Underground water multi-pollutant rapid prediction and diagnosis system

Abstract

The invention discloses an underground water multi-pollutant rapid prediction and diagnosis system, which comprises a multi-parameter sensing and data acquisition module used for acquiring underground water quality parameters in real time; the edge pretreatment unit is used for carrying out edge pretreatment on the underground water quality parameters; the off-line synthetic data engine is used for generating synthetic training data of transient extreme working conditions; the feature engineering processing unit is used for constructing interaction features based on a geochemical mechanism; the AutoML reasoning unit is used for deducing and outputting a pollutant concentration predicted value through a model; and the physical-data dual-drive fault diagnosis module is used for diagnosing the authenticity of pollution according to three-dimensional calculation results of the physical residual calculated by the physical residual calculation unit, the mean square error of the sensor data calculated by the data reconstruction error unit and the difference between the observed value and the predicted value calculated by the statistical test unit. And judging whether to give an alarm according to the diagnosis result. According to the scheme, the specific pollutants can be accurately identified, and the authenticity of pollution can be accurately identified.

Description

Technical Field

[0001] This invention belongs to the technical field of groundwater pollution detection, specifically relating to a rapid prediction and diagnosis system for multiple pollutants in groundwater. Background Technology

[0002] Currently, the most advanced groundwater online monitoring and early warning systems in the industry typically employ an architecture of "multi-parameter sensors + data transmission unit + cloud data analysis." The specific approach involves deploying integrated multi-parameter water quality sensors within the monitoring well to collect real-time data on conventional physicochemical indicators such as conductivity (EC), pH, dissolved oxygen (DO), and turbidity. This data is transmitted to a cloud platform via 4G / 5G or NB-IoT networks. The cloud platform utilizes statistical methods (such as setting upper thresholds) or basic machine learning algorithms (such as multiple linear regression and support vector machines, SVM) to establish statistical correlations between conventional indicators and target pollutants (such as heavy metals and specific organic compounds) (i.e., "soft measurement" technology), thereby predicting changes in pollutant concentrations. When the monitored or predicted values ​​exceed national standard limits, the system triggers an alarm.

[0003] Although existing technologies have achieved automation of monitoring to some extent, the following unavoidable shortcomings still exist in practical engineering applications: First, existing systems struggle to directly monitor characteristic pollutants in real time, and alternative monitoring methods suffer from low accuracy. For high-risk heavy metals (such as arsenic and cadmium) or specific organic compounds in groundwater, there is a lack of mature, low-cost sensors capable of long-term online operation (existing sensors require frequent maintenance, are susceptible to poisoning, and are extremely expensive). Current technologies typically use EC or pH as proxies for estimation, but this estimation often ignores geochemical reaction mechanisms (such as shifts in the heavy metal release / adsorption equilibrium caused by changes in redox potential). Simple correlations alone cannot accurately reflect complex chemical transformations, resulting in extremely low accuracy in estimating pollutant concentrations.

[0004] Secondly, existing systems rely on a single dimension for pollution diagnosis, making it difficult to distinguish between sensor malfunction and actual pollution. When monitoring data shows anomalies (such as sudden numerical changes), existing systems struggle to determine whether the sensor itself has drifted / damaged or the water body has actually been polluted. Current technologies largely depend on single statistical data characteristics (such as variance tests) and lack physical consistency checks. This leads to false pollution reports when sensors malfunction, or misjudgments and rejection of data when actual illegal discharges (from unknown sources) occur. Summary of the Invention

[0005] This invention provides a rapid prediction and diagnosis system for multiple pollutants in groundwater, solving the aforementioned technical problems. Specifically, the technical solution is as follows: A rapid prediction and diagnosis system for multiple pollutants in groundwater includes: The multi-parameter sensing and data acquisition module is used to collect groundwater quality parameters in real time through multi-dimensional water quality sensors. The edge preprocessing unit is used to perform edge preprocessing of groundwater quality parameters. The offline synthetic data engine stores historical hydrogeological parameters and generates synthetic training data by driving numerical models, and uses the synthetic training data to pre-train machine learning models. The feature engineering processing unit is used to receive real-time data preprocessed by the edge preprocessing unit, construct interactive features based on geochemical mechanisms, and calculate semi-mechanistic interactive features; The AutoML inference unit is used to load a machine learning model pre-trained with synthetic training data, input semi-mechanistic interaction features into the machine learning model, and finally output the predicted value of pollutant concentration. The physical-data dual-drive fault diagnosis module is activated when the predicted pollutant concentration exceeds the threshold. It diagnoses the authenticity of pollution based on three dimensions: the physical residual calculated by the physical residual calculation unit, the mean square error of the sensor data calculated by the data reconstruction error unit, and the difference between the observed and predicted values ​​calculated by the statistical test unit. Based on the diagnosis results, it determines whether to trigger an alarm. The real-time feedback and closed-loop management module is used to generate early warning signals and push them to the user terminal based on the judgment results of the physical-data dual-drive fault diagnosis module.

[0006] Furthermore, the multi-parameter sensing and data acquisition module also integrates: The disturbance covariate monitor is used to collect the pumping flow rate, the water pump cleaning signal and the vertical flow velocity in the well in real time as non-chemical disturbance signals. Hydraulic monitoring instruments are used to collect groundwater levels in real time. The edge preprocessing unit performs perturbation and noise reduction processing on the collected data of pumping flow rate, pump cleaning signal, vertical flow velocity, and groundwater level, and uses perturbation covariates to adjust the suspended solids concentration (SS). raw Make corrections: SS corr =SS raw -α·v well -β·I clean , Among them, SS corr v represents the corrected suspended solids concentration. well The vertical flow velocity within the well; I clean α represents the water pump cleaning signal, which is 1 during cleaning and 0 otherwise; α and β are empirical regression coefficients.

