Digital twin soil pollution remediation whole-process management and control system

Abstract

The application relates to a digital twin soil pollution remediation whole-process management and control system. The system comprises a twin modeling module, a belief generation module, an intelligent decision-making module and a closed-loop control module. The twin modeling module constructs a multi-physical field constrained Fourier neural operator proxy model to realize efficient and accurate pollution evolution simulation; the belief generation module generates a belief state containing an attribution confidence based on Bayesian inference of the posterior probability of each competitive remediation hypothesis; the intelligent decision-making module drives reinforcement learning decision-making with the belief state to output remediation actions; and the closed-loop control module iteratively updates model parameters and belief states according to new monitoring data, and uses the updated posterior distribution as the prior of the next round to form a continuously evolving adaptive closed-loop management and control. Through the fusion of multi-physical field constraint intelligent agent modeling, Bayesian mechanism uncertainty reasoning and other means, the system can effectively improve the accuracy of soil pollution remediation mechanism attribution and the reliability of dynamic simulation, and realize the whole-process intelligentization of remediation operation and monitoring regulation.

Description

Technical Field

[0001] This invention belongs to the field of computer technology, and in particular relates to a digital twin soil pollution remediation whole-process control system. Background Technology

[0002] Soil in-situ remediation projects are characterized by strong spatial heterogeneity, complex pollution migration and evolution processes, and difficulty in predicting remediation effects. However, existing digital twin management solutions typically rely on a single numerical model to simulate pollutants, enabling only static or quasi-dynamic predictions of pollutant concentration fields. They struggle to differentiate between different causes of concentration declines, such as microbial degradation, physical dilution, spatial redistribution of pollutants, or monitoring noise, and cannot quantitatively identify the true remediation mechanism. This can easily lead to false compliance issues, where apparent concentrations meet standards but underlying pollution risks persist. Furthermore, traditional digital twin systems mostly employ an open-loop architecture, with model parameters calibrated only once based on initial site survey data. Newly acquired monitoring data during the remediation process cannot continuously provide feedback to correct the proxy model. As remediation progresses, model prediction biases gradually accumulate, causing simulation results to continuously deviate from the actual site evolution.

[0003] Furthermore, most existing remediation decision optimization methods employ conventional reinforcement learning frameworks, using only observed pollutant concentrations as input and failing to incorporate uncertainties arising from remediation mechanism identification. These decision strategies lack the ability to perceive the confidence level of mechanism attribution, and decision schemes typically only consider pollution removal effectiveness and treatment costs, lacking quantitative assessment of monitoring information gain. They cannot dynamically allocate monitoring resources and struggle to achieve an adaptive trade-off between "conducting monitoring to obtain mechanism information" and "implementing remediation to reduce pollution." In addition, conventional numerical solutions to multiphysics coupled convection-dispersion equations are computationally extremely expensive, making it difficult to support large-scale interactive simulations of intelligent decision-making modules. Moreover, existing physical information neural network models often employ fixed network structures, lacking mechanisms for integrated modeling to quantify uncertainties, making it difficult to transmit model prediction biases to probabilistic inference modules, further restricting the reliability of competitive hypothesis testing for remediation mechanisms.

[0004] In summary, existing digital twin management technologies for soil pollution remediation have shortcomings such as insufficient attribution of remediation mechanisms, lack of online adaptive update loops in models, failure to integrate uncertainty and evidentiary information value in the decision-making process, and high cost of high-precision simulation calculations. These shortcomings make it difficult to meet the practical engineering needs for refined and adaptive management of the entire soil pollution remediation process. Summary of the Invention

[0005] Based on this, it is necessary to provide a digital twin soil pollution remediation whole-process management and control system to address the above-mentioned technical problems. The system aims to improve the ability to identify the causes of remediation mechanisms and the ability to perceive the value of evidence, as well as to enhance the ability to quantify the uncertainty of the causes of remediation mechanisms and the ability to update the model online.

[0006] Firstly, this application provides a digital twin soil pollution remediation end-to-end management and control system, including:

[0007] The twin modeling module is used to execute S1, which includes: establishing a multi-physics coupled partial differential equation system based on historical site survey data; constructing a deep neural network based on the multi-physics coupled partial differential equation system; constructing a loss function based on historical site survey data; training the deep neural network based on the loss function to obtain the trained deep neural network; generating a training dataset based on the trained deep neural network; and training a Fourier NeuralOperator surrogate model based on the training dataset to obtain a physical information neural network surrogate model.

[0008] The belief generation module is used to execute S2, which includes: obtaining the prior probability of each competing repair hypothesis; based on the physical information neural network surrogate model and real-time monitoring data, and combining the prior probability of each competing repair hypothesis, calculating the posterior probability of each competing repair hypothesis through Bayes' theorem, and generating the belief state.

[0009] The intelligent decision-making module is used to execute S3, which includes: constructing the state space of the Markov decision process based on the belief state; constructing an environment simulator of the Markov decision process based on the physical information neural network agent model; performing meta-reinforcement learning training in the environment simulator to obtain a decision agent; processing the current state based on the decision agent and outputting repair actions; wherein, the current state is constructed based on real-time monitoring data and the belief state; the repair actions include instructions for adjusting the injection rate, extraction rate, and monitoring frequency of the repair agent.

[0010] The closed-loop control module is used to execute S4. S4 includes: performing soil pollution remediation operations according to the remediation action, collecting new monitoring data after execution, updating the parameters and belief states of the physical information neural network surrogate model according to the new monitoring data, obtaining the updated physical information neural network surrogate model and the updated belief state, determining whether the preset remediation termination condition is met according to the updated belief state, terminating the remediation if it is met, and if it is not met, using the posterior probability distribution in the updated belief state as the prior probability of each competing remediation hypothesis in the next round S2, and returning to S2 for the next round of iteration.

[0011] In one embodiment, the physical information neural network proxy model is obtained through the following steps:

[0012] S11. Based on the geological parameters, initial pollution concentration distribution, and boundary conditions in the site's historical survey data, establish a set of multi-physics coupled partial differential equations; among which, the set of multi-physics coupled partial differential equations includes the convection-diffusion reaction equation, the multi-component reaction kinetic equation, and the Richards equation;

[0013] S12. Based on the multiphysics field coupled partial differential equations, construct a deep neural network with spatial coordinates and time coordinates as inputs and the predicted concentration values ​​of each pollutant component as outputs.

[0014] S13. Calculate the partial derivatives of the deep neural network output with respect to spatial and temporal coordinates using automatic differentiation; substitute these partial derivatives into a system of multiphysics coupled partial differential equations to calculate the equation residuals; use the equation residuals as physical constraint terms and the monitoring data from the site's historical survey data as data fitting terms; construct a loss function based on the physical constraint terms and the data fitting terms; train the deep neural network based on the loss function to obtain the trained deep neural network.

[0015] S14. Using different preset combinations of remediation operation parameters as input, perform forward inference calculations based on the trained deep neural network to generate predicted concentration values ​​of each pollutant component corresponding to the remediation operation parameter combination; collect the predicted concentration values ​​of each remediation operation parameter combination and its corresponding pollutant component to generate a training dataset; train a Fourier Neural Operator surrogate model based on the training dataset to obtain the trained Fourier Neural Operator surrogate model, and use the trained Fourier Neural Operator surrogate model as the initial physical information neural network surrogate model;

[0016] S15. Perform multiple random initialization training on the initial physical information neural network surrogate model to obtain multiple optimized Fourier Neural Operator surrogate models. Calculate the prediction variance based on the output of each optimized Fourier Neural Operator surrogate model, and use the prediction variance as a measure of prediction uncertainty. Construct an ensemble model based on each optimized Fourier Neural Operator surrogate model, and use the ensemble model as the final physical information neural network surrogate model.

[0017] In one embodiment, the belief state is obtained through the following steps:

[0018] S21. Obtain a set of competitive remediation hypotheses, including multiple competitive remediation hypotheses; wherein, the multiple competitive remediation hypotheses include the biodegradation hypothesis, the physical dilution hypothesis, the pollutant redistribution hypothesis, and the monitoring error hypothesis;

[0019] S22. Based on the physical information neural network surrogate model, set a corresponding kinetic parameter vector for each competitive remediation hypothesis; based on the kinetic parameter vector, generate the predicted concentration values ​​of each pollutant component under the corresponding competitive remediation hypothesis through the physical information neural network surrogate model.

[0020] S23. Based on the predicted concentration values ​​of each pollutant component, the prediction uncertainty measure, and the observed pollutant concentration values ​​in the real-time monitoring data, calculate the likelihood function of each competitive remediation hypothesis; take the logarithm of the likelihood function to calculate the log-likelihood function of each competitive remediation hypothesis; wherein, the variance term of the log-likelihood function is composed of the sum of the preset observation noise variance and the prediction uncertainty measure.

[0021] S24. Based on the prior probability of each competing repair hypothesis and the corresponding log-likelihood function, calculate the posterior probability of each competing repair hypothesis using Bayes' theorem.

[0022] S25. Calculate the belief entropy based on the posterior probability of each competing repair hypothesis; use the posterior probability of each competing repair hypothesis and the belief entropy together as the belief state.

[0023] In one embodiment, the system further includes a belief optimization module, which is used to:

[0024] Identify the types of monitoring data in real-time monitoring data, including concentration observation data, isotope ratio data, functional gene abundance data, and geochemical parameter data;

[0025] For each type of monitoring data, the evidence weight of the monitoring data type is calculated based on the logarithm of the ratio of the likelihood functions of the monitoring data type under each competing repair hypothesis.

[0026] The evidence values ​​are sorted in descending order based on the absolute values ​​of the evidence weights of each monitored data type to generate a ranking list of evidence values.

