Sewage heavy metal recovery system and method based on generative artificial intelligence and edge intelligence

Abstract

The invention provides a sewage heavy metal recovery system and method based on generative artificial intelligence and edge intelligence, the system introduces heavy metal-containing sewage to be treated through a sensing and input layer and performs data acquisition, and performs real-time state prediction on the future state of the system through an intelligent decision and control layer to obtain future state data; generating a corresponding control strategy according to the future state data, and performing simulation verification on the control strategy by adopting a mode of combining physics with virtuality to generate a final control strategy; the at least two adsorption reactors are arranged on the core treatment and execution layer in a parallel connection and / or series connection mode, and the adsorption reactors are filled with the adsorption filler, so that the to-be-treated heavy metal-containing sewage is continuously adsorbed and regenerated, and the heavy metal in the sewage is recycled. Through cooperative work of the sensing layer, the decision-making layer and the execution layer, efficient, continuous and self-adaptive recovery of heavy metal ions in sewage is realized.

Description

Technical Field

[0001] This invention relates to the field of wastewater treatment, and in particular to a wastewater heavy metal recovery system and method based on generative artificial intelligence and edge intelligence. Background Technology

[0002] The presence of heavy metal ions such as lead (Pb) and cadmium (Cd) in industrial wastewater poses a serious threat to the ecological environment and human health. Therefore, developing efficient and economical heavy metal removal technologies is crucial. Among numerous water treatment technologies, adsorption is widely used due to its simplicity, cost-effectiveness, and renewability. The core of this method is to utilize porous materials with high specific surface area and abundant active sites to capture heavy metal ions in water through physical or chemical processes.

[0003] Metal-Organic Frameworks (MOFs) and Adsorption Principles: Metal-Organic Frameworks (MOFs) are a class of crystalline porous materials formed by the self-assembly of metal ions or clusters with organic ligands. Benefiting from their extremely high specific surface area (up to several thousand m² / g), tunable pore structure, and easily functionalized chemical properties, MOFs are widely recognized as ideal candidate materials for the selective adsorption of heavy metal ions. Their adsorption mechanisms are typically based on ion exchange, Lewis acid-base interactions, or coordination chelation with organic ligand functional groups, enabling the efficient capture of specific metal ions.

[0004] To achieve continuous and uninterrupted wastewater treatment, fixed adsorption reactor (ADR) systems are commonly used in industrial practice. To address the problem of interrupted treatment processes caused by the need to shut down and regenerate a single ADR once the adsorption packing reaches saturation, a mature and well-known technology is to operate dual or multiple ADRs in parallel. In this mode, at least two ADRs operate in parallel. While one ADR (e.g., ADR) is performing its adsorption task online, the other, saturated ADR (e.g., BDR) is regenerated offline. When ADR approaches saturation, the system uses a circulating pump and valves to switch the influent to the regenerated BDR, simultaneously regenerating ADR. This alternating cycle ensures the continuity of the overall effluent from the system.

[0005] Modern chemical processes widely employ automated control systems to improve operational efficiency and stability. In adsorption systems, the core of the control logic lies in accurately determining the switching timing of the adsorption reactor and optimizing the regeneration process. Traditional control methods are mostly based on switching at fixed time periods or with fixed water volumes, which cannot adapt to dynamic fluctuations in influent concentration and flow rate. Therefore, introducing more advanced control algorithms has become an industry consensus. Model Predictive Control (MPC) is a well-known advanced process control technology that uses a dynamic model of the process to predict future system behavior and calculates the optimal control operation (such as flow regulation) through online rolling optimization to minimize energy consumption or maximize treatment efficiency while meeting constraints.

[0006] With the development of Industry 4.0, digital twin technology has been introduced into the management of complex industrial processes. A digital twin refers to creating a high-fidelity, dynamic digital twin model of a physical system. This model is synchronized with the physical entity in real time and can be used for condition monitoring, fault diagnosis, and simulation verification of control strategies. Meanwhile, artificial intelligence (AI) and machine learning, especially deep reinforcement learning (DRL), a technology capable of autonomously learning optimal strategies through interaction with the environment, are also being explored for application in process control to cope with highly nonlinear and uncertain complex operating conditions and to seek globally optimal long-term operating strategies. Edge computing, on the other hand, is a technology that executes computational tasks on local devices close to the data source, effectively reducing decision latency and meeting the real-time requirements of industrial control.

[0007] Existing technical solutions can be divided into two categories: one is an automated adsorption system that uses traditional immobilized control logic, and the other is an intelligent adsorption system that initially introduces predictive models at the control level. These two types of solutions differ fundamentally from this invention in terms of material sourcing, the adaptability of control logic, decision security, and system integration.

[0008] ① An automated parallel adsorption system based on a fixed threshold or period; This is currently the most common continuous water treatment technology solution in industrial applications.

[0009] Method and Procedure: First, based on experience or literature, a commercially available, broad-spectrum adsorbent (such as activated carbon, zeolite, or MOF with a known structure) is manually selected and loaded into two or more parallel adsorption reactors. Then, the wastewater containing heavy metals to be treated is fed into one adsorption reactor (adsorption reactor A), while the other adsorption reactor (adsorption reactor B) remains in standby mode. During operation, the concentration of heavy metal ions is monitored by an online sensor at the outlet, or the treated water volume is accumulated via a flow meter. When the outlet concentration reaches a preset, fixed alarm threshold (i.e., the breakthrough point), or when the system operation time reaches a fixed cycle, or when the treated water volume reaches a fixed total volume, the system automatically triggers a switchover. At this time, the controller switches the influent flow to the standby adsorption reactor B by controlling the circulating pump and valves, and simultaneously connects the saturated adsorption reactor A to the regeneration unit to initiate a standardized regeneration procedure (e.g., pumping in a fixed concentration of eluent at a fixed flow rate for a fixed duration). Finally, while adsorption reactor B performs adsorption, the regenerated adsorption reactor A returns to standby mode, awaiting the next switchover command, and this cycle repeats continuously.

[0010] Structural composition and its relationships: The system mainly consists of: parallel adsorption reactors (usually two units), filled with general-purpose adsorption packing material; online sensors (concentration, flow, and pressure sensors) for data acquisition; a central controller, typically a programmable logic controller (PLC), used to receive sensor signals and control the circulating pumps and valves to perform switching and regeneration operations according to preset fixed logic (IF-THEN rules); and a regeneration unit. The core of the entire system's control is the fixed program in the PLC, lacking dynamic prediction and optimization capabilities.

[0011] Limitations: The core flaw of this scheme lies in the rigidity of its control logic. It cannot adapt to dynamic changes in influent water quality (such as concentration and composition) and operating conditions (such as flow rate). If the influent concentration suddenly increases, switching at a fixed cycle may result in effluent exceeding standards; if the concentration decreases, the adsorption reactor may be switched to regenerate before it is saturated, resulting in a waste of adsorption capacity and regeneration reagents. This scheme is essentially a "passive response" control rather than an "active predictive" optimization.

[0012] ② Intelligent adsorption control system based on data-driven model prediction; This is a more advanced solution that attempts to introduce artificial intelligence to optimize processes based on traditional automation.

[0013] Method and Procedure: This method first employs a general-purpose adsorption packing material. During the initial system operation or laboratory phase, a large amount of operational data (including inlet and outlet concentrations, flow rate, pressure, temperature, etc.) is collected. This historical data is used to train an offline predictive model (such as an Artificial Neural Network (ANN) or Support Vector Machine (SVM) to predict the breakthrough curve or remaining adsorption lifetime of the adsorption reactor. During actual system operation, real-time sensor data is input into the deployed predictive model for online prediction. The model outputs a predicted value for the effluent concentration within a specific future time period. The controller then calculates the theoretically optimal switching time based on this prediction to maximize the single-cycle adsorption capacity. Once the predicted effluent concentration is about to exceed the limit, the controller issues a command to execute the adsorption reactor switching and regeneration operations.

[0014] Structural composition and relationships: The device structure is similar to scheme ①, but its controller is usually upgraded to an industrial computer (IPC) or an edge device with stronger computing power, capable of supporting and running pre-trained AI prediction models. The system's decision-making basis is upgraded from "whether the current value exceeds the standard" to "whether the predicted value is about to exceed the standard".

[0015] Limitations: The fundamental flaws of this solution are reflected in three aspects: First, the materials and control are separated. Its intelligent control optimizes a fixed-performance, non-customized adsorption material, failing to achieve full-chain optimization from the source (material design) to the process (intelligent control). Second, there is a lack of decision safety verification. The control strategy output by the AI ​​model is directly executed. If the model's predictions become inaccurate due to operating condition drift, it may lead to incorrect decisions and operational risks. It lacks the crucial safety closed loop of "digital twin pre-simulation and risk assessment" in this invention. Third, the human-computer interaction and system iteration capabilities are insufficient. The AI ​​model of this type of system is usually a "black box," making it difficult for maintenance personnel to understand the decision-making basis. Furthermore, the system lacks the ability to learn and continuously optimize online. Once the model's performance deteriorates, costly manual retraining and redeployment are required.