[0007] Furthermore, the feature engineering processing unit constructs a redox release potential energy feature F based on the Nernst equation. redoxCharacterizing the release potential of redox-sensitive heavy metals: , Where EC is the conductivity, which characterizes the background value of ionic strength, ORP is the redox potential, DO is the dissolved oxygen, and pH is the acidity or alkalinity. Thermodynamic constants related to temperature: , Where R is the ideal gas constant, T is the real-time collected water temperature, F is the Faraday constant, and n is the number of electrons transferred in the redox reaction of the main target pollutant; The feature engineering processing unit is based on thermodynamics and constructs the thermodynamic reaction potential energy feature F. thermo : F thermo =T×ln(SS corr +e), Where e is the natural constant; The AutoML inference unit inputs the redox release potential energy characteristic Fredox and the thermodynamic reaction potential energy characteristic F. thermo Output predicted pollutant concentration values.

[0008] Furthermore, the physical residual calculation unit calculates the physical residual loss based on the 1D convection-diffusion-reaction equation. phy : , in, The output is the predicted concentration of pollutants, where t is the time variable, x is the spatial distance along the mainstream, and D is the longitudinal dispersion coefficient. For the reaction term, v(t) is the real-time flow rate; v(t) = K * i / n e , Where i is the real-time water level H collected. level The calculated hydraulic gradient, where K is the preset aquifer permeability coefficient, and n e Effective porosity; , Where, k deg (T) is the temperature-corrected first-order kinetic decay coefficient; The data reconstruction error unit calculates the mean square error (Loss) between the input vector and the reconstructed output vector of the sensor data using an autoencoder. data : Among them, X input X is the real-time normalized sensor data input vector. recon Reconstruct the output vector for the autoencoder; The statistical test unit calculates the difference S between the observed and predicted values ​​by computing the Kalman filter innovation sequence. inn : , Where, γ k Z represents the value of the innovation sequence at time k. k Let be the observed state vector at the current moment. H is the state vector for cross-dimensional prediction, and H is the normalized vector. Calculate the new information covariance matrix S at the current time. k : , Among them, P k Let R be the state prediction error covariance matrix, and let R be the measurement noise covariance matrix. Construct the normalized innovation square as the statistical test statistic S inn The calculated S inn The data is compared with a preset threshold range to determine whether it conforms to a chi-square distribution. , in, S represents the new information covariance matrix k The inverse matrix, T represents the innovation sequence γ k Transpose of; when and And S inn Normal indicates true contamination; when and This indicates a sensor malfunction; when and This indicates illegal discharge of pollutants. Where ε1 is the physical residual threshold, Set the data reconstruction error threshold.

[0009] Furthermore, the real-time feedback and closed-loop management module also includes: An online dynamic calibration unit is used to update the system status using observational data from diagnostic results; The rapid prediction and diagnosis system for multiple pollutants in groundwater is constructed with a full-state vector that includes both easily measurable parameters and difficult-to-measurable pollutant parameters. The predicted pollutant concentration values ​​include: values ​​of easily measurable parameters and values ​​of difficult-to-measurable pollutant parameters; When the easily measurable parameters acquired by the multi-parameter sensing and data acquisition module change, the online dynamic calibration unit calculates the correction value X of the difficult-to-measurable pollutant parameter sequence by running the ensemble Kalman filter (EnKF) algorithm. updated : , Among them, X forecast Z is the state vector for the predicted pollutant concentration. obs H is the state vector of the measured easily measurable parameters, H is the normalized vector, and K is the weighting coefficient of the Kalman gain.

[0010] Furthermore, the real-time feedback and closed-loop management module also includes: The inference computing unit is used to input the corrected sequence of difficult-to-measure pollutant parameters X within a preset time period into the pre-trained LSTM network. updated After time-series simulation using an LSTM network, the trend of pollutant concentration changes within a preset time period is output.

[0011] Furthermore, the real-time feedback and closed-loop management module also includes: The hierarchical dynamic early warning unit is used to generate multi-level early warning signals and push them to user terminals; The tiered dynamic early warning unit analyzes the pollutant concentration change trend line output by the inference calculation unit: A blue warning signal is sent when the pollutant concentration trend value approaches the standard threshold; a yellow warning signal is sent when the pollutant concentration trend value exceeds the standard threshold; and a red warning signal is sent when the pollutant concentration trend value exceeds the standard threshold and the difference between the pollutant concentration trend value and the standard threshold is greater than a preset value.

[0012] Furthermore, the offline synthetic data engine generates synthetic training data for normal operating conditions by driving the numerical model, and then overlays synthetic training data for transient extreme operating conditions onto the synthetic training data for normal operating conditions.

[0013] Furthermore, for conventional operating conditions, the upper boundary infiltration rate is set as the average rainfall infiltration rate Ravg for a predetermined number of years based on hydrological and meteorological data from the site. Extreme operating conditions include extreme rainstorm conditions, sudden leakage conditions, and pumping well failure conditions; The upper boundary infiltration rate for extreme rainstorm conditions is set at 10~50×Ravg, and the duration is set at 24-72 hours. In the event of a sudden leak, a transient high-concentration flux boundary C is superimposed on the grid nodes corresponding to the potential pollution source area. source C source >>C background C background The standard concentration of groundwater pollutants in the area where the grid node is located is set to a duration of 1-4 hours.

[0014] In the case of a pumping well failure, the flow rate sink Q of the specified pumping well will be forcibly set to zero; The numerical model generates corresponding synthetic training data based on data from normal and extreme operating conditions. The machine learning model is pre-trained using synthetic training data to learn the pollutant concentration response patterns under physical constraints.

[0015] Furthermore, the multidimensional water quality sensor integrates: conductivity sensor, pH sensor, oxidation-reduction potential sensor, dissolved oxygen sensor, temperature sensor, and suspended solids sensor. The dissolved oxygen sensor adopts a dual-sensor redundancy configuration.

[0016] The advantage of this invention lies in the fact that the provided rapid prediction and diagnosis system for multiple pollutants in groundwater is based on geochemical mechanisms to construct interactive features. It transforms conventionally measurable parameters into semi-mechanistic features and inputs them into machine learning models for deduction. Compared with the simple linear statistical inference in the prior art, this can more accurately capture the complex nonlinear coupling relationship between conventional indicators and heavy metals (such as arsenic and cadmium) in the groundwater chemical field. Thus, without the need to deploy expensive and vulnerable online heavy metal analyzers, it achieves rapid and high-precision measurement of high-risk pollutants, ensuring detection accuracy while effectively reducing construction and maintenance costs.

