weight sequence. S2: Obtain the spatial connection vector and input it into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix. Then, call the three-dimensional spatial coordinate data and the connection structure matrix to perform matrix operations to generate node representations. S3: By collecting pollutant concentration sequences and detecting blocking factors, the pollutant concentration sequences are screened according to node expression to obtain screened segments. Time span is judged based on the screened segments to generate time segments. S4: Generate reverse-order segments by performing reverse processing on the time tags of time segments, collect diffusion field information and groundwater flow direction information, and perform calculations based on the diffusion field information and groundwater flow direction information to obtain transfer expressions. Then, fuse the reverse-order segments and transfer expressions to generate spatiotemporal aggregated content. S5: By calling the spatiotemporal aggregated content and node expression to perform spatial mapping calculation to obtain aggregated coordinates, collect the coordinates and thresholds of suspected leakage points from the aquifer hydrodynamic parameter group, and perform distance filtering based on the aggregated coordinates and the coordinates of suspected leakage points to generate groundwater pollution source identification results.

[0020] The spatial connection vector includes spatial correlation, attenuation coefficient, and directional weight; the node representation includes node spatial features, node connection weight, and node structural factor; the time segment includes start time, end time, and duration; the spatiotemporal aggregation content includes reverse time information, migration feature information, and fusion expression parameters; and the groundwater pollution source identification results include source location, distance threshold judgment item, and risk assessment index.

[0021] Please see Figure 2 The specific steps of S1 are as follows: S101: Obtain three-dimensional spatial coordinate data and water flow gradient vector, construct a spatial grid based on the coordinate data, calculate the difference between adjacent points according to the grid and correct the water flow gradient direction, extract the direction change amount and form a direction consistency factor, and generate a coordinate difference matrix; The process involves using high-precision global positioning to obtain three-dimensional spatial coordinate data for groundwater monitoring wells in various locations. For example, the x-coordinate of a monitoring well in the regional coordinate system is 34512.23 meters, the y-coordinate is 45612.85 meters, and the bottom elevation is -12.50 meters. Simultaneously, a thermal pulse current meter is used to collect water flow gradient vectors at different depths within the aquifer, with a sampling interval of 0.5 meters, to obtain a dataset containing flow velocity amplitude and flow direction azimuth. For example, at a depth of 15.5 meters, the flow velocity is measured to be 0.85 meters per day, and the flow direction azimuth is 125.4 degrees. This discrete coordinate data is then input into a spatial interpolation program, with a grid resolution of 10 meters horizontally, 10 meters vertically, and 2 meters in the vertical direction, constructing a three-dimensional cuboid grid covering the survey area. For each node in the grid, the coordinate differences between it and its six adjacent nodes in the horizontal, vertical, and height axes are calculated. For example, the difference between the coordinates of the center node and the node to the east on the horizontal axis is 10.05 meters, the difference with the node to the north on the vertical axis is 9.98 meters, and the difference with the node below on the vertical axis is 2.01 meters. The spatial vector formed by these differences is then compared with the water flow gradient vector at that location using cosine similarity calculation to obtain the directional deviation. If the cosine similarity value is below 0.85, it is determined that the water flow at that point is significantly affected by heterogeneity, and correction based on local permeability changes is required, calculating the gradient direction deflection angle. The extracted directional deflection angle is converted to a value between 0 and 1 using a Gaussian function, forming a directional consistency factor. For example, when the deflection angle is 5 degrees, the consistency factor is calculated to be 0.98. Finally, the coordinate difference components of all adjacent nodes in the grid are arranged in a three-dimensional tensor form to generate a coordinate difference matrix containing geometric spacing information, where the element values ​​of the matrix represent the physical distance and relative orientation between spatial nodes.

[0022] Table 1. Three-dimensional spatial parameters of monitoring points; As shown in Table 1, collecting geometric and dynamic data from different locations provides a foundation for subsequent spatial grid construction. The advantage of this step is that by introducing a directional consistency factor, the deviation of water flow paths in complex strata is effectively quantified, enabling the coordinate difference matrix to describe the dynamic characteristics of the water flow.

