A method for evaluating the early warning level of soil and groundwater pollution in industrial parks

By calculating the time-series-driven vertical flux of water level and constructing a time-series causal relationship graph, the problem of lagging groundwater pollution monitoring and early warning in existing technologies has been solved, and the accurate risk location and quantitative classification of groundwater pollution in industrial parks have been realized.

CN123119652APending Publication Date: 2026-11-17JIANGXI HUAGAN GEFENG ECOLOGICAL ENVIRONMENTAL PROTECTION CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611284985.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-24
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies for monitoring groundwater pollution in industrial parks rely on concentration data from fixed-point sampling, which cannot reveal the migration dynamics and diffusion trends of pollutants, resulting in serious delays in early warning. Static threshold determination methods frequently fail under non-stationary conditions, leading to high false alarm and false negative rates.

Method used

By calculating the vertical flux sequence and hydraulic gradient direction driven by water level time series, a time series causal relationship graph is constructed. The sliding window traverses the topological structure change trajectory, marks abnormal points, calculates the spatial migration distance of the pollution front, and adjusts the early warning parameters to achieve accurate early warning.

Benefits of technology

It enables precise pollution risk location and quantitative classification under non-stationary conditions, avoids frequent failure of static thresholds, and improves the predictability and accuracy of early warning.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123119652A_ABST
    Figure CN123119652A_ABST
Patent Text Reader

Abstract

This invention discloses a method for evaluating the early warning level of soil and groundwater pollution in industrial parks, relating to the field of groundwater pollution monitoring and early warning technology. The method includes calculating the vertical flux sequence and hydraulic gradient direction driven by water level fluctuations within each time step based on the water level and concentration time series of monitoring points; using continuous time steps as time slices, determining the driving-response time sequence relationship between monitoring points according to the order of flux changes on the time axis; constructing a time-series causal relationship graph with the direction from the driving point to the response point as the edge and the monitoring point as the vertex; and recording the topological change trajectory of the time-series causal relationship graph on the time slice sequence. This invention organizes dispersed monitoring points into a graph structure reflecting the hydraulic conduction path through the construction of the time-series causal relationship graph, and combines sliding window multi-scale collaborative detection and local direction field consistency screening to avoid frequent failure of static thresholds under non-stationary conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of groundwater pollution monitoring and early warning technology, specifically a method for evaluating the early warning level of soil and groundwater pollution in industrial parks. Background Technology

[0002] For a long time, groundwater pollution monitoring in industrial parks has relied on concentration data from fixed-point sampling. Concentration data reflects the pollution outcome but cannot reveal the migration dynamics and diffusion trends of pollutants. By the time monitoring points detect concentrations exceeding standards, the pollution front has often already migrated a considerable distance, resulting in a severe delay in early warning.

[0003] Existing static threshold determination methods assume that the groundwater system is in a stable flow state. However, the aquifer in industrial parks is subject to multiple disturbances such as seasonal pumping, rainfall infiltration, and engineering activities, resulting in a continuously non-stationary hydraulic field. The static threshold frequently fails under non-stationary conditions, leading to high false alarm and false negative rates. Summary of the Invention

[0004] The purpose of this invention is to provide a method for evaluating the early warning level of soil and groundwater pollution in industrial parks, so as to solve the problems mentioned in the background art.

[0005] A method for evaluating the early warning level of soil and groundwater pollution in industrial parks, comprising: A1. Calculate the vertical flux sequence and hydraulic gradient direction driven by water level fluctuations within each time step based on the water level time series and concentration time series of the monitoring points; A2. Using continuous time steps as time slices, determine the driving-response time sequence relationship between monitoring points based on the order of flux changes at each monitoring point on the time axis. Construct a time-series causal relationship graph with the direction from the driving point to the response point as the edge and the monitoring point as the vertex, and record the topological structure change trajectory of the time-series causal relationship graph on the time slice sequence. A3. A sliding window traverses the topological change trajectory. When a directed edge of a preset type reverses or disappears in the temporal causal relationship graph within the window, the monitoring points associated with the reversed or disappeared directed edges are marked as abnormal points. The proportion of abnormal points to all points is counted. If the proportion exceeds a preset proportion threshold, it is marked as a hydraulic connectivity anomaly. The preset type is automatically determined by the system based on the frequency of occurrence of each directed edge in the temporal causal relationship graph during historical stable periods. A4. After marking the anomaly, starting from the anomaly point, trace the influence range along the directed edges of the time-series causal relationship graph where no anomaly has occurred, and estimate the spatial migration distance of the pollution front based on the flux change amplitude of each anomaly point. A5. Delineate the risk boundary based on the spatial migration distance, overlay it onto a three-dimensional stratigraphic grid, determine the comprehensive early warning level based on the cumulative flux value of the nodes, and output the result. A6. Based on the measured offset between the pollution boundary and the risk boundary, adjust the preset percentage threshold and the sliding window width according to the preset step feedback rule.

[0006] In some possible implementations, the driver-response timing relationship described in step A2 is determined as follows: For any two monitoring points, extract the times when the extreme values ​​of flux changes occur within the time slice; If the extreme point of the first point occurs earlier than the extreme point of the second point, and the time difference between the two is less than the preset time difference threshold, then the first point is determined as the driving point and the second point is determined as the response point.

[0007] In some possible implementations, the determination method for whether the directed edge of the preset type is reversed or disappears in step A3 is as follows: For each directed edge in the temporal causal relationship graph, record the frequency of the directed edge appearing in all time slices during the historical stationary period; If the frequency of a directed edge is lower than a preset frequency threshold, then the directed edge is marked as a low-frequency edge. When traversing the sliding window, if a low-frequency edge appears continuously in multiple consecutive time slices, or a non-low-frequency edge disappears continuously in multiple consecutive time slices, or a non-low-frequency edge reverses direction in multiple consecutive time slices, these are respectively treated as abnormal events of the corresponding preset types.

[0008] In some possible implementations, after marking the hydraulic connectivity anomaly in step A3, the method further includes continuous tracking and dynamic maintenance of the anomaly marker: After the current time slice that marks an anomaly, the anomaly point screening and percentage statistics are repeated for each subsequent consecutive time slice; Calculate the statistical representative value of the percentage of outlier points in each time slice within a statistical window that takes the current time slice as the endpoint and covers the current slice and multiple consecutive time slices before it; If the statistical representative value is consistently lower than the preset percentage threshold within multiple consecutive time slices, the hydraulic connectivity anomaly will be downgraded to a suspected hydraulic connectivity anomaly and a re-inspection process will be initiated. If the statistical representative value rises above the preset percentage threshold within multiple consecutive time slices, the suspected hydraulic connectivity anomaly will be restored to a hydraulic connectivity anomaly.

[0009] In some possible implementations, after marking the monitoring points associated with the directed edges that have reversed or disappeared as anomalous points in step A3, the method further includes: For each monitoring point marked as an anomaly, trace the upstream and downstream related points of the anomaly point in the time-series causal relationship diagram, and extract the flux change direction of the upstream and downstream related points before and after the jump moment. If the directed edge between an abnormal point and its upstream associated point disappears, the abnormal point is determined to be a pollution source release point. If the directed edge between an abnormal point and its downstream associated point reverses, the abnormal point is determined to be a pollution migration path deflection point. After distinguishing between the pollution source release points and the pollution migration path deflection points, the spatial migration distance of the pollution front is estimated using different propagation coefficients.

[0010] In some possible implementations, the tracing of the influence range along directed edges in the temporal causal relationship graph where no anomalies have occurred, as described in step A4, specifically includes: Starting from the marked abnormal points, extend downstream level by level along all directed edges in the time-series causal relationship graph that have not reversed or disappeared, and include all monitoring points passed through during the extension process into the set of influence ranges. For each point included in the set of affected areas, obtain the deviation of the flux change after the time of the anomaly occurrence relative to the flux change before the time of the anomaly occurrence, and calculate the deviation ratio of the deviation relative to the flux change before the time of the anomaly occurrence. The propagation attenuation characteristics of each point are determined based on the deviation ratio, and the spatial migration distance of the pollution front along the directed edge is calculated step by step based on the propagation attenuation characteristics.

[0011] In some possible implementations, the determination of the drive-response timing relationship in step A2 further includes verification of the physical constraints of the hydraulic conduction path: For any two monitoring points that are preliminarily determined to have a driving-response relationship based on the order of extreme point times, obtain the spatial distribution of the formation permeability coefficient and the groundwater flow direction vector field between the two points; Calculate the shortest effective hydraulic conduction path length between the two points along the groundwater flow direction, and divide the shortest effective hydraulic conduction path length by the Darcy velocity calculated based on the average permeability coefficient and average hydraulic gradient between the two points to obtain the theoretical hydraulic propagation time; If the relative deviation between the theoretical hydraulic propagation time and the time difference of the extreme point exceeds a preset deviation threshold, the driving-response relationship of the pair of points is marked as a relationship to be verified and is not included in the construction of the time-series causal relationship graph.

[0012] In some possible implementations, the sliding window traversal in step A3 further includes multi-timescale collaborative detection: Multiple sliding windows of different widths are set up to traverse the topological change trajectory in parallel, with each sliding window corresponding to a different time response scale; For the same directed edge, when the directed edge is detected to reverse or disappear within a narrow window, and the cumulative occurrence frequency of the directed edge in the corresponding time period within a wide window is lower than the set proportion of its cumulative occurrence frequency in historical stable periods, the abnormal event of the directed edge is marked as a cross-scale confirmed anomaly. Only monitoring points associated with cross-scale confirmed anomalies are included in the anomaly point statistics.

[0013] In some possible implementations, before the step A3 describes the percentage of outlier points among all points, a local orientation field consistency screening of the outlier points is also included: For each monitoring point marked as an anomaly, extract the hydraulic gradient direction change vector of that point in one time slice before and after the jump moment; Using the anomaly point as the center, search for adjacent monitoring points according to the preset spatial neighborhood radius, and calculate the angle between the hydraulic gradient direction change vector of the adjacent points at the corresponding time and the direction change vector of the center point. If the proportion of the number of adjacent points with an included angle less than a preset included angle threshold to the total number of adjacent points is higher than a preset neighborhood consistency threshold, then the center point is retained in the abnormal points. If the ratio is lower than the preset neighborhood consistency threshold, the center point will be removed from the abnormal points.

[0014] In some possible implementations, the tracing of the influence range along directed edges in the temporal causal relationship graph where no anomalies have occurred, as described in step A4, also includes segmented calculation of the attenuation mechanism of the propagation path: For each directed edge that did not exhibit any anomalies during the tracing process, the type of propagation medium to which the directed edge belongs is determined based on the difference in lithology between the upstream and downstream points connected by the directed edge. If the propagation medium is a fractured medium, the equivalent hydraulic conductivity coefficient corresponding to the directed edge is calculated using the cubic law of fractured networks, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. If the propagation medium is a porous medium, the measured permeability coefficient of the region corresponding to the directed edge is used as the equivalent hydraulic conductivity coefficient, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. The deviation ratio described in step A4 is corrected according to the propagation correction factor for each segment, and the spatial migration distance of the pollution front is calculated step by step based on the corrected deviation ratio.