[0017] The advantage of this invention lies in the fact that the groundwater multi-pollutant rapid prediction and diagnosis system innovatively combines three dimensions of evidence—physical equation residuals, data autocorrelation reconstruction errors, and statistical information verification—to make logical fusion judgments. This solves the problem that existing technologies cannot distinguish the causes of anomalies by relying solely on data thresholds, thereby effectively distinguishing the authenticity of pollution and avoiding false or missed reports of pollution. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the rapid prediction and diagnosis system for multiple pollutants in groundwater according to this application; Figure 2 This is a schematic diagram of the workflow of the rapid prediction and diagnosis system for multiple pollutants in groundwater according to this application.

[0020] The system includes a multi-parameter sensing and data acquisition module 100, a multi-dimensional water quality sensor 101, a disturbance covariate monitor 102, a hydraulic monitor 103, a self-cleaning maintenance unit 104, an edge preprocessing unit 105, a mechanism-enhanced rapid prediction module 200, an offline synthetic data engine 201, a feature engineering processing unit 202, an AutoML inference unit 203, a fault diagnosis module 300, a physical residual calculation unit 301, a data reconstruction error unit 302, a statistical testing unit 303, a joint diagnostic logic judgment unit 304, a real-time feedback and closed-loop control module 400, an online dynamic calibration unit 401, an inference calculation unit 402, and a hierarchical dynamic early warning unit 403. Detailed Implementation

[0021] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0022] In the description of this application, it should be noted that, unless otherwise specified and limited, the terms "installation", "connection" and "linkage" should be interpreted broadly, and can refer to mechanical or electrical connections, or internal connections between two components, or direct connections. "Up", "down", "left", "right", etc., are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may change.

[0023] The accompanying drawings of the embodiments disclosed in this invention only involve structures relevant to the embodiments disclosed in this invention. Other structures can be referred to with common designs. Unless otherwise specified, the same embodiment and different embodiments of this invention can be combined with each other.

[0024] like Figures 1 to 2 As shown, a rapid prediction and diagnosis system for multiple pollutants in groundwater according to this application includes: The multi-parameter sensing and data acquisition module 100 is used to collect groundwater quality parameters in real time through the multi-dimensional water quality sensor 101. This module, as the sensing front end of the system, is deployed inside the groundwater monitoring well and collects groundwater quality parameters in real time through the integrated multi-dimensional water quality sensor 101. The multi-dimensional water quality sensor 101 integrates: a conductivity sensor, a pH sensor, an oxidation-reduction potential sensor, a dissolved oxygen sensor, a temperature sensor, and a suspended solids sensor. The dissolved oxygen sensor adopts a dual-sensor redundancy configuration. Specific water quality parameters include conventional physicochemical indicators such as conductivity (EC), pH, oxidation-reduction potential (ORP), dissolved oxygen (DO), temperature (T), and suspended solids (SS). The dual-sensor redundancy configuration of the dissolved oxygen sensor allows for cross-checking, ensuring data reliability. This module enables in-situ real-time monitoring of the groundwater environment, solving the problems of long sampling cycles and strong lag in traditional manual sampling.

[0025] Edge preprocessing unit 105 is used to perform edge preprocessing of groundwater quality parameters. Deployed at the edge of the monitoring well, this unit incorporates signal conditioning circuitry and an edge computing chip, responsible for preprocessing the collected groundwater quality parameters. Preprocessing includes signal filtering, outlier removal, and data packaging, converting the raw sensor data into a standardized format before transmission to subsequent modules. This initial data cleaning at the edge effectively reduces data transmission bandwidth requirements, improves the system's real-time response capability, and enhances the efficiency of subsequent modules in processing the data.

[0026] The mechanism-enhanced fast prediction module 200 integrates: an offline synthetic data engine 201, a feature engineering processing unit 202, and an AutoML inference unit 203. Among them: The offline synthetic data engine 201 stores historical hydrogeological parameters and generates synthetic training data for conventional operating conditions by driving numerical models. It then overlays this conventional operating condition synthetic training data to generate synthetic training data for transient extreme operating conditions. The engine stores historical hydrogeological parameters of the site and generates synthetic training data by driving numerical models (such as PHT3D and MT3D-USGS). The numerical models simulate the transport process of groundwater pollutants based on preset boundary conditions for conventional and extreme operating conditions, outputting corresponding pollutant concentration response curves as synthetic training data. This synthetic training data is used to pre-train machine learning models, enabling them to learn the pollutant concentration response patterns under physical constraints, thus solving the problem of poor model generalization ability caused by the scarcity of measured data under extreme operating conditions.

[0027] The feature engineering processing unit 202 receives real-time data preprocessed by the edge preprocessing unit 105, constructs interactive features based on geochemical mechanisms, and calculates semi-mechanistic interactive features. Specifically, this unit introduces the Nernst equation and thermodynamic principles to transform conventionally measurable parameters (EC, pH, ORP, DO, T) into characteristic variables characterizing the release potential of high-risk pollutants such as heavy metals. This unit overcomes the limitations of traditional soft sensing techniques that rely solely on statistical correlation, enhancing the causal correlation strength between features and target pollutants through mechanistic embedding.

[0028] The AutoML inference unit 203 loads a machine learning model pre-trained with synthetic training data, inputs semi-mechanistic interaction features into the model, and ultimately outputs predicted pollutant concentrations. This unit loads the machine learning model pre-trained with synthetic training data, receives the semi-mechanistic interaction features output by the feature engineering processing unit 202 as input, and outputs predicted pollutant concentrations through model deduction. This unit enables rapid deduction from easily measurable parameters to difficult-to-measurable pollutant parameters, achieving minute-level monitoring of high-risk pollutants without deploying expensive and fragile dedicated sensors.