[0027] Adding the evidence value ranking list to the belief state yields an optimized belief state;

[0028] The optimized belief state is used as the final belief state for operations S3 and S4.

[0029] In one embodiment, the intelligent decision-making module includes an agent training unit, which is used for:

[0030] S31. Obtain the historical remediation case database and reward function; wherein, the reward function includes a pollution removal amount reward item, an economic cost penalty item, a belief entropy penalty item, and an information value reward item;

[0031] S32. Sample multiple repair scenario tasks from the historical repair case database; initialize the global policy parameters of the preset policy network to obtain the current policy;

[0032] S33. For each repair scenario task, interact with the environment simulator under the current policy to collect trajectory data; calculate the cumulative reward based on the trajectory data and the reward function; calculate the policy gradient loss based on the cumulative reward; perform multi-step inner gradient updates based on the policy gradient loss to obtain the adapted policy parameters.

[0033] S34. Calculate the meta-loss based on the adapted policy parameters; perform outer gradient update and update the global policy parameters based on the meta-loss.

[0034] S35. Repeat S32-S34 until the meta-loss reaches the preset meta-loss threshold, then stop repeating the operation. Use the last updated global policy parameters as the optimal global policy parameters, and construct a decision agent based on the policy network corresponding to the optimal global policy parameters.

[0035] In one embodiment, the current state is obtained through the following steps:

[0036] Extract the observed concentration vector of each monitoring point at the current moment based on the real-time monitoring data, and calculate the concentration gradient vector based on the observed concentration vector;

[0037] Get the repair process time; the repair process time is the time length between the current moment and the repair start moment.

[0038] The current state is generated based on the observed concentration vector, concentration gradient vector, belief state, parameters of the physical information neural network surrogate model, and repair process time.

[0039] In one embodiment, the mathematical expression for the log-likelihood function of each competing repair hypothesis is:

[0040]

[0041] in, In the first Competitive Repair Hypothesis and the corresponding dynamic parameter vector The observed evidence set The log-likelihood function value is dimensionless; For a moment The accumulated set of monitoring evidence; An index for competitive repair hypotheses; For the index of monitoring points, , The total number of monitoring points, dimensionless; For the index of time steps, , The total number of time steps is dimensionless. For the first The spatial coordinates of the monitoring points are in meters. For the first The moment of a time step is measured in seconds or days; The pre-defined observation noise variance is expressed in squares of concentration units. For physical information neural network surrogate models at location ,time The measurement of prediction uncertainty at a given location, with dimensions in the square of the concentration unit; For the first Each monitoring point at time The observed pollutant concentration values ​​in the real-time monitoring data, with the dimension of concentration units; In the first Competitive Repair Hypothesis and dynamic parameter vector Below, the location predicted by the physical information neural network surrogate model. ,time The pollutant concentration value at the location is expressed in units of concentration. It is the natural logarithm with base 1. .

[0042] Secondly, this application also provides a digital twin-based method for the whole-process management and control of soil pollution remediation, including:

[0043] S41. Establish a set of multi-physics coupled partial differential equations based on historical site survey data, construct a deep neural network based on the set of multi-physics coupled partial differential equations, construct a loss function based on historical site survey data, train the deep neural network based on the loss function to obtain the trained deep neural network, generate a training dataset based on the trained deep neural network, and train a Fourier Neural Operator surrogate model based on the training dataset to obtain a physical information neural network surrogate model.

[0044] S42. Obtain the prior probability of each competing repair hypothesis. Based on the physical information neural network surrogate model and real-time monitoring data, and combined with the prior probability of each competing repair hypothesis, calculate the posterior probability of each competing repair hypothesis using Bayes' theorem to generate the belief state.

[0045] S43. Construct the state space of the Markov decision process based on the belief state, construct an environment simulator of the Markov decision process based on the physical information neural network agent model, perform meta-reinforcement learning training in the environment simulator to obtain the decision agent, process the current state based on the decision agent, and output the repair action; wherein, the current state is constructed based on real-time monitoring data and belief state; the repair action includes instructions for adjusting the injection rate, extraction rate and monitoring frequency of the repair agent.

[0046] S44. Perform soil pollution remediation operations according to the remediation actions, collect new monitoring data after execution, update the parameters and belief states of the physical information neural network surrogate model according to the new monitoring data, and obtain the updated physical information neural network surrogate model and the updated belief states. Determine whether the preset remediation termination condition is met according to the updated belief states. If it is met, terminate the remediation. If it is not met, use the posterior probability distribution in the updated belief states as the prior probability of each competing remediation hypothesis in the next round S42, and return to S42 for the next round of iteration.

[0047] Thirdly, this application also provides a computer electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0048] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0049] The aforementioned digital twin soil pollution remediation end-to-end management system first constructs a surrogate model with both physical consistency and real-time reasoning capabilities through cascaded training of a physical information neural network and a Fourier Neural Operator. Combined with ensemble modeling, it outputs a predictive uncertainty metric, effectively addressing the high computational cost of high-precision multiphysics simulations and the difficulty in quantifying model cognitive biases, thus providing a transferable uncertainty boundary for subsequent reasoning. Secondly, based on the likelihood function generated by the surrogate model and multi-source real-time monitoring evidence, it uses Bayes' theorem to infer the posterior probabilities of competing remediation hypotheses online. Furthermore, it generates belief states with attribution confidence levels by combining evidence weight ranking, solving the problems of difficulty in distinguishing the causes of concentration decline and insufficient attribution ability for remediation mechanisms. This enables the decision-making module to perceive attribution uncertainty. Finally, a state space and reward function are constructed based on belief states. The decision agent is trained through meta-reinforcement learning to output repair actions. The closed-loop control module continuously updates the agent model parameters and belief states based on new monitoring data. The updated posterior distribution is used as the prior for the next round to achieve Bayesian online evolution. Combined with preset termination conditions, the repair is judged to meet the standard. This effectively solves the problems of model prediction bias accumulation and difficulty in quantifying and evaluating monitoring information gain under the open-loop architecture, and realizes adaptive closed-loop control of mechanism cognition and repair decision-making. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 A schematic diagram of a digital twin soil pollution remediation whole-process control system is provided as an exemplary embodiment of the present invention;

[0052] Figure 2 A flowchart of a digital twin method for the whole-process control of soil pollution remediation is provided as an exemplary embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] In one embodiment, such as Figure 1As shown, a digital twin soil pollution remediation whole-process management and control system 100 is provided. This embodiment illustrates the system's application to a terminal. It is understood that this system can also be applied to a server, or to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the system includes the following steps:

[0055] The twin modeling module 101 is used to execute S1, which includes: establishing a multi-physics coupled partial differential equation system based on historical site survey data; constructing a deep neural network based on the multi-physics coupled partial differential equation system; constructing a loss function based on historical site survey data; training the deep neural network based on the loss function to obtain the trained deep neural network; generating a training dataset based on the trained deep neural network; and training a Fourier NeuralOperator surrogate model based on the training dataset to obtain a physical information neural network surrogate model.

[0056] Specifically, the twin modeling module 101, based on fundamental information such as site geological structure, initial pollution distribution, and hydrological boundaries covered by historical site survey data, can construct a set of multi-physics coupled partial differential equations adapted to the migration and transformation patterns of soil pollutants. This model characterizes the multi-field coupled evolution mechanism of pollutant convection and dispersion, component reaction, and hydrological infiltration in the soil environment, overcoming the shortcomings of traditional single numerical equations that cannot adapt to the complex heterogeneous pollution evolution of sites. Based on the constraints of these physical equations, this module can construct a corresponding deep neural network structure, using spatial location and evolution time as input dimensions and multi-component pollutant concentrations as output dimensions to build an intelligent fitting network framework adapted to the physical evolution patterns. Furthermore, to avoid the problems of distortion of the physical mechanism and poor generalization ability of purely data-driven neural networks, this module can also construct a loss function that integrates physical residuals and measured data, combining historical site survey data. The physical equation residuals constrain the physical rationality of the network fitting, while the measured site monitoring data constrains the numerical accuracy of the network fitting. Iterative training of the deep neural network is completed through backpropagation of the loss function, resulting in a trained deep neural network that accurately fits the physical laws of the site. Based on the trained deep neural network, multi-condition forward inference simulation can be performed. By traversing different combinations of remediation operation parameters, pollution evolution simulation data is generated, constructing a training dataset covering multiple remediation scenarios. A Fourier NeuralOperator surrogate model is trained based on this dataset, ultimately yielding a physical information neural network surrogate model that combines physical consistency, high accuracy, and high generalization. This lightweight intelligent surrogate model replaces the traditional high-computation numerical solution method, significantly reducing the computational cost of subsequent full-process simulation and decision-making interactions.

[0057] The belief generation module 102 is used to execute S2, which includes: obtaining the prior probability of each competing repair hypothesis, and generating a belief state by combining the prior probability of each competing repair hypothesis with the prior probability of each competing repair hypothesis based on the physical information neural network surrogate model and real-time monitoring data and Bayes' theorem.

[0058] Specifically, the belief generation module 102 can first define or acquire various competing evolutionary hypotheses in the soil pollution remediation process, and initialize the prior probability distributions corresponding to each remediation hypothesis, thus serving as the initial mechanism cognition basis. After the physical information neural network surrogate model completes its modeling training, this module, based on the site pollution evolution physical mechanism carried by the surrogate model and combined with real-time monitoring data collected during the remediation process, can perform probability update deductions of multiple competing remediation hypotheses based on Bayes' theorem. This deduction process differs from the traditional remediation control model, which relies solely on concentration values ​​to directly determine the remediation effect. By combining model simulation predictions with real-time site observations, it can quantify the credibility of different pollution causes and remediation mechanisms, achieving a technological upgrade from judging single numerical results to multi-mechanism probabilistic attribution. Ultimately, it can output a belief state characterizing the current site pollution evolution mechanism and cognitive uncertainty, providing core input for the uncertainty dimension of subsequent intelligent decision-making, and compensating for the technical deficiencies of traditional decision-making systems that lack mechanism cognition and cannot quantify remediation uncertainty.