[0016] (3) Similar patent documents The following patents or published documents represent technical solutions similar to those in this invention, mainly focusing on the artificial intelligence design of adsorption materials, advanced control of wastewater treatment processes, and applications of digital twins: A Multi-Objective Control Method for Wastewater Treatment Processes Based on Nonlinear Model Prediction (CN103197544A): This patent discloses a method for controlling dissolved oxygen and nitrate nitrogen concentrations in wastewater treatment processes. Its core is establishing a predictive model of the wastewater treatment process and utilizing nonlinear model predictive control (MPC) for multi-objective optimization. It represents a technological direction for applying advanced process control (MPC) to the field of wastewater treatment to address the poor adaptive capability of traditional PID control.

[0017] A Smart Control Method for Wastewater Treatment (Publication No. CN120097501A): This patent application discloses a smart control method for wastewater treatment, characterized by the use of a deep reinforcement learning (DRL) model. This model enables the system to autonomously learn the optimal control strategy through interaction with the environment and trial and error, thereby adapting to changes in operating conditions in real time. This document represents a cutting-edge exploration of applying deep reinforcement learning to wastewater treatment process control.

[0018] A water resources information management method and system based on digital twin technology (Publication No. CN118822307A): This patent application relates to a water resources management system based on digital twin technology, which realizes the monitoring, simulation and management of water resources systems by constructing a real-time mapping between the physical world and a digital twin model. It represents the application of digital twin technology in the broad fields of water treatment and water resources management.

[0019] A Drug Molecule Generator Training Method Based on Domain Knowledge and Deep Reinforcement Learning (CN113223637A): This patent discloses a method for generating novel drug molecules with specific objectives (such as pharmacodynamics) using deep learning and reinforcement learning. Although the application fields are different (drug design vs. materials design), its core idea—using generative artificial intelligence combined with reinforcement learning to create novel molecular structures with specific functions—represents the application direction of AI in the field of molecular / material "reverse design".

[0020] As can be seen from the above, existing heavy metal wastewater adsorption treatment systems can be mainly divided into automated systems using fixed control logic and intelligent systems that initially incorporate data model prediction. The core drawback of the former lies in its inherently "universal materials" and "rigid control." These systems typically use commercially available broad-spectrum adsorption packing materials, whose adsorption performance is not optimized for specific wastewater components, thus limiting treatment efficiency and selectivity from the source. Simultaneously, their control logic is mostly based on fixed time periods or simple concentration thresholds for adsorption / regeneration switching. This rigid control method cannot adapt to the real-time dynamic fluctuations in the concentration and flow rate of industrial wastewater. The consequence is that under low-load conditions, the adsorption reactor is switched before it is saturated, resulting in a serious waste of adsorption capacity and regeneration reagents, increasing operating costs; while under high-load shocks, untimely switching may lead to effluent exceeding standards, posing environmental risks.

[0021] On the other hand, while intelligent systems employing data model prediction introduce predictive capabilities, several deep-seated technical bottlenecks remain. First, these systems still suffer from a disconnect between materials and processes. Their intelligent algorithms optimize a fixed-performance, non-customized adsorbent material, resulting in a low ceiling for overall system performance. Second, and most critically, these systems generally lack secure verification mechanisms for AI decisions. The control strategies output by their predictive models are directly implemented; if operational conditions drift and model predictions become inaccurate, catastrophic erroneous decisions may occur, leading to high system operational risk and poor robustness. Furthermore, the AI ​​models in these systems are often "black boxes," making it difficult for maintenance personnel to understand their decision-making rationale, resulting in difficulties in human-machine collaboration. Moreover, the models are typically trained offline, lacking the ability for continuous online learning and self-optimization, and are unable to cope with performance degradation caused by equipment aging or environmental changes.

[0022] In summary, existing technologies generally suffer from several technical deficiencies, including a disconnect between materials and processes, rigid or unsafe control strategies, weak system adaptability, opaque decision-making processes, and a lack of continuous self-optimization mechanisms. These problems collectively result in existing heavy metal adsorption systems failing to meet the urgent demands of modern industry for efficient, low-consumption, and stable operation in terms of processing efficiency, operating costs, reliability, and intelligence. Summary of the Invention

[0023] In view of the above, the main objective of this invention is to propose a wastewater heavy metal recovery system and method based on generative artificial intelligence and edge intelligence to solve the aforementioned technical problems.

[0024] This invention proposes a wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence, the system comprising: The perception and input layer is used for: The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. The intelligent decision-making and control layer is used for: Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; The system support layer is used for: It provides communication and power, and switches to a preset PID control strategy when a failure occurs in the intelligent decision-making and control layer; The core processing and execution layer includes at least two adsorption reactors connected in parallel and / or series. The adsorption reactors are filled with highly selective metal-organic framework adsorption materials designed and screened using generative artificial intelligence based on the target heavy metal ions. This is used for: Based on the final control strategy or the preset PID control strategy, the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor are controlled to continuously adsorb and regenerate the wastewater containing heavy metals, so as to achieve the recovery of heavy metals from the wastewater.

[0025] This invention also proposes a method for heavy metal recovery from wastewater based on generative artificial intelligence and edge intelligence. The method is used to implement the aforementioned wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence, and includes the following steps: At least two adsorption reactors are set up in parallel and / or in series, and the adsorption reactors are filled with adsorption packing material of highly selective metal-organic framework adsorption material designed and screened by generative artificial intelligence based on the target heavy metal ions. The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; A preset PID control strategy is used, and the final control strategy or the preset PID control strategy is selected according to the operating status to control the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor, so as to continuously adsorb and regenerate the wastewater containing heavy metals to be treated, thereby realizing the recovery of heavy metals from wastewater.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention utilizes generative artificial intelligence to directly target heavy metal ions (Pb). 2+ Cd 2+ This method generates MOF candidate structures and uses graph attention networks to quickly predict their isothermal adsorption parameters and renewability scores, forming a closed loop of material design → prediction → operating condition adaptation. This mechanism directly links the design of adsorbent materials with the characteristics of wastewater influent, creating an innovative approach that links customized materials with processes, unlike traditional methods that rely on fixed material libraries or manual screening.

[0027] 2. This invention employs an A / B parallel adsorption reactor structure, achieving alternating cycles of adsorption and regeneration in a single adsorption reactor under the scheduling of an edge control unit. The control logic is based on the collaborative optimization of MPC (Model Predictive Control) and DRL (Deep Reinforcement Learning), dynamically calculating the switching timing and regeneration mode of the adsorption reactors to ensure continuous compliance with effluent standards, optimal energy consumption, and minimum regeneration cost. This control method breaks through the traditional periodic switching operation mode of adsorption systems, realizing intelligent adaptive operation under high-frequency disturbance conditions.

[0028] 3. This invention deploys a hybrid digital twin model (mechanism model and residual data-driven model) at the edge and in the cloud, enabling parallel operation of the physical system and the virtual twin. All control strategies undergo 24-hour simulation verification in the digital twin environment before execution, calculating the risk index Rrisk. Strategies that do not meet the safety threshold are intercepted, ensuring the safety and robustness of the control strategies. This pre-simulation mechanism significantly reduces the operational risks caused by control strategy errors.

[0029] 4. This invention introduces XAI (using methods such as SHAP and LIME) into the control layer to provide real-time interpretation and visualization of the key feature contributions of each MPC or DRL decision. Through the XAI decision support terminal, operations and maintenance personnel can clearly understand the basis, weight ranking, and potential impact of control decisions, and can directly intervene and verify strategies, forming a model that integrates AI control and human decision-making.

[0030] 5. This invention integrates a safety backoff controller into the control system. When an edge inference link timeout, communication interruption, or sensor malfunction is detected, the controller automatically takes over the system and switches to a PID control strategy to maintain stable flow and safe effluent. This mechanism achieves controllable redundancy throughout the entire process through safety flow limiting and PID backoff, ensuring that the entire wastewater treatment process can still operate stably even if the AI ​​system fails.

[0031] 6. This invention constructs a closed-loop self-healing mechanism between anomaly detection (based on hard threshold + residual monitoring), rapid rollback control, historical data accumulation, and continuous model optimization (GNN bias correction + MPC weight fine-tuning + DRL incremental training). This mechanism enables the system to have adaptive and performance evolution capabilities during long-term operation without relying on frequent external manual intervention.

[0032] 7. This invention constructs a complete "detection-rollback-optimization" closed-loop self-healing system. A safe rollback controller ensures basic safe operation when the AI ​​link fails, and by accumulating all operational and anomaly data, it triggers continuous online optimization of the model, enabling the system to possess adaptive and performance evolution capabilities under long-term operation. In summary, this invention has significant technical advantages in achieving truly "efficient, intelligent, safe, and economical" heavy metal recycling.