[0029] The physical-data dual-drive fault diagnosis module 300 is activated when the predicted pollutant concentration exceeds a threshold. It diagnoses the pollution's authenticity based on three dimensions: the physical residual calculated by the physical residual calculation unit 301, the mean square error of the sensor data calculated by the data reconstruction error unit 302, and the difference between the observed and predicted values ​​calculated by the statistical test unit 303. Based on the diagnostic results, it determines whether to trigger an alarm. This module automatically activates when the predicted pollutant concentration exceeds a preset threshold, integrating three sub-modules: the physical residual calculation unit 301, the data reconstruction error unit 302, and the statistical test unit 303. The physical residual calculation unit 301 calculates the physical residual based on the 1D convection-diffusion-reaction equation; the data reconstruction error unit 302 calculates the mean square error of the sensor data using an autoencoder; and the statistical test unit 303 calculates the difference between the observed and predicted values ​​using a Kalman filter innovation sequence. The module includes a joint diagnostic logic judgment unit 304, which performs a logical fusion judgment based on the three-dimensional calculation results from the above three units to diagnose the pollution's authenticity and determine whether to trigger an alarm. This module solves the problem that existing technologies cannot distinguish between sensor malfunctions, actual pollution, and illegal discharge.

[0030] The real-time feedback and closed-loop control module 400 generates early warning signals and pushes them to the user terminal based on the judgment results of the physical-data dual-drive fault diagnosis module 300. This module generates early warning signals of corresponding levels and pushes them to the user terminal based on the diagnostic results of the physical-data dual-drive fault diagnosis module 300. Simultaneously, the module utilizes diagnostically confirmed valid observation data to update the system state through an ensemble Kalman filter algorithm, achieving online dynamic calibration of the predictive model and forming a closed-loop control mechanism from monitoring to prediction, to diagnosis, and then to feedback.

[0031] This invention achieves rapid and accurate prediction and intelligent diagnosis of multiple pollutants in groundwater by constructing a closed-loop technical system comprising a multi-parameter sensing and data acquisition module 100, a mechanism-enhanced rapid prediction module 200, a physical-data dual-driven fault diagnosis module 300, and a real-time feedback and closed-loop control module 400. Compared with existing technologies, this invention has the following advantages: This invention introduces extreme working condition data generated by physical numerical simulation through an offline synthetic data engine 201, and combines it with semi-mechanistic interaction features constructed based on geochemical mechanisms by a feature engineering processing unit 202. Conventional easily measurable parameters are transformed into semi-mechanistic features and input into a machine learning model for inference. Compared with the simple linear statistical inference in the prior art, this can more accurately capture the complex nonlinear coupling relationship between conventional indicators and heavy metals (such as arsenic and cadmium) in the groundwater chemical field. Thus, without the need to deploy expensive and vulnerable online heavy metal analyzers, it achieves rapid and high-precision measurement of high-risk pollutants through mechanism embedding, ensuring prediction accuracy while effectively reducing construction and maintenance costs.

[0032] This invention innovatively integrates three dimensions of evidence—physical equation residuals, data reconstruction errors, and statistical innovation tests—to construct a physical-data dual-driven fault diagnosis logic. This multi-dimensional evidence fusion enables accurate diagnosis, effectively distinguishing between real pollution, sensor malfunctions, and covert illegal discharges, thus solving the problem of false alarms and missed alarms caused by single data threshold diagnosis.

[0033] As a further solution, the multi-parameter sensing and data acquisition module 100 also integrates a disturbance covariate monitor 102 and a hydraulic monitor 103. The disturbance covariate monitor 102 is used to collect in real time the pumping flow rate (Q) in the wellbore, the pump cleaning signal in the wellbore, and the vertical flow velocity (v) in the wellbore. well This serves as a non-chemical disturbance signal. The hydraulic monitor 103 is used to collect groundwater levels in real time.

[0034] Based on the detection data from the disturbance covariate monitor 102 and the hydraulic monitor 103, the edge preprocessing unit 105 performs disturbance denoising processing on the collected data of pumping flow rate, pump cleaning signal, vertical flow velocity, and groundwater level, and uses the disturbance covariate to analyze the suspended solids concentration (SS). raw Make corrections: SS corr =SS raw -α·v well -β·I clean , Among them, SS corr v represents the corrected suspended solids concentration. well The vertical flow velocity within the well; I clean The signal is 1 for pump cleaning and 0 otherwise. α and β are empirical regression coefficients, where α represents the hydraulic disturbance coefficient and β represents the cleaning disturbance coefficient. Specifically, using background monitoring data from non-pollution periods, the SS (Self-Regulation Scheme) is established using the least squares method. raw With v well A linear regression model was used to determine α; SS was determined by statistical analysis during multiple cleaning events. raw The average spike peak of the readings is used to determine β.

[0035] This solution introduces non-water quality disturbance covariates and establishes a quantitative correction model. By combining disturbance covariates with a real-time noise reduction algorithm, it deducts the contribution of physical turbidity caused by hydraulic disturbances and cleaning operations. This effectively solves the problem of falsely high suspended solids readings caused by downhole physical operations (pumping start-up and shutdown, sensor cleaning) and misjudging the physical disturbance as water pollution. It reduces the false alarm rate from the source, effectively solving the problem of false alarms caused by downhole physical operations and reducing the ineffective on-site verification work of maintenance personnel.

[0036] As a specific implementation scheme, the feature engineering processing unit 202 constructs the redox release potential energy feature F based on the Nernst equation. redox Characterizing the release potential of redox-sensitive heavy metals: , Where EC is the conductivity, which characterizes the background value of ionic strength, ORP is the redox potential, DO is the dissolved oxygen, and pH is the acidity or alkalinity. Thermodynamic constants related to temperature: , Where R is the ideal gas constant (8.314 J / (mol·K)), T is the real-time collected water temperature (°C), F is the Faraday constant (96485 C / mol), and n is the number of electrons transferred in the redox reaction of the main target pollutant (e.g., for the arsenic As(V) / As(III) system, n is assumed to be 2). From F redoxThe calculation formula shows that low pH (acidity) promotes desorption, and low ORP (reduction) promotes the reduction and dissolution of iron and manganese oxides, thereby releasing heavy metals.

[0037] The feature engineering processing unit 202 constructs the thermodynamic reaction potential energy feature F based on thermodynamics. thermo It uses the corrected suspended solids concentration SS corr Participating in the calculation: F thermo =T×ln(SS corr +e), where e is the natural constant, which is 2.718 here.