[0059] The intelligent decision-making module 103 is used to execute S3, which includes: constructing the state space of the Markov decision process based on the belief state; constructing an environment simulator of the Markov decision process based on the physical information neural network agent model; performing meta-reinforcement learning training in the environment simulator to obtain a decision agent; processing the current state based on the decision agent; and outputting repair actions. The current state is constructed based on real-time monitoring data and the belief state. The repair actions include instructions for adjusting the injection rate, extraction rate, and monitoring frequency of the repair agent.

[0060] Specifically, the intelligent decision-making module 103 uses the belief state output by the belief generation module 102 as its core foundation to construct the state space of a Markov decision-making process. This allows it to incorporate the uncertain cognitive characteristics of pollution evolution into the decision state dimension, overcoming the limitation of traditional reinforcement learning which relies solely on pollutant concentration data to construct the state space. Simultaneously, this module builds an environmental simulator of the Markov decision-making process based on a trained physical information neural network surrogate model. This allows it to leverage the efficient and accurate pollution evolution simulation capabilities of the surrogate model, providing a high-fidelity, low-latency environmental interaction platform for reinforcement learning training and overcoming the computational bottleneck of traditional numerical simulations that cannot support large-scale iterative training. Furthermore, this module can perform iterative training of meta-reinforcement learning within the environmental simulator, obtaining a generalized decision agent through multi-scenario task adaptation learning, enabling it to adapt to different site pollution evolution conditions and uncertain scenarios. The trained decision agent can then perform feature analysis and policy reasoning on the real-time constructed current state. The current state integrates real-time monitoring data and belief state information, taking into account both real-time site pollution conditions and the uncertainty of mechanistic cognition, effectively ensuring the comprehensiveness and authenticity of the decision input. Ultimately, the trained decision-making agent can output multi-dimensional remediation actions, including instructions on the injection rate of remediation agents, the extraction rate of polluted water, and the adjustment of monitoring frequency. This enables the dynamic allocation of pollution remediation control and monitoring resources, achieving synergistic optimization of remediation effectiveness, governance costs, and information gain.

[0061] The closed-loop control module 104 is used to execute S4. S4 includes: performing soil pollution remediation operations according to the remediation action, collecting new monitoring data after execution, updating the parameters and belief states of the physical information neural network surrogate model according to the new monitoring data, obtaining the updated physical information neural network surrogate model and the updated belief state, determining whether the preset remediation termination condition is met according to the updated belief state, terminating the remediation if it is met, and if it is not met, using the posterior probability distribution in the updated belief state as the prior probability of each competing remediation hypothesis in the next round S2, and returning to S2 for the next round of iteration.

[0062] Specifically, the closed-loop control module 104 can receive the remediation actions output by the intelligent decision-making module 103 and drive / instruct the completion of the corresponding soil pollution remediation operation. After the operation is completed, it can collect new monitoring data of the site in real time to achieve dynamic perception of the remediation status. Furthermore, this module can also perform dual iterative updates based on the newly collected monitoring data. For example, on the one hand, it can correct the model parameters of the physical information neural network proxy model online and continuously optimize the model simulation accuracy using real-time operating data, thereby effectively solving the open-loop defects of traditional digital twin models, which involve one-time modeling and long-term accumulation of operational deviations. On the other hand, it can update the posterior probability distribution of each competing remediation hypothesis by combining the new monitoring data, thereby achieving dynamic iteration of the belief state and continuously optimizing the accuracy of the system's understanding of the site remediation mechanism. In addition, this module can determine whether the current remediation process has reached the preset remediation termination condition through the updated belief state. If the termination condition is met, the remediation control process is directly terminated. If the termination condition is not met, the updated posterior probability distribution can be used as the prior probability for the next round of mechanism inference, realizing the temporal connection of Bayesian inference cognition and driving the system to return to the upstream process to carry out the next round of iterative control. Through a continuous closed-loop iterative process of decision-making, execution, monitoring, and updating, dynamic adaptive optimization of model accuracy, mechanism understanding, and decision-making strategies can be achieved throughout the entire remediation cycle, ensuring the continuity and accuracy of remediation and control in complex heterogeneous soil pollution scenarios.

[0063] The aforementioned system first achieves high-precision, low-overhead simulation modeling of site pollution evolution by constructing a Fourier neural operator surrogate model with multi-physics constraints, solving the problems of high computing power and insufficient prediction accuracy in traditional numerical simulations. Second, it solves the probabilistics of multiple competing remediation hypotheses based on Bayesian inference, achieving accurate attribution of remediation mechanisms and quantification of uncertainties, effectively avoiding false judgments of remediation compliance. Furthermore, it integrates real-time monitoring data and belief states to construct a decision state space, and combines meta-reinforcement learning intelligent training to achieve adaptive intelligent decision-making that balances remediation effectiveness, governance costs, and the value of monitoring information. Finally, it iteratively updates model parameters and belief states in real time using monitoring data, constructing a complete closed-loop iterative optimization mechanism, achieving continuous adaptive updates to model accuracy and temporal cognition, significantly improving the precision and intelligence of dynamic management throughout the entire soil pollution remediation process.

[0064] In one embodiment, the physical information neural network proxy model in the twin modeling module 101 can be obtained through the following steps:

[0065] S11. Based on the geological parameters, initial pollution concentration distribution, and boundary conditions in the site's historical survey data, establish a set of multi-physics coupled partial differential equations; among which, the set of multi-physics coupled partial differential equations includes the convection-diffusion reaction equation, the multi-component reaction kinetic equation, and the Richards equation;

[0066] S12. Based on the multiphysics field coupled partial differential equations, construct a deep neural network with spatial coordinates and time coordinates as inputs and the predicted concentration values ​​of each pollutant component as outputs.

[0067] S13. Calculate the partial derivatives of the deep neural network output with respect to spatial and temporal coordinates using automatic differentiation; substitute these partial derivatives into a system of multiphysics coupled partial differential equations to calculate the equation residuals; use the equation residuals as physical constraint terms and the monitoring data from the site's historical survey data as data fitting terms; construct a loss function based on the physical constraint terms and the data fitting terms; train the deep neural network based on the loss function to obtain the trained deep neural network.

[0068] S14. Using different preset combinations of remediation operation parameters as input, perform forward inference calculations based on the trained deep neural network to generate predicted concentration values ​​of each pollutant component corresponding to the remediation operation parameter combination; collect the predicted concentration values ​​of each remediation operation parameter combination and its corresponding pollutant component to generate a training dataset; train a Fourier Neural Operator surrogate model based on the training dataset to obtain the trained Fourier Neural Operator surrogate model, and use the trained Fourier Neural Operator surrogate model as the initial physical information neural network surrogate model;

[0069] S15. Perform multiple random initialization training on the initial physical information neural network surrogate model to obtain multiple optimized Fourier Neural Operator surrogate models. Calculate the prediction variance based on the output of each optimized Fourier Neural Operator surrogate model, and use the prediction variance as a measure of prediction uncertainty. Construct an ensemble model based on each optimized Fourier Neural Operator surrogate model, and use the ensemble model as the final physical information neural network surrogate model.

[0070] Specifically, geological parameters in the site's historical survey data can include permeability tensor, porosity, soil medium density, longitudinal dispersion, and lateral dispersion. These parameters are obtained through site borehole sampling, laboratory geotechnical tests, and field hydrological tests. Furthermore, the concentration values ​​of each pollutant component can be determined by laboratory chemical analysis of samples collected from the site's monitoring well network. This analysis, combined with spatial interpolation methods, generates an initial concentration field across the entire site, yielding the initial pollutant concentration distribution. Boundary conditions can be determined based on the actual hydrogeological conditions of the site, including constant concentration boundaries, constant flux boundaries, and constant head boundaries. These geological parameters, the initial concentration field, and the boundary conditions together constitute the fundamental inputs for solving the partial differential equation system.

[0071] For example, when establishing a system of partial differential equations, one can first write out the convection-dispersion reaction equation independently for each pollutant component, which is:

[0072]

[0073] in, For the first Spatial location of pollutant components and time The concentration at that point is in units of mol. or mg ; Let be the hydrodynamic dispersion coefficient tensor, with dimensions . This tensor is composed of two superpositions: the molecular diffusion coefficient and the mechanical dispersion coefficient, wherein the mechanical dispersion coefficient is linearly related to the pore water flow velocity. Let m be the pore water velocity vector, with dimensions m. ; For the first The components in the first The attenuation coefficient in the reaction pathway, with dimensions of ; This represents the total number of reaction pathways. For the first The source and sink terms of each component, with dimensions in mol. This term is used to describe the impact of human interventions such as the injection of repair agents and extraction wells on the concentration.

[0074] For remediation scenarios involving the interconversion of multiple pollutant components, a multi-component reaction kinetic equation can be constructed. This equation expresses the reaction rate as the rate of change of each component's concentration, and represents the net formation rate of each component as the algebraic sum of all chemical reaction rate functions involving that component:

[0075]

[0076] in, For the first The components in the first The stoichiometric coefficients of a reaction, when When it is a reactant When it is negative, When it is a product It is a positive value with one dimension; For the first The rate function of a reaction, with dimensions in mol. The specific form of this rate function is determined according to the reaction type; This represents the total number of pollutant components. This is a vector of reaction kinetic parameters, including parameters such as maximum specific degradation rate, half-saturation constant, and yield coefficient. The initial values ​​of these parameters are preset.

