A method and system for real-time monitoring of groundwater pollution concentration

Abstract

The present application relates to a kind of real-time monitoring methods of groundwater pollution concentration, it includes: a set of vertical layered in-situ sensing probe is laid out in target monitoring area, the probe is divided into multiple independent sampling units along the depth direction, each sampling unit corresponds to one aquifer or aquiclude interface, and is equipped with independent pollutant sensing module, fluid micro-disturbance excitation device, environmental parameter sensing unit and local data processing unit.The present application realizes independent, synchronous, in-situ monitoring to multilayer aquifer structure by vertical layered in-situ sensing probe, effectively captures the vertical migration characteristics of pollution plume, solves the problem that prior art cannot support fine risk assessment due to lack of layered sensing capability.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and sensing technology, specifically to a method and system for real-time monitoring of groundwater pollution concentration. Background Technology

[0002] Real-time, accurate, and continuous monitoring of groundwater pollution concentrations has become a core requirement for environmental risk early warning and pollution source tracing and remediation. However, current mainstream water pollution monitoring technologies are mostly concentrated on surface water bodies or artificial water treatment facilities. For groundwater, a medium with low flow velocity, high concealment, and complex geological coupling characteristics, there is still a lack of in-situ monitoring systems that are highly adaptable, stable, and possess multi-dimensional sensing capabilities.

[0003] Real-time monitoring of groundwater pollution concentration requires highly sensitive identification and dynamic tracking of trace pollutants under extreme conditions such as no clear water flow direction, long-term uninterrupted operation, and the coexistence of multiple aquifer structures. While existing technologies have achieved online detection in specific closed pipelines or industrial wastewater scenarios, they rely on prerequisites such as stable flow fields, clear boundaries, and abrupt changes in high concentrations. This makes them unsuitable for the actual conditions in groundwater environments where pollutants diffuse slowly, have complex compositions (such as heavy metals, organic solvents, and nitrates), generally low concentrations, and are easily affected by geological background interference. Furthermore, conventional sensors face problems such as biofouling, chemical deposition, and signal drift during long-term in-situ deployment, lacking effective self-cleaning, self-calibration, and low-power operation mechanisms, leading to a significant decrease in the reliability of monitoring data over time.

[0004] Existing technologies generally suffer from drawbacks such as limited monitoring dimensions, insufficient spatial resolution, and weak resistance to environmental interference. For example, while some solutions introduce multi-point sensing and data fusion strategies, their algorithm models are based on the assumption of controllable flow patterns. In static or micro-seepage groundwater environments, the lack of reliable flow velocity and concentration gradient support leads to the failure of the fusion logic and an increased false positive rate. Other systems neglect the ability to perform stratified sampling and independent identification of aquifers at different depths, failing to capture the vertical migration characteristics of pollution plumes and making it difficult to support refined risk assessment and remediation decisions. These problems collectively hinder the leap from "fixed-point sampling" to "continuous sensing across the entire region" in groundwater pollution monitoring. There is an urgent need to construct an integrated real-time monitoring method and system that is oriented towards real aquifer environments, integrates high-sensitivity sensing, a stratified sensing architecture, an adaptive calibration mechanism, and intelligent data fusion capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for real-time monitoring of groundwater pollution concentration to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for real-time monitoring of groundwater pollution concentration, comprising: A set of vertically layered in-situ sensing probes is deployed in the target monitoring area. The probes are divided into multiple independent sampling units along the depth direction. Each sampling unit corresponds to an aquifer or impermeable layer interface and is equipped with an independent pollutant sensing module, a fluid micro-disturbance excitation device, an environmental parameter sensing unit, and a local data processing unit. The pollutant sensing module collects pollutant characteristic signals from each depth layer in real time. The pollutant characteristic signals include heavy metal ion concentration, organic solvent characteristic absorption spectrum, nitrate electrochemical response current and total dissolved solids content. The fluid micro-perturbation excitation device generates controllable microscale flow field perturbations locally in each sampling unit to enhance the mass transfer efficiency of pollutant molecules to the sensing surface, and simultaneously records the dynamic response curves of pollutant signals before and after the perturbation. The environmental parameter sensing unit synchronously acquires the temperature, pressure, conductivity, redox potential and pH value of each depth layer, which are used as geological background interference factors for subsequent signal correction. Adaptive baseline calibration is performed on the raw pollutant signals collected by each sampling unit. The adaptive baseline calibration is based on the nonlinear mapping relationship between the steady-state mean of the signal during historical undisturbed periods and the current environmental parameters, and dynamically corrects the signal offset caused by bioattachment, chemical deposition or temperature drift. Based on the dynamic response curves of pollutant signals and synchronous environmental parameters of each sampling unit, a multidimensional spatiotemporal feature tensor is constructed. The multidimensional spatiotemporal feature tensor includes depth dimension, time dimension, pollutant type dimension and environmental interference dimension. The multidimensional spatiotemporal feature tensor is input into a geologically constrained graph neural network model. The nodes of the geologically constrained graph neural network model represent each sampling unit. The edge weights are predetermined by the permeability coefficient, lithological similarity and vertical hydraulic gradient between adjacent aquifers. The model aggregates neighborhood node information through a message passing mechanism, suppresses isolated outliers and enhances the ability to identify the vertical migration path of the pollution plume. The geologically constrained graph neural network model outputs the estimated pollution concentration of each depth layer at the current moment, and combines it with the historical concentration time series to generate a pollution diffusion trend prediction result; The estimated pollution concentration and the predicted pollution diffusion trend are uploaded to the remote monitoring center via a low-power wide-area communication module to complete the real-time monitoring and early warning of groundwater pollution concentration.