[0023] S102: Call the coordinate difference matrix and permeability data, convert the difference components according to the permeability coefficient set and obtain the attenuation factor sequence, perform corresponding correction on the attenuation factor sequence and coordinate difference matrix and update the directional attenuation relationship to obtain the attenuation weight sequence. The process retrieves the distance component from the obtained coordinate difference matrix and simultaneously retrieves the permeability data of the corresponding aquifer from the laboratory permeability experiment database. The permeability data includes the permeability coefficients of different lithological layers; for example, the permeability coefficient of fine sand is 5.5 m / day, and that of medium sand is 12.8 m / day. This process performs a conversion on each coordinate difference component based on the permeability coefficient set, transforming the geometric distance into equivalent hydraulic resistance. For example, when the physical distance between two points is 10 meters and they are located in an area with a permeability coefficient of 5 m / day, the equivalent resistance value is calculated as the ratio of distance to permeability coefficient, i.e., 2.0 days. An exponential decay operation is performed on the resistance distribution across the entire grid to obtain a decay factor sequence. The decay factor is set between 0.2 and 0.9; for example, in the low-permeability clay impermeable layer region, the decay factor is set to 0.35, while in the high-permeability gravel layer region, the decay factor is set to 0.82. Each value in the decay factor sequence is multiplied by the corresponding coordinate component in the coordinate difference matrix to correct the original geometric connection strength. If a path passes through a low-permeability region, its corresponding matrix components will decrease due to multiplication by a down-adjusted attenuation factor, updating the spatial directional attenuation relationship. Finally, row normalization is performed on the corrected matrix, scaling the sum of elements in each row to 1, thus obtaining the attenuation weight sequence. For example, the weights of the three adjacent directions of a node are 0.45, 0.35, and 0.20, reflecting the ease with which pollutants migrate in that direction.

[0024] S103: Establish a weighted correlation value based on the attenuation weight sequence and the coordinate difference matrix, aggregate the directional components based on the correlation value to obtain the directional synthesis rate, and call the directional synthesis rate to perform concatenation and normalization processing on the directional sequence of the sampling points to generate a spatial connection vector. Based on the weight values ​​in the generated attenuation weight sequence, a weighted multiplication is performed with the corresponding displacement components in the coordinate difference matrix, and the summation establishes a weighted correlation value. For example, if the weight from node 1 to node 2 is 0.6 and the coordinate displacement component is 10 meters, then its weighted correlation value is 6.0. This process performs vector synthesis on the components of each spatial direction based on the weighted correlation value, calculates the magnitude and direction of the synthesized displacement vector, and thus obtains the directional synthesis rate. The directional synthesis rate represents the intensity of the dominant direction of water flow and solute transport at that point; the closer its value is to 1, the more obvious the dominant direction. Subsequently, the obtained directional synthesis rate is used to perform concatenation processing on the time-series directional data collected from the sampling points, connecting the discrete time-point directions into a continuous path trajectory. For the concatenated trajectory sequence, minimum-maximum normalization processing is performed, mapping all coordinate components to the interval between -1 and +1 to ensure the comparability of displacements of different orders of magnitude. Finally, the normalized directional components are arranged into a vector form according to time sequence to generate a spatial connection vector. This vector not only includes the geometric topological relationship between monitoring points but also incorporates hydrodynamic attenuation characteristics.

[0025] Please see Figure 3 The specific steps of S2 are as follows: S201: Obtain the spatial connection vector and input it into the graph structure construction process. Perform inter-segment difference on the connection vector and retrieve the direction change of adjacent segments. Filter data components that exceed the preset direction deviation threshold based on the fixed difference reference value and rearrange the remaining components to generate the connection structure matrix. The spatial connectivity vector generated in the preceding steps is obtained and passed as input parameters to the graph structure construction process. This process performs a difference operation on adjacent time-segment components in the spatial connectivity vector, calculating their angle change in three-dimensional space. For example, if the vector direction in the previous time segment is 45 degrees and in the next time segment is 48 degrees, the change in direction is 3 degrees. The direction change characteristics of adjacent segments are retrieved, and filtering is performed based on a pre-set fixed difference benchmark value, which is set to 15 degrees according to the aquifer heterogeneity coefficient. If the direction change of a segment exceeds 15 degrees, it is determined that the segment is subject to abnormal interference such as local fissures or artificial pumping, and data components exceeding a preset direction deviation threshold are removed. For the retained vector components that meet the smooth migration characteristics, they are rearranged in chronological order to ensure path continuity. The rearranged vector components are filled into the non-zero positions of the adjacency matrix to generate the connectivity structure matrix. For example, a value of 0.88 in the 2nd row and 5th column of the matrix indicates a strong spatial topological connection between the 2nd monitoring node and the 5th potential node; this process improves the reliability of the connectivity structure by filtering noisy data.