[0077] For remediation scenarios involving unsaturated water flow, the Richards equation can be constructed to describe the movement of water in unsaturated soil. The general form of the Richards equation is:

[0078]

[0079] in, The volumetric water content is dimensionless. It is a function of the unsaturated permeability coefficient, with dimensions m. The change of this function with water content is described by the soil moisture characteristic curve; The pressure head is measured in meters (m). The coordinates are elevation coordinates, with dimensions in meters. The constitutive relationships between the unsaturated permeability coefficient and water content, as well as between the pressure head and water content, in this equation are established using the van Genuchten model or the Brooks-Corey model. The model parameters were determined by site soil texture and indoor soil column experiments.

[0080] Specifically, there is a coupling relationship between the three equations mentioned above. The pore water velocity in the convection-dispersion reaction equation is related to the head distribution obtained from the Richards equation through Darcy's law; this relationship is as follows: ,in For unsaturated permeability tensor, Porosity. The decay coefficients in the convection-dispersion reaction equations can be further expanded within the multi-component reaction kinetic equations, replacing the single decay coefficient with a combined expression related to each reaction rate function. This allows the total decay rate of pollutants to be determined by the superposition of rate functions from each reaction pathway. The water content distribution obtained from the Richards equations directly affects the unsaturated permeability coefficient in Darcy's law, thus altering the pore water velocity field in the convection-dispersion reaction equations. Ultimately, these three equations, through their respective variable and parameter coupling, can be combined to form a set of governing equations characterizing the multi-physics coupling evolution during soil pollution remediation. This set of equations serves as the mathematical foundation for subsequently constructing deep neural networks and physical information loss functions.

[0081] Based on the aforementioned multiphysics coupled partial differential equations, a deep neural network can be constructed. This deep neural network is used to approximate the solution function of the partial differential equations, that is, to establish a mapping relationship from any spatiotemporal point to the concentration of each pollutant component. For example, the input layer of this network contains four neurons, corresponding to the x, y, and z components in three-dimensional spatial coordinates and the time coordinate t, respectively. The number of neurons in the output layer is consistent with the number of pollutant components to be predicted, with one output neuron for each pollutant, and the output value is the predicted concentration of that component at a given spatiotemporal point. The hidden layers of this network can adopt a multi-layer fully connected structure, with multiple neurons in each layer. Nonlinear activation functions are used between layers to introduce nonlinear transformation capabilities, enabling the network to fit complex concentration distribution surfaces.

[0082] During training, automatic differentiation techniques can be used to calculate the partial derivatives of the deep neural network output with respect to spatial and temporal coordinates. This automatic differentiation process uses the chain rule to backpropagate gradients on the computational graph, efficiently obtaining the higher-order partial derivatives of the network output with respect to the input at machine precision, avoiding the truncation error introduced by traditional numerical differentiation. Substituting the calculated partial derivatives into the aforementioned multiphysics coupled partial differential equations, the residuals can be calculated. These residuals reflect the degree to which the network output deviates from the constraints of the physical equations at spatiotemporal points; the closer the residuals are to zero, the more strictly the network output follows the physical laws. Using the equation residuals as physical constraint terms and the monitoring data from the site's historical survey data as data fitting terms, a loss function can be constructed. The specific form is:

[0083]

[0084] in, The parameter is Deep neural networks in spatial coordinates and time coordinates The predicted concentration values ​​at the location are in units of concentration, which may be mg / L or μmol / L depending on the specific pollutant type. This represents the concentration observation value at the k-th monitoring sample point, with dimensions equal to... Consistent; The total number of monitoring sample points, with a dimension of one; This represents the residual of the equations obtained by substituting the partial derivatives of each order from the network output into a system of multiphysics coupled partial differential equations. Its dimension depends on the specific equation form; for the convection-dispersion equation, its dimension is concentration units divided by time units, such as mg. ; The total number of physical constraint configuration points, with a dimension of one; The weight coefficients for the physical constraint term, with dimensions of one, are used to adjust the relative contributions of the physical constraint term and the data fitting term to the total loss. In this loss function, the data fitting term constrains the network output to be consistent with the measured monitoring data of the site to ensure prediction accuracy, while the physical constraint term constrains the network output to satisfy a system of partial differential equations to ensure that the prediction results conform to the laws of physical evolution. Based on this loss function, the gradient of the loss function with respect to the network parameters can be calculated using the backpropagation algorithm, and the network parameters can be iteratively updated using a gradient descent optimization algorithm until the loss function converges, resulting in the trained deep neural network.

[0085] Furthermore, the remediation operation parameters can include the injection rate of the remediation agent, the extraction rate of the extraction well, and the spatial layout parameters of the injection and extraction wells. Different parameter combinations are obtained through equally spaced sampling or Latin hypercube sampling within a preset range to cover typical operating conditions in the remediation operation space. Using different preset remediation operation parameter combinations as input, forward inference calculations can be performed through a trained deep neural network to generate the concentration prediction values ​​of each pollutant component corresponding to each remediation operation parameter combination. By aggregating the concentration prediction values ​​of each remediation operation parameter combination and its corresponding pollutant component, a training dataset can be constructed. Each sample in this dataset contains a set of remediation operation parameters and the concentration prediction value of the entire field under that parameter combination. Based on this training dataset, a Fourier Neural Operator surrogate model can be further trained. The Fourier Neural Operator is a neural operator architecture that directly learns the mapping relationship from remediation operation parameters and spatiotemporal coordinates to the concentration field in Fourier space through learnable global convolutional kernels. During training, the Fourier Neural Operator optimizes network parameters by minimizing the difference between the predicted concentration field and the concentration field output by the deep neural network in the training dataset, using mean squared error as the loss function and updating network weights through backpropagation. After training, the trained Fourier Neural Operator surrogate model can be used as the initial physical information neural network surrogate model. Compared to the original deep neural network, this surrogate model has resolution invariance and computational efficiency, and after training, it can complete concentration field inference under different combinations of repair operation parameters at a speed of seconds.

[0086] Furthermore, the initial physical information neural network surrogate model can be trained multiple times using random initialization. For example, for each training iteration, the network parameters can be independently and randomly initialized under the same network architecture, using the same training dataset and optimization strategy, until each model converges. Due to the different parameter initializations and the randomness of the training process in each training iteration, the parameters of each converged model will differ, and the model's prediction output for the same input will also be slightly different. Subsequently, the prediction variance can be calculated based on the output of each optimized Fourier Neural Operator surrogate model after training, and this prediction variance can be used as a measure of prediction uncertainty. The prediction variance is calculated as follows:

[0087]

[0088] in, To optimize the total number of Fourier Neural Operator agent models, the dimension is one; This indicates that the k-th optimized Fourier Neural Operator surrogate model is at a specific point in time and space. The predicted concentration at the location is expressed in units of concentration. The arithmetic mean of the predicted values ​​from all models is expressed in units of concentration. The prediction variance is expressed in squared units of concentration. This variance reflects the uncertainty in prediction cognition caused by the uncertainty of model parameters. A large prediction variance indicates a lack of reliable prediction confidence in a given region when the predictions of different models at a given spatiotemporal point differ significantly. Conversely, a small prediction variance indicates more reliable predictions in that region when the predictions of different models are consistent. Integrating the optimized Fourier Neural Operator surrogate models yields an ensemble model, which serves as the final physical information neural network surrogate model. In subsequent actual predictions, the final output of the ensemble model is the average of the predictions from each model. It also outputs the prediction variance as a measure of the uncertainty of the prediction result, which is then used for subsequent Bayesian inference steps.

[0089] In one embodiment, the belief state can be obtained through the following steps:

[0090] S21. Obtain a set of competitive remediation hypotheses, including multiple competitive remediation hypotheses; wherein, the multiple competitive remediation hypotheses include the biodegradation hypothesis, the physical dilution hypothesis, the pollutant redistribution hypothesis, and the monitoring error hypothesis.

[0091] Specifically, the set of competing remediation hypotheses covers all possible mechanisms leading to changes in observed site contaminant concentrations, including the biodegradation hypothesis, physical dilution hypothesis, contaminant redistribution hypothesis, and monitoring error hypothesis. The biodegradation hypothesis posits that the observed decrease in contaminant concentration is primarily due to microbial-mediated biodegradation, which is the desired remediation mechanism for remediation projects. The physical dilution hypothesis attributes the concentration decrease mainly to hydrodynamic dispersion and molecular diffusion, a non-remediation mechanism. The contaminant redistribution hypothesis suggests that the concentration decrease is primarily due to the migration of contaminants with groundwater flow to unmonitored areas, a risk migration mechanism rather than genuine contaminant removal. The monitoring error hypothesis argues that the observed concentration changes are mainly caused by sampling errors, laboratory analysis errors, or natural fluctuations, a noise mechanism. These hypotheses constitute a complete and mutually exclusive hypothesis space, meaning that the true mechanism of site contamination evolution necessarily belongs to one and only one of these hypotheses.

[0092] S22. Based on the physical information neural network surrogate model, set a corresponding kinetic parameter vector for each competitive remediation hypothesis; based on the kinetic parameter vector, generate the predicted concentration values ​​of each pollutant component under the corresponding competitive remediation hypothesis through the physical information neural network surrogate model.

[0093] Specifically, the types of parameters included in the kinetic parameter vector are directly related to the type of hypothesis. For the biodegradation hypothesis, the kinetic parameter vector can include the first-order decay coefficients or Monod kinetic parameters for each degradation pathway. For the physical dilution hypothesis, the decay coefficients in the kinetic parameter vector can be set to zero or near zero. For the pollutant redistribution hypothesis, the convective velocity component in the kinetic parameter vector can be adjusted to enhance the horizontal migration flux. For the monitoring error hypothesis, the kinetic parameter vector only includes parameters describing the variance of the observation noise, which is a Gaussian white noise variance independent of the physical process. Since the physical information neural network surrogate model takes spatial and temporal coordinates as input and outputs the predicted concentration values ​​of each pollutant component, given the kinetic parameter vector, the model outputs the concentration field prediction results over the entire spatiotemporal domain under that hypothesis. That is, for each competitive remediation hypothesis, the corresponding kinetic parameter vector is input into the physical information neural network surrogate model, and through one forward inference calculation, the predicted concentration values ​​of each pollutant component at all monitoring spatiotemporal points under that hypothesis can be obtained.