[0007] Furthermore, the vertically layered in-situ sensing probe has no fewer than three sampling units. The vertical spacing between adjacent sampling units is set according to the thickness of the target aquifer, with a minimum spacing of no less than 0.5 meters and a maximum spacing of no more than 5 meters. Each sampling unit's pollutant sensing module integrates an electrochemical sensor array, a UV-Vis absorption spectroscopy microprobe, and an ion-selective field-effect transistor. The electrochemical sensor array uses a three-electrode system: a gold-modified glassy carbon electrode as the working electrode, a silver / silver chloride solid electrode as the reference electrode, and a platinum wire mesh electrode as the counter electrode. The UV-Vis absorption spectroscopy microprobe operates in the wavelength range of 200 to 800 nanometers, has an optical path length of 10 millimeters, uses a deep ultraviolet light-emitting diode as the light source, and employs a silicon photodiode array as the detector. The sensitive membrane material of the ion-selective field-effect transistor is a polyvinyl chloride carrier membrane, which is loaded with crown ethers, thiols, and quaternary ammonium salt ion carriers for lead ions, cadmium ions, and nitrate ions, respectively.

[0008] Furthermore, the fluid micro-disturbance excitation device includes a micro piezoelectric pump, an annular micro-orifice nozzle, and a reflux chamber. The micro piezoelectric pump is located at the bottom of the sampling unit and drives a small amount of groundwater to be sprayed in a pulse manner through the annular micro-orifice nozzle, forming a vortex disturbance zone with a diameter of no more than 20 mm in front of the sensing module. The disturbance lasts for 5 to 15 seconds, with an interval of 30 minutes to 2 hours. The disturbance intensity is dynamically adjusted by the rate of change of the current pollutant signal.

[0009] Furthermore, the specific steps of the adaptive baseline calibration include: First, performing sliding window statistics on the pollutant signal of each sampling unit under undisturbed conditions, calculating the signal mean and standard deviation over a window length of 24 hours; Second, establishing a multivariate nonlinear regression model of environmental parameters and signal baseline offset, wherein the multivariate nonlinear regression model adopts a radial basis function neural network structure, with temperature, pressure, conductivity, redox potential, and pH value as inputs, and baseline offset as output; Third, before each perturbation sampling, predicting the baseline offset using the current environmental parameters through the multivariate nonlinear regression model, and subtracting the offset from the original signal to obtain the calibrated pollutant characteristic signal.

[0010] Furthermore, the construction process of the geologically constrained graph neural network model includes: First, based on borehole core data and geophysical logging data, determining the formation type, porosity, and permeability coefficient of each sampling unit; second, calculating the vertical hydraulic conductivity coefficient between any two adjacent sampling units as the initial weight of the corresponding edge in the graph neural network; third, during the model training phase, introducing contamination injection experimental data as a supervision signal to optimize the message passing function and node update rules of the graph neural network, so that the model output can accurately reflect the vertical distribution pattern of the known contamination plume; finally, during the actual operation phase, fixing the model parameters and performing forward inference only based on the real-time input multidimensional spatiotemporal feature tensor.

[0011] Furthermore, the low-power wide-area communication module adopts a narrowband IoT communication protocol, operates at a frequency of 800 MHz, and has a transmission power of no more than 20 milliwatts. The data upload cycle is dynamically adjusted according to the rate of change of pollution concentration. When the rate of change of concentration exceeds a preset threshold, the upload cycle is shortened to five minutes; otherwise, it is extended to four hours. After receiving the data, the remote monitoring center, in conjunction with the geographic information system platform, visualizes the vertical profile of pollution concentration, the time evolution heat map, and the diffusion direction vector field.