[0015] The technical solutions provided by the embodiments of this disclosure bring at least the following beneficial effects: This invention transforms water level time series into flux time series through vertical flux calculation driven by water level fluctuations, achieving early warning by leveraging the pre-indicative role of hydraulic gradient changes. It organizes dispersed monitoring points into a graph structure reflecting hydraulic transmission paths through the construction of a time-series causal relationship graph, and combines multi-scale collaborative detection with sliding window and local orientation field consistency screening to avoid frequent failures of static thresholds under non-stationary conditions. By distinguishing between pollution source release points and pollution migration path deflection points and using differentiated propagation coefficients and attenuation mechanisms to segmentally correct pollution front calculations, it avoids distance estimation biases caused by unified propagation models. Through three-dimensional stratigraphic grid risk assessment, it quantitatively correlates flux intensity with pollution duration, achieving accurate spatial positioning and quantitative classification of pollution risk. Finally, it adaptively adjusts early warning parameters based on the offset between measured pollution boundaries and risk boundaries, avoiding systematic deviations of fixed parameters under time-varying conditions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] The invention will be further described in detail below with reference to a specific embodiment of a chemical industrial park. This chemical industrial park covers an area of ​​approximately 5 square kilometers and has 20 multi-layer monitoring wells as monitoring points. Each monitoring well is 30 meters deep and includes three monitoring depths: a shallow depth of 5 meters, a medium depth of 15 meters, and a deep depth of 30 meters. Water level and concentration data are collected hourly at each monitoring point, forming water level and concentration time series data.

[0019] Please see Figure 1 This application provides a method for evaluating the early warning level of soil and groundwater pollution in industrial parks, including: A1. Calculate the vertical flux sequence and hydraulic gradient direction driven by water level fluctuations within each time step from the water level time series of monitoring points, and identify the active periods of vertical migration of pollutants by combining the concentration time series.

[0020] Vertical flux refers to the sum of vertical hydraulic flux and pollution flux passing through a unit horizontal area per unit time. The calculation of vertical hydraulic flux relies solely on water level time series, directly obtained from the water level gradient using Darcy's law. Concentration time series identifies active periods of vertical pollutant migration, and within these active periods, the vertical hydraulic flux is multiplied by the pollutant concentration to obtain the pollution flux; the two are then superimposed to form a complete vertical flux sequence. Water level fluctuation refers to the change in head difference between different depth layers at a monitoring point over time, i.e., the rate of change of the vertical head gradient over time. Let the shallow water level be h1, the deep water level be h2, and the vertical distance be Δz, then the vertical head gradient iz is equal to the difference between h1 and h2 divided by Δz. Water level fluctuation, i.e., the rate of change of iz over time, is equal to the difference between the current time step iz value and the previous time step iz value divided by the time step size Δt.

[0021] The vertical hydraulic flux is calculated according to Darcy's law and is equal to the vertical permeability coefficient multiplied by the vertical head gradient, i.e., qz equals Kz multiplied by iz. The vertical flux sequence driven by water level fluctuations refers to a time series formed by arranging the vertical hydraulic flux values ​​at each time step in chronological order. Water level fluctuations change the vertical head gradient, thereby changing the vertical hydraulic flux and driving the vertical flux sequence.

[0022] The concentration time series is used to identify active periods of vertical pollutant migration and to calculate vertical pollution flux. Vertical pollution flux equals the product of vertical hydraulic flux and the pollutant concentration at the corresponding depth. When the concentration time series shows that the pollutant concentration at that depth is higher than the background value, it is determined to be an active period of vertical pollutant migration, and the vertical flux sequence within this period is included in the key analysis; when the concentration is lower than the background value, it is determined to be an inactive period, and the vertical flux sequence within this period is only used for hydraulic connectivity background analysis.

[0023] The background value refers to the baseline pollutant concentration at the depth of the monitoring point before the pollution source enters the park, or in the upstream control area of ​​the park. Specifically, it is determined as follows: three control monitoring wells are selected upstream of the park boundary, and the median concentration data at each sampling time within six consecutive months prior to the start of monitoring is taken as the background value for that depth. If there are no control wells outside the park boundary, the Class III water standard limit in the "Groundwater Quality Standard" (GB / T14848-2017) is used as the background value for that depth. If there are fewer than three but at least one control well outside the park boundary, the median of the existing control well data is used as the background value. A risk warning regarding the insufficient number of control wells is also included in the early warning report, prompting technical personnel to supplement the control well deployment in subsequent monitoring.

[0024] The hydraulic gradient direction is the direction in which the water head decreases along the flow of groundwater. In three-dimensional space, the hydraulic gradient vector is composed of a horizontal hydraulic gradient component and a vertical hydraulic gradient component. The horizontal component is calculated by dividing the water level difference between the monitoring point and adjacent monitoring points by the horizontal distance, while the vertical component is calculated by dividing the water level difference between different depth layers at the monitoring point by the vertical distance. The hydraulic gradient direction is represented by azimuth and dip angle. The azimuth is the angle between the horizontal projection and true north, ranging from 0 degrees to 360 degrees. The dip angle is the angle between the vector and the horizontal plane, with upward being positive and downward being negative.

[0025] When water level data or concentration data is missing or abnormal at a monitoring point within a certain time step, median interpolation is performed using valid data from three time steps before and after that point. If there are fewer than three valid data points before and after, the data for that time step is marked as missing, will not be included in subsequent calculations, and a data missing log will be recorded. The data missing log includes the missing time step number, the missing data type, and the missing reason flag. The log is stored in CSV format in the system data directory for periodic review by technical personnel.

[0026] The concentration data anomaly refers to the concentration value at a given time step meeting any of the following conditions: (1) exceeding the 99th percentile of the historical concentration data for that location, where the historical concentration data refers to all valid concentration data for that location from the start of monitoring to the current time, and the 99th percentile is calculated using linear interpolation; (2) the absolute value of the concentration change rate between adjacent time steps exceeds 300%, where the concentration change rate is equal to the difference between the current time step concentration value and the previous time step concentration value divided by the time step length; (3) the concentration value is negative, or exceeds the upper limit of the instrument's range (e.g., the upper limit of the range of a hexavalent chromium detector is 10 mg / L), or is below the instrument's detection limit (e.g., the detection limit of hexavalent chromium is 0.004 mg / L). When the concentration data meets any of the above anomaly conditions, the concentration data at that time step is marked as anomaly, and the median of the valid data from the preceding and following three time steps is used for interpolation and replacement, and the anomaly type is recorded in the data missing log.

[0027] Taking monitoring point W1 as an example, the shallow water level is 15.2 meters, the deep water level is 14.8 meters, the vertical distance is 25 meters, and the time step is 1 hour. The vertical head gradient is equal to the difference between 15.2 and 14.8 divided by 25, resulting in 0.016. In the previous time step, the shallow water level was 15.0 meters, the deep water level was 14.9 meters, and the vertical head gradient in the previous time step was equal to the difference between 15.0 and 14.9 divided by 25, resulting in 0.004. The water level fluctuation is equal to the difference between 0.016 and 0.004 divided by 1, resulting in 0.012 per hour. The vertical permeability coefficient is determined based on pumping test data from the hydrogeological survey report of the industrial park. The aquifer permeability coefficient is obtained through steady-flow pumping tests, and then calculated using the empirical ratio of vertical to horizontal permeability coefficients. This empirical ratio is typically between 0.1 and 0.5 in gravel layers; in this embodiment, it is determined to be 0.3 based on the measured ratio of vertical to horizontal permeability coefficients from the industrial park's pumping tests. The vertical hydraulic flux is calculated by multiplying the vertical permeability coefficient by the vertical head gradient. The vertical flux sequence is the time series of vertical hydraulic flux values ​​at each time step.

[0028] Pollution flux is the product of vertical hydraulic flux and the pollutant concentration at the corresponding depth. Taking the middle layer of W1 as an example, the pollutant concentration in this layer is 0.5 mg / L. The vertical pollution flux is equal to the vertical hydraulic flux multiplied by 0.5, and the vertical pollution flux value is calculated.

[0029] Existing technologies rely on concentration thresholds to determine pollution, but concentration data reflects the pollution outcome rather than the migration dynamics, resulting in significant delays in early warning. This invention uses a vertical flux sequence driven by water level fluctuations to capture the dynamic changes in the hydraulic field in real time. By applying Darcy's law, the water level time series is transformed into a flux time series, quantifying hydraulic gradient changes before concentration changes occur. This proactive indication of hydraulic connectivity changes avoids the lag blind spots of traditional concentration monitoring. According to groundwater dynamics theory, the response time of hydraulic gradient changes is typically on the order of hours to days, while the response time of concentration changes is controlled by convection and diffusion processes, typically on the order of weeks to months. Therefore, hydraulic connectivity early warning can be several time scales ahead of concentration early warning.

[0030] A2. Using continuous time steps as time slices, determine the driving-response time sequence relationship between monitoring points based on the order of flux changes at each monitoring point on the time axis. Construct a time-series causal relationship graph with the direction from the driving point to the response point as the edge and the monitoring point as the vertex, and record the topological structure change trajectory of the time-series causal relationship graph on the time slice sequence.

[0031] The time slice refers to dividing a continuous monitoring time series into several analysis units of fixed length. Each time slice contains monitoring data from several consecutive time steps, used to analyze the temporal characteristics of flux changes at each monitoring point at a uniform time granularity. In this embodiment, the time slice consists of 24 consecutive time steps, corresponding to one day of monitoring data.

[0032] Within each time slice, the vertical flux change sequence of each monitoring point is smoothed using a three-point moving average to eliminate high-frequency noise interference. The vertical flux change sequence refers to the time series formed by the difference in vertical flux values ​​between adjacent time steps. After smoothing, local maxima and local minima of the vertical flux change at each point are extracted. A local maximum is defined as a point whose value is greater than the values ​​of all other points in a given time neighborhood, determined by the vertical flux change value at that time step being greater than the vertical flux change values ​​of the two time steps before and after it. A local minimum is defined as a point whose value is less than the values ​​of all other points in a given time neighborhood, determined by the vertical flux change value at that time step being less than the vertical flux change values ​​of the two time steps before and after it.

[0033] For any two monitoring points, extract the times when the extreme points of flux changes occur within a time slice. Taking points W1 and W2 as examples, assume that within time slice T1, the local maximum of the vertical flux change of W1 occurs at hour 8, and the local maximum of the vertical flux change of W2 occurs at hour 14. If the extreme point of W1 occurs earlier than the extreme point of W2, and the time difference between them is less than a preset time difference threshold, then W1 is determined as the driving point and W2 as the response point, and a directed edge W1 pointing to W2 is established in the time-series causal relationship graph. If the extreme point of W2 occurs earlier than W1, then W2 is determined as the driving point and W1 as the response point, and a directed edge W2 pointing to W1 is established. The preset time difference threshold is determined based on the average hydraulic conduction characteristics of the aquifer in the park. The propagation speed of groundwater in gravel aquifers is typically on the order of 10⁻⁶ to 10⁻⁴ meters per second. When the monitoring points are spaced 80 to 150 meters apart, the hydraulic propagation time is usually in the range of several hours to tens of hours. Therefore, setting the time difference threshold between 6 and 24 hours can cover most hydraulic conduction scenarios while avoiding the inclusion of irrelevant delayed responses due to an excessively large threshold. In this embodiment, the time difference threshold is determined to be 12 hours based on the average hydraulic conduction speed measured by the pumping test of the aquifer in the park.

[0034] The time-series causal relationship graph is constructed with monitoring points as vertices and driver-response relationships as edges. Each directed edge records the driver point number, response point number, time difference of extreme points, and flux change ratio. For 20 monitoring points, the time-series causal relationship graph can contain a maximum of 380 directed edges.