[0026] S202: Call the connection structure matrix and three-dimensional spatial coordinate data, perform component splitting and linear superposition on the coordinates corresponding to the matrix index, form a coordinate association set based on the superposition result, and obtain the node coordinate mapping value; The generated connection structure matrix and the initially acquired 3D spatial coordinate data are used. For each node pair corresponding to a non-zero index in the matrix, coordinate component decomposition is performed. The horizontal, vertical, and vertical components of the starting and ending coordinates are extracted and subjected to a weighted linear superposition operation. The superposition coefficient is determined based on the values ​​in the connection structure matrix. For example, if the connection strength is 0.8, the starting coordinate contributes a weight of 0.2, and the ending coordinate contributes a weight of 0.8. By performing this weighted summation on the coordinates of multiple adjacent nodes, a coordinate association set is formed. This process calculates the centroid projection of the coordinate values ​​within the association set to obtain the node coordinate mapping values. For example, the mapped horizontal coordinate of a certain cluster node is 34520.15 meters. This value is not the original monitoring point coordinate, but the spatial centroid position derived based on the flow field connection relationship.

[0027] S203: Based on the node coordinate mapping values ​​and the connection structure matrix, perform vector aggregation on the matrix node positions and filter out components with amplitudes lower than the preset vector amplitude threshold. Perform directional merging and sequence recombination on the aggregation results to generate node representations. Based on the calculated node coordinate mapping values ​​and the connection structure matrix, vector aggregation is performed on the positions of all active nodes in the matrix. All spatial connection vectors pointing to the same node are summed to calculate their combined pointing strength. For the aggregated vector magnitude, a screening threshold of 0.15 is set for components with amplitudes lower than a preset vector amplitude threshold, eliminating invalid connections caused by weak water flow fluctuations. The retained high-amplitude aggregation results are then subjected to direction merging, combining vectors with similar angles into a single representative vector. Subsequently, the merged vector sequence is reorganized and arranged according to spatial topological hierarchy to generate node representations. Node representations exist in the form of multi-dimensional feature vectors, containing the node's geometric position information and topological importance features in the flow field. For example, the first digit of a node's feature vector represents its projection length in the mainstream direction as 12.5 units.

[0028] Please see Figure 4 The specific steps of S3 are as follows: S301: Collect pollutant concentration sequences and detect retardation factors, retrieve time components of concentration sequences and determine the range of numerical changes based on retardation factors, remove numerical components that are lower than the fixed change benchmark value based on a fixed change benchmark value, and generate a concentration screening matrix. Pollutant concentration sequence data is retrieved from automatic monitoring stations at a frequency of once per hour. Key monitoring indicators include chloride ion concentration and total dissolved solids. Simultaneously, the blocking factor of the soil and aquifer media for pollutants is measured. This factor is obtained through soil column experiments; for example, the blocking factor for a certain heavy metal pollutant in sandy soil is 1.5. The concentration components at each time point in the concentration sequence are retrieved, and the magnitude of the numerical change between adjacent time points is calculated. A reasonable boundary for the range of numerical change is set based on the blocking factor; if the blocking factor is large, the concentration change should be relatively gradual. A fixed baseline value of 5.0 mg / L per hour is set. If the measured concentration change exceeds this baseline value and does not conform to the blocking characteristics, it is determined to be an instrument error or abnormal fluctuation. After removing these abnormal data points that do not meet the criteria, the remaining effective concentration data sequence is rearranged. The rearranged concentration values ​​are combined with their corresponding time indices to generate a concentration screening matrix. For example, row 10 of the matrix records an effective concentration value of 25.6 mg / L for the 10th hour.

[0029] Table 2. Pollutant Concentration Screening Parameters; As shown in Table 2, a retardation factor was introduced to verify the rationality of concentration fluctuations. The advantage of this process is that it eliminates spurious peak data caused by external interference, ensuring that the concentration series entering subsequent analysis can truly reflect the actual transport patterns of pollutants in groundwater.