[0012] According to another aspect of the present invention, a real-time monitoring system for groundwater pollution concentration is provided, comprising: A vertically layered in-situ sensing probe is divided into multiple independent sampling units along the depth direction. Each sampling unit is equipped with a pollutant sensing module, a fluid micro-disturbance excitation device, an environmental parameter sensing unit, and a local data processing unit. An adaptive baseline calibration unit is used to dynamically correct the original pollutant signals of each sampling unit based on the nonlinear mapping relationship between the steady-state mean of the signal during historical undisturbed periods and the current environmental parameters. The multidimensional spatiotemporal feature tensor construction unit is used to integrate the pollutant signals, dynamic response curves and environmental parameters after calibration of each sampling unit to form a feature tensor containing depth, time, pollutant type and environmental disturbance dimensions. The geologically constrained graph neural network inference unit is used to receive the multidimensional spatiotemporal feature tensor and, based on the preset geological structure constraint relationship, output the pollution concentration estimate and diffusion trend prediction results of each depth layer. A low-power wide-area communication unit is used to upload the pollution concentration estimate and diffusion trend prediction results to a remote monitoring center.

[0013] Furthermore, the local data processing unit integrates a microcontroller, non-volatile memory, and a real-time clock chip. The microcontroller is a 32-bit low-power processor with a main frequency of 48 MHz and a built-in floating-point arithmetic unit for executing signal filtering, disturbance control logic, and baseline calibration algorithms. The non-volatile memory has a capacity of no less than 128 megabytes and is used to cache raw and calibrated data within 72 hours. The real-time clock chip has an accuracy of no more than five seconds per day to ensure time synchronization between multiple sampling units.

[0014] Furthermore, the environmental parameter sensing unit includes a digital temperature sensor, a piezoresistive pressure sensor, a four-electrode conductivity sensor, a platinum redox potential electrode, and a glass composite pH electrode. All sensors are encapsulated in a corrosion-resistant titanium alloy sheath with a protection rating of not less than IP68 and an operating temperature range of -10°C to 70°C.

[0015] Furthermore, the geological constraint-type graph neural network inference unit is deployed in the server cluster of the remote monitoring center, adopts a distributed computing architecture, and supports the simultaneous processing of concurrent data streams from no less than one hundred sets of sensor probes; the model inference latency does not exceed ten seconds, the relative error of the concentration estimate is within five percent, and the vertical migration path identification accuracy is not less than ninety percent.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention achieves independent, synchronous, and in-situ monitoring of multi-layered aquifer structures through a vertically layered in-situ sensing probe, effectively capturing the vertical migration characteristics of pollution plumes and solving the problem that existing technologies cannot support refined risk assessment due to a lack of layered sensing capabilities; it actively enhances pollutant mass transfer efficiency through a fluid micro-disturbance excitation device, significantly improving the detection sensitivity and response speed of trace pollutants, overcoming the signal hysteresis and weak response defects caused by passive diffusion in low-flow-velocity groundwater environments; and through an adaptive baseline calibration mechanism, it establishes a nonlinear mapping relationship between environmental parameters and historical steady-state signals, dynamically... The system compensates for signal shifts caused by biofouling, chemical deposition, and temperature drift, ensuring the reliability and timeliness of monitoring data under long-term uninterrupted conditions. By constructing a geologically constrained graph neural network model, real geological structure information is encoded as graph topological constraints, so that the data fusion logic no longer relies on the assumed stable flow field. Even under static or micro-seepage conditions, it can effectively suppress noise, identify anomalies, and accurately analyze pollution diffusion paths, significantly reducing the false positive rate. Through low-power wide-area communication and dynamic upload strategies, the system's field deployment lifespan is significantly extended while ensuring monitoring continuity, meeting the engineering requirements of "continuous sensing across the entire groundwater monitoring domain." In summary, this invention achieves a technological leap from "fixed-point sampling" to "three-dimensional continuous sensing," providing high-precision, highly robust, and spatially resolved real-time monitoring capabilities for groundwater pollution early warning, source tracing, and remediation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0019] Please see Figure 1 This invention provides a technical solution: a method and system for real-time monitoring of groundwater pollution concentration. Its core lies in constructing an integrated monitoring system with multi-dimensional sensing, adaptive calibration, and geological constraint data fusion capabilities through vertically layered in-situ sensing probes. This system achieves trace-level, continuous, and high spatial resolution in-situ monitoring of pollutants such as heavy metal ions, organic solvents, nitrates, and total dissolved solids in the target aquifer area, effectively overcoming the technical challenges posed by low groundwater flow velocity, high concealment, long-term uninterrupted operation, and interference from complex geological backgrounds.