[0035] The topological change trajectory is a record of the evolution of the temporal causal relationship graph on a time-slice sequence. For each time slice, the existence and direction states of all directed edges in the temporal causal relationship graph within that slice are recorded. The existence state includes two values: existence and non-existence. The direction state includes two values: positive (consistent with the mainstream direction of historical stable periods) and negative (opposite to the mainstream direction of historical stable periods). The topological change trajectory is stored in a three-dimensional data structure. The first dimension is the time slice number, the second dimension is the vertex pair number (arranged in the order of pairwise combinations of monitoring point numbers), and the third dimension is the state attribute, which contains four fields: existence state field (Boolean, 0 indicates non-existence, 1 indicates existence), direction state field (Boolean, 0 indicates negative, 1 indicates positive), extreme point time difference field (floating-point, in hours), and flux change amplitude ratio field (floating-point, dimensionless). Arranging the temporal causal relationship graphs of consecutive time slices in chronological order forms the topological change trajectory.

[0036] If no valid extreme points are detected for a pair of monitoring points within a certain time slice, no directed edge is established between the pair of points, and it is recorded as unrelated. If multiple extreme points are detected for a pair of points within a time slice, the extreme point with the largest vertical flux change is taken as the representative.

[0037] Existing technologies treat monitoring points as independent entities, neglecting the hydraulic connections between them. This invention establishes a drive-response relationship between extreme points sequentially, organizing the dispersed monitoring points into a graph structure that reflects the hydraulic transmission path. By leveraging the topological sensitivity of the graph structure, the non-stationary hydraulic field is transformed into a traceable graph evolution, providing structured input for subsequent anomaly detection.

[0038] For any two monitoring points initially determined to have a drive-response relationship based on the order of their extreme points, the spatial distribution of the stratigraphic permeability coefficient and the groundwater flow direction vector field between the two points are obtained. Taking W1 and W2 as examples, the shortest effective hydraulic conduction path length between the two points along the groundwater flow direction is obtained using hydrogeological survey data from the park. The shortest effective hydraulic conduction path length is not a simple straight-line distance, but rather the actual conduction path length along the groundwater flow direction after considering the stratigraphic lithology distribution. Specifically, the underground space of the park is divided into a regular hexahedral grid of 50m x 50m x 2m, and the area between the two points is divided into several segments. Each segment is assigned a different permeability coefficient according to the lithology. The shortest conduction path from W1 to W2 is traced along the groundwater flow direction vector field, and the lengths of each segment are summed to obtain the shortest effective hydraulic conduction path length. The tracing uses Dijkstra's shortest path algorithm, with the reciprocal of the permeability coefficient as the edge weight, to solve for the shortest effective hydraulic conduction path from the drive point to the response point on the above-mentioned gridded stratigraphic model. The algorithm can be implemented using existing graph theory algorithm libraries or programming methods familiar to those skilled in the art.

[0039] The average permeability coefficient between two points is calculated as the harmonic mean of the permeability coefficients of each segment. The average hydraulic gradient is the difference in hydraulic head between the two points divided by the shortest effective hydraulic conduction path length. Based on Darcy's law, the Darcy velocity equals the average permeability coefficient multiplied by the average hydraulic gradient and then divided by the effective porosity. The effective porosity is determined based on the aquifer lithology; for gravel aquifers, it is typically between 0.15 and 0.35, a range derived from typical statistics on effective porosity of gravel aquifers in the *Hydrogeological Handbook*. In this embodiment, it is determined to be 0.25 based on pumping test data from the industrial park. The theoretical hydraulic propagation time equals the shortest effective hydraulic conduction path length divided by the Darcy velocity.

[0040] If the relative deviation between the theoretical hydraulic propagation time and the time difference at the extreme point exceeds a preset deviation threshold, the drive-response relationship of that pair of points is marked as a relationship to be verified and will not be included in the construction of the time-series causal relationship graph. The relative deviation is calculated using absolute values; that is, regardless of whether the theoretical propagation time is greater than or less than the time difference at the extreme point, as long as the absolute value of the deviation exceeds the threshold, it is marked as a relationship to be verified. The preset deviation threshold is typically set between 20% and 40%. This range is determined based on the discussion of the applicability conditions of Darcy's law in "Groundwater Dynamics." The calculation error of the theoretical propagation time is affected by formation heterogeneity, simplified boundary conditions, and measurement errors, and is usually between 20% and 50%. Taking the median value as the threshold can filter out most of the spurious correlations caused by model simplification, while retaining the true hydraulic conduction path. In this embodiment, based on the degree of formation heterogeneity in the park and the quality of the pumping test data, the preset deviation threshold is determined to be 30%. Relationships to be verified are stored separately in the system's relationship to be verified table for technical personnel to refer to in subsequent analysis, and may indicate the existence of unexplored preferential flow channels or data anomalies.

[0041] A3. A sliding window traverses the topological change trajectory. When a directed edge of a preset type reverses or disappears in the temporal causal relationship graph within the window, the monitoring points associated with the reversed or disappeared directed edges are marked as abnormal points. The proportion of abnormal points to all points is calculated. If the proportion exceeds a preset proportion threshold, it is marked as a hydraulic connectivity anomaly. The preset type is automatically determined by the early warning evaluation system based on the frequency of occurrence of each directed edge in the temporal causal relationship graph during historical stable periods. After automatic determination, a frequency distribution report is generated for technical personnel to verify. The verification results are only used for recording and do not affect the automatic marking results.

[0042] The identification of historical stable periods employs a combination of sliding scanning and statistical testing. Specifically, starting from the earliest time slice, the scanning proceeds in a window of 30 time slices, sliding one time slice at a time. For each scanning window, the percentage of outliers (i.e., the maximum percentage of points marked as outliers in each time slice of the window relative to the total number of monitoring points) and the magnitude of change in the hydraulic gradient direction (i.e., the maximum angle between the hydraulic gradient direction vectors of each monitoring point in the window across different time slices) are calculated. If a scanning window satisfies both an outlier percentage below 5% and the magnitude of change in the hydraulic gradient direction below 15 degrees, then the starting time slice of that window is marked as the beginning of a historical stable period. The window continues to expand until the conditions are no longer met, and the last time slice that meets the conditions is marked as the end of a historical stable period. If no 30 consecutive time slices satisfying the above conditions are found in the entire historical data, the requirements are automatically lowered, the window length is shortened to 20 time slices, the outlier percentage threshold is relaxed to 10%, and the scanning is repeated. If the requirements are still not met, the window length will be further shortened to 15 time slices, the threshold for the proportion of outliers will be relaxed to 15%, and the threshold for the change in hydraulic gradient direction will be relaxed to 20 degrees, and a third scan will be performed. If the requirements are still not met after three scans, the system will prompt technicians to manually set a historical stable period. During manual setting, technicians will input the start and end time slice numbers of the historical stable period through the system's human-computer interaction interface. The input format is two positive integers separated by a comma, such as "1,30" representing the 1st to the 30th time slice. The system will validate the input values, requiring that the start number is less than the end number and the end number does not exceed the total number of time slices currently collected. After the validation is passed, the manually input values ​​will replace the automatic scan results as the historical stable period. The termination condition for the above automatic reduction requirements is that the window length is not less than 15 time slices. If it is less than this length, the system will directly prompt manual setting to avoid frequency statistics distortion caused by an excessively short window.

[0043] After identifying historically stable periods, the early warning evaluation system automatically calculates the frequency of each directed edge in the time-series causal relationship graph within these periods, without manual intervention. The frequency is equal to the number of time slices containing that directed edge within the historically stable period divided by the total number of time slices within that period, ranging from 0 to 1. Based on this frequency, a preset type is automatically determined: if the frequency of a directed edge is below a preset frequency threshold, it is marked as a low-frequency edge, belonging to the unstable association type within the preset types; if the frequency of a directed edge is above or equal to the preset frequency threshold, it is marked as a non-low-frequency edge, belonging to the stable association type within the preset types. The frequency statistics and type marking processes are executed automatically in the background. After the statistics are completed, a frequency distribution report is generated for technical personnel to verify, but the type marking results are directly used for subsequent sliding window anomaly detection, and the verification results do not affect the automatic marking results.

[0044] For each directed edge in the time-series causal relationship graph, the frequency of its occurrence across all time slices during historical stable periods is recorded. If the frequency of a directed edge is lower than a preset frequency threshold, it is marked as a low-frequency edge. The preset frequency threshold is determined based on the stability of the park's hydraulic system, typically ranging from 10% to 30%. This range is determined according to the definition of sparse graphs in graph theory; when the frequency of an edge is lower than this range, it can be considered that the edge does not constitute a stable connection in the graph structure and belongs to a weak association driven by occasional or noise. In this embodiment, based on the statistical characteristics of graph sparsity in the historical monitoring data of the park, the preset frequency threshold is set to 20%. Directed edges with a frequency higher than or equal to the preset frequency threshold are marked as non-low-frequency edges.

[0045] During sliding window traversal, if a low-frequency edge appears continuously across multiple consecutive time slices, or a non-low-frequency edge disappears continuously across multiple consecutive time slices, or a non-low-frequency edge reverses direction across multiple consecutive time slices, these are respectively considered as abnormal events of the corresponding preset types. The setting of three consecutive time slices (approximately three days) is based on the following: the hydraulic response of groundwater systems has inertia; short-term data fluctuations (one to two time slices) are usually caused by instantaneous disturbances such as rainfall and pumping, while topological changes lasting more than three time slices are more likely to reflect real changes in the hydraulic connectivity of the aquifer. This time span matches the response timescale determined by the elastic release coefficient of the aquifer. Specific judgment rules are as follows: If a low-frequency edge occurs at a frequency below a preset frequency threshold during a stable period, but appears in all three consecutive time slices of the current sliding window, it is classified as an anomaly type one, indicating a sudden increase in previously unstable hydraulic connections. If a non-low-frequency edge occurs at a frequency higher than or equal to a preset frequency threshold during a stable period, but does not appear in all three consecutive time slices of the current sliding window, it is classified as an anomaly type two, indicating a sudden interruption of previously stable hydraulic conduction paths. If a non-low-frequency edge maintains a fixed direction during a stable period, but reverses direction in all three consecutive time slices of the current sliding window, it is classified as an anomaly type three, indicating a reversal of groundwater flow direction.

[0046] When a directed edge of a preset type reverses or disappears in the time-series causal relationship graph within the window, the monitoring points associated with that reversed or disappeared directed edge are marked as anomalous points. A monitoring point associated with a reversed or disappeared directed edge means that both endpoints of the directed edge, i.e., the start and end points, are marked as anomalous points. For example, if a directed edge W5 points to W6 and disappears, then both W5 and W6 are marked as anomalous points. If a directed edge W7 points to W8 and reverses (W8 points to W7), then both W7 and W8 are marked as anomalous points. When calculating the percentage of anomalous points out of all points, duplicate markings of the same monitoring point are removed, and each monitoring point is counted only once in the statistics. For example, if W5 is marked twice due to both W5 pointing to W6 and W4 pointing to W5 and reversing, W5 is only counted as one anomalous point in the percentage calculation.