[0038] AutoML inference unit 203 inputs redox release potential energy feature F redox and thermodynamic reaction potential energy characteristics F thermo Output predicted pollutant concentration values.

[0039] This scheme constructs intermediate features by explicitly introducing the Nernst equation and thermodynamic principles, transforming the nonlinear coupling relationship of conventional parameters such as temperature, conductivity, pH, and redox potential into feature variables with clear geochemical significance. Compared with directly inputting raw data into a black-box model, it can more accurately capture the release / adsorption of redox-sensitive heavy metals. Without the need to deploy an online heavy metal analyzer, it achieves minute-level high-precision soft measurement of high-risk pollutants, significantly reducing system construction and maintenance costs.

[0040] As a specific implementation, the physical residual calculation unit 301 described above calculates the physical residual loss based on the 1D convection-diffusion-reaction equation. phy : , in, The output is the predicted concentration of pollutants, where t is the time variable, x is the spatial distance along the mainstream, and D is the longitudinal dispersion coefficient. During the system initialization phase, D is obtained through PINN inversion using field tracer experiments or historical long-term monitoring data, and is stored as a fixed parameter. is the reaction term, characterizing the attenuation or adsorption of pollutants. v(t) is the real-time flow rate.

[0041] v(t) = K * i / n e , Where i is the real-time water level H collected. level The calculated hydraulic gradient, where K is the preset aquifer permeability coefficient, and n e The effective porosity is K, which was obtained from previous pumping tests.

[0042] , Where, k deg (T) is the temperature-corrected first-order kinetic decay coefficient, the reference value of which was obtained through laboratory batch experiments.

[0043] As a specific implementation, the data reconstruction error unit 302 described above calculates the mean square error Loss of the input vector and the reconstructed output vector of the sensing data using an autoencoder. data : Among them, X input This is a real-time normalized sensor data input vector, which originates from real-time multidimensional data collected by the module and is composed after standardization processing. , ' indicates the normalized value. X recon The output vector is reconstructed for the autoencoder; it is obtained by: converting X... input The input is fed into an autoencoder network trained with uncontaminated historical data. The encoder compresses the data into a low-dimensional latent space, and the decoder then reconstructs the output vector. .

[0044] As a specific implementation method, the statistical testing unit 303 described above obtains the difference between the observed and predicted values ​​by calculating the Kalman filter innovation sequence: First, calculate the innovation sequence γ at discrete time k. k (i.e., single-step observation difference): , Where, γ k Z represents the innovation sequence value at time k, the residual between the data actually measured by the sensor and the data originally predicted by the model to be measured at that time. k The current observation state vector is the cleaning data (such as the corrected suspended matter SS) after the edge preprocessing unit 105 removes physical disturbances such as water pump cleaning and flow velocity. corr ). This is a cross-dimensional predicted state vector, representing the current theoretical prediction of groundwater derived from the model. It includes both easily measurable parameters (EC, pH, etc.) and difficult-to-measurable pollutant parameters (heavy metals As, Cd, etc.) in its entirety. H is a normalized vector, which... Normalize the full state vector in the vector so that its dimension Z is normalized. k Consistent.

[0045] Subsequently, the new information covariance matrix S at the current time is calculated. k : , Among them, P kR is the state prediction error covariance matrix, representing the physical inference uncertainty of the prediction model. It is calculated in real time from the dynamic variance of the set of inference samples during system operation. R is the measurement noise covariance matrix, representing the inherent physical measurement error of the sensor hardware. It is determined by the factory nominal accuracy or field calibration error of the multidimensional water quality sensor 101.

[0046] Based on the above results, the normalized innovation square (NIS) is constructed as the statistical test statistic S. inn : ,in, S represents the new information covariance matrix k The inverse matrix, T represents the innovation sequence γ k The transpose of the sequence is calculated by transposing the sequence horizontally and vertically. Finally, the calculated S inn The data is compared with a preset threshold range to determine whether it conforms to a chi-square distribution x with m degrees of freedom (dimension of the observed variable). 2 .

[0047] The judgment logic of the joint diagnostic logic judgment unit 304 is as follows: when (Conforms to the laws of physics) and (Data structure is normal) and S inn Normal indicates true contamination; when and This indicates a sensor malfunction; when and This indicates illegal discharge of pollutants.

[0048] The above This represents the physical residual threshold, indicating the maximum permissible deviation between groundwater dynamics and chemical reaction patterns. If... This indicates that the drastic changes in the current water quality concentration cannot be explained by normal water flow convection, diffusion, or chemical decay in the physical space. This means that there is an unknown external input (such as covert illegal discharge) or the data provided by the system is distorted.

[0049] The above The data reconstruction error threshold represents whether the chemical coupling relationship between multidimensional data within the sensor is within a reasonable range. In real groundwater bodies, there are objective geochemical correlations between parameters such as conductivity, pH, and ORP. If... This indicates that the data structures returned by these sensors do not corroborate each other, leading to the conclusion that the sensors are faulty / drifting / damaged.

[0050] The physical residual calculation unit 301 of the above scheme calculates the loss based on the 1D convection-diffusion-reaction equation. phyThe flow velocity v(t) is determined by combining the hydraulic gradient calculated from the real-time water level with a preset permeability coefficient and effective porosity. The reaction term uses a temperature-corrected first-order kinetic attenuation coefficient. The data reconstruction error unit 302 calculates the mean square error Loss between the input vector and the reconstructed vector using an autoencoder. data。 Statistical test unit 303 calculates the difference S between the observed and predicted values ​​using the Kalman filter innovation sequence. inn Thus, this solution, by clearly defining the specific algorithms and joint decision-making logic of the three-dimensional calculation units, achieves fine-grained identification of the causes of anomalies: normal performance in all three dimensions indicates genuine pollution; when data violates physical equations but the data structure is normal, it can be determined as a hidden, unknown source of illegal discharge; when data violates both physical equations and cannot be reconstructed, it can be determined as a sensor malfunction. This three-evidence fusion diagnostic capability provides solid evidence for environmental supervision and precise guidance for equipment maintenance, solving the problem that existing technologies cannot distinguish the causes of anomalies by relying solely on data thresholds.