[0020] The method for real-time monitoring of groundwater pollution concentration includes the following steps: S1. A set of vertically layered in-situ sensing probes is deployed in the target monitoring area. The probes are divided into multiple independent sampling units along the depth direction. Each sampling unit corresponds to an aquifer or impermeable layer interface and is equipped with an independent pollutant sensing module, a fluid micro-disturbance excitation device, an environmental parameter sensing unit, and a local data processing unit. The vertically layered in-situ sensing probes are made of high-strength, corrosion-resistant titanium alloy tubing with an outer diameter of 50 mm. The interior is divided into several sealed chambers, each constituting an independent sampling unit. There are no fewer than three sampling units. The vertical spacing between adjacent sampling units is set according to the thickness of the target aquifer, with a minimum spacing of 0.5 meters and a maximum spacing of 5 meters. Each sampling unit has an independent inlet, outlet, and return channel to ensure that the fluid circulates only within the unit, avoiding cross-layer interference. A microporous filter membrane with a pore size of 0.2 micrometers is installed at the inlet to prevent suspended particles from entering the sensing area and causing blockage or contamination. The probe base integrates an anchoring structure, allowing it to penetrate more than 30 meters underground, suitable for monitoring shallow to medium-deep aquifers.

[0021] S2, the pollutant sensing module collects pollutant characteristic signals from each depth layer in real time. These signals include heavy metal ion concentration, characteristic absorption spectra of organic solvents, electrochemical response current of nitrates, and total dissolved solids content. The pollutant sensing module integrates an electrochemical sensor array, a UV-Vis absorption spectroscopy microprobe, and an ion-selective field-effect transistor. The electrochemical sensor array employs a three-electrode system: a gold-modified glassy carbon electrode as the working electrode, a silver / silver chloride solid electrode as the reference electrode, and a platinum wire mesh electrode as the counter electrode, used to detect the anodic stripping voltammetric signal of heavy metal ions such as lead, cadmium, and arsenic. The UV-Vis absorption spectroscopy microprobe operates in the wavelength range of 200 nm to 800 nm, with an optical path length of 10 mm. The light source is a deep ultraviolet light-emitting diode, and the detector is a silicon photodiode array, used to identify the characteristic absorption peaks of organic solvents such as benzene compounds and chlorinated hydrocarbons. The sensitive membrane material of the ion-selective field-effect transistor is a polyvinyl chloride carrier membrane, loaded with crown ethers, thiols, and quaternary ammonium salt ion carriers for lead, cadmium, and nitrate ions, respectively, and the ion activity is inverted through threshold voltage offset. All sensing elements are encapsulated on an inert ceramic substrate and coated with an anti-biodegradation fluoropolymer coating, with a service life of no less than two years.

[0022] S3. Controllable micro-scale flow field disturbances are generated locally in each sampling unit using the fluid micro-perturbation excitation device to enhance the mass transfer efficiency of pollutant molecules to the sensing surface. The dynamic response curves of the pollutant signals before and after the perturbation are recorded simultaneously. The fluid micro-perturbation excitation device includes a micro piezoelectric pump, an annular micro-orifice nozzle, and a reflux chamber. The micro piezoelectric pump is located at the bottom of the sampling unit and is controlled by the local data processing unit. It drives a small amount of groundwater to be pulsed through the annular micro-orifice nozzle, forming a vortex disturbance zone with a diameter not exceeding 20 mm in front of the sensing module. The perturbation duration is five to fifteen seconds, with an interval of thirty minutes to two hours. The perturbation intensity is dynamically adjusted by the current pollutant signal change rate: when the signal change rate is less than 0.05% per hour, the perturbation intensity is set to the maximum value; when the change rate is greater than 2% per hour, the perturbation intensity is reduced to the minimum value to avoid excessive perturbation masking the true concentration gradient. During the perturbation process, the time series of the pollutant signal is recorded simultaneously, forming a dynamic response curve including the rising edge, peak plateau, and decay tail, which is used for subsequent mass transfer efficiency modeling and concentration inversion.