[0047] Building upon this, the sliding window traversal also includes multi-timescale collaborative detection. Multiple sliding windows of varying widths are set up to traverse the topological change trajectory in parallel, each corresponding to a different time response scale. In this embodiment, two sliding windows are used: a narrow window with a width of 3 time slices, corresponding to short-term hydraulic fluctuation responses, and a wide window with a width of 15 time slices, corresponding to long-term hydraulic trend changes. According to the multi-scale decomposition principle in time series analysis, the narrow window width should be less than the shortest response period of the groundwater system to typical disturbances (such as a single rainfall event), while the wide window width should cover the complete response period of the aquifer to seasonal changes. Three time slices (approximately 3 days) are shorter than the typical groundwater response period (5 to 7 days) during rainfall, capturing instantaneous fluctuations; 15 time slices (approximately 15 days) are close to a half-month scale, filtering out short-term noise. Parallel traversal refers to calculating the detection results of each window separately at the logical level. At the computational implementation level, multi-threaded parallel processing or sequential processing can be used; both methods are technically equivalent.

[0048] For the same directed edge, the specific determination method for cross-scale confirmed anomalies is as follows: When the narrow window detects that the directed edge has reversed or disappeared, the start time slice number and end time slice number of the anomaly event in the narrow window are recorded. During the wide window traversal, based on the end time slice number of the anomaly event in the narrow window, the 15 most recent time slices ending at that time slice in the wide window are extracted as the corresponding time periods, and the cumulative occurrence frequency of the directed edge in the corresponding time period is calculated. The above alignment rules ensure that the short-term anomalies detected by the narrow window and the long-term trend of the wide window statistics are strictly corresponding in time: the end time slice of the narrow window and the end time slice of the wide window are at the same moment, and the time period covered by the 15 time slices going back in the wide window has an inclusion relationship with the time periods of the 3 time slices in the narrow window, avoiding misjudgment due to time period misalignment. If the cumulative occurrence frequency is lower than the set proportion of the cumulative occurrence frequency of its historical stable period, the anomaly event of the directed edge is marked as a cross-scale confirmed anomaly.

[0049] The set percentage is typically between 40% and 60%, determined based on the fundamental principles of statistical significance testing. When the cumulative frequency within a wide window is less than half of the historical frequency, it can be considered that the stability of that side has significantly degraded statistically, and the anomalies detected by the narrow window are not caused by local noise. In this embodiment, based on the response attenuation characteristics of the park's hydraulic system to typical disturbances, the set percentage is determined to be 50%. Only monitoring points associated with cross-scale confirmed anomalies are included in the anomaly point statistics. Non-cross-scale confirmed anomalies are recorded as events to be observed and are not included in the anomaly point statistics to avoid false alarms caused by short-term noise interference.

[0050] Existing static thresholding methods assume a steady-flow groundwater system, but aquifers in industrial parks are subject to multiple disturbances, resulting in a continuously non-stationary hydraulic field. This invention uses a narrow window to capture short-term fluctuations and a wide window to verify long-term trends, only including anomalies in statistics when both scales are consistently confirmed. This multi-scale collaborative detection avoids the failure of static thresholds under non-stationary conditions. Based on time series analysis theory, let the false alarm rate of independent detection with a narrow window be Pnarrow, and the false alarm rate of independent detection with a wide window be Pwide. Then, the false alarm rate of cross-scale confirmation detection does not exceed Pnarrow multiplied by Pwide (assuming the two windows are approximately independent). Since both Pnarrow and Pwide are less than 1, their product is necessarily less than the false alarm rate of either single window. Therefore, the false alarm rate of multi-scale collaborative detection is theoretically lower than that of single-scale detection.

[0051] As a specific implementation method, a chemical industrial park conducted background analysis using multi-timescale collaborative monitoring in March 2024. The park deployed 20 multi-layer monitoring wells, with the aquifer being a gravel-porous aquifer. The historically stable period was from March 1st to March 31st, 2024. During this period, the frequency of the directed edge W3 pointing to W7 was 22%, higher than the preset frequency threshold of 20%, and was marked as a non-low-frequency edge. From April 5th to April 7th, a narrow window detected the disappearance of the W3-to-W7 direction in three consecutive time slices; during the same period, the cumulative frequency of the W3-to-W7 direction in the corresponding time frame within a wide window was 8%, lower than 50% (i.e., 11%) of the historical frequency of 22%, meeting the cross-scale confirmation criteria. W3 and W7 were included in the anomaly point statistics. If not confirmed by cross-scale, the disappearance of the W3-to-W7 direction might have been due to a short-term rainfall and pumping disturbance on April 5th, which would have led to a false alarm if included in the statistics.

[0052] Furthermore, before calculating the proportion of outlier points to all points, the method also includes screening for local directional field consistency of outlier points.

[0053] For each monitoring point marked as an anomaly, the hydraulic gradient direction change vector is extracted for one time slice before and after the jump moment. The jump moment is the moment corresponding to the first time slice in which the directed edge reverses or disappears. Specifically, it is determined by the earlier of the time slice number where the directed edge first reverses or disappears (detected by a narrow window) and the time slice number where the anomaly event is confirmed (confirmed by a wide window). The hydraulic gradient direction change vector is equal to the hydraulic gradient direction vector of the time slice after the jump minus the hydraulic gradient direction vector of the time slice before the jump, and is expressed as the change in azimuth and dip angle.

[0054] Taking anomaly point W9 as an example, before the jump, the hydraulic gradient direction was azimuth 120 degrees and dip angle -5 degrees; after the jump, it was azimuth 150 degrees and dip angle -8 degrees, with a direction change vector of +30 degrees and -3 degrees. Using W9 as the center, adjacent monitoring points are searched according to a preset spatial neighborhood radius. The preset spatial neighborhood radius is typically between 50 and 150 meters. This range is determined based on the 100-200 meter spacing of monitoring wells in the "Technical Specification for Groundwater Environmental Monitoring" (HJ164-2020). Taking the middle value can cover 1 to 2 adjacent monitoring points, ensuring sufficient samples within the neighborhood for consistency judgment while avoiding the inclusion of irrelevant points due to an excessively large radius. In this embodiment, based on the actual 100-meter spacing of monitoring wells in the park, the preset spatial neighborhood radius is determined to be 100 meters. If there are no other monitoring points within the search range, the local direction field consistency screening of the anomaly point is skipped, and it is directly retained as an anomaly point. Calculate the angle between the hydraulic gradient direction change vector and the W9 direction change vector at each adjacent point at the corresponding time. The angle is calculated using the three-dimensional vector dot product formula, and the result is taken as the absolute value.

[0055] If the proportion of adjacent points with an angle smaller than a preset angle threshold to the total number of adjacent points is higher than a preset neighborhood consistency threshold, then the center point is retained among the abnormal points. The preset angle threshold is typically between 20 and 45 degrees. This range is determined based on the principle of hydraulic gradient continuity in groundwater dynamics. The hydraulic gradient direction of adjacent points within the same aquifer is spatially continuous. Due to local heterogeneity, the direction deviation is typically between 15 and 30 degrees. Taking the middle value as the threshold can tolerate reasonable spatial variation while eliminating isolated anomalies caused by instrument errors or local disturbances. In this embodiment, based on the degree of heterogeneity of the aquifer in the park and the historical statistical deviation of the direction changes of adjacent monitoring wells, the preset angle threshold is determined to be 30 degrees. The preset neighborhood consistency threshold is typically between 50% and 70%. This range is determined based on the majority voting principle. When more than half of the adjacent points show a consistent direction change, the change of the center point can be considered spatially representative, rather than isolated noise. In this embodiment, the preset neighborhood consistency threshold is determined to be 60%. If the proportion is lower than the preset neighborhood consistency threshold, the center point will be removed from the list of abnormal points to avoid isolated data anomalies being misjudged as hydraulic connectivity anomalies.

[0056] The number of anomalous points after local orientation field consistency screening is counted, and their proportion of the total number of monitoring points is calculated. If the proportion exceeds a preset threshold, it is marked as a hydraulic connectivity anomaly. The preset threshold is usually set between 20% and 40%. This range is determined based on the principle of connected graph segmentation in graph theory. When more than a certain proportion of vertices in the graph show anomalies simultaneously, the connectivity of the graph undergoes a structural change, which can be determined as an overall anomaly of the hydraulic system rather than a local disturbance. Below 20%, false alarms are easily caused by individual point failures, while above 40%, early changes in hydraulic connectivity may be missed. Therefore, the middle value is taken as the threshold to achieve a balance between sensitivity and robustness. In this embodiment, based on the density of monitoring points in the park and the connectivity characteristics of the aquifer, the preset threshold is determined to be 30%. Hydraulic connectivity anomalies refer to significant changes in the hydraulic conduction characteristics of the aquifer in the park, which may lead to the migration path and diffusion range of pollutants exceeding expectations.

[0057] As a specific implementation method, a chemical industrial park applied local orientation field consistency screening in April 2024. The anomalous point W8 experienced a jump on April 15th. Before the jump, the hydraulic gradient direction was 95 degrees azimuth and -3 degrees dip; after the jump, it was 125 degrees azimuth and -6 degrees dip, with a direction change vector of +30 degrees and -3 degrees. Using W8 as the center, neighboring points W7, W9, and W11 were searched within a 100-meter radius. The direction change vector of W7 was +28 degrees and -2 degrees, forming an angle of approximately 3.3 degrees with W8; the direction change vector of W9 was +32 degrees and -4 degrees, forming an angle of approximately 3.6 degrees with W8; and the direction change vector of W11 was -10 degrees and +2 degrees, forming an angle of approximately 42 degrees with W8. With a preset angle threshold of 30 degrees, the angle between W7 and W9 was less than 30 degrees, while the angle between W11 and W8 was greater than 30 degrees. There are 3 adjacent points. Two of them have an angle less than the threshold, accounting for 66.7%, which is higher than the preset neighborhood consistency threshold of 60%. Therefore, W8 is retained among the abnormal points. If the direction change of W11 is consistent with W8 (e.g., +29 degrees, -3 degrees), then all 3 adjacent points are less than 30 degrees, accounting for 100%, indicating stronger consistency. If only W7 is consistent while W9 and W11 deviate, then the proportion is 33.3%, which is lower than 60%, and W8 will be removed.

[0058] After marking an anomaly as hydraulic connectivity, the process also includes continuous tracking and dynamic maintenance of the anomaly marker.

[0059] After marking the current time slice as an anomaly, the anomaly point screening and percentage statistics are repeated for each subsequent consecutive time slice. A statistical representative value is calculated for the percentage of anomalies in each time slice within a statistical window that extends from the current time slice to several preceding consecutive time slices. This representative value is the median of the percentages of anomalies in each time slice within the statistical window. The median is chosen over the mean because the median is less sensitive to extreme values ​​and better reflects the central trend of anomaly percentages, avoiding misjudgments caused by sudden high percentages in individual time slices. The statistical window width is typically set between 5 and 10 time slices. This range is determined based on the moving average principle in time series analysis. The window width should cover the typical cycle of a hydraulic system from disturbance to recovery. Seven time slices (approximately 7 days) match the weekly variation cycle of groundwater levels, effectively smoothing intraday fluctuations and single-point noise. In this embodiment, the statistical window width is set to seven time slices.