[0094] S23. Based on the predicted concentration values ​​of each pollutant component, the measurement of prediction uncertainty, and the observed pollutant concentration values ​​in the real-time monitoring data, calculate the likelihood function of each competitive remediation hypothesis; take the logarithm of the likelihood function to calculate the log-likelihood function of each competitive remediation hypothesis; wherein, the variance term of the log-likelihood function is composed of the sum of the preset observation noise variance and the measurement of prediction uncertainty.

[0095] Specifically, in calculating the log-likelihood function of each competing remediation hypothesis based on the predicted concentrations of each pollutant component, the measurement of prediction uncertainty, and the observed concentrations in real-time monitoring data, it is first necessary to determine the form of the likelihood function for each type of monitoring data under each hypothesis. For concentration observation data, under given assumptions, the difference between the observed values ​​and the model predictions follows a normal distribution with a mean of zero and a variance equal to the sum of the pre-defined observation noise variance and the measurement of prediction uncertainty. The reason for including the measurement of prediction uncertainty in the variance term is that it reflects the cognitive uncertainty of the physical information neural network surrogate model at a specific spatiotemporal point. When the model's predictions in that region are unreliable, the deviation between the observed and predicted values ​​should not be over-interpreted as evidence supporting or opposing a particular hypothesis. Therefore, the mathematical expression for the log-likelihood function of each competing remediation hypothesis can be:

[0096]

[0097] in, In the first Competitive Repair Hypothesis and the corresponding dynamic parameter vector The observed evidence set The log-likelihood function value has one dimension; An index for competitive repair hypotheses. These correspond to the biodegradation hypothesis, the physical dilution hypothesis, the pollutant redistribution hypothesis, and the monitoring error hypothesis, respectively. For the index of monitoring points, , The total number of monitoring points is defined as having one dimension; For the index of time steps, , The total number of time steps, dimensionless; For the first The spatial coordinates of the monitoring points are in meters. For the first The moment of a time step is measured in seconds or days; The pre-defined observation noise variance is expressed in squares of concentration units. For physical information neural network surrogate models at location ,time The measurement of prediction uncertainty at a given location, with dimensions in the square of the concentration unit; For the first Each monitoring point at time The observed pollutant concentration values ​​in the real-time monitoring data, with the dimension of concentration units; In the first Competitive Repair Hypothesis and dynamic parameter vector Below, the location predicted by the physical information neural network surrogate model. ,time The pollutant concentration value at the location is expressed in units of concentration. It is the natural logarithm with base 1. The variance term in the above log-likelihood function It is composed of the sum of the pre-set observation noise variance and the prediction uncertainty measure, reflecting the combined impact of observation randomness and model cognitive uncertainty on likelihood estimation.

[0098] S24. Based on the prior probability and corresponding log-likelihood function of each competing repair hypothesis, calculate the posterior probability of each competing repair hypothesis using Bayes' theorem.

[0099] Specifically, the posterior probability of each competing repair hypothesis can be calculated based on its prior probability and the corresponding log-likelihood function. The mathematical expression of Bayes' theorem is as follows:

[0100]

[0101] in, In the evidence set Next The posterior probability of a competing repair hypothesis, with a dimension of one; For the first The prior probabilities of the competing remediation hypotheses are dimensionless and are set based on the site conceptual model and remediation engineering experience. The denominator is the sum of the products of the prior probabilities and likelihood functions of all four hypotheses, serving as a normalization constant to ensure that the sum of all posterior probabilities equals 1. In practical calculations, to avoid numerical underflow, the posterior probabilities can be obtained by taking the exponent after calculation in the logarithmic field.

[0102] S25. Calculate the belief entropy based on the posterior probability of each competing repair hypothesis; use the posterior probability of each competing repair hypothesis and the belief entropy together as the belief state.

[0103] Specifically, the mathematical expression for belief entropy is:

[0104]

[0105] in, State of belief The belief entropy, with dimensions in knights; For a moment state of belief Belief entropy is a quantitative indicator that measures the uncertainty of decision-makers' understanding of the current repair mechanism. Belief entropy reaches its maximum value when the posterior probability distributions of the four hypotheses are uniform. At this point, the decision-maker is completely unable to distinguish which hypothesis is more credible, and the attribution of the repair mechanism is in a state of maximum uncertainty. When the posterior probability of a certain hypothesis approaches 1 while the others approach 0, the belief entropy approaches 0, indicating that the decision-maker has a high degree of confidence in the attribution of the repair mechanism. The posterior probabilities of each competing repair hypothesis and the belief entropy are used together as the belief state output. The belief state includes both the current distribution of confidence in the attribution of the repair mechanism and the degree of certainty of the attribution. Both serve as inputs to subsequent intelligent decision-making modules, enabling the decision-making process to perceive the uncertainty of the attribution of the repair mechanism.

[0106] In one embodiment, the system further includes a belief optimization module. This module optimizes the belief state output by the belief generation module. Specifically, it calculates the evidence weights for each type of monitored data and generates an evidence value ranking list. This ranking list serves as supplementary information to the belief state, enabling the subsequent decision-making module to identify which types of monitored data are most informative in distinguishing between different competing repair hypotheses. This allows for the priority collection of high-value evidence when monitoring resources are limited, improving monitoring efficiency and decision reliability. For example, the optimization process is as follows:

[0107] Identify the types of monitoring data in real-time monitoring data, including concentration observation data, isotope ratio data, functional gene abundance data, and geochemical parameter data;

[0108] For each type of monitoring data, the evidence weight of the monitoring data type is calculated based on the logarithm of the ratio of the likelihood functions of the monitoring data type under each competing repair hypothesis.

[0109] The evidence values ​​are sorted in descending order based on the absolute values ​​of the evidence weights of each monitored data type to generate a ranking list of evidence values.

[0110] Adding the evidence value ranking list to the belief state yields an optimized belief state;

[0111] The optimized belief state is used as the final belief state for operations S3 and S4.

[0112] Specifically, the belief optimization module can first identify the data types in real-time monitoring data, such as concentration observation data, isotope ratio data, functional gene abundance data, and geochemical parameter data. Concentration observation data refers to the concentration values ​​of various pollutant components obtained from laboratory analysis of groundwater samples collected from site monitoring wells, usually expressed in milligrams per liter or micromoles per liter. Isotope ratio data refers to the compound-specific stable isotope ratios, such as the ratio of carbon-13 to carbon-12 relative to international standard reference materials, expressed in parts per thousand. During biodegradation, lighter isotopes (carbon-12) are preferentially converted due to their higher reactivity, leading to a relative enrichment of heavier isotopes (carbon-13) in the remaining pollutants. Therefore, changes in isotope ratios are specific evidence for determining whether biodegradation has occurred. Functional gene abundance data refers to the copy number concentration of genes encoding specific degradation enzymes in soil or groundwater samples, usually expressed in copies per milliliter or copies per gram of soil. Measured using quantitative polymerase chain reaction (PCR), its abundance level directly characterizes the size of the microbial population capable of degradation. Geochemical parameters typically include dissolved oxygen concentration, redox potential, pH, ferrous ion concentration, and sulfate ion concentration. These parameters can reflect the suitability of the soil groundwater environment for biodegradation reactions from different perspectives.

[0113] For each type of monitoring data, the evidence weight can be calculated by taking the logarithm of the ratio of the likelihood functions of that monitoring data type under each competing remediation hypothesis. This evidence weight quantifies the information content of a particular monitoring data type in distinguishing between two competing remediation hypotheses. For example, the ratio of the likelihood function value of that monitoring data type under the biodegradation hypothesis to the likelihood function value under the physical dilution hypothesis can be taken, and then the natural logarithm of that ratio can be calculated. The mathematical expression for the definition of evidence weight is:

[0114]

[0115] in, For the first The evidence weight for each type of monitoring data is of one dimension. For the first Observations of various monitoring data types; To support the biodegradation hypothesis Observed evidence The likelihood function value; To implement the physical dilution hypothesis Observed evidence The likelihood function value. When A value greater than zero indicates that the evidence supports the biodegradation hypothesis better than the physical dilution hypothesis, and the larger the value, the stronger the support; when... A value less than zero indicates that the evidence supports the physical dilution hypothesis better than the biodegradation hypothesis, and the larger the absolute value, the stronger the support; when... A value close to zero indicates that the evidence has no informational value in distinguishing between the biodegradation and physical dilution hypotheses. The biodegradation and physical dilution hypotheses were chosen as the baseline hypothesis pair for calculating the weight of evidence because these two hypotheses represent the most critical decision in remediation projects—whether the concentration decrease is due to actual pollution removal or physical migration. This decision directly determines whether the current remediation strategy should be maintained or adjusted.

[0116] Specifically, the larger the absolute value of the evidence weight, the stronger the ability of the monitored data type to distinguish between the two hypotheses of biodegradation and physical dilution, and the higher its information value in decision-making. Therefore, the evidence value ranking list can be generated by sorting the monitored data types in descending order based on the absolute value of their evidence weights. The monitored data types that rank higher in this ranking list should be given higher sampling priority in subsequent monitoring scheme optimization. Adding the evidence value ranking list to the belief state yields the optimized belief state. This optimized belief state retains the original belief state information while adding the information value ranking dimension for each monitored data type. This ranking list provides a clear priority basis for the subsequent intelligent decision-making module when formulating monitoring frequency adjustment instructions, enabling the allocation of monitoring resources to focus on the most discriminative evidence types and avoiding excessive monitoring costs on low-information-value data. The optimized belief state is used as the final belief state for the subsequent state space construction of the Markov decision process and the iterative update operations of the closed-loop control module.