[0023] S4. The environmental parameter sensing unit synchronously acquires the temperature, pressure, conductivity, redox potential, and pH value of each depth layer as geological background interference factors for subsequent signal correction. The environmental parameter sensing unit includes a digital temperature sensor, a piezoresistive pressure sensor, a four-electrode conductivity sensor, a platinum redox potential electrode, and a glass composite pH electrode. All sensors are encapsulated in a corrosion-resistant titanium alloy sheath with a protection rating of not less than IP68 and an operating temperature range of -10°C to 70°C. The temperature sensor has an accuracy of ±0.1°C, the pressure sensor has a range of 0-10 MPa and an accuracy of 0.05% of full scale, the conductivity sensor has a measurement range of 0-2000 mS / cm (conductivity unit: millisiemens per centimeter) and a resolution of not less than 0.1 μS / cm (conductivity unit: microsiemens per centimeter), the redox potential electrode has a response time of less than 10 s, and the pH electrode has a linear error of no more than ±0.05 in the range of 4-10. The above parameters are sampled synchronously every thirty seconds and strictly time-aligned with the pollutant signal to form a multi-source synchronous data stream.

[0024] S5. Adaptive baseline calibration is performed on the raw pollutant signals collected by each sampling unit. This adaptive baseline calibration is based on the nonlinear mapping relationship between the steady-state mean of the signal during historical undisturbed periods and the current environmental parameters, dynamically correcting signal offsets caused by bioattachment, chemical deposition, or temperature drift. The specific execution process of the adaptive baseline calibration is as follows: First, a sliding window statistical analysis is performed on the pollutant signals of each sampling unit under undisturbed conditions. The window length is 24 hours, and the mean μ and standard deviation σ of the signal within this window are calculated. Second, a multivariate nonlinear regression model is established between environmental parameters and the signal baseline offset ΔB. This model uses a radial basis function neural network structure with five input layer nodes corresponding to temperature, pressure, conductivity, redox potential, and pH value, 20 hidden layer nodes, and one output layer node, outputting the baseline offset ΔB. Third, before each perturbation sampling, the baseline offset ΔB is predicted using the current environmental parameters through the radial basis function neural network, and the result is obtained from the raw pollutant signals. Subtracting this offset from the value yields the calibrated pollutant characteristic signal. = - ΔB. The regression model is updated online monthly, with the updated data derived from steady-state signals and corresponding environmental parameters during all undisturbed periods over the past thirty days, ensuring that the model continuously adapts to long-term drift trends.

[0025] S6. Based on the dynamic response curves of pollutant signals and synchronous environmental parameters of each sampling unit, a multidimensional spatiotemporal feature tensor is constructed. This multidimensional spatiotemporal feature tensor includes depth, time, pollutant type, and environmental interference dimensions. The structure is constructed as follows: D represents the depth dimension, which is the total number of sampling units N; T represents the time dimension, taking sampling points from the most recent 72 hours, with a sampling frequency of once per minute, hence T=4320; C represents the pollutant type dimension, including four categories: heavy metal ions, organic solvents, nitrates, and total dissolved solids, hence C=4; E represents the environmental disturbance dimension, including five items: temperature, pressure, conductivity, redox potential, and pH value, hence E=5. For each depth d and each time t, the pollutant characteristic signal is calibrated to form a four-dimensional vector, and the environmental parameters form a five-dimensional vector. The two are concatenated to form a slice of the tensor at the position (d, t). This tensor fully preserves the vertical distribution, temporal evolution, component differences, and environmental coupling relationships of pollutant concentrations, providing structured input for subsequent graph neural network inference.

[0026] S7, the multidimensional spatiotemporal feature tensor is input into a geologically constrained graph neural network model. The nodes of this model represent sampling units, and the edge weights are pre-determined based on the permeability coefficient, lithological similarity, and vertical hydraulic gradient between adjacent aquifers. The model aggregates neighborhood node information through a message passing mechanism, suppresses isolated outliers, and enhances the ability to identify the vertical migration path of the pollution plume. The geologically constrained graph neural network model G = (V, E, W) is constructed based on real geological data. The node set V = { , , ..., }, each node The edge set E corresponds to the i-th sampling unit; it is determined by the hydraulic connectivity between adjacent sampling units. If two units are located in the same aquifer or adjacent aquifers without a continuous aquitard in between, then an edge exists; edge weight. Calculated using the following formula: ;in, , These are the permeability coefficients of the layers containing units i and j, respectively. The vertical cross-sectional area of ​​the water passage is... The distance between the centers of the two units. , The lithological similarity score (derived from the coded borehole core data, with a value ranging from 0 to 1) is used. This is the lithological attenuation coefficient, with a value of 2. This weight comprehensively reflects the hydraulic conductivity and geological continuity.