[0060] If the statistical representative value remains below a preset percentage threshold for three consecutive time slices, the hydraulic connectivity anomaly is downgraded to a suspected hydraulic connectivity anomaly, and a re-inspection process is initiated. The preset percentage is typically between 50% and 70%, determined based on the statistical characteristics of the hydraulic system recovery process. When the statistical representative value is below 60% of the original threshold, the clustering of anomaly points has significantly decreased, and the connectivity of the hydraulic system tends to recover. Downgrading at this point avoids maintaining a false alarm state for an extended period. In this embodiment, the preset percentage is set to 60%. The re-inspection process includes the following specific steps: (1) Recheck the quality of the original monitoring data and check whether the data integrity rate of each monitoring point is less than 90% during the abnormal period. If it is less than 90%, the data quality is marked as suspicious. (2) Check whether the monitoring well has suffered physical damage. Record the changes in wellhead elevation, the integrity of the well pipe, and the blockage of the filter pipe through on-site inspection. If the well pipe is found to be broken or the blockage rate of the filter pipe exceeds 30%, the monitoring well is marked as abnormal. (3) Compare the surrounding meteorological and hydrological data and retrieve the rainfall, evaporation, and extraction data of surrounding pumping wells within 7 days before and after the abnormal period. If the rainfall exceeds 50 mm during the abnormal period or the extraction volume of surrounding pumping wells changes by more than 20%, the external interference factor is marked. If any of the above three checks confirms the existence of abnormal factors, the suspected hydraulic connectivity abnormality mark is maintained and the technicians are prompted to manually review it. If all three checks confirm that there are no abnormal factors, the hydraulic connectivity abnormality mark is removed and the monitoring status is restored to normal.

[0061] If the statistical representative value rises above the preset percentage threshold within three consecutive time slices, the suspected hydraulic connectivity anomaly recovery is marked as a hydraulic connectivity anomaly. The dynamic maintenance mechanism avoids false alarms caused by single data fluctuations, while ensuring that genuine changes in hydraulic connectivity are not masked by temporary data dips.

[0062] A4. After marking the anomaly, starting from the anomaly point, trace the influence range along the directed edges of the time-series causal relationship graph where no anomaly has occurred, and estimate the spatial migration distance of the pollution front based on the flux change amplitude of each anomaly point.

[0063] Starting from the marked anomaly point, the process extends downstream level by level along all directed edges in the temporal causal relationship graph that have not reversed or disappeared, including all monitoring points passed during the extension. The downstream direction refers to the direction of the hydraulic gradient, i.e., from high to low water head. During the tracing process, starting from the anomaly point, the process extends downstream level by level along the directed edge (i.e., the outgoing edge direction) that serves as the driving point, skipping directed edges that have reversed or disappeared, and only extending along directed edges that maintain a normal direction and have a stable state. The set of visited vertices is recorded during the tracing process; when a visited vertex is encountered, the tracing of that branch stops to avoid infinite traversal caused by loops. The tracing depth is typically set between levels 2 and 5. This range is determined based on the theory of convection dispersion of pollutants in groundwater. When the migration distance of the pollution front is on the same order of magnitude as the distance between monitoring points, level 3 tracing can cover the influence range from the anomaly source to 2 to 3 downstream points, corresponding to a distance of 200 to 450 meters. This is sufficient to cover the emergency response needs at the park scale, while avoiding the inclusion of irrelevant remote points due to excessive tracing depth. In this embodiment, the tracing depth is determined to be level 3.

[0064] For each monitoring point marked as an anomaly, trace its upstream and downstream related points in the time-series causal relationship diagram, and extract the flux change direction of the upstream and downstream related points before and after the transition moment. Taking anomaly point W13 as an example, assume that in the time-series causal relationship diagram, W12 pointing to W13 is the upstream edge, and W13 pointing to W14 is the downstream edge. Before the transition moment, the direction of W12 pointing to W13 is W12 pointing to W13, and after the transition moment, this edge disappears; at the same time, the direction of W13 pointing to W14 is reversed after the transition moment, becoming W14 pointing to W13. If there are multiple upstream related points for an anomaly point, and some upstream edges disappear while others remain normal, the determination is based on the proportion of the number of disappeared upstream edges to the total number of upstream edges: if the disappearance ratio exceeds 50%, the anomaly point is determined to be a pollution source release point; If the disappearance rate is less than or equal to 50%, the original type determination of the anomalous point is maintained, based on the normal flux change direction of the upstream edge. If the anomalous point has multiple downstream related points, with some downstream edges reversing and others remaining normal, the determination is based on the proportion of the number of reversing downstream edges to the total number of downstream edges: if the reversal proportion exceeds 50%, the anomalous point is determined to be a pollution migration path deflection point; if the reversal proportion is less than or equal to 50%, the original type determination of the anomalous point is maintained, based on the normal flux change direction of the downstream edge. The above 50% threshold is determined based on the majority voting principle: when more than half of the upstream or downstream edges are anomalous, the anomalous mechanism of the anomalous point can be considered to be dominated by the corresponding type, thus allowing for type determination; this threshold is statistically robust, preventing misjudgment due to occasional anomalies of individual edges, and also preventing the omission of true mechanism changes due to requiring all edges to be anomalous.

[0065] The technical essence of the above judgment logic lies in the following: the disappearance of the upstream edge means that the original hydraulic transmission path flowing to the anomaly point is interrupted, and a new pollution source may be released at the anomaly point, leading to the reconstruction of the local hydraulic field; the reversal of the downstream edge means that the original outward hydraulic channel of the anomaly point has reversed, and the pollution migration path has been deflected. By distinguishing between these two anomaly mechanisms, differentiated physical constraints can be provided for subsequent pollution front estimation.

[0066] After distinguishing between pollution source release points and pollution migration path deflection points, the spatial migration distance of the pollution front is estimated using different propagation coefficients. For pollution source release points, the propagation coefficient is typically between 1.0 and 1.5. This range is determined based on the source term enhancement effect in solute transport theory. When pollution is directly released into the aquifer, the diffusion flux driven by the concentration gradient is superimposed on hydraulic convection, making the apparent migration velocity of the pollution front higher than the pure Darcy velocity. Bear, in "Porous Media Fluid Dynamics," points out that the source enhancement coefficient is typically in the range of 1.0 to 1.5. In this embodiment, based on the type of pollution source in the industrial park (electroplating wastewater leakage) and the characteristics of the aquifer, the propagation coefficient of the pollution source release point is determined to be 1.2. For pollution migration path deflection points, the propagation coefficient typically ranges from 0.6 to 1.0. This range is determined based on the theory of local head loss in hydraulics. Path deflection means that groundwater needs to take a longer conduction path or cross a low-permeability medium. Energy dissipation leads to a decrease in migration velocity, and additional resistance typically reduces the effective conduction coefficient by 20% to 40%. In this embodiment, the propagation coefficient at the pollution migration path deflection point is determined to be 0.8.

[0067] The spatial migration distance of the pollution front is estimated based on the flux change amplitude at each anomaly point. The flux change amplitude at each anomaly point refers to the absolute value of the difference between the vertical flux value at that point in the time slice after the jump and the vertical flux value in the time slice before the jump. Taking anomaly point W13 as an example, the vertical flux before the jump was 2.5 x 10⁻⁷ m / s, and the vertical flux after the jump was 4.0 x 10⁻⁷ m / s. Therefore, the flux change amplitude of W13 is 1.5 x 10⁻⁷ m / s. The anomaly occurrence time refers to the moment when the monitoring point is first marked as an anomaly point, that is, the moment corresponding to the time slice when the directed edge associated with that point first reverses or disappears, not the moment when the hydraulic connectivity anomaly is confirmed as a whole.

[0068] For each point included in the affected area set, the deviation of the flux change after the anomaly occurred from the flux change before the anomaly occurred is obtained, and the proportion of this deviation relative to the flux change before the anomaly is calculated. Taking point W15 in the affected area set as an example, the vertical flux change of W15 before the anomaly occurred was 2 x 10⁻⁷ m / s, and after the anomaly occurred, it was 3.5 x 10⁻⁷ m / s. The deviation is equal to 1.5 x 10⁻⁷ m / s. The deviation proportion is equal to the deviation divided by the flux change before the anomaly occurred, resulting in 75%.

[0069] The propagation attenuation characteristics of each location are determined based on the deviation ratio. The spatial migration distance of the pollution front along the directed edge is then calculated step-by-step based on these characteristics. The relationship between the deviation ratio and the migration distance is determined using the analytical solution of the convection-dispersion equation: In the one-dimensional convection-dispersion model, the migration distance of the pollution front is positively correlated with the flux change, but due to the influence of aquifer dispersion, the relationship is not linear. When the deviation ratio is below 50%, dispersion dominates, the front advances slowly, and the migration distance is approximately 80% to 100% of the distance between adjacent locations. When the deviation ratio is between 50% and 100%, convection and dispersion are roughly equal, and the migration distance is 50% to 80% of the distance between adjacent locations. When the deviation ratio exceeds 100%, convection dominates but is limited by the hydraulic gradient saturation effect, and the migration distance is restricted to 20% to 50% of the distance between adjacent locations. The above three-segment division refers to the discussion on the propagation characteristics of the solute front in Bear's "Porous Media Fluid Dynamics". The deviation ratio of 50% corresponds to the critical point of transition between dispersion-dominated and convection-dominated modes, and the deviation ratio of 100% corresponds to the critical point of gradient saturation under convection-dominated mode.

[0070] The deviation rate of W15 is 75%, which is considered moderate propagation attenuation, corresponding to a migration distance of 50% to 80% of the distance between adjacent points. Assuming the horizontal distance between W15 and the upstream point W14 is 100 meters, the spatial migration distance of the pollution front from W14 to W15 is 50 to 80 meters, with the median value of 65 meters taken as the representative value.

[0071] The tracing process also includes segmented calculation of the attenuation mechanism of the propagation path.

[0072] For each directed edge that did not exhibit any anomalies during the tracing process, the type of propagation medium to which the directed edge belongs is determined based on the lithological differences between the upstream and downstream points connected by the directed edge. If the strata between the upstream and downstream points are mainly composed of fractured bedrock with a fracture density greater than 5 fractures per meter, it is classified as a fractured medium. If the strata are mainly composed of gravel or silt with a porosity greater than 15%, it is classified as a porous medium.

[0073] If the propagation medium is a fractured medium, the equivalent hydraulic conductivity coefficient corresponding to the directed edge is calculated using the cubic law of fracture networks, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. The formula for the cubic law of fracture networks is: the equivalent hydraulic conductivity coefficient equals the density of water multiplied by the gravitational acceleration, multiplied by the cube of the fracture aperture, divided by 12 times the dynamic viscosity coefficient. The density of water is taken as 1000 kg / m³, the gravitational acceleration as 9.8 m / s², and the fracture aperture is determined based on fracture measurement data, typically between 0.1 mm and 1 mm. If there is no measured fracture aperture data in the park, a default value of 0.5 mm is used. This default value is determined based on the recommended range of hydraulic parameters for fractured media in the "Hydrogeological Handbook," and the median fracture aperture of the typical fracture aperture at the interface between gravel aquifers and bedrock fractures is between 0.4 and 0.6 mm. The dynamic viscosity coefficient is taken as 1.002 multiplied by 10 to the power of -3 Pascals per second. The normalized baseline value is the maximum value of the equivalent hydraulic conductivity coefficient of each segment in all tracing paths within the park.

[0074] If the propagation medium is porous, the measured permeability coefficient of the region corresponding to the directed edge is used as the equivalent hydraulic conductivity coefficient, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. The measured permeability coefficient of the porous medium is determined based on the hydrogeological survey report of the park. The normalized baseline value is also the maximum value of the equivalent hydraulic conductivity coefficient of each segment in all tracing paths within the park.