[0117] In one embodiment, the intelligent decision-making module 103 includes an agent training unit, which can be used for:

[0118] S31. Obtain the historical remediation case database and reward function; wherein, the reward function includes a pollution removal amount reward item, an economic cost penalty item, a belief entropy penalty item, and an information value reward item;

[0119] S32. Sample multiple repair scenario tasks from the historical repair case database; initialize the global policy parameters of the preset policy network to obtain the current policy;

[0120] S33. For each repair scenario task, interact with the environment simulator under the current policy to collect trajectory data; calculate the cumulative reward based on the trajectory data and the reward function; calculate the policy gradient loss based on the cumulative reward; perform multi-step inner gradient updates based on the policy gradient loss to obtain the adapted policy parameters.

[0121] S34. Calculate the meta-loss based on the adapted policy parameters; perform outer gradient update and update the global policy parameters based on the meta-loss.

[0122] S35. Repeat S32-S34 until the meta-loss reaches the preset meta-loss threshold, then stop repeating the operation. Use the last updated global policy parameters as the optimal global policy parameters, and construct a decision agent based on the policy network corresponding to the optimal global policy parameters.

[0123] Specifically, the historical remediation case database contains historical remediation decision trajectories from multiple different contaminated sites. Each trajectory records a complete state sequence, action sequence, and reward sequence from the initial state to the completion of remediation in a given scenario. The reward function can be used to quantify the immediate benefit of performing a specific action in a specific state; its specific form is as follows:

[0124]

[0125] in, For a moment The Markov decision process state includes information such as observed concentration vector, concentration gradient vector, belief state, surrogate model parameters, and repair process time; For a moment The performed repair operations include instructions for adjusting the injection rate, extraction rate, and monitoring frequency of the repair agent; In the state Next action The instant reward value obtained is dimensionless and, after normalization, ranges from 0 to 1. To perform the action The amount of pollutants removed is obtained by predicting the difference in concentration field before and after the action using a physical information neural network surrogate model and integrating it over the entire site space. The dimension is the product of concentration units and volume units, and after normalization, it is used in the reward calculation in a dimensionless form; Cost To perform the action The required economic costs include the cost of drug consumption, the cost of equipment operation energy consumption, and the cost of labor operation. The various costs are weighted and summed according to the market price and the unit consumption, and then normalized to obtain the dimensionless cost value. State of belief Belief entropy; IV To perform the action The expected value of the information obtained is calculated by traversing the evidence value ranking list. The top-ranking evidence types, based on the current belief state, predict the expected decrease in belief entropy after strengthening the collection of these evidence types; , , , The four weighting coefficients, each with a dimension of one, are all greater than zero and their sum equals 1. They are used to adjust the relative importance of pollution removal efficiency, economic cost, cognitive uncertainty, and information value in the total reward. In the reward function described above, the pollution removal amount reward can guide the agent to prioritize actions that can effectively reduce pollution; the economic cost penalty can inhibit the overuse of high-cost actions; the belief entropy penalty can prompt the agent to consider the cognitive uncertainty of the repair mechanism when choosing actions; and the information value reward can encourage the agent to prioritize monitoring actions to obtain highly discriminative evidence when the information value is high enough.

[0126] Subsequently, multiple remediation scenario tasks can be sampled from the historical remediation case database. Each remediation scenario task corresponds to a specific contaminated site condition, including site hydrogeological parameters, pollutant type and initial concentration distribution, and remediation technology type. The global policy parameters of the preset policy network are initialized to obtain the current policy. These global policy parameters are the initial values ​​of the neural network weights of the decision-making agent. The network structure can contain two hidden layers, each with 256 neurons, using modified linear units as the activation function. The output layer outputs the mean and standard deviation of continuous actions, as well as the selection probability of discrete actions.

[0127] Specifically, the environmental simulator is built on a physical information neural network agent model, capable of predicting the pollutant concentration distribution at the next moment based on the current state and actions. Therefore, for each remediation scenario task, during interaction with the environmental simulator under the current policy, the agent selects an action at each time step according to the current policy. The environmental simulator can then predict the next state based on the action and return an immediate reward. This process repeats until a preset number of interaction steps is reached, generating trajectory data containing a state sequence, action sequence, and reward sequence. The cumulative reward is calculated based on the trajectory data and the reward function. The cumulative reward is the discounted sum of the immediate rewards at each time step in the trajectory, with the discount factor ranging from 0.95 to 0.99.

[0128] Specifically, cumulative reward measures the total expected gain an agent can obtain by executing subsequent actions starting from the current time step. Based on the cumulative reward, policy gradient loss can be calculated, which estimates the superiority of the current policy relative to the average policy through the advantage function. The advantage function is calculated by comparing the expected cumulative reward of a given action with the average cumulative reward obtainable under the current policy in that state; the difference is the advantage function value. When the expected cumulative reward of an action is higher than the average level of the current policy in that state, the advantage function is positive, increasing the probability of selecting that action during policy gradient updates; when the expected cumulative reward is lower than the average level, the advantage function is negative, decreasing the probability of selecting that action during policy gradient updates. Therefore, cumulative reward provides an absolute measure of action superiority, while the advantage function provides a relative measure. By driving policy gradient updates through the advantage function, the agent can continuously improve the quality of its decisions.

[0129] Based on the policy gradient loss, gradient descent can be used to perform multi-step inner gradient updates. Each update step can use a batch of samples sampled from the trajectory data to calculate the gradient. The learning rate is set to a value between 0.001 and 0.01, and the number of update steps is set to 5 to 10. After completing the inner update of one repair scene task, the adapted policy parameters can be obtained, and the process can move on to the next repair scene task until the inner update of all sampled repair scene tasks has been completed.

[0130] Furthermore, a meta-loss can be calculated based on the adapted policy parameters. This meta-loss measures the sum of the performance of each repair scenario task under the adapted policy parameters, reflecting the applicability of the current global policy parameters as the starting point for training each task. For example, under the adapted policy parameters for each task, the policy gradient loss can be calculated using the validation set trajectory data of that task, and the meta-loss is obtained by summing the validation losses of all tasks. The meta-loss can be used to update the second-order gradient of the initial global policy parameters using the outer gradient, so that the updated global policy parameters can achieve better performance on each task after a small number of inner gradient updates. The outer learning rate can be set to a value between 0.0001 and 0.001.

[0131] The meta-loss threshold is pre-set based on the decreasing trend of meta-loss during training. When the meta-loss continuously decreases and stabilizes below the threshold, it indicates that the global policy parameters have the ability to generalize across tasks. Therefore, the process of repeatedly sampling repair scenario tasks from the historical repair case database, initializing global policy parameters, interacting with the environment simulator to collect trajectory data, performing inner gradient updates, calculating meta-loss, and performing outer gradient updates is stopped when the meta-loss reaches the preset meta-loss threshold. After stopping the repeated execution, the last updated global policy parameters can be used as the optimal global policy parameters. The policy network corresponding to the optimal global policy parameters is used as the policy part of the decision agent. This policy network, together with the current state input interface and the environment interaction interface, constitutes the complete decision agent. During the deployment phase, the decision agent can receive the current state as input and calculate the output repair action through forward inference of the policy network, without needing to continue performing gradient updates of the policy network during the deployment phase.

[0132] In one embodiment, the intelligent decision-making module 104 can also process the current state based on the trained decision-making agent to output a repair action. Illustratively, the observed concentration vector of each monitoring point at the current moment can be extracted from real-time monitoring data. The dimension of this observed concentration vector is equal to the product of the total number of monitoring points and the total number of pollutant components. Assuming M monitoring points are deployed in the site, and each monitoring point detects the concentration of P pollutant components, the observed concentration vector is a vector of dimension M×P, where each element is the observed concentration value of the corresponding component at the corresponding monitoring point. Subsequently, a concentration gradient vector can be calculated based on the observed concentration vector using spatial interpolation methods. For example, the discrete concentration observation values ​​of each monitoring point are first used as spatial discrete sample points. Inverse distance weighted interpolation or Kriging interpolation is used to map the discrete points to a continuous concentration field over the entire site space. Then, the first-order partial derivatives of this continuous concentration field are calculated in the three directional components of the spatial coordinates to obtain the gradient components of the concentration in the x, y, and z directions. The dimension of the gradient components is concentration per meter, reflecting the rate of change of pollutant concentration in space. By extracting the gradient components at each monitoring point, a concentration gradient vector with the same dimension as the observed concentration vector can be constructed.

[0133] Furthermore, the parameters of the physical information neural network surrogate model can be obtained. These parameters include all trainable parameters in the network structure, specifically the weight matrices and bias vectors between each hidden layer. For example, assuming the surrogate model contains L Fourier Neural Operator layers, each containing a physical space transformation matrix and a Fourier space transformation matrix, flattening all parameters of all layers forms a high-dimensional parameter vector, which is used to construct the current state. The repair process time is obtained; this repair process time is the length of time between the current moment and the repair start time, measured in seconds, minutes, hours, or days, depending on the time scale of the repair project. In state construction, the repair process time participates in subsequent concatenation as a scalar.