[0027] The message passing mechanism of the model is defined as: hi(l+1)=σW(l)⋅CONCAT(hi(l),j∈N(i)wij⋅hj(l)); where, Let be the embedding vector of node i in the l-th layer. Let σ be the learnable weight matrix, and σ be the ReLU activation function. Let be the set of neighboring nodes of i. After three layers of message passing, the embedding vectors of each node are input into a fully connected decoder, which outputs the estimated concentration of each pollutant at the current time. During the model training phase, manually injected pollution experimental data are used as supervision labels, and the loss function is the weighted mean squared error, with the weights proportional to the toxicity level of the pollutants.

[0028] S8. The geologically constrained graph neural network model outputs the estimated pollution concentration of each depth layer at the current moment, and combines it with the historical concentration time series to generate a pollution diffusion trend prediction result. The pollution diffusion trend prediction result is realized through a long short-term memory network. The concentration estimation value sequence of the past 72 hours is input into a single-layer LSTM with 180 hidden units, and the output is the concentration prediction sequence for the next 24 hours. The prediction result and the current concentration together constitute the early warning criteria: if the concentration of any depth layer exceeds the national groundwater quality Class III standard limit, or the vertical concentration gradient exceeds 0.5 mg / L per meter, a Level I early warning is triggered; if the concentration shows a continuous upward trend and the predicted value will exceed the standard within 12 hours, a Level II early warning is triggered.

[0029] S9, the estimated pollution concentration and the predicted pollution diffusion trend are uploaded to the remote monitoring center via a low-power wide-area communication module to complete real-time monitoring and early warning of groundwater pollution concentration. The low-power wide-area communication module uses a narrowband IoT communication protocol, operates at 800MHz, has a transmit power not exceeding 20mW, and a standby current of less than 5μA. The data upload cycle is dynamically adjusted according to the pollution concentration change rate: the concentration change rate is calculated within the most recent hour. If r > 0.05%, the upload period is set to five minutes; if r ≤ 0.05%, the upload period is extended to four hours. All data is transmitted after being encrypted with AES-128 to ensure information security. After receiving the data, the remote monitoring center, in conjunction with the geographic information system platform, generates a vertical profile of pollution concentration, a time evolution heat map, and a diffusion direction vector field, supporting multi-probe data fusion and regional pollution situation assessment.

[0030] The real-time groundwater pollution concentration monitoring system includes a vertically layered in-situ sensing probe, an adaptive baseline calibration unit, a multi-dimensional spatiotemporal feature tensor construction unit, a geologically constrained graph neural network inference unit, and a low-power wide-area communication unit.

[0031] As described above, each sampling unit of the vertically layered in-situ sensing probe has independent sensing and local disturbance capabilities.

[0032] The adaptive baseline calibration unit is deployed in the local data processing unit of each sampling unit and is executed by a microcontroller. The microcontroller is a 32-bit low-power processor with a clock frequency of 48MHz and a built-in floating-point unit for real-time execution of sliding window statistics and baseline offset prediction algorithms. Non-volatile memory with a capacity of at least 128 megabytes is used to cache raw and calibrated data from the past 72 hours, ensuring no data loss during network interruptions. The real-time clock chip has a daily error of no more than five seconds, and hardware synchronization signals ensure that the timestamp consistency error of all sampling units is less than ten milliseconds.

[0033] The multidimensional spatiotemporal feature tensor construction unit is completed collaboratively by the local data processing unit and the remote monitoring center. The local unit is responsible for data alignment and initial packaging, while the remote unit is responsible for tensor recombination and dimensional standardization.

[0034] The geologically constrained graph neural network inference unit is deployed in a server cluster at a remote monitoring center. It employs a distributed computing architecture and supports processing concurrent data streams from at least one hundred sensor probes. The model inference latency is no more than ten seconds, the relative error of the concentration estimate is within five percent, and the vertical migration path identification accuracy is no less than 90%. The server cluster is equipped with GPU accelerator cards for efficient execution of graph convolution operations.

[0035] The low-power wide-area communication unit is integrated into the electronic compartment at the top of the probe. The antenna uses a spiral structure embedded in the outer wall of the probe to avoid protruding parts affecting the installation in the well. The communication protocol supports breakpoint resume and data compression, with a compression ratio of no less than 70%, significantly reducing energy consumption.

[0036] In summary, this embodiment utilizes five core technologies—vertical layered sensing, active micro-perturbation mass transfer, environmental parameter-driven adaptive calibration, geological structure-constrained graph neural network fusion, and dynamic low-power communication—to construct a highly robust, high-precision, and long-life real-time monitoring system suitable for real groundwater environments, achieving a technological leap from "fixed-point sampling" to "three-dimensional continuous sensing."