[0051] As a further solution, the real-time feedback and closed-loop control module 400 also includes an online dynamic calibration unit 401. The online dynamic calibration unit 401 updates the system status using the observation data from the aforementioned diagnostic results. Specifically: The rapid prediction and diagnosis system for multiple pollutants in groundwater is constructed with a full-state vector that includes both easily measurable parameters and difficult-to-measurable pollutant parameters.

[0052] The predicted pollutant concentration values ​​include: values ​​of easily measurable parameters and values ​​of difficult-to-measurable pollutant parameters.

[0053] When the easily measurable parameters acquired by the multi-parameter sensing and data acquisition module 100 change, the online dynamic calibration unit 401 calculates the correction value X of the difficult-to-measurable pollutant parameter sequence by running the ensemble Kalman filter (EnKF) algorithm. updated : , Among them, X forecast Z represents the state vector for the predicted pollutant concentration. The prediction results calculated by the AutoML model in the previous step include both easily measurable parameter values ​​and difficult-to-measurable pollutant parameter values. obs The state vector represents the measured, easily measurable parameters, which are the real values ​​transmitted back from the sensors in real time (containing only the easily measurable parameters). H is the normalized vector used to extract... The easily measurable parameters in This term refers to the difference between the measured and predicted values. K is the weighting coefficient of the Kalman gain, which determines how the deviation of easily measurable parameters is converted into correction values ​​for difficult-to-measurable pollutant parameters. X updated The calibrated pollutant parameter status represents the most accurate groundwater data at the current moment.

[0054] The online dynamic calibration unit 401 addresses the challenge of error correction for target pollutants (such as arsenic and cadmium) due to the lack of real-time online sensors. Employing the ensemble Kalman filter (EnKF) algorithm, the system pre-constructs a full-state vector containing easily measurable parameters (EC, pH, ORP) and difficult-to-measurable pollutant parameters (As, Cd). Leveraging the strong geochemical correlation (covariance) between these two parameters, the algorithm automatically calculates the correction amount for the difficult-to-measurable pollutant parameters when the sensor detects a change in the easily measurable parameters. This ensures that the system's prediction baseline remains synchronized with the actual hydrogeological environment even under long-term unattended operation. The use of ensemble Kalman filtering for state assimilation solves the drift and accumulated error problems that occur during long-term model operation, ensuring the system's monitoring maintains continuous accuracy.

[0055] As a further implementation, the real-time feedback and closed-loop control module 400 also includes an inference computing unit 402. This inference computing unit 402 is used to input the corrected sequence of difficult-to-detect pollutant parameters X within a preset time period into the pre-trained LSTM network. updated After time-series simulation using an LSTM network, the system outputs the pollutant concentration change trend within a preset time period. This scheme, by introducing an LSTM network to perform time-series simulation on the calibrated state sequence, enables the prediction of the pollution plume distribution over the next 24-72 hours, expanding the system's function from current status monitoring to trend early warning. This provides a basis for pollution prevention and control decisions and supports the shift from a reactive to a proactive control model.

[0056] As a specific implementation plan, the real-time feedback and closed-loop control module 400 also includes a hierarchical dynamic early warning unit 403. The hierarchical dynamic early warning unit 403 is used to generate multi-level early warning signals and push them to the user terminal. Specifically, the hierarchical dynamic early warning unit 403 analyzes the pollutant concentration change trend line output by the inference calculation unit 402: when the pollutant concentration change trend value approaches the standard threshold, a blue early warning signal is sent, the system monitors the pollutant concentration change trend, and automatically increases the sampling frequency; when the pollutant concentration change trend value exceeds the standard threshold, a yellow early warning signal is sent, pushing a manual review instruction and a list of suspected pollution sources for verification. When the pollutant concentration change trend value exceeds the standard threshold, and the difference from the standard threshold is greater than a preset value, a red early warning signal is sent, pushing a high-level alarm, and automatically generating a groundwater pollution control recommendation plan (such as recommending which well to pump and the amount of pumping). This solution achieves optimized allocation of early warning resources by establishing a hierarchical early warning system that matches the degree of deviation from the predicted concentration: a blue early warning triggers automatic increased monitoring, a yellow early warning pushes a manual review instruction, and a red early warning initiates an emergency response and automatically generates a pollution control recommendation plan. This tiered mechanism avoids the waste of response resources caused by a one-size-fits-all approach to alarms and improves the precision of groundwater pollution control.

[0057] As a further solution, the offline synthetic data engine 201 generates synthetic training data for normal operating conditions by driving the numerical model, and then overlays this synthetic training data for transient extreme operating conditions to generate synthetic training data. Specifically: Under normal operating conditions, the upper boundary infiltration rate is set as the average rainfall infiltration rate Ravg for the number of years based on hydrological and meteorological data of the site over a preset number of years. Specifically, it can be set based on the site's historical hydrological and meteorological data of 5-10 years. The source and sink terms are pumping wells operating at the rated frequency, and the pollution source is continuously generalized according to the background release rate. The side boundary is set as a constant head boundary or a flow boundary, and the value is taken as the average rainfall of each month.

[0058] Extreme operating conditions include extreme rainstorm conditions, sudden leakage conditions, and pumping well failure conditions; The upper boundary infiltration rate for extreme rainstorm conditions is set to 10~50×Ravg, and the duration is set to 24-72 hours. This can simulate the surge in water level and rapid leaching of pollutants caused by once-in-a-century rainfall.

[0059] In the event of a sudden leak, a transient high-concentration flux boundary C is superimposed on the grid nodes corresponding to the potential pollution source area. source C source >>C background The duration is set to 1-4 hours to simulate a tank rupture. Among them, C... background This represents the normal concentration (or background concentration) of groundwater pollutants in the area where the grid node is located.

[0060] The pumping well failure condition will force the flow rate sink Q of the specified pumping well to zero, which can simulate the plume escape caused by hydraulic interception failure.

[0061] The numerical model generates corresponding synthetic training data based on data from both normal and extreme operating conditions. The numerical model uses RT3D / FEFLOW / MT3D-USGS / MT3DMS / TOUGH2 / PHT3D (choose one) to generate corresponding pollutant concentration response curves (synthetic training data). This synthetic training data is then used to pre-train the machine learning model, enabling it to learn the pollutant concentration response patterns under physical constraints.