[0033] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence proposed in this invention. Figure 2 This is a schematic diagram of the sensing and entry layer in the wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence proposed in this invention. Figure 3 This is a schematic diagram of the intelligent decision-making and control layer in the wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence proposed in this invention. Figure 4 This is a flowchart of the wastewater heavy metal recovery method based on generative artificial intelligence and edge intelligence proposed in this invention. Detailed Implementation

[0035] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0036] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0037] Please see Figure 1 This embodiment provides a wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence. The system includes: The perception and input layer is used for: The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. Please see Figure 2 In a preferred embodiment of the present invention, the sensing and input layer includes: The raw water inlet is used to introduce wastewater containing heavy metals to be treated; Pretreatment units, including screens, flocculation or grit chambers, are used for preliminary filtration to remove large particulate suspended solids from the raw water and reduce the risk of clogging in subsequent adsorption reactors. An online heavy metal analysis sensor is installed at the outlet and inlet of the adsorption reactor to monitor the concentration of target heavy metal ions such as lead (Pb) and cadmium (Cd) in the influent and effluent in real time. The sampling frequency is not less than 0.2Hz to ensure that it can capture sudden changes in the concentration of influent. Flow meters and pressure sensors are used to collect real-time data on liquid flow rate in the pipeline and pressure difference in the adsorption reactor, in order to monitor the system's operating status and determine whether the adsorption reactor is blocked. The adaptive sensor calibration module is used to monitor the drift of online heavy metal analysis sensors, as well as flow meters and pressure sensors, and make real-time adjustments to ensure sensor accuracy and data accuracy, thereby improving the system's adaptability to environmental changes. The multimodal anomaly detection and feature fusion unit, located between the edge computing unit and the digital twin server, is used to fuse data from online heavy metal analysis sensors, flow meters and pressure sensors, actuators and control logs using a temporal convolutional network (TCN) and a multi-head attention mechanism to achieve feature alignment and fusion, generating fused features. Based on the fused features, anomaly detection is performed to generate preliminary anomaly detection results, providing the system with high-confidence early warning input.

[0038] The intelligent decision-making and control layer is used for: Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; Please see Figure 3 In a preferred embodiment of the present invention, the intelligent decision-making and control layer includes: The edge computing unit, as the local intelligent core, has dedicated computing hardware such as TPU / NPU, carries a lightweight artificial intelligence model, performs real-time state prediction and inference, and sends control strategies to the main controller, with inference latency controlled within 200 milliseconds. The edge computing unit embeds a lightweight self-learning module, which is used to perform model drift detection and parameter fine-tuning on the lightweight artificial intelligence model based on real-time data. When a long-term deviation of the prediction residual is detected (>10% threshold), the system automatically triggers local retraining and updates the digital twin model parameters simultaneously to achieve distributed continuous learning. The main controller integrates a model predictive control module (MPC) and a deep reinforcement learning module (DRL). The model predictive control module receives instructions from the edge computing unit and, based on the future predicted state, calculates the optimal flow rate and regeneration switching time with the optimization objectives of minimizing the deviation from the target effluent concentration, minimizing energy consumption, and minimizing the number of regenerations, in order to obtain the first strategy. The optimal regeneration method is calculated with the optimization objectives of minimizing regeneration energy consumption, maximizing regeneration rate, and maximizing adsorption recovery rate to obtain the second strategy; the first strategy and the second strategy constitute the control strategy. The deep reinforcement learning module is used to learn the long-term optimal control policy by using reinforcement learning, guided by the control policy, and thus obtain the final control policy. The digital twin server communicates with the edge computing unit in real time to build and run a digital twin model synchronized with the physical system. This model is used for "hypothesis-simulation-verification" of control strategies, optimization of system parameters, and prediction of long-term operational risks. The XAI decision support terminal provides a graphical human-computer interaction interface. It uses explainable artificial intelligence technology to explain the control decisions made based on the final control strategy or the preset PID control strategy, and displays the contribution of key parameters to achieve visual output, providing operation and maintenance personnel with an intervention entry point and decision support. The explainable artificial intelligence technology includes, but is not limited to, one or a combination of the following methods: SHAP: Based on cooperative game theory, it assigns a Shapley value to each input feature (such as real-time concentration, flow rate, and pressure difference), quantifies the contribution of the feature to the current control decision (such as switching adsorption reactors and selecting regeneration mode), and visualizes and sorts it in the form of force-directed graphs or bar charts.

[0039] LIME: By building local alternative models near specific decision points, it generates easy-to-understand "if-then" rules that explain the main influencing factors of decision-making under specific conditions.

[0040] Integrated Gradients: Particularly suitable for deep neural networks, it calculates the gradient from the baseline to the current input through path integral, accurately attributing it to each input variable, and is especially good at handling continuous sensor data.

[0041] Attention visualization: If the underlying AI model contains an attention mechanism, the learning weights of that mechanism are directly visualized using a heatmap to reveal the historical time points or data modalities that the model focuses on when making decisions.

[0042] Through the single or combined application of the above methods, the terminal can transform complex AI decision-making processes into trustworthy, verifiable, and controllable transparent information.

[0043] The digital twin self-testing unit is used to monitor the differences between the physical system and the digital twin model in real time, and to adjust or reconstruct the digital twin model to ensure the coordinated operation of hardware and software, detect potential fault risks in advance, and ensure the stability and reliability of the system. The risk decision and fault tolerance engine is located between the main controller and the digital twin server, operating across layers. This engine has built-in Monte Carlo disturbance analysis, metastable risk estimation and fault tolerance logic modules. It calculates risk indicators for each control strategy and compares them with preset risk thresholds. When a high risk or abnormal trend is detected, the safety backoff controller is triggered to enter the safety mode and execute the preset PID control strategy to achieve self-auditing and autonomous fault tolerance of the strategy. The graph-based intelligent root cause analysis module, located between the digital twin server and the main controller, is used to construct a system dependency graph based on the preliminary anomaly detection results, using sensors, actuators, and control logic units as nodes and temporal dependencies between nodes as edges. By constructing the system dependency graph and graph attention network, it analyzes the anomaly propagation path, outputs the root cause nodes and their impact range, and realizes anomaly explanation and decision support.

[0044] The decision and control layer also includes an online adaptive learning module, which continuously collects real-time data during system operation and uses online learning algorithms to automatically update and optimize the Model Predictive Control (MPC) module and the Deep Reinforcement Learning (DRL) module. At the same time, it works with the digital twin server to perform safety verification of new strategies, so as to achieve continuous adaptation of control strategies and autonomous evolution of long-term performance, ensuring that the system always maintains the optimal operating state when facing changes in operating conditions or equipment aging. The digital twin server runs a hybrid model, which includes a mechanism model based on adsorption kinetics and hydraulics, as well as a residual data-driven model that compensates for prediction bias. The digital twin server is configured to perform parallel simulations for a predetermined duration based on the control policy before the control policy is issued and executed, in order to generate simulation results.

[0045] The digital twin server also incorporates a multi-fidelity simulation coordination unit, which automatically selects the simulation accuracy based on real-time load and operating conditions through a model switching function. This enables dynamic coordination between high-fidelity CFD mechanism models and low-fidelity AI approximation models, and performs parameter recalibration. Its output serves as the basis for multi-scale simulations for strategy verification and risk assessment. The digital twin server can be deployed in an edge-cloud hybrid, pure cloud, or pure edge manner.

[0046] The system support layer is used for: It provides communication and power, and switches to a preset PID control strategy when a failure occurs in the intelligent decision-making and control layer; The core processing and execution layer includes at least two adsorption reactors connected in parallel and / or series. The adsorption reactors are filled with highly selective metal-organic framework adsorption materials designed and screened using generative artificial intelligence based on the target heavy metal ions. This is used for: Based on the final control strategy or the preset PID control strategy, the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor are controlled to continuously adsorb and regenerate the wastewater containing heavy metals, so as to achieve the recovery of heavy metals from the wastewater.

[0047] In a preferred embodiment of the present invention, the system support layer includes: The power management unit is used to provide stable power to all electrical components; Data bus and communication gateway are used to connect all sensing, decision-making and execution components using industry standard protocols such as OPC UA, MQTT, Modbus, LoRa or NB-IoT to build a stable, low-latency system intranet; The safety backoff controller, as a safety redundancy design of the system, is used to switch to a preset PID control strategy when an edge computing unit inference timeout, communication link interruption, or main controller failure is detected, in order to maintain the basic operation of the system, ensure that the effluent water quality always meets safety standards, and guarantee the basic operational capability of the system under extreme conditions.