[0075] The aforementioned deviation ratios are corrected using propagation correction factors for each segment, and the spatial migration distance of the pollution front is calculated step by step based on the corrected deviation ratios. The correction formula is that the corrected deviation ratio equals the original deviation ratio multiplied by the propagation correction factor for that segment. By calculating in segments using attenuation mechanisms, the hindering or promoting effects of different media types on pollution migration can be reflected more accurately.

[0076] Existing technologies treat anomalous locations uniformly, but pollution source release leads to a surge in pollution flux and accelerated migration, while path deflection causes changes in migration direction and localized pollution accumulation. This invention determines pollution source release by identifying the disappearance of the upstream edge and path deflection by identifying the reverse of the downstream edge, quantifying both effects with a differentiated propagation coefficient. This precise differentiation of location types avoids the distance estimation bias caused by the uniform propagation model. According to Bear's solute transport theory, when both source strength and path changes exist simultaneously, the concentration prediction error of the uniform propagation model is positively correlated with the product of the dispersion coefficient and the rate of change of source strength. In contrast, the differentiated propagation model decouples the error sources by separating the source and path terms, resulting in higher theoretical prediction accuracy than the uniform model.

[0077] As a specific implementation method, a chemical industrial park applied the attenuation mechanism for segmented calculations during a designated period. The strata between anomaly point W16 and downstream point W17 are mainly bedrock with well-developed fractures, with a fracture density of 8 fractures per meter, and are classified as fractured media. The measured fracture aperture is 0.3 mm. The equivalent hydraulic conductivity is calculated as 1000 multiplied by 9.8 multiplied by 0.3 mm cubed, divided by 12 multiplied by 1.002 multiplied by 10 to the power of -3, resulting in approximately 7.35 multiplied by 10 to the power of -5 meters per second. The maximum equivalent hydraulic conductivity among all tracing paths within the park is 2.1 multiplied by 10 to the power of -4 meters per second (located in the W3-W4 segment of porous media). The normalized propagation correction factor is calculated as 7.35 multiplied by 10 to the power of -5 divided by 2.1 multiplied by 10 to the power of -4, resulting in 0.35. The original deviation ratio for W16 was 80%. The corrected deviation ratio was 80% multiplied by 0.35, resulting in 28%. A corrected deviation ratio below 50% corresponds to a migration distance of 80% to 100% of the distance between adjacent points. With a distance of 120 meters between W16 and W17, the spatial migration distance of the pollution front from W16 to W17 is 96 to 120 meters, taking the median of 108 meters. Without segmented correction, calculated using the original deviation ratio of 80%, the migration distance would be 60 to 96 meters, taking the median of 78 meters, which is significantly underestimated compared to the rapid propagation characteristics in actual fractured media. Therefore, segmented calculation using attenuation mechanisms avoids distance estimation errors caused by misjudgment of the media type.

[0078] A5. Delineate the risk boundary based on the spatial migration distance, overlay it onto the three-dimensional stratigraphic grid, determine the comprehensive early warning level based on the cumulative flux value of the nodes, and output it.

[0079] The risk boundary refers to the envelope of a circular area of ​​influence drawn on a horizontal plane, with each anomaly point as its center and the corresponding spatial migration distance as its radius. Taking anomaly point W13 as an example, its estimated spatial migration distance is 120 meters, so a circular area of ​​influence is drawn with W13 as its center and 120 meters as its radius. Taking the union of the influence areas of all anomaly points yields the projection range of the risk boundary on the horizontal plane.

[0080] The three-dimensional stratigraphic mesh discretizes the underground space of the park into regular hexahedral grid units. The horizontal grid spacing is typically between 30 and 100 meters, a range determined by the 100-200 meter well spacing specified in the "Technical Specification for Groundwater Environmental Monitoring" (HJ164-2020). Using an intermediate value allows for reasonable interpolation between wells while ensuring computational efficiency. Vertically, the aquifer is divided into multiple layers based on its structure, with layer thickness determined by lithological variations, typically between 2 and 10 meters. In this embodiment, the horizontal grid spacing is set to 50 meters, and the vertical aquifer is divided into 5 layers based on its structure: shallow layer (0-5 meters, 5 meters thick), middle layer (5-15 meters, 10 meters thick), deep layer (15-30 meters, 15 meters thick), and two transition layers. Each grid node records coordinates, lithology, permeability coefficient, porosity, current water level, and current concentration attributes.

[0081] The risk boundary on the horizontal plane is projected and superimposed onto each layer of the 3D stratigraphic grid. For grid nodes located inside the risk boundary, their cumulative flux value is calculated. The cumulative flux value is the vertical flux time integral value of the grid node's location from the time of anomaly occurrence to the current time. Specifically, it is obtained by numerically integrating the vertical flux value of the node at each time step using the trapezoidal rule, summing the product of the vertical flux value and the duration of each time step. The trapezoidal rule numerical integration formula is as follows: for two adjacent time steps tk and tk+1, the integral term equals the sum of qz(tk) and qz(tk+1) multiplied by Δt and divided by 2, where Δt is the time step length; summing the integral terms of all time steps yields the cumulative flux value.

[0082] For grid nodes not containing monitoring wells, the vertical flux value of the node is calculated using inverse distance weighted interpolation. Specifically, with the grid node as the center, all monitoring wells within a 200-meter radius are searched. The vertical flux value of each monitoring well is weighted and averaged using the reciprocal of its distance from the node to the node, thus obtaining the vertical flux value of the grid node. The inverse distance weighted interpolation formula is: the vertical flux value of the node is equal to the sum of the vertical flux values ​​of each monitoring well multiplied by their corresponding weights, divided by the total weights; where the weight of the i-th monitoring well is equal to 1 divided by the horizontal distance from the monitoring well to the target node. The 200-meter search radius is set based on the following: in this embodiment, the spacing between monitoring wells is 80 to 150 meters. A 200-meter search radius can cover at least 2 to 3 monitoring wells around each grid node, ensuring the spatial representativeness of the interpolation results; if there are no monitoring wells within the search range, the vertical flux value of the node is marked as missing and is not included in the cumulative flux calculation. In the early warning level determination, the node inherits the early warning level of its vertical nearest neighbor node. The vertical nearest neighbor node refers to the adjacent layer node in the three-dimensional stratigraphic grid that is located at the same horizontal coordinate as the node and has the smallest vertical distance.

[0083] The comprehensive early warning level is determined based on the cumulative flux value at each node. The comprehensive early warning level is divided into four levels, with thresholds derived quantitatively from Darcy's Law. The specific derivation is as follows: Assume the typical Darcy velocity v equals K multiplied by i and divided by ne. Taking the typical vertical permeability coefficient K of this park as 2 x 10⁻⁵ m / s, the typical hydraulic gradient i as 0.01, and the effective porosity ne as 0.25, then v equals 2 x 10⁻⁵ multiplied by 0.01 and divided by 0.25, resulting in 8 x 10⁻⁷ m / s, or approximately 0.069 m / s per day. A cumulative flux value of 0.5 m corresponds to a continuous flux time of approximately 7.2 days, corresponding to approximately one week of continuous low-flux leakage; a cumulative flux value of 1.0 m corresponds to a continuous flux time of approximately 14.5 days, corresponding to approximately two weeks of continuous moderate-intensity leakage; and a cumulative flux value of 2.0 m corresponds to a continuous flux time of approximately 29 days, corresponding to approximately one month of continuous high-intensity leakage. This threshold system quantitatively correlates flux intensity with pollution duration, and is consistent with the graded management approach for pollutant concentration in the Groundwater Quality Standard (GB / T14848-2017).

[0084] The comprehensive early warning levels are specifically classified as follows: Level 1 Blue Warning corresponds to a cumulative flux value of less than 0.5 meters, indicating a low-intensity leakage risk within about 1 week; Level 2 Yellow Warning corresponds to 0.5 to 1.0 meters, indicating a moderate-intensity leakage risk within about 1 to 2 weeks; Level 3 Orange Warning corresponds to 1.0 to 2.0 meters, indicating a high-intensity leakage risk within about 2 weeks to 1 month; and Level 4 Red Warning corresponds to greater than 2.0 meters, indicating a continuous high-intensity leakage risk exceeding 1 month.

[0085] Simultaneously, adjustments are made based on the distance of the pollution front from the park boundary or sensitive targets. Sensitive targets include drinking water source wells, surface water intakes, and ecological protection zone boundaries surrounding the park, with their coordinates determined by the park's environmental management records. If the pollution front is less than 100 meters from a sensitive target, the warning level is raised by one level; if it is greater than 500 meters, the warning level is lowered by one level (but not lower than level one). The aforementioned 100-meter and 500-meter adjustments are determined according to the classification of groundwater sensitive target protection distances in the "Technical Guidelines for the Division of Groundwater Pollution Prevention and Control Zones," where 100 meters corresponds to the boundary distance of the Level I protection zone (key protection), and 500 meters corresponds to the boundary distance of the Level II protection zone (general protection).

[0086] The comprehensive early warning level is output in a 3D visualization format. Different colors are used to render the areas within the risk boundary on the park's floor plan, and the distribution of warning levels across vertical layers is displayed on the cross-sectional view. The output interface is the standard data interface of the park's environmental management platform. The output format includes a JSON structured data package and a Shapefile spatial data file. The JSON data package contains the warning level, a list of abnormal locations, coordinates of the affected area, the migration distance of the pollution front, and recommended remedial measures. The Shapefile file contains the spatial vector data of the risk boundary, which can be directly imported into a GIS system for overlay analysis.

[0087] Existing two-dimensional planar assessments neglect vertical pollution migration and aquifer heterogeneity. This invention superimposes risk boundaries in three-dimensional space, allowing the cumulative flux values ​​at nodes to comprehensively reflect the intensity and duration of pollution flux, and quantitatively corresponding the four warning levels to the cumulative flux values. By leveraging the fine characterization of the three-dimensional grid, it avoids the coarseness of two-dimensional assessments, achieving precise spatial positioning and quantitative classification of pollution risks.

[0088] As a specific implementation method, a complete calculation is explained using grid node G(100,200,10) as an example. The horizontal coordinates of this node are X=100 meters and Y=200 meters in the park coordinate system, and the vertical coordinates are the 10-meter underground depth layer. The anomaly occurred at 00:00 on April 18, 2024, and the current time is 00:00 on April 25, for a total of 168 time steps (one per hour). There are no monitoring wells at the location of this node, so the vertical flux value of each time step needs to be calculated by inverse distance weighted interpolation. Taking the first time step on April 18 as an example, monitoring wells within a 200-meter radius of this node are searched, and three monitoring wells W5 (horizontal distance 80 meters), W6 (horizontal distance 120 meters), and W9 (horizontal distance 150 meters) are found. The vertical flux at time step W5 is 3.5 x 10⁻⁷ m / s, W6 is 2.8 x 10⁻⁷ m / s, and W9 is 2.2 x 10⁻⁷ m / s. The inverse distance weights are: W5 weight = 1 divided by 80 (0.0125), W6 weight = 1 divided by 120 (0.0083), and W9 weight = 1 divided by 150 (0.0067), with a total weight of 0.0275. The vertical flux at this node at this time step is calculated as (3.5 x 10⁻⁷ x 0.0125 + 2.8 x 10⁻⁷ x 0.0083 + 2.2 x 10⁻⁷ x 0.0067) divided by 0.0275, which is approximately 2.97 x 10⁻⁷ m / s. After calculating the vertical flux values ​​for 168 time steps using this method, numerical integration was performed using the trapezoidal rule: the first integral term equals (first-step flux 2.97 x 10⁻⁷ + second-step flux 3.02 x 10⁻⁷) x 3600 seconds divided by 2, resulting in 1.08 x 10⁻³ meters; the second integral term equals (second-step flux 3.02 x 10⁻⁷ + third-step flux 3.15 x 10⁻⁷) x 3600 divided by 2, resulting in 1.11 x 10⁻³ meters; and so on, accumulating the 167 integral terms for the 168 time steps to obtain a cumulative flux value of 1.35 meters at the node. This value falls within the range of 1.0 to 2.0 meters, corresponding to a Level 3 orange alert. The node is 320 meters horizontally from the nearest sensitive target (a drinking water source well on the east side of the park, coordinates X=350, Y=180). This does not meet the condition of upgrading the alert level by one level if the distance is less than 100 meters or downgrading it by one level if the distance is greater than 500 meters. Therefore, the alert level remains unchanged at Level 3 Orange. The alert level result for this node is written to a JSON data packet, and the risk boundary range it occupies is written to a Shapefile file.