[0030] S302: Call the node expression and concentration screening matrix, perform numerical comparison on the node expression corresponding to the screening matrix index, extract the segment interval that meets the preset numerical difference threshold based on the comparison difference, and aggregate the continuous interval to obtain the screened segment; For each time point in the screening matrix corresponding to the concentration value, the associated node expression features are retrieved. A numerical comparison process is performed to calculate the concentration gradient between adjacent nodes. A concentration change threshold of 10% is set; if the concentration difference between two associated nodes exceeds 10%, the segment is determined to be in the main front region of pollutant migration. Segment intervals that meet this threshold condition are extracted, and spatially continuous and temporally connected segments are aggregated. For example, continuous high-gradient segments from hour 5 to hour 12 are extracted and merged into a single screening segment. After obtaining the screening segment, the set of node numbers it covers and the corresponding concentration change trend are recorded to provide spatiotemporal constraints for subsequent source tracing analysis.

[0031] S303: Based on the selected segments, retrieve the segment time range and calculate the start and end time difference. Compare the time difference with the preset fixed time span threshold and retain the matching segments. Perform time series recombination on the retained segments to generate time segments. Based on the metadata of each selected segment, its corresponding time range label is retrieved. The duration of each segment is calculated by subtracting the start time from the end time. This time difference is compared to a preset fixed time span threshold, which is set to 24 hours based on groundwater flow velocity and regional scale. Segments with a duration exceeding 24 hours are retained as representative pollution migration events, while short, impulsive interference segments are discarded. Time series reconstruction is performed on all eligible retained segments, rearranging their concentration and location information in chronological order. Time segment data is generated, describing the complete evolution of the pollution plume within a specific time period.

[0032] Please see Figure 5 The specific steps of S4 are as follows: S401: Obtain time segments and retrieve time tags, perform reverse order rearrangement of time tags while maintaining the start and end correspondence, recombine segment intervals according to the reverse order position and remove inconsistent intervals to generate reverse order segments; The process involves acquiring generated time segments and retrieving their associated time tags. Because reverse tracing is required, all time tags are rearranged in reverse order; for example, data from January 7th is placed first, while earlier data from January 1st is placed last. During this rearrangement, the start and end correspondences of coordinates, concentrations, and other attributes at each moment are strictly maintained to ensure the physical meaning remains intact. The segment intervals are then combined based on their reversed positions, and the spatial continuity of adjacent reversed intervals is checked, eliminating isolated intervals that exhibit logical inconsistencies in the reverse flow calculation. Finally, reverse-order segments are generated, laying the foundation for subsequent calculations of the pollutant's trajectory from the monitoring point to the potential source.

[0033] S402: Collect diffuse field information and groundwater flow direction information, detect the numerical gradient of diffuse field information and compare it with the directional component of groundwater flow direction information, distinguish the effective components according to the preset directional consistency threshold, and perform weighted superposition of the effective components to obtain the transfer expression. Dispersion field information, including longitudinal and lateral dispersion, was collected through field dispersion experiments. For example, the measured longitudinal dispersion value was 2.5 meters. Simultaneously, groundwater flow direction information was retrieved to obtain the average flow direction vector at each measuring point. This process involved real-time monitoring of the spatial distribution of the concentration gradient within the dispersion field and performing a cosine comparison with the directional component of the groundwater flow direction. A directional consistency threshold of 0.75 was set; if the cosine value of the deviation angle between the dispersion gradient direction and the water flow direction was greater than 0.75, it was defined as an effective migration component. All effective components were weighted and superimposed, with weights allocated according to the magnitude of the dispersion coefficient; the larger the coefficient, the higher the weight. Through superposition processing, a migration expression was obtained. The migration expression is a tensor reflecting the combined effects of solute diffusion and convection in a complex flow field. For example, a migration intensity value of 0.42 at a certain point represents the degree of contribution of that location to pollutant diffusion.

[0034] S403: Based on the reverse sequence fragments and the transfer expression, match the corresponding time period values ​​of the transfer expression for the time segment of the reverse sequence fragments and perform numerical fusion, retrieve the component differences of the fused sequence and filter out abnormal components according to the fixed amplitude benchmark value, reconstruct the remaining sequence and establish the time and transfer relationship structure to generate spatiotemporal aggregated content; Based on the generated reverse-order fragments and migration expression data, for each discrete time interval in the reverse-order fragments, the corresponding field value in the migration expression is matched. Numerical fusion processing is performed, multiplying the displacement vector of the reverse order by the migration intensity to estimate the possible location of the pollutant at historical moments. Component differences in the fused sequence are retrieved, and the displacement deviation between adjacent backtracking steps is calculated. Screening is performed based on a fixed amplitude benchmark value; if the backtracking displacement exceeds 50 meters and exceeds the upper limit of the combined force of flow velocity and dispersion, it is identified as an anomalous component. Reconstruction is performed on the remaining sequences to establish a structured relationship between the time dimension and the migration trajectory. Spatiotemporal aggregated content is generated, which is a four-dimensional dataset that records in detail the spatial distribution probability of pollutants at different historical moments and their correlation with flow field characteristics.