[0048] In a preferred embodiment of the present invention, the core processing and execution layer further includes: The adsorbent regeneration unit, connected to the parallel adsorption reactor, regenerates the saturated adsorption packing material after receiving the control strategy to restore its adsorption capacity. The adsorbent regeneration unit supports at least one of the following regeneration modes: electrochemical regeneration, chemical elution regeneration, low-temperature thermal regeneration, vacuum desorption regeneration, ultrasonic regeneration, and microwave regeneration. The selection of the regeneration mode is intelligently decided by the final control strategy or a preset PID control strategy. The circulating pumps and valves, controlled by the main controller, are used to switch between adsorption reactors, reflux, and regulate the flow rate within the system. The AI-GenMOF design unit, based on a generative artificial intelligence model, generates MOF candidate structures according to the coordination characteristics of the target heavy metal ions. Reinforcement learning is then used to optimize the specific surface area, pore size distribution, and charge density distribution of the MOF candidate structures. Finally, density functional theory (DFT) calculations are used to screen and obtain the optimal MOF structure. The predictive maintenance and fault-tolerant coordination unit, located between the main controller and the safety rollback controller, is used to monitor equipment degradation trends, generate proactive protection strategies through the predictive maintenance model, and actively coordinate the rollback controller to enter steady-state mode when the risk exceeds the threshold, thereby achieving proactive fault tolerance and continuous operation assurance.

[0049] This embodiment also provides a wastewater heavy metal recovery method based on generative artificial intelligence and edge intelligence, wherein the method is used to implement the above-mentioned wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence, and the method includes the following steps: At least two adsorption reactors are set up in parallel and / or in series, and the adsorption reactors are filled with adsorption packing material of highly selective metal-organic framework adsorption material designed and screened by generative artificial intelligence based on the target heavy metal ions. The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; A preset PID control strategy is used, and the final control strategy or the preset PID control strategy is selected according to the operating status to control the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor, so as to continuously adsorb and regenerate the wastewater containing heavy metals to be treated, thereby realizing the recovery of heavy metals from wastewater.

[0050] Please refer to Figure 4 The workflow of this invention is as follows: S10, System power-on initialization and self-test, specifically including S101 to S106.

[0051] This step is completed collaboratively by the raw water inlet, online heavy metal analysis sensor, flow meter and pressure sensor, edge computing unit, main controller, digital twin server, XAI decision support terminal, power management unit, data bus and communication gateway, safety rollback controller, digital twin self-test unit, and adaptive sensor calibration module. Its purpose is to enable the system of this invention to perform environmental and hardware self-tests, software image integrity verification, communication link and time synchronization, key sensor calibration, electrical circuit continuity testing, and operating clock alignment after power-on. This process specifically introduces a simulation self-test performed by the digital twin self-test unit, and realizes intelligent calibration and adaptive self-test processes through the adaptive sensor calibration module, laying a stable and low-error operational foundation for the subsequent adsorption-control-regeneration process.

[0052] S101. Software image integrity verification and digital twin pre-inspection, the specific steps are as follows: After the system is powered on, the main controller loads the core control firmware file. The hash digest is then calculated using the SHA-256 algorithm, and the corresponding process relationship is as follows: ; in, This refers to the core control firmware file. Represents a hash digest. This represents the SHA-256 algorithm; The data bus and communication gateway synchronously read the cloud signature digest and compare it with the hash digest.

[0053] like: ; in, This represents the cloud-based signature digest; If the firmware integrity is deemed satisfactory, the safety rollback controller will not intervene, and the system will continue to execute. Meanwhile, before the physical system starts, the digital twin self-test unit, in conjunction with the digital twin server, performs an initialization check. By simulating and verifying the logical states of sensors, circulating pumps, valves, and flow paths, potential fault risks can be detected in advance, ensuring the stability and reliability of the system.

[0054] If firmware verification fails, the safety rollback controller takes over, the system switches to PID control mode and displays "Firmware Abnormal Warning" on the XAI decision support terminal, and records the event log "IntegrityFlag=0" in the digital twin server.

[0055] This process ensures the source security and reliable execution of MPC / DRL (Model Predictive Control / Deep Reinforcement Learning) models, preventing malicious code from causing operational anomalies.

[0056] Model health (MHI) is calculated by verifying metrics such as the integrity of the model weight file, parameter drift, and historical validation accuracy. When MHI < 0.8, local retraining or safe mode is triggered.

[0057] S102. Power circuit and continuity test, the specific steps are as follows: The power management unit performs power-on continuity testing on each power supply branch, with the test current set as follows: ; in, Indicates the test current. The unit of electric current is the ampere.

[0058] Measured voltage drop The corrected on-resistance is calculated as follows: ; in, This indicates a correction for the on-resistance. Indicates voltage drop. Indicates the measurement line resistance. , The unit of resistance is ohm; like: ; in, Indicates the maximum permissible on-resistance threshold; If the branch circuit is deemed to be functioning correctly, it is determined that the circuit is in good condition. If the circuit exceeds the limit, the safety backoff controller enters degraded control, shutting down the regeneration unit and maintaining a safe flow state in the adsorption reactor. Log recording is performed by a digital twin server for subsequent safety traceability and status compensation.

[0059] S103. Communication link handshake and time synchronization, the specific steps are as follows: The system establishes a secure, encrypted communication channel (OPC UA or MQTT QoS=1) with the digital twin server via a data bus and communication gateway, and performs standard time synchronization. Let the request and response timestamps be... Then the network round-trip delay and clock skew are respectively: ; in, Indicates network round-trip delay. Indicates clock skew; The system filters out the set of servers that meet the requirements: ; in, This indicates the maximum allowed round-trip time threshold. The unit of measurement for time is milliseconds; Choose the server with the smallest deviation from the set of servers that meet the requirements. .

[0060] like: ; in, This represents the clock offset of the slave device relative to the master device at the k-th measurement. This represents the clock drift rate or clock skew of the slave device relative to the master device during the k-th measurement period. This indicates the maximum permissible synchronization error threshold. The unit of time is microsecond; If synchronization is successful, then synchronization is successful; otherwise, the safety backoff controller triggers degraded operation, and the system enters protection mode based on the local clock reference. After successful synchronization, the edge computing unit, main controller, online heavy metal analysis sensor, flow meter, and pressure sensor are unified with a unified timestamp reference to ensure that the adsorption reactor operation and predictive control timing are aligned.

[0061] S104. Sensor baseline calibration and zero-point drift detection, the specific steps are as follows: Online heavy metal sensors and flow / pressure sensors perform zero-point drift and linearity verification during the system initialization phase.

[0062] ① Adaptive self-test process The edge computing unit uses AI algorithms to detect anomalies in the initial collected data (such as based on statistical thresholds or initial pattern recognition). This algorithm can identify data inconsistencies and automatically adjust the focus of the self-inspection process for different operating conditions (such as dynamically selecting the correction order), thereby improving the intelligence level of the startup phase.

[0063] ② Intelligent sensor calibration and zero-point drift detection: Record 60 seconds of zero signal output and calculate the average value of the zero signal output. with standard deviation .like ,and ,in, A dimensionless unit for expressing concentration. The maximum static error between the sensor measurement and the true value that the system can tolerate indicates that the sensor is in good condition. If drift is detected, the adaptive sensor calibration module (18) will automatically detect and adjust in real time to ensure sensor accuracy and data accuracy. If it cannot be corrected, the safety backoff controller 16 will take over and record the fault type.

[0064] ③ Linearity verification Standard Pb / Cd concentration solutions (low, medium, and high) were injected into the online heavy metal analysis sensor, and the response curves were fitted. Calculate the goodness of fit ,like The linearity is then acceptable. The linearity error of the flow meter and pressure sensor is considered. Pressure sensor error The calibration parameters are cached in the edge computing unit for MPC initial prediction and DRL state normalization.

[0065] S105. Communication delay and data jitter detection, the specific steps are as follows: If the data acquisition trigger time and sampling completion time are obtained, the average delay is: ; in, Indicates the average delay. Indicates the time of data acquisition triggering. Indicates the sampling completion time. This represents the total number of sample data packets used to calculate the average latency. j Indicates the summation index; The jitter is: ; in, Indicates shaking; like If the timing acquisition link is qualified, then an alarm will be automatically triggered and the system will enter safe flow-limiting operation.

[0066] S106. Status registration and initialization are complete. The specific steps are as follows: After all S101–S105 passes, the main controller sets the status flags: ; in, Indicates a system-level global readiness status flag; The digital twin server records the device initialization parameters, including: each branch. Resistance value; Pb / Cd sensor zero point and linear fitting parameters; time synchronization delay and jitter; firmware hash verification result; current AI model version number and loading timestamp.

[0067] The XAI decision support terminal automatically generates an initialization log report, indicating whether each status is qualified. If all are normal, the system enters S20 (adsorbent material initialization and control strategy loading); if some are unqualified, the safety backoff controller takes over and enters a restricted safety mode to await manual handling.

[0068] Add an "Environmental Status Registration" sub-step to S106 to record environmental parameters such as temperature, humidity, and conductivity, and write them into the initialization database of the digital twin server for subsequent model bias correction and environmental compensation.