[0062] This solution addresses the issues of uneven distribution of training data and scarcity of extreme scenario samples in machine learning models by overlaying extreme scenario data onto conventional operating condition data. This enables the pre-trained model to respond to extreme events such as once-in-a-century rainstorms and sudden leaks, significantly improving the system's predictive generalization ability in zero-sample or rare scenarios and avoiding the failure risk of purely data-driven models when encountering out-of-distribution data. In other words, by clearly defining quantitative boundary condition setting methods for three typical extreme operating conditions, this solution provides reproducible and parameterizable operating condition configuration specifications for the offline synthetic data engine 201. This ensures that the generated synthetic training data covers the most critical sudden risk scenarios in groundwater pollution prevention and control (extreme weather, accident leaks, facility failures), guaranteeing that the pre-trained model possesses the physical response knowledge to cope with the most destructive events in actual engineering, thus enhancing the system's engineering practical value. Simultaneously, this solution, through extreme operating condition data combined with semi-mechanistic interaction features constructed based on geochemical mechanisms by the feature engineering processing unit 202, enables the machine learning model to possess physical consistency, significantly improving its generalization ability and prediction accuracy in zero-sample extreme scenarios.

[0063] The process of the above scheme is as follows: Figure 2 As shown in the figure, the operation steps for the offline mechanism pre-training part are as follows: S1: Set up normal working conditions in the offline synthetic data engine that stores historical hydrogeological parameters, and superimpose boundary conditions for various working conditions such as extreme rainstorm, sudden leakage and pumping well failure on the basis of normal working conditions. S2: The offline synthetic data engine drives the numerical model to perform numerical simulation calculations and generate response samples under physical constraints, i.e., synthetic training data. S3: Use the synthetic training data generated by S2 to train the learning model in the AutoML inference unit. By receiving the semi-mechanistic interaction features output by the feature engineering processing unit, the model pre-training based on physical constraints is realized, and the pre-trained model weights are output.

[0064] The operating steps for the online real-time process flow section shown in the diagram are as follows: S1: Real-time acquisition of groundwater quality parameters through multi-dimensional water quality sensors, and edge preprocessing such as disturbance reduction and noise reduction of water quality parameters; S2: The AutoML inference unit receives groundwater quality parameters after edge preprocessing and constructs characteristic values ​​representing geochemical mechanisms (such as redox release potential F). redox and thermodynamic reaction potential F thermoThe system loads pre-trained model weights into the AutoML inference unit for inference and outputs predicted pollutant concentration values. It then determines whether the predicted pollutant concentration values ​​exceed preset limits (thresholds set according to national groundwater environmental standards) or exhibit characteristic anomalies. If the predicted pollutant concentration values ​​do not exceed the limits or exhibit characteristic anomalies, the system returns to step 1. If the predicted concentration exceeds the limits or exhibits characteristic anomalies, the system proceeds to step 3. S3: Based on physical residual calculation, data reconstruction error calculation and statistical test, realize the three evidence fusion fault diagnosis; if the diagnosis conclusion is sensor fault, then provide sensor fault or maintenance prompt and return to step 1; if the diagnosis conclusion is actual pollution or illegal discharge, then proceed to step 4. S4: Diagnostic information on actual pollution or illegal discharge triggers real-time feedback and closed-loop control modules to perform trend analysis, achieve hierarchical state assimilation and closed-loop control, and trigger hierarchical dynamic early warning based on the trend analysis results, thereby sending hierarchical dynamic early warning signals.

[0065] As a further solution, the multi-parameter sensing and data acquisition module 100 is also equipped with a self-cleaning maintenance unit 104 for cleaning the multi-dimensional water quality sensor 101, thereby ensuring the accuracy of data acquisition. The cleaning and maintenance device can use an electrically telescopic rod to drive the cleaning arm for surface cleaning.

[0066] The inference unit used in the prediction model architecture of the above scheme, namely the AutoML / ensemble learning model, can also be replaced by deep neural networks (DNN), long short-term memory networks (LSTM) or support vector regression (SVR). These machine learning regression models that accept feature inputs and output concentration prediction values ​​can all be replaced.

[0067] The physical constraint equations in the above scheme adopt the 1D convection-diffusion-reaction equation (ADRE). When calculating the physical residuals, it can be extended to the 2D / 3D ADRE equations, or simplified to the convection equation or the convection-diffusion equation (ignoring any one or both of the dispersion and adsorption terms), depending on the specific hydrological conditions of the monitoring well. The core algorithm is to use the fluid dynamics equations as part of the loss function.

[0068] The above scheme reconstructs data through an autoencoder, and can also replace reconstruction algorithms such as principal component analysis (PCA) to reconstruct errors or the isolated forest algorithm. Any unsupervised learning method that judges anomalies by calculating the structural correlation of the data itself is acceptable.

[0069] The ensemble Kalman filter (EnKF) used in the above scheme can also be replaced by extended Kalman filter (EKF) or particle filter. These standard recursive estimation algorithms that use observations to correct predictions can all be used for state assimilation correction.

[0070] The early warning decision-making grading scheme in the above plan can be based on continuous risk probability index output (0-100%) or decision output based on fuzzy logic. Both of these are control recommendations based on model prediction results.

[0071] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the present invention in any way, and all technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present invention.

Claims

1. A rapid prediction and diagnosis system for multiple pollutants in groundwater, characterized in that, include: The multi-parameter sensing and data acquisition module is used to collect groundwater quality parameters in real time through multi-dimensional water quality sensors. An edge preprocessing unit is used to perform edge preprocessing on the groundwater quality parameters. An offline synthetic data engine stores historical hydrogeological parameters and generates synthetic training data by driving a numerical model, and uses the synthetic training data to pre-train a machine learning model. The feature engineering processing unit is used to receive real-time data preprocessed by the edge preprocessing unit, construct interactive features based on geochemical mechanisms, and calculate semi-mechanistic interactive features. The AutoML inference unit is used to load the machine learning model, input the semi-mechanistic interaction features into the machine learning model, and finally output the predicted value of pollutant concentration. The physical-data dual-drive fault diagnosis module is activated when the predicted pollutant concentration exceeds the threshold. It diagnoses the authenticity of pollution based on three dimensions: the physical residual calculated by the physical residual calculation unit, the mean square error of the sensor data calculated by the data reconstruction error unit, and the difference between the observed and predicted values ​​calculated by the statistical test unit. Based on the diagnosis results, it determines whether to trigger an alarm. The real-time feedback and closed-loop management module is used to generate early warning signals and push them to the user terminal based on the judgment results of the physical-data dual-drive fault diagnosis module.

2. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 1, characterized in that, The multi-parameter sensing and data acquisition module also integrates: The disturbance covariate monitor is used to collect the pumping flow rate, the water pump cleaning signal and the vertical flow velocity in the well in real time as non-chemical disturbance signals. Hydraulic monitoring instruments are used to collect groundwater levels in real time. The edge preprocessing unit performs perturbation and noise reduction processing on the collected data of pumping flow rate, pump cleaning signal, vertical flow velocity, and groundwater level, and uses perturbation covariates to adjust the suspended solids concentration (SS). raw Make corrections: SS corr =SS raw -α·v well -β·I clean , Among them, SS corr v represents the corrected suspended solids concentration. well The vertical flow velocity within the well; I clean α represents the water pump cleaning signal, which is 1 during cleaning and 0 otherwise; α and β are empirical regression coefficients.

3. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 2, characterized in that, The feature engineering processing unit constructs the redox release potential energy feature F based on the Nernst equation. redox Characterizing the release potential of redox-sensitive heavy metals: ， Where EC is the conductivity, which characterizes the background value of ionic strength, ORP is the redox potential, DO is the dissolved oxygen, and pH is the acidity or alkalinity. Thermodynamic constants related to temperature: ， Where R is the ideal gas constant, T is the real-time collected water temperature, F is the Faraday constant, and n is the number of electrons transferred in the redox reaction of the main target pollutant; The feature engineering processing unit is based on thermodynamics and constructs the thermodynamic reaction potential energy feature F. thermo : F thermo =T×ln(SS corr +e), Where e is the natural constant; The AutoML inference unit is input with the redox release potential energy feature Fredox and the thermodynamic reaction potential energy feature F. thermo Output predicted pollutant concentration values.

4. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 2, characterized in that, The physical residual calculation unit calculates the physical residual loss based on the 1D convection-diffusion-reaction equation. phy ; The data reconstruction error unit calculates the mean square error (Loss) between the input vector and the reconstructed output vector of the sensing data using an autoencoder. data ; The statistical testing unit calculates the difference S between the observed and predicted values ​​by measuring the Kalman filter innovation sequence. inn ; when and And S inn Normal indicates true contamination; when and This indicates a sensor malfunction; when and This indicates illegal discharge of pollutants. Where ε1 is the physical residual threshold, Set the data reconstruction error threshold.

5. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 1, characterized in that, The real-time feedback and closed-loop control module also includes: An online dynamic calibration unit is used to update the system status using observational data from diagnostic results; The rapid prediction and diagnosis system for multiple pollutants in groundwater is constructed with a full-state vector that includes both easily measurable parameters and difficult-to-measurable pollutant parameters. The predicted pollutant concentration values ​​include: easily measurable parameter values ​​and difficult-to-measurable pollutant parameter values; When the easily measurable parameters acquired by the multi-parameter sensing and data acquisition module change, the online dynamic calibration unit calculates the correction value X of the difficult-to-measurable pollutant parameter sequence by running the ensemble Kalman filter (EnKF) algorithm. updated : ， Among them, X forecast Z is the state vector of the predicted pollutant concentration. obs H is the state vector of the measured easily measurable parameters, H is the normalized vector, and K is the weighting coefficient of the Kalman gain.

6. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 5, characterized in that, The real-time feedback and closed-loop control module also includes: The inference computing unit is used to input the corrected sequence of difficult-to-measure pollutant parameters X within a preset time period into the pre-trained LSTM network. updated After time-series simulation using an LSTM network, the trend of pollutant concentration changes within a preset time period is output.

7. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 6, characterized in that, The real-time feedback and closed-loop control module also includes: The hierarchical dynamic early warning unit is used to generate multi-level early warning signals and push them to user terminals; The hierarchical dynamic early warning unit analyzes the pollutant concentration change trend line output by the inference calculation unit: A blue warning signal is sent when the pollutant concentration trend value approaches the standard threshold; a yellow warning signal is sent when the pollutant concentration trend value exceeds the standard threshold; and a red warning signal is sent when the pollutant concentration trend value exceeds the standard threshold and the difference between the pollutant concentration trend value and the standard threshold is greater than a preset value.

8. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 1, characterized in that, The offline synthetic data engine generates synthetic training data for normal operating conditions by driving the numerical model, and then overlays synthetic training data for transient extreme operating conditions onto the synthetic training data for normal operating conditions.

9. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 8, characterized in that, The conventional operating condition is based on hydrological and meteorological data from a preset number of years at the site, with the upper boundary infiltration rate set as the average rainfall infiltration rate Ravg for those years. The extreme operating conditions include extreme rainstorm conditions, sudden leakage conditions, and pumping well failure conditions; The upper boundary infiltration rate of the extreme rainstorm condition is set to 10~50×Ravg, and the duration is set to 24-72 hours. In the case of a sudden leak, a high-concentration flux boundary C is superimposed on the grid node corresponding to the potential pollution source area. source C source >>C background C background The typical concentration of groundwater pollutants in the area where the grid node is located is set to a duration of 1-4 hours; The pumping well failure condition will force the flow rate sink Q of the specified pumping well to be set to zero; The numerical model generates corresponding synthetic training data based on the data from the normal operating conditions and the extreme operating conditions. The machine learning model is pre-trained using the synthetic training data to learn the pollutant concentration response patterns under physical constraints.

10. The rapid prediction and diagnosis system for multiple pollutants in groundwater according to claim 1, characterized in that, The multidimensional water quality sensor integrates a conductivity sensor, a pH sensor, a redox potential sensor, a dissolved oxygen sensor, a temperature sensor, and a suspended solids sensor. The dissolved oxygen sensor adopts a dual-sensor redundancy configuration.

Images