[0035] Please see Figure 6 The specific steps of S5 are as follows: S501: Call the spatiotemporal aggregated content and node representation, retrieve the positional components in the spatiotemporal aggregated content and perform mapping based on the spatial index of the node representation, perform linear transformation on the mapping result based on a fixed weight factor and perform directional merging on the transformation sequence to generate aggregated coordinates; The generated spatiotemporal aggregated content and node representation data are retrieved, and the spatial components of each possible location within the spatiotemporal aggregated content are obtained. Based on the spatial indexes in the node representations, these locations are mapped to the actual geographic grid. For each mapped grid cell, a linear transformation is performed using a fixed weight factor. The weight factor is set according to the reliability weight of the monitoring points; for example, the weight of a core monitoring point is 1.2, and that of an auxiliary monitoring point is 0.8. The transformed sequence is then directionally merged, and the endpoints of all backtracking paths are spatially averaged. Aggregated coordinates are generated, representing the potential pollution source center location derived from the monitoring data; for example, the aggregated predicted center coordinates are (34580.45 meters, 45680.12 meters).

[0036] S502: Based on aggregated coordinates, the coordinates of suspected leakage points are collected from the hydrodynamic parameter set of the aquifer. The spatial components of the coordinates of the suspected leakage points are detected and the distance reference interval is defined according to the preset threshold. The distance reference interval is compared with the numerical range of the aggregated coordinates to form a distance judgment sequence and obtain the leakage distance screening value. Based on the generated aggregated coordinates, coordinates of all suspected leak points are collected from the aquifer hydrodynamic parameter set. These points typically correspond to chemical plants, gas stations, and sewage network nodes within the area. The spatial components of the coordinates of each suspected leak point are detected, and its Euclidean distance to the aggregated coordinates is calculated. A distance baseline interval is defined based on a preset error tolerance value, set to 0 to 150 meters. This distance baseline interval is compared with the calculated distance values ​​of each suspected point to generate a distance judgment sequence. For example, a chemical plant is 85 meters away from the aggregated coordinates, within the baseline interval, and has a judgment value of 1; while a gas station is 320 meters away, outside the interval, and has a judgment value of 0. The resulting leak distance screening values ​​are used for preliminary screening of potential pollution sources.

[0037] S503: Based on the leakage distance screening value and the coordinates of suspected leakage points, perform coordinate aggregation on the points that meet the benchmark interval in the distance screening value and remove those that do not meet the benchmark. Reorganize the remaining coordinate sequence and establish a point association set to generate groundwater pollution source identification results. Based on the obtained leakage distance screening values ​​and the coordinates of suspected leakage points, coordinate aggregation is performed on points with a screening value of 1. These points that meet the baseline interval are extracted, and those with a screening value of 0 are removed. The remaining coordinate sequences are then reorganized to establish a point association set. This process calculates the matching degree between the points in the association set and the pollutant components by performing a secondary probability check. For example, if the characteristic pollutant stored at a point matches the detected pollutant, its confidence level is increased to 0.95. Finally, the point with the highest confidence level and its geographical information are output, generating the groundwater pollution source identification result.

[0038] Table 3 Evaluation Table of Pollution Source Identification Results; As shown in Table 3, high-risk pollution sources were accurately identified through multi-criteria screening. The advantage of this step is that it greatly reduces the ambiguity of the source tracing results through dual verification of spatial distance and attribute matching.