[0069] S20, Preparation and initialization of adsorption materials, specifically including S201 to S205.

[0070] In this embodiment, two adsorption reactors are configured, namely adsorption reactor A and adsorption reactor B, which are connected in parallel.

[0071] This step is completed collaboratively by adsorption reactor A, adsorption reactor B, adsorption packing material, online heavy metal analysis sensor, flow meter and pressure sensor, edge computing unit, main controller, digital twin server, circulation pump and valves, and AI-GenMOF design unit. Its purpose is to first utilize the AI-GenMOF design unit to complete the generative AI design and multi-objective optimization of the adsorption material, and then to achieve precise filling, structural state confirmation, and pre-commissioning adsorption capacity parameter calibration of the material, providing a material basis for subsequent continuous operation.

[0072] S201. Detection of the packing status of the adsorption reactor, the specific steps are as follows: This step assumes that the AI-GenMOF design unit has already completed the design of the adsorption packing material. This design employs generative artificial intelligence to automatically mine MOF structures, generating candidate ligands and nodes through deep learning to achieve adsorption targeting of Pb. 2+ / Cd 2+ The MOF intelligent design incorporates lattice topological constraints and synthesizability evaluation to ensure the effectiveness and fabrication of the new material. Subsequently, the designed adsorption packing material is precisely packed into parallel adsorption reactors A and B. The main controller initiates a self-test process, triggering flow meters and pressure sensors to monitor the static pressure difference.

[0073] Assume the static pressure drop of the liquid after the adsorption reactor is filled is The pressure drop during clear water flow is Judgment threshold: ; in, The pressure unit is kilopascal (kPa). If the pressure drop exceeds the threshold, the system determines that the filling is too dense or blocked, records the abnormality, and prohibits further steps.

[0074] The loading mass m and the adsorption reactor volume V are recorded in the main controller, and the edge computing unit is synchronized with the digital twin server.

[0075] S202. Material parameter calibration and isothermal adsorption curve loading: The specific steps are as follows: Before commissioning, AI-GenMOF adsorbents require the loading of corresponding isothermal adsorption curves and kinetic parameters. These parameters are derived from the multi-objective optimization results of the AI-GenMOF design unit during the design process. This optimization process combines generative models and reinforcement learning, considering indicators such as adsorption capacity, regeneration efficiency, and stability. A reward mechanism guides the generative process, making it more likely to produce MOF configurations that perform well in simulation screening.

[0076] Let the maximum adsorption capacity (mg / g) in the parameter package corresponding to the MOF configuration be... The Langmuir isotherm constant (L / mg) is... Mass transfer rate constant (s) -1 )for ; The main controller reads the parameter package of the corresponding batch of materials through the edge computing unit and loads it into the digital twin model of the digital twin server. In this embodiment, the digital twin model adopts a multi-layered coupled modeling approach, specifically a three-layer model. The bottom layer is a fluid dynamics model based on the Navier-Stokes equations, the middle layer is an isothermal adsorption model of MOFs, and the top layer is a residual data-driven model. These three layers interact through serial coupling and residual compensation. The specific coordination logic is as follows: Bottom layer (fluid dynamics model based on Navier-Stokes equations): Based on the Navier-Stokes equations or Darcy's law, the flow field distribution and velocity within the adsorption bed are calculated and used as boundary conditions for the middle layer.

[0077] Middle layer (Isothermal adsorption model and mass transfer model of MOF): The isothermal adsorption model of MOF receives the flow rate data of the bottom layer and calculates the equilibrium adsorption amount on the outer surface of the particles by combining the Langmuir equation; the mass transfer model calculates the theoretical mass transfer and adsorption reaction rate based on the equilibrium adsorption amount on the outer surface of the particles and the average adsorption amount of the particles, and obtains a "theoretical prediction value".

[0078] Upper layer (residual data driven model): Receives the "theoretical prediction value" from the middle layer and compares it with actual sensor data. This layer (usually LSTM or TCN) corrects and compensates for the nonlinear errors (i.e., residuals) that the prediction mechanism model cannot capture, ultimately outputting a high-precision "final prediction value".

[0079] The isothermal adsorption model can be any one or a combination of Langmuir, Freundlich, Sips, etc., to describe the isothermal adsorption behavior of candidate materials; the mass transfer model can be replaced between LDF, surface diffusion, or multi-rate models depending on the real-time load and operating conditions. In this embodiment, the isothermal adsorption model used is Langmuir, and the specific formula is as follows: ; in, Indicates the target metal ion concentration. This represents the equilibrium adsorption capacity function.

[0080] Material parameters such as These are recommended values ​​and can be updated based on actual batches.

[0081] To ensure the consistency and traceability of the adsorption material parameters, the system introduces a material batch management mechanism in this step.

[0082] Each batch of AI-GenMOF material is generated with a unique identification code during production, and the corresponding process relationship is as follows: ; in, This indicates the identifier for the corresponding batch of adsorbent material. Represents a hash function. This indicates the precise chemical formula of the corresponding adsorbent material. Indicates the specific synthesis date of the corresponding batch of adsorbent material; A mapping relationship is established between the parameter and its corresponding parameter package ID. This relationship table is stored in the digital twin server in the following key-value format, and the corresponding process relationship is as follows: ; This index enables the system to trace material properties and verify model reproduction during operation. When a deviation in adsorption performance is detected, the parameter package version can be quickly compared to determine whether it is caused by batch differences in materials.

[0083] The edge computing unit completes the fusion calculation of parameters and actual loading amount to obtain the initial total adsorption capacity of the single adsorption reactor: ; in, This represents the initial total adsorption capacity of a single adsorption reactor. Indicates the maximum adsorption capacity. Indicates the actual filling amount; This value is then synchronized to the main controller for threshold determination in subsequent control strategies.

[0084] Simultaneously, to achieve adaptive modeling of the dynamic performance of AI-GenMOF, this step also introduces adaptive calibration of adsorption kinetics. That is, in the initial stage of adsorption, the edge computing unit adjusts the calibration based on the real-time Pb / Cd removal rate. Simulated predicted removal rate of Pb / Cd heavy metal ions The deviation dynamically updates the mass transfer rate constant. : ; in, This represents the updated mass transfer rate constant. This represents the mass transfer rate constant before the update. Indicates the correction factor. .

[0085] Updated parameters are synchronized to the digital twin server to achieve real-time adaptation of model parameters.

[0086] S203. Initial rinsing and background signal removal, the specific steps are as follows: To avoid interference with measurements caused by material dust and background leaching during the initial commissioning phase, the system performs background rinsing on adsorption reactors A and B.

[0087] Circulating pumps and valves at low flow rates (Recommended 0.5L / min) Rinse for 5 minutes.

[0088] The flow meter and pressure sensor monitor the pressure difference and flow stability in real time.

[0089] Online heavy metal analysis sensor checks background concentration in effluent ,Require Otherwise, extend the rinsing time or determine that the material pretreatment is abnormal.

[0090] After rinsing is completed, the main controller records the process parameters in the digital twin server for subsequent stability comparison.

[0091] S204. Pre-wetting and initial freezing of the adsorption reactor, the specific steps are as follows: After rinsing, the adsorption packing material enters a wetted and saturated state. The edge computing unit collects the flow rate / pressure difference curve to assess the initial hydraulic resistance of the adsorption reactor. With adsorption reactor volume utilization rate Perform the calculation.

[0092] Initial hydraulic resistance As a benchmark for subsequent clogging assessment; adsorption reactor volume utilization rate Used to predict the effective adsorption cycle.

[0093] The main controller freezes this state as the initial operating condition and writes it into the digital twin server's initial twin state.

[0094] S205. Simulation of initial degradation of adsorption packing and establishment of performance baseline, the specific steps are as follows: To assess the performance degradation characteristics of adsorption materials during long-term operation, degradation simulations were performed on the pore size distribution and specific surface area of ​​the adsorption packing.

[0095] Simulated temperature range: T = 300–340 K; Number of cycles: N = 10.

[0096] The pore size change rate was calculated. With the rate of change of specific surface area Then, based on the rate of change of pore size and the rate of change of specific surface area, the expected degradation coefficient is calculated: ; in, Indicates the expected degradation coefficient. Indicates the rate of change of aperture. This represents the rate of change of specific surface area. Indicates the initial specific surface area. Indicates the initial average particle size; The expected degradation coefficient is used as a benchmark input to the digital twin database for the degradation of material properties, and is used as a reference for the S50 predictive maintenance and life model.

[0097] like If the system then enables degradation compensation mode in subsequent operation, the edge computing unit will correct the adsorption kinetic parameters in real time to achieve performance drift self-compensation.

[0098] After S205 is completed, the system has the ability to trace material parameters, predict structural performance, and initialize adsorption models, and can enter the parallel adsorption reactor adsorption and intelligent control stage of S30.

[0099] S30, Operation and Intelligent Control of Adsorption Reactor, specifically including S301 to S306.