Claims

1. A method for real-time monitoring of groundwater pollution concentration, characterized in that, include: S1: A set of vertically layered in-situ sensing probes is deployed in the target monitoring area. The probes are divided into multiple independent sampling units along the depth direction. Each sampling unit corresponds to an aquifer or impermeable layer interface and is equipped with an independent pollutant sensing module, a fluid micro-disturbance excitation device, an environmental parameter sensing unit, and a local data processing unit. S2: The pollutant characteristic signals of each depth layer are collected in real time through the pollutant sensing module. The pollutant characteristic signals include heavy metal ion concentration, organic solvent characteristic absorption spectrum, nitrate electrochemical response current and total dissolved solids content. S3: The fluid micro-perturbation excitation device generates controllable microscale flow field perturbation in each sampling unit to enhance the mass transfer efficiency of pollutant molecules to the sensing surface, and simultaneously records the dynamic response curve of pollutant signals before and after the perturbation. S4: The environmental parameter sensing unit synchronously acquires the temperature, pressure, conductivity, redox potential and pH value of each depth layer, which are used as geological background interference factors for subsequent signal correction. S5: Perform adaptive baseline calibration on the raw pollutant signals collected by each sampling unit. The adaptive baseline calibration is based on the nonlinear mapping relationship between the steady-state mean of the signal during historical undisturbed periods and the current environmental parameters, and dynamically corrects the signal offset caused by bioattachment, chemical deposition or temperature drift. S6: Based on the dynamic response curve of pollutant signals and synchronous environmental parameters of each sampling unit, a multidimensional spatiotemporal feature tensor is constructed. The multidimensional spatiotemporal feature tensor includes depth dimension, time dimension, pollutant type dimension and environmental interference dimension. S7: Input the multidimensional spatiotemporal feature tensor into a geologically constrained graph neural network model. The nodes of the geologically constrained graph neural network model represent each sampling unit. The edge weights are predetermined by the permeability coefficient, lithological similarity and vertical hydraulic gradient between adjacent aquifers. The model aggregates neighborhood node information through a message passing mechanism, suppresses isolated outliers and enhances the ability to identify the vertical migration path of the pollution plume. S8: The estimated pollution concentration of each depth layer at the current moment is output by the geologically constrained graph neural network model, and the pollution diffusion trend prediction result is generated by combining the historical concentration time series. S9: The estimated pollution concentration and the predicted pollution diffusion trend are uploaded to the remote monitoring center through a low-power wide-area communication module to complete the real-time monitoring and early warning of groundwater pollution concentration.

2. The method for real-time monitoring of groundwater pollution concentration according to claim 1, characterized in that, Adaptive baseline calibration is performed on the raw pollutant signals collected by each sampling unit, including: For each sampling unit, a sliding window statistical analysis of the pollutant signal under undisturbed conditions is performed, and the mean and standard deviation of the signal over a 24-hour window are calculated. A multivariate nonlinear regression model is established for environmental parameters and signal baseline offset. The multivariate nonlinear regression model adopts a radial basis function neural network structure. The inputs are temperature, pressure, conductivity, redox potential and pH value, and the output is the baseline offset. Before each perturbation sampling, the baseline offset is predicted using the current environmental parameters through the multivariate nonlinear regression model, and the offset is subtracted from the original signal to obtain the calibrated pollutant characteristic signal.

3. The method for real-time monitoring of groundwater pollution concentration according to claim 2, characterized in that, Based on the dynamic response curves of pollutant signals from each sampling unit and synchronous environmental parameters, a multidimensional spatiotemporal feature tensor is constructed, including: The depth dimension is set to the total number of sampling units, the time dimension is set to the sampling points in the most recent 72 hours, the pollutant type dimension includes four categories: heavy metal ions, organic solvents, nitrates and total dissolved solids, and the environmental disturbance dimension includes five items: temperature, pressure, conductivity, redox potential and pH value. For each depth and each time point, the calibrated pollutant characteristic signal is spliced ​​with the synchronous environmental parameters to form the multidimensional spatiotemporal feature tensor.

4. The method for real-time monitoring of groundwater pollution concentration according to claim 3, characterized in that, The multidimensional spatiotemporal feature tensor is input into a geologically constrained graph neural network model, including: Based on borehole core data and geophysical logging data, the formation type, porosity, and permeability coefficient of each sampling unit were determined. Calculate the vertical hydraulic conduction coefficient between any two adjacent sampling units and use it as the initial weight of the corresponding edge in the graph neural network; During the model training phase, contamination injection experimental data is introduced as a supervision signal to optimize the message passing function and node update rules of the graph neural network. In the actual operation phase, the model parameters are fixed, and forward inference is performed only based on the multidimensional spatiotemporal feature tensors input in real time.