[0089] A6. Based on the measured offset between the pollution boundary and the risk boundary, adjust the preset percentage threshold and the sliding window width according to the preset step feedback rule.

[0090] The measured pollution boundary is obtained through actual sampling analysis or geophysical exploration. The specific methods for obtaining this boundary are as follows: (1) Sampling and analysis method: Supplementary sampling points are set up inside and outside the projected range of the risk boundary. The inner sampling points are set up at 50-meter intervals, 20 meters away from the risk boundary into the park, and the outer sampling points are set up at 50-meter intervals, 20 meters away from the risk boundary outward. Groundwater samples are collected at each sampling point, and the detection indicators are consistent with those of the monitoring well (hexavalent chromium concentration in this embodiment). The sampling depth is consistent with the shallow depth of the corresponding monitoring well (5 meters underground). The criteria for determining the measured pollution boundary are: sampling points whose concentration exceeds twice the background value of the layer at the depth of the point, or exceed the Class III water standard limit in the "Groundwater Quality Standard" (GB / T14848-2017), are determined to be pollution points; adjacent pollution points are connected by the minimum convex hull to form the measured pollution boundary. If the number of pollution points in both the inner and outer sampling points is less than 3, the sampling interval is increased to 25 meters and the sampling points are redone.

[0091] (2) Geophysical Exploration Acquisition Method (Optional Auxiliary): High-density resistivity method is used for detection, with an electrode spacing of 10 meters. A parallel survey line is laid out both inside and outside the risk boundary, and the length of the survey line covers the projected range of the risk boundary. Data processing adopts the least squares inversion method. The specific inversion parameters are set as follows: the initial model resistivity is taken as the average value of the background resistivity of the survey area, the regularization factor is between 0.1 and 1.0 (dynamically adjusted according to data quality, with a larger value when the data noise is high), the number of iterations is set to 5 to 10, and the convergence condition is that the model fitting difference between two adjacent iterations is less than 5% or the maximum number of iterations is reached. After the inversion is completed, the resistivity values ​​of each grid node are extracted. Areas with resistivity more than 30% lower than the surrounding background resistivity are identified as low-resistivity anomaly areas. The outer envelope of the low-resistivity anomaly area is used as the pollution boundary for geophysical exploration. When the spatial deviation between the sampling analysis results and the geophysical exploration results exceeds 20% of the average distance between monitoring points, the sampling analysis results shall prevail, and the geophysical exploration results shall only be used as auxiliary verification; when the deviation is within 20%, the arithmetic mean of the two boundaries shall be taken as the comprehensive measured pollution boundary.

[0092] (3) Data quality control: During the sampling and analysis process, parallel double samples are collected at each sampling point, and the relative deviation is controlled within 20%; standard substances are used for quality control in laboratory analysis, and the recovery rate is controlled between 80% and 120%. Data from sampling points that do not meet the above quality control requirements are discarded, and resampling is carried out at the same point. A measured pollution boundary survey is conducted once a month, and the survey cycle is synchronized with the park's environmental monitoring plan.

[0093] The measured results are compared with the risk boundary calculated using A5. The offset is defined as the average distance difference between the measured pollution boundary and the risk boundary. Specifically, at least 36 sampling points (one every 10 degrees) are taken on the risk boundary according to the principle of equal arc length. The vertical distance between the measured boundary and the risk boundary at each sampling point is calculated, and the arithmetic mean is taken as the offset. When the boundary shape is complex or there are branches, 36 sampling points are taken for each independent branch to calculate the local branch offset, and then the offset is obtained by weighting the average according to the branch length.

[0094] As a specific step-feedback rule, if the offset is positive, the measured pollution boundary is outside the risk boundary, meaning the actual pollution range is larger than the predicted range. This indicates that the current warning parameters are too conservative, and the risk of missed detection is high. In this case, the preset percentage threshold is reduced by 10%, and the sliding window width is reduced by one time slice, making the system more sensitive to changes in hydraulic connectivity. If the offset is negative, the measured pollution boundary is inside the risk boundary, meaning the actual pollution range is smaller than the predicted range. This indicates that the current warning parameters are too aggressive, and the risk of false alarms is high. In this case, the preset percentage threshold is increased by 10%, and the sliding window width is increased by one time slice, making the system more tolerant of short-term fluctuations.

[0095] The preset percentage threshold and sliding window width are adjusted using a linkage mechanism, meaning both parameters are adjusted simultaneously. If the offset direction changes after two consecutive adjustments (from positive to negative or vice versa), it indicates the system is overly sensitive to parameter adjustments. In this case, only the preset percentage threshold is adjusted, keeping the sliding window width unchanged, until the offset stabilizes in the same direction before resuming linkage adjustment. If the absolute value of the offset increases after two consecutive adjustments, it indicates an incorrect adjustment direction. In this case, the adjustment direction is reversed (threshold and window are adjusted in reverse), and the adjustment step size is halved. This adaptive logic originates from the feedback regulation principle in control theory: when the system oscillates (the offset direction changes repeatedly), the regulation dimension should be reduced to stabilize the system; when the regulation effect is opposite to the expectation, the regulation direction should be reversed and the step size reduced to avoid divergence.

[0096] The adjusted parameters take effect within the next three time slices, during which the offset changes continue to be monitored. If a new measurement cycle (monthly pollution boundary survey) occurs during this effective period, the latest measurement result will be used, and the current parameter adjustment will be paused until the measurement is completed and the adjustment amount is recalculated and executed based on the new offset. If three consecutive adjustments are made in the same direction, the adjustment step size will be doubled to accelerate parameter convergence. If the absolute value of the offset is less than 10% of the average spacing between monitoring points, the current parameter is considered to have converged, the adjustment will stop, and the current parameter will be maintained until the next measurement cycle. The 10% convergence tolerance is consistent with the relative error allowable range commonly found in engineering surveying. In the field of groundwater environmental monitoring, when the deviation between the predicted boundary and the measured boundary is less than 10% of the monitoring well spacing, the prediction accuracy can be considered to have met the requirements for engineering applications. This value refers to the accuracy requirements for monitoring data quality control in the "Technical Specification for Groundwater Environmental Monitoring" (HJ164-2020).

[0097] When measured pollution boundary data is missing or of substandard quality, the system maintains the current parameters and pushes a data missing alarm, prompting technicians to conduct supplementary surveys. The data missing alarm is sent through the alarm push interface of the park's environmental management platform, and the alarm information includes the time of the missing data, the reason for the missing data, and a list of recommended supplementary measurement points. Substandard quality refers to a relative deviation of more than 20% between parallel samples in the sampling analysis, a standard substance recovery rate exceeding 80% to 120%, or a signal-to-noise ratio of geophysical exploration data below 10 dB.

[0098] Existing technologies use fixed early warning parameters once set, failing to adapt to seasonal changes in park hydrogeological conditions or long-term evolution caused by engineering activities. This invention quantifies the prediction deviation by measuring the offset between the measured pollution boundary and the risk boundary, adaptively adjusting the preset percentage threshold and sliding window width according to the offset direction, and dynamically doubling the adjustment step size with continuous adjustment. This feedback correction using measured data avoids systematic deviations of fixed parameters under time-varying conditions. According to control theory, when the feedback adjustment satisfies the Lyapunov stability condition, the system parameters converge to near the optimal value that minimizes the prediction error. This condition requires the offset as a feedback signal to satisfy a monotonically decreasing characteristic; this invention ensures this characteristic is maintained during the iteration process through an adaptive step size mechanism.

[0099] As a specific implementation method, it is assumed that a chemical industrial park applied this method for groundwater pollution early warning from March to June 2024. The park deployed 20 multi-layer monitoring wells, with a well spacing of 80 to 150 meters. The aquifer was a gravel-sand porous aquifer, and the main pollutant was hexavalent chromium. The historical stable period was set from March 1st to March 31st, 2024, comprising 30 time slices.

[0100] On April 15th, the system detected that within a narrow window, the direction of the directed edge W7 pointing to W8 was reversed to W8 pointing to W7. Within a wide window, the cumulative frequency of W7 pointing to W8 decreased to 5%, below 50% of the historical stable period frequency of 25%, confirming an anomaly across scales. In the local orientation field consistency screening of W8, two of its three adjacent points had an angle less than 30 degrees, a proportion of 66.7%, higher than 60%, and were retained as anomaly points. During the same period, two other anomaly points, W3 and W12, were also detected, accounting for 15% of the total, which did not exceed the 30% threshold.

[0101] On April 18th, W5 disappeared pointing to W6, and W9 reversed its pointing direction to W10, adding four new anomalous points: W5, W6, W9, and W10. After direction field consistency filtering, W5, W6, and W9 were retained. A total of six anomalous points were identified, representing 30%, which exactly reached the threshold and were marked as hydraulic connectivity anomalies.

[0102] The system traces the impact range starting from points W5, W6, W7, W8, and W9. The upstream edge of W5 disappears, indicating a pollution source release point with a propagation coefficient of 1.2. The downstream edge of W8 reverses direction, indicating a deflection point in the pollution migration path with a propagation coefficient of 0.8. After deviation ratio calculation and segmented correction based on attenuation mechanisms, the impact radius of W5 is 144 meters, and the impact radius of W8 is 96 meters. The combined risk boundary covers an area of ​​approximately 0.12 square kilometers.

[0103] In the 3D stratigraphic grid, the maximum cumulative flux value at each node is 1.35 meters, corresponding to a Level 3 orange alert. The system outputs the alert information to the park's environmental management department. On April 20, manual sampling confirmed that the concentration of hexavalent chromium within 30 meters of W5 increased from 0.3 mg / L to 1.2 mg / L. The actual location of the pollution front deviated from the predicted location by approximately 18 meters, a positive deviation.

[0104] Based on the offset, the system reduced the preset percentage threshold from 30% to 27% and the sliding window width from 7 days to 6 days. After the adjustment, on April 25, the system detected a slight change in hydraulic connectivity in another area two days earlier, which was subsequently confirmed as a real pollution migration event.