[0100] This step is completed collaboratively by modules including adsorption reactor A, adsorption reactor B, adsorption packing material, online heavy metal analysis sensor, flow meter and pressure sensor, edge computing unit, main controller, digital twin server, XAI decision support terminal, circulating pump and valves, safety back-off controller, and online adaptive learning module. Its purpose is to achieve stable flow, maximize adsorption efficiency, and automatic regeneration scheduling during the adsorption stage. The operating strategy is adaptively adjusted through MPC+DRL control logic to ensure continuous and efficient removal of Pb. 2+ Cd 2+ Heavy metal ions, etc. During this process, the online adaptive learning module continuously updates the control model, enabling system adaptation and performance evolution. During adsorption operation, the edge computing unit automatically adjusts the influent concentration C... in effluent concentration Cout Pressure difference Δ in the adsorption reactor P Pump energy consumption E pump The operational data is labeled as "optimal adsorption cycle" or "non-optimal cycle", and the value function of the DRL strategy is updated through a self-supervised comparative learning mechanism to improve the accuracy of bed changing decisions.

[0101] S301. Water inlet switching and initial operation loading, the specific steps are as follows: After S20 pretreatment, the main controller controls the circulating pump and valves to introduce the raw water from the inlet through the pretreatment unit into adsorption reactor A, while adsorption reactor B is in standby or regeneration state.

[0102] Flow meter and pressure sensor collect influent flow rate in real time Pressure difference with adsorption reactor The pressure difference in the adsorption reactor is compared with a set threshold. ; If the limit is exceeded, the safety back-off logic of the safety back-off controller will be triggered immediately, switching to the operation of adsorption reactor B.

[0103] S302. Adsorption dynamic control and breakthrough curve monitoring, the specific steps are as follows: During the adsorption operation, an online heavy metal analysis sensor monitored the Pb / Cd concentration in the effluent at a frequency of 0.2 Hz. .

[0104] When the Pb / Cd concentration in the effluent Approaching or exceeding the preset threshold , The inlet concentration indicates that the system determines that adsorption reactor A is close to saturation.

[0105] Edge computing units calculate the remaining effective adsorption capacity, i.e., future state data, based on prediction models: ; in, Indicates the remaining effective adsorption capacity. Indicates instantaneous inbound traffic. Indicates the instantaneous inlet concentration. Indicates the instantaneous outlet concentration. Represents the integral variable; Once the remaining effective adsorption capacity falls below the set lower limit, the edge computing unit sends a switching and regeneration command to the main controller.

[0106] S303, Control Strategy Optimization (MPC+DRL), the specific steps are as follows: The main controller executes the Model Predictive Control (MPC) module to calculate the optimal flow rate and regeneration switching time, obtaining the first strategy. The first multi-objective function is: ; in, This represents the weight coefficients corresponding to different terms in the first multi-objective function; Indicates pump and regeneration energy consumption; Indicates the number of regenerations. Indicates the target value of effluent concentration. Represents the system's control input variables. This represents the objective function for model predictive control.

[0107] Meanwhile, the online adaptive learning module is responsible for the DRL agent strategy. Continuous training and optimization are performed. This module interacts with the parallel simulation environment of the digital twin server, adjusting the control strategy based on the twin simulation feedback to achieve long-term optimality.

[0108] The digital twin server runs parallel simulations in real time, while edge computing units perform inference, with an inference latency of no more than 200ms. The training and updating of the DRL agent are continuously performed in the background by an online adaptive learning module, ensuring that the control strategy can adapt to changes in operating conditions. If the MPC or DRL module malfunctions, the safety fallback controller automatically takes over PID control.

[0109] Adsorption performance degradation compensation coefficient The calculation formula is: ; in, This represents the adsorption performance degradation compensation coefficient. This represents the actual specific adsorption capacity calculated in real time. This represents the specific adsorption capacity predicted by the simulation. when At that time, the digital twin server is triggered to update the dynamic parameters of the material model to maintain prediction accuracy.

[0110] S304. Adsorption reactor switching and regeneration triggering: The specific steps are as follows: When adsorption reactor A reaches the breakthrough point, the main controller switches the circulating pump and valves to supply water to adsorption reactor B. Adsorption reactor A is then connected to the adsorbent regeneration unit for regeneration. During the regeneration phase, the regenerated liquid flow rate is maintained. ,satisfy .

[0111] And monitor the recovery of heavy metal concentrations at the export level. This is considered a complete regeneration process. The entire switching process is completed within 30 seconds, ensuring continuous water output.

[0112] The MPC control module is based on a multi-objective function (regenerative energy consumption). Regeneration rate Adsorption recovery rate The optimal regeneration method is dynamically selected to obtain the second strategy. The first and second strategies constitute a complete control strategy.

[0113] The second multi-objective function is defined as follows: ; in, This represents the value of the second multi-objective function. Let represent the weight coefficients corresponding to different terms in the second multi-objective function, and ; This represents a reference value for regenerative energy consumption. Indicates the reference regeneration rate. This represents the ideal adsorption recovery rate.

[0114] S305, Real-time Interpretation and Manual Intervention Interface, the specific steps are as follows: The XAI decision support terminal analyzes in real time the control decisions made based on the final control strategy or the preset PID control strategy (i.e., all decisions made by the edge computing unit and the main controller), and uses any or a combination of methods such as SHAP, LIME, Integrated Gradients, and Attention visualization to perform interpretable decomposition of key decisions such as flow adjustment, regeneration timing, and energy consumption scheduling.

[0115] If manual intervention is required, the XAI decision support terminal will transmit the instructions back to the main controller and immediately update the control strategy to maintain closed-loop consistency.

[0116] After manual intervention and confirmation, the XAI decision support terminal records the intervention operation, its effect, and explanatory text into the knowledge base for subsequent DRL policy reward function correction.

[0117] S306. Safety redundancy protection, the specific steps are as follows: If communication interruption, edge computing unit or main controller malfunction, or data loss of online heavy metal analysis sensor, flow meter and pressure sensor exceeds 30 seconds, the safety backoff controller will automatically take over and enable fixed PID control to maintain stable flow and switching cycle, ensuring that the system can still maintain qualified water output in the event of failure.

[0118] Meanwhile, the online adaptive learning module continuously collects all operational data during normal system operation and automatically updates and optimizes MPC model parameters and DRL policy networks through online learning algorithms. All updated policies must undergo security verification via a digital twin server before deployment, forming a closed-loop self-learning and adaptive system of "data acquisition → model update → twin verification → policy deployment," ensuring the system's long-term performance self-evolution.

[0119] After the adsorption operation and regeneration scheduling in the S30 stage, the system status information and control strategy candidates will be transmitted to the digital twin server, entering the parallel simulation and interpretable decision support stage in S40.

[0120] S40, Digital Twins and Explainable Decision Support, specifically including S401 to S406.

[0121] This step is completed collaboratively by the edge computing unit, main controller, digital twin server, XAI decision support terminal, multi-fidelity simulation coordination unit, risk decision and fault tolerance engine, data bus and communication gateway, and security rollback controller. Its core objective is to achieve a closed loop of pre-decision verification, risk visualization, and human intervention through parallel simulation and interpretable artificial intelligence of digital twins, ensuring high transparency, high reliability, and controllability in the system operation process.

[0122] S401, Digital Twin Model Synchronization and Real-Time Mirroring, the specific steps are as follows: During system operation, the data sampling frequency is ≥0.2Hz. The edge computing unit pushes the state vector to the digital twin server in real time through the data bus and communication gateway. The corresponding process has the following relationship: ; in, This represents the state vector at the current discrete time step t; The digital twin server performs dynamic predictions based on a hybrid model, and the corresponding process has the following relationship: ; in, This indicates the predictive operation of the mechanism model (adsorption kinetics + hydraulic model). This indicates the prediction operation of the residual compensation model (LSTM / TCN). This represents the simulation prediction result of the system at the next discrete time step t+1. This represents the input vector of the system at the current discrete time step t. This represents the set of parameters for the model.

[0123] The "adsorption kinetics + hydraulic model" constitutes the mechanistic model part of the digital twin hybrid model, which achieves prediction by mathematically describing the physicochemical nature of the adsorption process: Adsorption kinetics model: Based on the principles of mass transfer and surface reaction, this model describes the rate of diffusion and capture of target heavy metal ions from the bulk liquid phase to the surface and interior of the adsorption packing particles. It can be characterized by a combination of pseudo-second-order kinetic equations and an intraparticle diffusion model.

[0124] Hydraulic model: Based on the Navier-Stokes equations or their simplified forms, it describes the flow state of wastewater in a packed bed adsorption reactor, including velocity distribution, pressure drop changes, and residence time distribution, in order to determine the convection and diffusion transport of pollutants.