5. The method for real-time monitoring of groundwater pollution concentration according to claim 4, characterized in that, Calculate the vertical hydraulic conductivity coefficient between any two adjacent sampling units as the initial weight of the corresponding edge in the graph neural network, including: According to the formula wij=ki+kj2⋅AijLij⋅exp(−α⋅|φi−φj|); Where ki and kj are the permeability coefficients of the layers where the i and j units are located, respectively, Aij is the vertical cross-sectional area of ​​water flow, Lij is the center-to-center distance between the two units, φi and φj are the lithological similarity scores (derived from the borehole core data encoding, with values ​​ranging from 0 to 1), and α is the lithological attenuation coefficient.

6. The method for real-time monitoring of groundwater pollution concentration according to claim 5, characterized in that, The geologically constrained graph neural network model outputs estimated pollution concentrations at each depth layer at the current moment, and combines these with historical concentration time series to generate pollution diffusion trend prediction results, including: The concentration estimate sequence of the past 72 hours is input into a single-layer long short-term memory network with 128 hidden units, and the output is the concentration prediction sequence for the next 24 hours. The warning criteria are based on the current concentration and the predicted concentration. If the concentration at any depth exceeds the national Class III standard limit for groundwater quality, or if the vertical concentration gradient exceeds 0.5 mg / L per meter, a Level I warning will be triggered. If the concentration shows a continuous upward trend and the predicted value will exceed the standard within twelve hours, a Level II warning will be triggered.

7. The method for real-time monitoring of groundwater pollution concentration according to claim 6, characterized in that, The fluid micro-perturbation excitation device generates controllable microscale flow field perturbations locally in each sampling unit, including: A small amount of groundwater is driven by a micro piezoelectric pump and sprayed in a pulse manner through an annular micro-orifice nozzle, forming a vortex disturbance zone with a diameter of no more than 20 mm in front of the sensing module. The duration of the disturbance is five to fifteen seconds, with an interval of thirty minutes to two hours. The intensity of the disturbance is dynamically adjusted by the rate of change of the current pollutant signal.

8. The method for real-time monitoring of groundwater pollution concentration according to claim 7, characterized in that, The estimated pollution concentration and the predicted pollution diffusion trend are uploaded to the remote monitoring center via a low-power wide-area communication module, including: It adopts a narrowband IoT communication protocol, operates at a frequency of 800MHz, and has a transmission power of no more than 20mW; The data upload cycle is dynamically adjusted according to the rate of change of pollution concentration. When the rate of change of concentration exceeds the preset threshold, the upload cycle is shortened to five minutes; otherwise, it is extended to four hours.

9. A real-time monitoring system for groundwater pollution concentration, characterized in that, include: A vertically layered in-situ sensing probe is divided into multiple independent sampling units along the depth direction. Each sampling unit is equipped with a pollutant sensing module, a fluid micro-disturbance excitation device, an environmental parameter sensing unit, and a local data processing unit. An adaptive baseline calibration unit is used to dynamically correct the original pollutant signals of each sampling unit based on the nonlinear mapping relationship between the steady-state mean of the signal during historical undisturbed periods and the current environmental parameters. The multidimensional spatiotemporal feature tensor construction unit is used to integrate the pollutant signals, dynamic response curves and environmental parameters after calibration of each sampling unit to form a feature tensor containing depth, time, pollutant type and environmental disturbance dimensions. The geologically constrained graph neural network inference unit is used to receive the multidimensional spatiotemporal feature tensor and, based on the preset geological structure constraint relationship, output the pollution concentration estimate and diffusion trend prediction results of each depth layer. A low-power wide-area communication unit is used to upload the pollution concentration estimate and diffusion trend prediction results to a remote monitoring center.

10. The real-time monitoring system for groundwater pollution concentration according to claim 9, characterized in that, The adaptive baseline calibration unit is used for: For each sampling unit, a sliding window statistical analysis of the pollutant signal under undisturbed conditions is performed, and the mean and standard deviation of the signal over a 24-hour window are calculated. A multivariate nonlinear regression model is established for environmental parameters and signal baseline offset. The multivariate nonlinear regression model adopts a radial basis function neural network structure. The inputs are temperature, pressure, conductivity, redox potential and pH value, and the output is the baseline offset. Before each perturbation sampling, the baseline offset is predicted using the current environmental parameters through the multivariate nonlinear regression model, and the offset is subtracted from the original signal to obtain the calibrated pollutant characteristic signal.

Images