[0134] Specifically, the five components—observed concentration vector, concentration gradient vector, belief state, parameter vector of the physical information neural network surrogate model, and remediation process time—can be concatenated. Before concatenation, each element in the observed concentration vector and concentration gradient vector can be standardized by dividing each concentration value by its corresponding initial or reference concentration value to obtain a dimensionless relative concentration value. This eliminates the impact of differences in the magnitude of different pollutant concentrations on the stability of the decision model training. It's worth noting that the posterior probability distribution in the belief state can be left unstandardized, as its posterior probability naturally falls within the range of 0 to 1, and the sum of the posterior probabilities of each hypothesis is 1. Similarly, the parameter vector of the physical information neural network surrogate model can be left unstandardized because these parameters may shift overall after each iteration of the closed-loop control module, and the magnitude of the shift accumulates with the number of iterations. The original values ​​themselves are valid information, reflecting the model's evolutionary state in the current remediation process. The remediation process time is normalized by dividing it by the estimated total remediation period to obtain a dimensionless relative remediation time, with a value between 0 and 1.

[0135] The five parts mentioned above are concatenated into a complete vector in a preset order, and this concatenated vector is the current state. The total dimension of the current state is equal to the sum of the dimensions of the observed concentration vector, the concentration gradient vector, the belief state, the parameter vector of the physical information neural network surrogate model, and the remediation process time. Since the parameter vector dimension of the physical information neural network surrogate model is usually large, principal component analysis can be used to reduce the dimensionality of this high-dimensional parameter in actual storage and computation. The first few principal components are extracted as the parameter inputs to the state space to reduce the dimensionality of the state space while retaining the main information in the parameter evolution direction. The final generated current state includes information on the spatial distribution of site pollution and its spatial rate of change, the attribution confidence of the current remediation mechanism and its uncertainty, the current parameter state of the surrogate model, and the remediation process time. This provides comprehensive input information for the trained decision-making agent, enabling the decision-making process to simultaneously perceive the current site pollution status, the cognitive state of the remediation mechanism, and the model's own confidence level, thereby making adaptive decisions that take into account pollution removal, cost control, and information acquisition.

[0136] In one embodiment, the closed-loop control module 104 can perform soil pollution remediation operations based on the remediation actions output by the intelligent decision module 103. The remediation actions can be sent to the site remediation equipment control system via control commands. This control system can adjust the rotation speed or valve opening of the injection pump according to the remediation agent injection rate command, adjust the operating frequency of the extraction pump according to the extraction rate command, and reconfigure the sampling cycle of each monitoring device according to the monitoring frequency adjustment command. After the remediation operation is executed, the closed-loop control module 104 can collect new monitoring data. For example, new monitoring data can be automatically acquired through a site-in-situ sensor network. This sensor network includes multi-parameter water quality probes, soil vapor phase monitors, and surface displacement monitors arranged in each monitoring well. Each sensor synchronously collects data according to a set sampling cycle and collects the data to a data acquisition terminal via wireless transmission. The data types collected are consistent with the aforementioned monitoring data types, including pollutant concentration data, isotope ratio data, functional gene abundance data, and geochemical parameter data. After data collection, the raw data can be subjected to quality control and preprocessing to remove invalid data caused by sensor failure or communication anomalies, and missing values ​​can be imputed to obtain new monitoring data that can be updated.

[0137] Subsequently, the closed-loop control module 104 can update the parameters and belief states of the physical information neural network surrogate model based on the new monitoring data. During the parameter update process, the new monitoring data can be added to the existing training dataset to form an expanded training dataset. Then, using the current model parameters as initial values, a finite number of gradient descent fine-tuning steps are performed on the expanded training dataset. During fine-tuning, the loss function remains the same as the weighted sum of the physical constraint loss and data fitting loss during initial training; only the size of the training sample data increases. The number of fine-tuning steps can be set according to the batch size of the new monitoring data and the model convergence speed, typically 50 to 200 steps. The learning rate is one-tenth to one-hundredth of the initial training learning rate to ensure that the model parameters are appropriately adjusted towards the new data while maintaining the original physical law memory. After fine-tuning, the values ​​of each element in the weight matrix and bias vector of the physical information neural network surrogate model are updated, resulting in the updated physical information neural network surrogate model. During the belief state update process, the new monitoring data can be added as new evidence to the total evidence set, and the aforementioned Bayesian inference calculation is re-executed. Specifically, when recalculating the likelihood function of each competing repair hypothesis, the new monitoring data is included in the cumulative term of the log-likelihood function, that is, the term corresponding to each monitoring point at the current time is added. Then, combined with the posterior probability distribution which is the prior probability after the previous iteration, the posterior probability of each hypothesis in the new round is calculated by Bayes' theorem, and then the belief entropy is updated to obtain the updated belief state.

[0138] Schematic, the closed-loop control module 104 can also determine whether the preset remediation termination conditions are met based on the updated belief state. These preset remediation termination conditions are a three-dimensional set of conditions. The first dimension is the pollutant concentration compliance condition, meaning that the concentrations of each pollutant component at all monitoring points on the site are lower than the corresponding remediation target concentration values. The remediation target concentration values ​​are preset based on the site's future land use function and relevant environmental quality standards. The second dimension is the attribution confidence condition, meaning that the posterior probability value of the biodegradation hypothesis in the updated belief state is greater than a preset first threshold, which can be 0.90. This condition ensures that the concentration decrease is clearly attributed to the biodegradation remediation mechanism rather than physical dilution or redistribution. The third dimension is the uncertainty sufficiently reduced condition, meaning that the belief entropy value in the updated belief state is less than a preset second threshold, which can be 0.10. This condition ensures that decision-makers have sufficient certainty in attributing the current remediation mechanism. When all three conditions are met simultaneously, the closed-loop control module 104 can determine that the preset remediation termination conditions are met, output a remediation compliance signal, and terminate the entire remediation control process.

[0139] If the preset repair termination condition is not met, the closed-loop control module 104 can use the posterior probability distribution in the updated belief state as the prior probabilities of each competing repair hypothesis used in the Bayesian inference in the next iteration, and return to the belief generation module to execute the next iteration. For example, when the posterior probability of the biodegradation hypothesis is 0.75, the physical dilution hypothesis is 0.15, the pollutant redistribution hypothesis is 0.07, and the monitoring error hypothesis is 0.03 in the updated belief state, at the beginning of the next iteration, the prior probabilities of each competing repair hypothesis can be set to 0.75, 0.15, 0.07, and 0.03, respectively, and then a new round of posterior probability updates can be performed by combining the new monitoring data collected at the next moment. This operation realizes the temporal connection of Bayesian inference cognition, so that the prior probabilities can continuously evolve with the progress of the repair process. The reasoning results of each iteration are based on the reasoning results of the previous round, avoiding repeated calculations and information waste caused by starting the reasoning from a zero-information prior each time. Through the closed-loop iterative process of decision-making, execution, monitoring, updating, and feedback described above, the prediction accuracy of the physical information neural network surrogate model can be continuously improved with the accumulation of monitoring data, while the belief state can gradually converge the attribution confidence of the remediation mechanism with the accumulation of evidence. The remediation operation actions can be adaptively optimized with the improvement of model accuracy and cognitive state, ultimately achieving refined closed-loop management of the entire process of soil pollution remediation.

[0140] Based on the same inventive concept, this application also provides a method for implementing the aforementioned digital twin soil pollution remediation whole-process management system. The solution provided by this method is similar to the implementation scheme described in the above system. Therefore, the specific limitations of one or more embodiments of the digital twin soil pollution remediation whole-process management method provided below can be found in the above-described limitations of the digital twin soil pollution remediation whole-process management system, and will not be repeated here.

[0141] In one exemplary embodiment, such as Figure 2 As shown, a digital twin-based method for the whole-process management and control of soil pollution remediation is provided, including:

[0142] S41. Establish a set of multi-physics coupled partial differential equations based on historical site survey data, construct a deep neural network based on the set of multi-physics coupled partial differential equations, construct a loss function based on historical site survey data, train the deep neural network based on the loss function to obtain the trained deep neural network, generate a training dataset based on the trained deep neural network, and train a Fourier Neural Operator surrogate model based on the training dataset to obtain a physical information neural network surrogate model.

[0143] S42. Obtain the prior probability of each competing repair hypothesis. Based on the physical information neural network surrogate model and real-time monitoring data, and combined with the prior probability of each competing repair hypothesis, calculate the posterior probability of each competing repair hypothesis using Bayes' theorem to generate the belief state.

[0144] S43. Construct the state space of the Markov decision process based on the belief state, construct an environment simulator of the Markov decision process based on the physical information neural network agent model, perform meta-reinforcement learning training in the environment simulator to obtain the decision agent, process the current state based on the decision agent, and output the repair action; wherein, the current state is constructed based on real-time monitoring data and belief state; the repair action includes instructions for adjusting the injection rate, extraction rate and monitoring frequency of the repair agent.

[0145] S44. Perform soil pollution remediation operations according to the remediation actions, collect new monitoring data after execution, update the parameters and belief states of the physical information neural network surrogate model according to the new monitoring data, and obtain the updated physical information neural network surrogate model and the updated belief states. Determine whether the preset remediation termination condition is met according to the updated belief states. If it is met, terminate the remediation. If it is not met, use the posterior probability distribution in the updated belief states as the prior probability of each competing remediation hypothesis in the next round S42, and return to S42 for the next round of iteration.

[0146] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the digital twin soil pollution remediation whole-process control system of this application. A multi-core processor is preferred to improve the system's parallel processing capability. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate large amounts of data and computational tasks.

[0147] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the digital twin soil pollution remediation whole-process control system of this application. The computer-readable storage medium may include: a read-only memory, a random access memory, a solid-state drive, or an optical disk, etc.