[0125] The two models described above are solved in a coupled spatiotemporal domain using partial differential equations, jointly outputting the pollutant concentration distribution and adsorption bed state of the system in future periods under given initial states and control strategies. This mechanistic model provides an interpretable physical basis for the entire prediction, ensuring that the system still possesses basic extrapolation capabilities even in the absence of historical data.

[0126] Maintaining the synchronization deviation between the physical and virtual dual tracks is governed by the following relationship: ; in, This indicates that the length of the vector is being calculated. This represents the predicted state vector of the system at the current discrete time step t. Indicates the synchronization deviation between the two tracks. Recommended value ; Based on the magnitude of the dual-track synchronization deviation, steady-state monitoring, strategy verification, or anomaly analysis operations are performed.

[0127] The multi-fidelity simulation coordination unit plays a core role in this stage: it adaptively selects high / medium / low fidelity models according to different operating scenarios (steady-state monitoring, strategy verification, or anomaly analysis); it adopts the "co-simulation assimilation algorithm" to fuse the mechanism model and the data-driven model, realizes cross-scale dynamic reproduction and real-time mirroring, and significantly improves simulation efficiency and response speed.

[0128] S402. Control strategy pre-simulation and risk assessment, the specific steps are as follows: When Model Predictive Control (MPC) or Deep Reinforcement Learning (DRL) agent outputs a new policy When (e.g., bed change time, flow rate adjustment, regeneration mode selection) occurs, the digital twin server invokes the multi-fidelity simulation coordination unit to execute the strategy simulation. The simulation duration is 24 hours, and the output indicators include... That is, the effluent concentration prediction curve, That is, total energy consumption, That is, the expected number of regenerations, That is, risk indicators, which are defined as follows: ; in, Indicates risk indicators, These represent the weights for performance / compliance risk, economic / cost risk, and equipment / stability risk, respectively. This represents the simulated predicted export concentration. The statutory or operational baseline risk value indicating the export concentration. Represents the variance function; like , If the preset risk threshold is met, the strategy passes the verification and the final control strategy is obtained; otherwise, the strategy is rejected and the current control is retained.

[0129] The risk decision-making and fault-tolerance engine further analyzes the simulation results from multiple dimensions, constructs a risk knowledge graph, and classifies risks by combining historical operating data and equipment health status. When a potential anomaly is detected (such as delayed response of circulating pumps and valves, or excessive flow fluctuations), the system automatically triggers a fault-tolerance degradation mechanism, activating a preset PID strategy to proactively avoid the risk of runaway.

[0130] S403, XAI decision interpretation and contribution analysis, the specific steps are as follows: Once the strategy passes simulation, the XAI decision support terminal calculates the contribution of the policy input features to the output decision. Input features: These represent real-time influent concentration, remaining effective adsorption capacity, pressure difference in the adsorption reactor, current electricity price, and historical operating trend data, respectively. Output decision: ; These represent the influent flow rate setpoint, adsorption reactor switching time, and regeneration mode selection, respectively; The contribution ranking results are visualized on the XAI decision support terminal. The XAI decision support terminal display interface also provides policy explanation text, which helps operation and maintenance personnel understand the control logic.

[0131] S404. Manual intervention and strategy modification, the specific steps are as follows: When maintenance personnel input policy modifications (such as switching the adsorption reactor in advance or changing the regeneration method) into the XAI decision support terminal, the digital twin server simultaneously executes the risk simulation of the manually implemented policy. It then invokes the risk decision and fault tolerance engine to perform a safety reassessment. If the conditions are met... , If the risk indicator of the manual intervention strategy is indicated, the edge computing unit and the main controller will receive and execute the correction strategy; otherwise, the XAI decision support terminal will prompt "high risk, operation rejected" and automatically restore the PID backup strategy provided by the safety rollback controller.

[0132] This mechanism ensures that AI control and human intervention always operate in a closed loop within the safety boundaries, avoiding system instability caused by human error.

[0133] S405. Operation Log and Knowledge Accumulation: The specific steps are as follows: Every strategic decision and intervention is automatically recorded as a structured log by the digital twin server, which includes: Strategy input parameters; simulation results and risk assessment; this embodiment uses the SHAP / LIME contribution vector; operation and maintenance personnel intervention content and results; risk classification and fault-tolerant execution records.

[0134] The risk decision engine transforms logs into structured knowledge entries, which are then incorporated into a risk knowledge graph database for subsequent DRL incremental training and MPC parameter adaptive fine-tuning, enabling continuous learning and long-term performance evolution of the system.

[0135] S406. Self-verification and model reconstruction, the specific steps are as follows: The digital twin self-test unit monitors the deviation between the physical system and the virtual digital twin model in real time. When a long-term deviation exceeding a threshold (>10%) is detected, it triggers: ① The multi-fidelity simulation coordination unit redistributes modeling fidelity; ②Risk decision-making and fault-tolerant engine perform security checks; ③ The system automatically reconstructs the local model or calls the most recent valid model snapshot.

[0136] This ensures that the digital twin model maintains high consistency and self-healing capabilities over the long term, supporting the robust operation of the system under complex working conditions.

[0137] S50, anomaly detection, safety rollback and continuous model optimization, specifically including S501 to S506.

[0138] This step is completed collaboratively by the edge computing unit, main controller, digital twin server, XAI decision support terminal, data bus and communication gateway, power management unit, safety rollback controller, multimodal anomaly detection and feature fusion unit, graph intelligence root cause analysis module, predictive maintenance and fault-tolerant coordination unit, and online adaptive learning module. Its core objective is to achieve early anomaly identification, source diagnosis, and self-healing response, ensuring long-term system stability and autonomous evolution through a closed-loop mechanism of "detection—location—prediction—rollback—optimization."

[0139] S501, Multimodal Data Acquisition and Fusion Modeling, the specific steps are as follows: During operation, the system continuously collects data from different sources through edge computing units, data buses, and communication gateways: ; in, This represents sensor data (flow rate, pressure, concentration). This indicates actuator data (circulating pump and valve opening, pump speed). Indicates the flow of control commands. This indicates the system event log.

[0140] The multimodal anomaly detection and feature fusion unit standardizes and aligns the data over time, and employs multi-resolution time-window sliding modeling. ; in, Encoder function It is a deep feature encoder (composed of LSTM+TCN+Transformer). This indicates the output fused feature representation. Represents a data matrix or tensor. Indicates the width or duration of the sliding time window.

[0141] The multimodal anomaly detection and feature fusion unit can extract periodic features through multi-channel convolution, capture abrupt changes or drift trends through temporal attention mechanisms, and establish a multimodal joint distribution. This is used for subsequent anomaly detection.

[0142] S502. Anomaly detection and classification, the specific steps are as follows: Anomaly detection can employ other specific algorithms, such as AE, VAE, and CUSUM. This embodiment uses a probabilistic graphical modeling and self-supervised comparative learning mechanism for anomaly detection. ; in, This represents the outlier score at time t. Represents the state prediction function, when When an abnormal alarm is triggered, This represents the set of parameters for the anomaly detection model.

[0143] Based on the pattern clustering and density assessment results of the anomaly feature vectors, the anomalies are divided into: ① Transient anomalies (short-term fluctuations, such as sensor noise) ② Continuous abnormalities (system drift, equipment aging) ③ Structural anomaly (control logic failure or communication interruption) After detecting an anomaly, the digital twin server automatically calls the digital twin model for comparative simulation to confirm whether it is a physical anomaly or virtual drift (i.e., digital twin error). If the digital twin model responds normally, it is marked as a "physical anomaly"; otherwise, it enters the self-calibration process (see S506).

[0144] S503, Graph Intelligent Root Cause Analysis and Impact Scope Assessment, the specific steps are as follows: When a persistent or structural anomaly is detected, the system invokes the graph-based intelligent root cause analysis module.

[0145] The graph-based intelligent root cause analysis module first constructs a system dependency graph: ; Where V represents the node set (sensors, actuators, control logic units), E represents the temporal dependency edges between nodes, the edge weights are calculated from mutual information and control influence matrix, and G represents the system dependency graph.

[0146] Calculate node importance using Graph Attention Network (GAT) or Process Control Graph Attention Network (PCGAT): ; in, Let i represent the input feature vector of the target node i. Let i represent the set of all neighboring nodes directly connected to node i in the knowledge graph. j To be able to refer to the summation index of the node, Let N(i) represent the input feature vector of neighbor node j, where j is an element in the set N(i). This represents a learnable weight matrix. This represents the attention weight, reflecting the strength of the influence of node j on node i.

[0147] Finally, the root cause node of the anomaly and its scope of influence are located through anomaly propagation path scoring. The anomaly propagation path scoring calculation process is as follows: ; in, This represents the root cause score of node i. Represents an observed anomalous node. Or symptom nodes, This represents the propagation path from node i to node o; The results are visualized on the XAI decision support terminal, supporting manual verification and intervention.