[0105] It should be noted that the calculation process for the 144-meter influence radius of W5 in the above scenario example is as follows: W5 is the pollution source release point, and the propagation coefficient is taken as 1.2. Before the jump moment of W5, the vertical flux was 2.0 x 10^-7 m / s, and after the jump moment, it was 5.0 x 10^-7 m / s, with a flux change range of 3.0 x 10^-7 m / s. The horizontal distance between W5 and the downstream adjacent point W6 is 120 meters. Before the anomaly occurred, the vertical flux change range of W6 was 1.5 x 10^-7 m / s, and after the anomaly occurred, it was 4.5 x 10^-7 m / s, with a deviation of 3.0 x 10^-7 m / s. The deviation ratio is equal to 3.0 divided by 1.5, resulting in 200%. If the deviation ratio exceeds 100%, the corresponding migration distance is 20% to 50% of the distance between adjacent points, that is, 24 meters to 60 meters, with the median value of 42 meters.

[0106] After a propagation coefficient correction of 1.2, the migration distance is 42 meters multiplied by 1.2, resulting in 50.4 meters. Tracing back to W6, W6 continues to be traced downstream. The distance between W6 and W7 is 100 meters, and the deviation ratio of W7 is 85%, corresponding to a migration distance of 50 to 80 meters. Taking the median of 65 meters, this is corrected to 78 meters by the propagation coefficient correction of 1.2. The cumulative migration distance from W5 to W7 is 50.4 meters plus 78 meters, approximately 128 meters. Considering the tracing results in other directions around W5, the maximum influence radius of 144 meters is taken as the representative value for W5.

[0107] The core innovation of this technical solution lies in transforming groundwater pollution early warning from static concentration monitoring to dynamic hydraulic connectivity early warning. Traditional methods treat monitoring points as isolated individuals, determining pollution status based on concentration thresholds. This invention, however, recognizes that the groundwater system is a continuous hydraulic medium, and the hydraulic connections between monitoring points are the fundamental driving force for pollution migration. A time-series causal relationship diagram makes the hydraulic connections between monitoring points explicit; a sliding window detection quantifies abrupt changes in hydraulic connections; a differentiated propagation coefficient distinguishes the migration effects of different anomaly types; a three-dimensional stratigraphic grid refines the spatial distribution of risk; and an adaptive closed loop automates the continuous optimization of early warning parameters. These five interconnected steps form a complete technical closed loop from data acquisition to risk assessment and parameter optimization.

[0108] In summary, this invention transforms water level time series into flux time series through vertical flux calculation driven by water level fluctuations, achieving early warning by leveraging the pre-indicative role of hydraulic gradient changes. It organizes dispersed monitoring points into a graph structure reflecting hydraulic transmission paths through the construction of a time-series causal relationship graph, and combines multi-scale collaborative detection with sliding window and local orientation field consistency screening to avoid frequent failures of static thresholds under non-stationary conditions. It avoids distance estimation bias caused by a unified propagation model by distinguishing between pollution source release points and pollution migration path deflection points and using differentiated propagation coefficients and attenuation mechanisms to segmentally correct pollution front calculations. It achieves precise spatial positioning and quantitative classification of pollution risk through three-dimensional stratigraphic grid risk assessment, and adaptively adjusts early warning parameters based on the offset between measured pollution boundaries and risk boundaries to avoid systematic deviations of fixed parameters under time-varying conditions. Those skilled in the art can make adaptive adjustments to the number of monitoring points, time slice length, sliding window width, and early warning level threshold parameters without departing from the principles of this invention; all such adjustments should be considered within the scope of protection of this invention.

Claims

1. A method for evaluating the early warning level of soil and groundwater pollution in industrial parks, characterized in that, include: A1. Calculate the vertical flux sequence and hydraulic gradient direction driven by water level fluctuations within each time step based on the water level time series and concentration time series of the monitoring points; A2. Using continuous time steps as time slices, determine the driving-response time sequence relationship between monitoring points based on the order of flux changes at each monitoring point on the time axis. Construct a time-series causal relationship graph with the direction from the driving point to the response point as the edge and the monitoring point as the vertex, and record the topological structure change trajectory of the time-series causal relationship graph on the time slice sequence. A3. A sliding window traverses the topological change trajectory. When a directed edge of a preset type reverses or disappears in the temporal causal relationship graph within the window, the monitoring point associated with the reversed or disappeared directed edge is marked as an abnormal point. The proportion of abnormal points to all points is counted. If the proportion exceeds a preset proportion threshold, it is marked as a hydraulic connectivity anomaly. The preset type is automatically determined by the system based on the frequency of occurrence of each directed edge in the temporal causal relationship graph during historical stable periods. A4. After marking the anomaly, starting from the anomaly point, trace the influence range along the directed edges of the time-series causal relationship graph where no anomaly has occurred, and estimate the spatial migration distance of the pollution front based on the flux change amplitude of each anomaly point. A5. Delineate the risk boundary based on the spatial migration distance, overlay it onto a three-dimensional stratigraphic grid, determine the comprehensive early warning level based on the cumulative flux value of the nodes, and output the result. A6. Based on the measured offset between the pollution boundary and the risk boundary, adjust the preset percentage threshold and the sliding window width according to the preset step feedback rule.

2. The method according to claim 1, characterized in that, The method for determining the driver-response timing relationship in step A2 is as follows: For any two monitoring points, extract the times when the extreme values ​​of flux changes occur within the time slice; If the extreme point of the first point occurs earlier than the extreme point of the second point, and the time difference between the two is less than the preset time difference threshold, then the first point is determined as the driving point and the second point is determined as the response point.

3. The method according to claim 1, characterized in that, The method for determining whether a directed edge of the preset type is reversed or disappears in step A3 is as follows: For each directed edge in the temporal causal relationship graph, record the frequency of the directed edge appearing in all time slices during the historical stationary period; If the frequency of a directed edge is lower than a preset frequency threshold, then the directed edge is marked as a low-frequency edge. When traversing the sliding window, if a low-frequency edge appears continuously in multiple consecutive time slices, or a non-low-frequency edge disappears continuously in multiple consecutive time slices, or a non-low-frequency edge reverses direction in multiple consecutive time slices, these are respectively treated as abnormal events of the corresponding preset types.

4. The method according to claim 1, characterized in that, After marking the hydraulic connectivity anomaly in step A3, the process also includes continuous tracking and dynamic maintenance of the anomaly marker: After the current time slice that marks an anomaly, the anomaly point screening and percentage statistics are repeated for each subsequent consecutive time slice; Calculate the statistical representative value of the percentage of outlier points in each time slice within a statistical window that takes the current time slice as the endpoint and covers the current slice and multiple consecutive time slices before it; If the statistical representative value is consistently lower than the preset percentage threshold within multiple consecutive time slices, the hydraulic connectivity anomaly will be downgraded to a suspected hydraulic connectivity anomaly and a re-inspection process will be initiated. If the statistical representative value rises above the preset percentage threshold within multiple consecutive time slices, the suspected hydraulic connectivity anomaly will be restored to a hydraulic connectivity anomaly.

5. The method according to claim 1, characterized in that, After marking the monitoring points associated with directed edges that have reversed or disappeared as abnormal points in step A3, the following steps are also included: For each monitoring point marked as an anomaly, trace the upstream and downstream related points of the anomaly point in the time-series causal relationship diagram, and extract the flux change direction of the upstream and downstream related points before and after the jump moment. If the directed edge between an abnormal point and its upstream associated point disappears, the abnormal point is determined to be a pollution source release point. If the directed edge between an abnormal point and its downstream associated point reverses, the abnormal point is determined to be a pollution migration path deflection point. After distinguishing between the pollution source release points and the pollution migration path deflection points, the spatial migration distance of the pollution front is estimated using different propagation coefficients.

6. The method according to claim 1, characterized in that, The step A4, which involves tracing the influence range along directed edges in the temporal causal relationship graph where no anomalies have occurred, specifically includes: Starting from the marked abnormal points, extend downstream level by level along all directed edges in the time-series causal relationship graph that have not reversed or disappeared, and include all monitoring points passed through during the extension process into the set of influence ranges. For each point included in the set of affected areas, obtain the deviation of the flux change after the time of the anomaly occurrence relative to the flux change before the time of the anomaly occurrence, and calculate the deviation ratio of the deviation relative to the flux change before the time of the anomaly occurrence. The propagation attenuation characteristics of each point are determined based on the deviation ratio, and the spatial migration distance of the pollution front along the directed edge is calculated step by step based on the propagation attenuation characteristics.

7. The method according to claim 1, characterized in that, The determination of the drive-response timing relationship in step A2 also includes verification of the physical constraints of the hydraulic conduction path: For any two monitoring points that are preliminarily determined to have a driving-response relationship based on the order of extreme point times, obtain the spatial distribution of the formation permeability coefficient and the groundwater flow direction vector field between the two points; Calculate the shortest effective hydraulic conduction path length between the two points along the groundwater flow direction, and divide the shortest effective hydraulic conduction path length by the Darcy velocity calculated based on the average permeability coefficient and average hydraulic gradient between the two points to obtain the theoretical hydraulic propagation time; If the relative deviation between the theoretical hydraulic propagation time and the time difference of the extreme point exceeds a preset deviation threshold, the driving-response relationship of the pair of points is marked as a relationship to be verified and is not included in the construction of the time-series causal relationship graph.

8. The method according to claim 1, characterized in that, The sliding window traversal described in step A3 also includes multi-timescale collaborative detection: Multiple sliding windows of different widths are set up to traverse the topological change trajectory in parallel, with each sliding window corresponding to a different time response scale; For the same directed edge, when the directed edge is detected to reverse or disappear within a narrow window, and the cumulative occurrence frequency of the directed edge in the corresponding time period within a wide window is lower than the set proportion of its cumulative occurrence frequency in historical stable periods, the abnormal event of the directed edge is marked as a cross-scale confirmed abnormality. Only monitoring points associated with cross-scale confirmed anomalies are included in the anomaly point statistics.

9. The method according to claim 1, characterized in that, Before calculating the percentage of outliers in step A3, the process also includes screening for local orientation field consistency of outliers: For each monitoring point marked as an anomaly, extract the hydraulic gradient direction change vector of that point in one time slice before and after the jump moment; Using the anomaly point as the center, search for adjacent monitoring points according to the preset spatial neighborhood radius, and calculate the angle between the hydraulic gradient direction change vector of the adjacent points at the corresponding time and the direction change vector of the center point. If the proportion of the number of adjacent points with an included angle less than a preset included angle threshold to the total number of adjacent points is higher than a preset neighborhood consistency threshold, then the center point is retained in the abnormal points. If the ratio is lower than the preset neighborhood consistency threshold, the center point will be removed from the abnormal points.

10. The method according to claim 1, characterized in that, Step A4, which involves tracing the influence range along directed edges in the temporal causal relationship graph where no anomalies have occurred, also includes segmented calculation of the attenuation mechanism of the propagation path: For each directed edge that did not exhibit any anomalies during the tracing process, the type of propagation medium to which the directed edge belongs is determined based on the difference in lithology between the upstream and downstream points connected by the directed edge. If the propagation medium is a fractured medium, the equivalent hydraulic conductivity coefficient corresponding to the directed edge is calculated using the cubic law of fractured networks, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. If the propagation medium is a porous medium, the measured permeability coefficient of the region corresponding to the directed edge is used as the equivalent hydraulic conductivity coefficient, and the normalized value of the equivalent hydraulic conductivity coefficient is used as the propagation correction factor for that segment. The deviation ratio described in step A4 is corrected according to the propagation correction factor of each segment, and the spatial migration distance of the pollution front is calculated step by step based on the corrected deviation ratio.