[0148] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A digital twin-based system for the entire process of soil pollution remediation, characterized in that, The system includes: The twin modeling module is used to execute S1, which includes: establishing a multi-physics coupled partial differential equation system based on historical site survey data; constructing a deep neural network based on the multi-physics coupled partial differential equation system; constructing a loss function based on the historical site survey data; training the deep neural network based on the loss function to obtain a trained deep neural network; generating a training dataset based on the trained deep neural network; and training a Fourier Neural Operator surrogate model based on the training dataset to obtain a physical information neural network surrogate model. The belief generation module is used to execute S2, which includes: obtaining the prior probability of each competing repair hypothesis; based on the physical information neural network surrogate model and real-time monitoring data, and combining the prior probability of each competing repair hypothesis, calculating the posterior probability of each competing repair hypothesis through Bayes' theorem, and generating a belief state. An intelligent decision-making module is used to execute S3, which includes: constructing a state space of a Markov decision process based on the belief state; constructing an environment simulator of the Markov decision process based on the physical information neural network agent model; performing meta-reinforcement learning training in the environment simulator to obtain a decision agent; and processing the current state based on the decision agent to output a repair action. The current state is constructed based on the real-time monitoring data and the belief state. The repair action includes instructions for adjusting the injection rate, extraction rate, and monitoring frequency of the repair agent. A closed-loop control module is used to execute S4, which includes: performing soil pollution remediation operations according to the remediation action; collecting new monitoring data after execution; updating the parameters of the physical information neural network surrogate model and the belief state according to the new monitoring data to obtain the updated physical information neural network surrogate model and the updated belief state; determining whether the preset remediation termination condition is met according to the updated belief state; if it is met, the remediation is terminated; if it is not met, the posterior probability distribution in the updated belief state is used as the prior probability of each of the competing remediation hypotheses in the next round of S2, and the process returns to S2 for the next iteration.

2. The system according to claim 1, characterized in that, The physical information neural network proxy model is obtained through the following steps: S11. Based on the geological parameters, initial pollution concentration distribution, and boundary conditions in the historical site survey data, establish the multiphysics coupled partial differential equation system; wherein, the multiphysics coupled partial differential equation system includes the convection-diffusion reaction equation, the multicomponent reaction kinetic equation, and the Richards equation; S12. Based on the multiphysics coupled partial differential equations, construct the deep neural network with spatial coordinates and time coordinates as inputs and the predicted concentration values ​​of each pollutant component as outputs. S13. Calculate the partial derivatives of the deep neural network output with respect to the spatial coordinates and the time coordinates using automatic differentiation; substitute the partial derivatives into the multiphysics coupled partial differential equations to calculate the equation residuals; use the equation residuals as physical constraint terms and the monitoring data from the site historical survey data as data fitting terms, and construct the loss function based on the physical constraint terms and the data fitting terms; train the deep neural network based on the loss function to obtain the trained deep neural network. S14. Using different preset combinations of remediation operation parameters as input, perform forward inference calculations based on the trained deep neural network to generate predicted concentration values ​​of each pollutant component corresponding to the remediation operation parameter combination; collect each combination of remediation operation parameters and the corresponding predicted concentration values ​​of each pollutant component to generate the training dataset; train a Fourier Neural Operator proxy model based on the training dataset to obtain the trained Fourier Neural Operator proxy model, and use the trained Fourier Neural Operator proxy model as the initial physical information neural network proxy model; S15. Perform multiple random initialization trainings on the initial physical information neural network proxy model to obtain multiple optimized Fourier Neural Operator proxy models. Calculate the prediction variance based on the output of each optimized Fourier Neural Operator proxy model, use the prediction variance as a measure of prediction uncertainty, construct an ensemble model based on each optimized Fourier Neural Operator proxy model, and use the ensemble model as the final physical information neural network proxy model.

3. The system according to claim 2, characterized in that, The belief state is obtained through the following steps: S21. Obtain a set of competitive remediation hypotheses, including a variety of such competitive remediation hypotheses; wherein the variety of such competitive remediation hypotheses includes the biodegradation hypothesis, the physical dilution hypothesis, the pollutant redistribution hypothesis, and the monitoring error hypothesis; S22. Based on the physical information neural network proxy model, set a corresponding kinetic parameter vector for each of the competitive remediation hypotheses; based on the kinetic parameter vector, generate the corresponding concentration prediction values ​​of each pollutant component under the competitive remediation hypotheses through the physical information neural network proxy model; S23. Based on the predicted concentration values ​​of each pollutant component, the prediction uncertainty measure, and the observed pollutant concentration values ​​in the real-time monitoring data, calculate the likelihood function of each competitive remediation hypothesis; take the logarithm of the likelihood function to calculate the log-likelihood function of each competitive remediation hypothesis; wherein, the variance term of the log-likelihood function is composed of the preset observation noise variance and the prediction uncertainty measure. S24. Based on the prior probability of each competitive repair hypothesis and the corresponding log-likelihood function, calculate the posterior probability of each competitive repair hypothesis using the Bayes theorem. S25. Calculate the belief entropy based on the posterior probability of each of the competing repair hypotheses; use the posterior probability of each of the competing repair hypotheses and the belief entropy together as the belief state.

4. The system according to claim 3, characterized in that, The system further includes a belief optimization module, which is used for: Identify the data types in the real-time monitoring data, including concentration observation data, isotope ratio data, functional gene abundance data, and geochemical parameter data; For each of the monitored data types, the evidence weight of the monitored data type is calculated based on the logarithm of the ratio of the likelihood functions of the monitored data type under each of the competing repair hypotheses. The evidence values ​​are sorted in descending order based on the absolute values ​​of the evidence weights of each monitored data type to generate an evidence value ranking list. Add the evidence value ranking list to the belief state to obtain the optimized belief state; The optimized belief state is used as the final belief state for operations S3 and S4.

5. The system according to claim 4, characterized in that, The intelligent decision-making module includes an agent training unit, which is used for: S31. Obtain a historical remediation case database and a reward function; wherein, the reward function includes a pollution removal amount reward item, an economic cost penalty item, a belief entropy penalty item, and an information value reward item; S32. Sample multiple repair scenario tasks from the historical repair case database; initialize the global policy parameters of the preset policy network to obtain the current policy; S33. For each repair scenario task, interact with the environment simulator under the current policy to collect trajectory data; calculate the cumulative reward based on the trajectory data and the reward function; calculate the policy gradient loss based on the cumulative reward; perform multi-step inner gradient update based on the policy gradient loss to obtain the adapted policy parameters; S34. Calculate the meta-loss based on the adapted policy parameters; perform outer gradient update and update the global policy parameters based on the meta-loss. S35. Repeat S32-S34 until the meta-loss reaches the preset meta-loss threshold, stop repeating the operation, take the last updated global policy parameter as the optimal global policy parameter, and construct the decision agent based on the policy network corresponding to the optimal global policy parameter.

6. The system according to claim 1, characterized in that, The current state is obtained through the following steps: Extract the observed concentration vector of each monitoring point at the current moment based on the real-time monitoring data, and calculate the concentration gradient vector based on the observed concentration vector; Obtain the repair process time; the repair process time is the time length between the current moment and the repair start moment. The current state is generated based on the observed concentration vector, the concentration gradient vector, the belief state, the parameters of the physical information neural network surrogate model, and the repair process time.

7. The system according to claim 3, characterized in that, The mathematical expression for the log-likelihood function of each of the aforementioned competitive repair hypotheses is as follows: in, In the first The aforementioned competitive repair hypothesis and the corresponding dynamic parameter vector The observed evidence set The log-likelihood function value is dimensionless; For a moment The accumulated set of monitoring evidence; For the index of the competing repair hypothesis; For the index of monitoring points, , The total number of monitoring points, dimensionless; For the index of time steps, , The total number of time steps is dimensionless. For the first The spatial coordinates of the monitoring points are in meters. For the first The moment of a time step is measured in seconds or days; The pre-defined observation noise variance is expressed in squares of concentration units. For the physical information neural network surrogate model at location ,time The measurement of prediction uncertainty at the location is in units of concentration squared; For the first Each monitoring point at time The observed pollutant concentration values ​​in the real-time monitoring data are expressed in units of concentration. For the in the first Competitive Repair Hypothesis and the dynamic parameter vector Below, the location predicted by the physical information neural network proxy model ,time The pollutant concentration value at the location is expressed in units of concentration. It is the natural logarithm with base 1. .

8. A digital twin-based method for the whole-process management and control of soil pollution remediation, characterized in that, The method includes: S41. Establish a set of multi-physics coupled partial differential equations based on historical site survey data, construct a deep neural network based on the set of multi-physics coupled partial differential equations, construct a loss function based on the historical site survey data, train the deep neural network based on the loss function to obtain the trained deep neural network, generate a training dataset based on the trained deep neural network, and train a Fourier Neural Operator proxy model based on the training dataset to obtain a physical information neural network proxy model. S42. Obtain the prior probability of each competing repair hypothesis. Based on the physical information neural network proxy model and real-time monitoring data, and combined with the prior probability of each competing repair hypothesis, calculate the posterior probability of each competing repair hypothesis using Bayes' theorem to generate a belief state. S43. Construct the state space of the Markov decision process based on the belief state, construct an environment simulator of the Markov decision process based on the physical information neural network agent model, perform meta-reinforcement learning training in the environment simulator to obtain a decision agent, process the current state based on the decision agent, and output a repair action; wherein, the current state is constructed based on the real-time monitoring data and the belief state; the repair action includes instructions for adjusting the injection rate, extraction rate, and monitoring frequency of the repair agent; S44. Perform soil pollution remediation operations according to the remediation action, collect new monitoring data after execution, update the parameters of the physical information neural network surrogate model and the belief state according to the new monitoring data, and obtain the updated physical information neural network surrogate model and the updated belief state. Determine whether the preset remediation termination condition is met according to the updated belief state. If it is met, terminate the remediation. If it is not met, use the posterior probability distribution in the updated belief state as the prior probability of each of the competing remediation hypotheses in the next round of S42, and return to S42 for the next round of iteration.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the system according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the system according to any one of claims 1 to 7.

Images