[0039] Please see Figure 7 An artificial intelligence-based groundwater pollution source identification system includes: The attenuation weight and spatial connection module is used to implement S1: acquire the concentration information of the monitoring well and acquire three-dimensional spatial coordinate data, permeability data and water flow gradient vector, perform differential analysis based on the three-dimensional spatial coordinate data and use the permeability data to obtain the attenuation weight sequence, and calculate and generate the spatial connection vector from the attenuation weight sequence. The node representation and structure matrix module is used to implement S2: obtain spatial connection vectors and input them into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix, and generate node representations based on the three-dimensional spatial coordinate data and the connection structure matrix. The pollutant screening and time segmentation module is used to implement S3: acquire pollutant concentration sequences and detect retardation factors, screen pollutant concentration sequences based on node expression to obtain screening segments, and generate time segments based on the time span determined by the screening segments. The spatiotemporal aggregation and migration fusion module is used to implement S4: perform time inversion calculation on time segments according to the inverse convection dispersion equation, obtain dispersion field information and groundwater flow direction information, calculate the migration expression, and fuse the inverse segment with the migration expression to generate spatiotemporal aggregated content; The pollution source identification and distance screening module is used to implement S5: using spatiotemporal aggregated content and node expression to perform spatial mapping to generate aggregated coordinates, obtain the coordinates of suspected leak points and distance screening thresholds, and perform distance screening based on aggregated coordinates and suspected leak point coordinates to generate groundwater pollution source identification results.

[0040] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the technical solution.

Claims

1. A method for identifying a source of groundwater pollution based on artificial intelligence, characterized by, Includes the following steps: S1: Obtain concentration information from monitoring wells and acquire three-dimensional spatial coordinate data, permeability data, and water flow gradient vector. Perform differential analysis based on the three-dimensional spatial coordinate data and use the permeability data to obtain an attenuation weight sequence. Calculate and generate a spatial connection vector from the attenuation weight sequence. S2: Obtain the spatial connection vector and input it into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix. Generate node representations based on the three-dimensional spatial coordinate data and the connection structure matrix. S3: Obtain pollutant concentration sequences and detect blocking factors; based on the node expression, filter pollutant concentration sequences to obtain filtered fragments; and based on the filtered fragments, determine the time span to generate time segments. S4: Perform time inversion calculation on the time segment according to the inverse convection dispersion equation, obtain dispersion field information and groundwater flow direction information, and calculate the migration expression. Then, merge the inverse segment with the migration expression to generate spatiotemporal aggregated content. S5: Utilize the spatiotemporal aggregated content and the node expression to perform spatial mapping to generate aggregated coordinates, obtain the coordinates of suspected leak points and distance screening thresholds, and perform distance screening based on the aggregated coordinates and the coordinates of suspected leak points to generate groundwater pollution source identification results.

2. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The spatial connection vector includes spatial correlation, attenuation coefficient, and directional weight; the node representation includes node spatial features, node connection weight, and node structural factor; the time segment includes start time, end time, and duration; the spatiotemporal aggregation content includes reverse time information, migration feature information, and fusion expression parameters; and the groundwater pollution source identification result includes source location, distance threshold determination item, and risk assessment index.

3. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The migration expression obtains the migration process and distribution characteristics of pollutants in groundwater by calculating dispersion field information and groundwater flow direction information.

4. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The selected segments are time periods selected from the pollutant concentration sequence based on the node expression, to determine the changes and temporal characteristics of pollutants.

5. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: Obtain three-dimensional spatial coordinate data and water flow gradient vector, construct a spatial grid based on the coordinate data, calculate the difference between adjacent points according to the grid and correct the water flow gradient direction, extract the direction change amount and form a direction consistency factor, and generate a coordinate difference matrix. S102: Call the coordinate difference matrix and permeability data, convert the difference components according to the permeability coefficient set and obtain the attenuation factor sequence, perform corresponding correction on the attenuation factor sequence and coordinate difference matrix and update the directional attenuation relationship to obtain the attenuation weight sequence; S103: Establish a weighted correlation value based on the attenuation weight sequence and the coordinate difference matrix, aggregate the directional components according to the correlation value to obtain the directional synthesis rate, and call the directional synthesis rate to perform concatenation and normalization processing on the directional sequence of the sampling points to generate a spatial connection vector.

6. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The specific steps of S2 are as follows: S201: Obtain the spatial connection vector and input it into the graph structure construction process, perform inter-segment difference on the connection vector and retrieve the direction change of adjacent segments, filter data components that exceed the preset direction deviation threshold according to the fixed difference reference value and rearrange the remaining components to generate a connection structure matrix. S202: Call the connection structure matrix and three-dimensional spatial coordinate data, perform component splitting and linear superposition on the coordinates corresponding to the matrix index, form a coordinate association set based on the superposition result, and obtain the node coordinate mapping value; S203: Based on the node coordinate mapping value and the connection structure matrix, perform vector aggregation on the matrix node positions and filter out components with amplitude values ​​lower than a preset vector amplitude threshold. Perform directional merging and sequence recombination on the aggregation results to generate node representations.

7. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The specific steps for S3 are as follows: S301: Collect pollutant concentration sequences and detect retardation factors, retrieve time components of concentration sequences and determine the range of numerical changes based on retardation factors, remove numerical components that are lower than the fixed change benchmark value based on a fixed change benchmark value, and generate a concentration screening matrix. S302: Call the node expression and the concentration screening matrix, perform numerical comparison on the node expression corresponding to the screening matrix index, extract the segment interval that meets the preset numerical difference threshold based on the comparison difference, and aggregate the continuous interval to obtain the screening segment; S303: Based on the selected segments, retrieve the segment time range and calculate the start and end time difference. Compare the time difference with a preset fixed time span threshold and retain the matching segments. Perform time series recombination on the retained segments to generate time segments.

8. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The specific steps of S4 are as follows: S401: Obtain the time segment and retrieve the time tag, perform reverse order rearrangement of the time tag while maintaining the start and end correspondence, recombine the segment intervals according to the reverse order position and remove inconsistent intervals to generate reverse order segments; S402: Collect diffuse field information and groundwater flow direction information, detect the numerical gradient of diffuse field information and compare it with the directional component of groundwater flow direction information, distinguish the effective components according to the preset directional consistency threshold, and perform weighted superposition of the effective components to obtain the transfer expression. S403: Based on the reverse sequence fragment and the migration expression, match the corresponding time period value of the migration expression for the time segment of the reverse sequence fragment and perform numerical fusion, retrieve the component differences of the fused sequence and filter out abnormal components according to the fixed amplitude benchmark value, reconstruct the remaining sequence and establish the time and migration relationship structure to generate spatiotemporal aggregated content.

9. The method for identifying groundwater pollution sources based on artificial intelligence according to claim 1, characterized in that, The specific steps of S5 are as follows: S501: Call the spatiotemporal aggregated content and the node expression, retrieve the positional component in the spatiotemporal aggregated content and perform mapping according to the spatial index of the node expression, perform linear conversion on the mapping result according to a fixed weight factor and perform directional merging on the conversion sequence to generate aggregated coordinates; S502: Based on the aggregated coordinates, collect the coordinates of suspected leakage points from the hydrodynamic parameter group of the aquifer, detect the spatial components of the coordinates of the suspected leakage points, define the distance reference interval according to the preset threshold, compare the distance reference interval with the numerical range of the aggregated coordinates and form a distance judgment sequence to obtain the leakage distance screening value. S503: Based on the leakage distance screening value and the coordinates of the suspected leakage points, perform coordinate aggregation on the points that meet the benchmark interval in the distance screening value and remove those that do not meet the benchmark interval, reorganize the remaining coordinate sequence and establish a point association set to generate groundwater pollution source identification results.

10. An artificial intelligence-based groundwater pollution source identification system, characterized in that, The system is used to implement the artificial intelligence-based groundwater pollution source identification method according to any one of claims 1-9, comprising: The attenuation weight and spatial connection module is used to implement S1: acquiring the concentration information of the monitoring well and acquiring three-dimensional spatial coordinate data, permeability data and water flow gradient vector; performing differential analysis based on the three-dimensional spatial coordinate data and using the permeability data to obtain the attenuation weight sequence; and calculating and generating a spatial connection vector from the attenuation weight sequence. The node representation and structure matrix module is used to implement S2: obtain the spatial connection vector and input it into the graph structure construction process to perform matrix difference processing to obtain the connection structure matrix, and generate node representation based on the three-dimensional spatial coordinate data and the connection structure matrix; The pollutant screening and time segmentation module is used to implement S3: acquire pollutant concentration sequences and detect retardation factors, screen pollutant concentration sequences based on the node expression to obtain screening segments, and determine the time span based on the screening segments to generate time segments. The spatiotemporal aggregation and migration fusion module is used to implement S4: performing time inversion calculation on the time segment according to the inverse convection dispersion equation, obtaining dispersion field information and groundwater flow direction information, and calculating the migration expression, and fusing the reverse segment with the migration expression to generate spatiotemporal aggregated content; The pollution source identification and distance screening module is used to implement S5: using the spatiotemporal aggregated content and the node expression to perform spatial mapping to generate aggregated coordinates, obtain the coordinates of suspected leakage points and distance screening thresholds, and perform distance screening based on the aggregated coordinates and the coordinates of suspected leakage points to generate groundwater pollution source identification results.

Images