[0148] S504. Proactive early warning and fault-tolerant execution, the specific steps are as follows: When the root cause analysis module outputs a high-risk conclusion (e.g., failure of critical circulating pumps and valves or adsorption reactor controls), the predictive maintenance and fault tolerance coordination unit activates the active protection mechanism. ① Soft degradation control: The edge computing unit temporarily switches to the redundant control model (linear MPC approximation).

[0149] ② Bypass switching: The main controller redistributes the flow, shuts down the abnormal path, and starts the backup adsorption reactor or regeneration path.

[0150] ③ Safety backoff control: If the risk persists, the safety backoff controller takes over the control logic and executes the PID steady-state mode to ensure the safety of the effluent water quality.

[0151] ④ Warning and notification: The system generates an anomaly report and pushes it through the XAI decision support terminal, including the anomaly type, root cause node, scope of impact and rollback level.

[0152] This step enables a shift from passive response to proactive fault tolerance, ensuring that the system retains its steady-state self-protection capability even if AI control fails.

[0153] S505, Model Drift Detection and Online Self-Learning: The specific steps are as follows: To prevent model aging from causing misjudgments or control shifts, the system periodically calculates the prediction residuals: ; in, Represents the instantaneous prediction residual at time t. This represents the measured value of heavy metal concentration in the effluent at time t. This represents the predicted value of the effluent heavy metal concentration at time t in the simulation prediction results; if Exceeding the threshold within a continuous window , This represents the residual statistics over a period of time. This represents the residual threshold, triggering a lightweight AI prediction module and a residual data-driven model retraining request. (Digital model outputs predicted values) The online adaptive learning module performs local fine-tuning based on the latest dataset and updates model parameters synchronously with the digital twin server.

[0154] The multi-fidelity simulation coordination unit is responsible for verifying the usability and risk safety of the new model during this process, ensuring that the new strategy passes simulation review before deployment.

[0155] S506, Abnormal Closed-Loop Learning and Knowledge Accumulation: The specific steps are as follows: Each anomaly detection, localization, and rollback process is recorded as a structured entry by the digital twin server, including: ① anomaly type and trigger time; ② root cause node and impact path; ③ fault-tolerant execution plan; ④ model update log; ⑤ decision feedback and recovery latency.

[0156] After being jointly processed by the modular graph intelligent root cause analysis module and the predictive maintenance and fault tolerance coordination unit, this data is incorporated into the "anomaly-response knowledge base" for subsequent model reinforcement learning (DRL) or adaptive controller (MPC) training, enabling relearning and continuous evolution in abnormal scenarios.

[0157] In summary, this invention directly addresses and solves the aforementioned technical problems by constructing an integrated closed-loop system of "materials-equipment-control-decision," thereby achieving significant advantages in treatment efficiency, operational economy, safety, and intelligence. First, to fundamentally solve the problem of "materials and processes being disconnected," this invention utilizes generative artificial intelligence to design adsorption packing material (AI-GenMOF) that precisely matches the characteristics of the wastewater to be treated. This directly couples the optimization of material performance with process requirements, breaking through the performance ceiling of traditional general-purpose adsorption packing material and significantly improving the selectivity and efficiency of treatment. Second, to overcome the deficiency of "rigid control strategies," this invention deploys advanced control logic (S303) that combines Model Predictive Control (MPC) and Deep Reinforcement Learning (DRL) on the edge computing unit. This logic dynamically optimizes the switching timing and regeneration strategy of the adsorption reactor based on real-time operating conditions, achieving intelligent adaptive operation under high-frequency disturbance conditions. Compared with fixed-cycle or threshold control, this significantly reduces operating costs and avoids the risk of effluent exceeding standards. Furthermore, addressing the key pain points of existing AI applications such as "lack of security verification" and "opaque decision-making process," this invention innovatively introduces a digital twin server to perform pre-simulation and risk assessment before executing all AI control strategies (S402), eliminating high-risk operations and ensuring the robustness of the system; at the same time, the XAI decision support terminal provides real-time interpretation of the decision basis (S305), opening the AI ​​"black box" and achieving efficient human-machine collaboration.

[0158] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0159] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0160] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0161] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence, characterized in that, The system includes: The perception and input layer is used for: The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. The intelligent decision-making and control layer is used for: Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; The system support layer is used for: It provides communication and power, and switches to a preset PID control strategy when a failure occurs in the intelligent decision-making and control layer; The core processing and execution layer includes at least two adsorption reactors connected in parallel and / or series. The adsorption reactors are filled with highly selective metal-organic framework adsorption materials designed and screened using generative artificial intelligence based on the target heavy metal ions. This is used for: Based on the final control strategy or the preset PID control strategy, the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor are controlled to continuously adsorb and regenerate the wastewater containing heavy metals, so as to achieve the recovery of heavy metals from the wastewater.

2. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 1, characterized in that, The perception and input layer includes: The raw water inlet is used to introduce wastewater containing heavy metals to be treated; Pretreatment units, including screens, flocculation or grit chambers, are used for preliminary filtration to remove large particulate suspended solids from the raw water; An online heavy metal analysis sensor is installed at the outlet and inlet of the adsorption reactor to monitor the concentration of target heavy metal ions in the influent and effluent in real time. Flow meters and pressure sensors are used to collect real-time data on liquid flow rate in the pipeline and pressure difference in the adsorption reactor. An adaptive sensor calibration module is used to monitor the drift of online heavy metal analysis sensors, as well as flow meters and pressure sensors, and to make real-time adjustments. The multimodal anomaly detection and feature fusion unit, located between the edge computing unit and the digital twin server, is used to fuse data from online heavy metal analysis sensors, flow meters and pressure sensors, actuators and control logs using a temporal convolutional network and a multi-head attention mechanism to achieve feature alignment and fusion, generate fused features, and perform anomaly detection based on the fused features to generate preliminary anomaly detection results.

3. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 2, characterized in that, The intelligent decision-making and control layer includes: The edge computing unit carries a lightweight artificial intelligence model for real-time prediction and reasoning of future states, and sends control strategies to the main controller. The main controller integrates a model predictive control module and a deep reinforcement learning module. The model predictive control module receives instructions from the edge computing unit and, based on the future predicted state, calculates the optimal flow rate and regeneration switching time with the optimization objectives of minimizing the deviation from the target effluent concentration, minimizing energy consumption, and minimizing the number of regenerations, in order to obtain the first strategy. The optimal regeneration method is calculated with the optimization objectives of minimizing regeneration energy consumption, maximizing regeneration rate, and maximizing adsorption recovery rate to obtain the second strategy; the first strategy and the second strategy constitute the control strategy. The deep reinforcement learning module is used to learn the long-term optimal control policy by using reinforcement learning, guided by the control policy, and thus obtain the final control policy. The digital twin server communicates with the edge computing unit in real time to build and run a digital twin model synchronized with the physical system. This model is used for "hypothesis-simulation-verification" of control strategies, optimization of system parameters, and prediction of long-term operational risks. The XAI decision support terminal is used to analyze the contribution of control decisions made based on the final control strategy or a preset PID control strategy and to visualize the output. The digital twin self-testing unit is used to monitor the differences between the physical system and the digital twin model in real time, and to adjust or reconstruct the digital twin model.

4. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 3, characterized in that, The intelligent decision-making and control layer also includes: The risk decision and fault tolerance engine is located between the main controller and the digital twin server, operating across layers. This engine has built-in Monte Carlo disturbance analysis, metastable risk estimation and fault tolerance logic modules. It calculates risk indicators for each control strategy and compares them with preset risk thresholds. When a high risk or abnormal trend is detected, the safety backoff controller is triggered to enter the safety mode and execute the preset PID control strategy. The graph-based intelligent root cause analysis module, located between the digital twin server and the main controller, is used to construct a system dependency graph based on the preliminary anomaly detection results, using sensors, actuators, and control logic units as nodes and temporal dependencies between nodes as edges. By constructing the system dependency graph and graph attention network, it analyzes the anomaly propagation path, outputs the root cause nodes and their impact range, and realizes anomaly explanation and decision support.

5. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 4, characterized in that, The edge computing unit embeds a lightweight self-learning module, which is used to perform model drift detection and parameter fine-tuning on the lightweight artificial intelligence model based on real-time data. When the prediction residual is detected to deviate from the threshold by more than 10% for a long time, local retraining is triggered and the digital twin model parameters are updated synchronously to achieve distributed continuous learning. The decision and control layer also includes an online adaptive learning module, which continuously collects real-time data during system operation and uses online learning algorithms to update and optimize the model prediction control module and deep reinforcement learning module, while coordinating with the digital twin server to perform security verification of new strategies.

6. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 5, characterized in that, The digital twin server runs a hybrid model, which includes a mechanism model based on adsorption kinetics and hydraulics, as well as a residual data-driven model that compensates for prediction bias. The digital twin server is configured to perform parallel simulations for a predetermined duration based on the control policy before the control policy is issued and executed, in order to generate simulation results.

7. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 6, characterized in that, The digital twin server is also equipped with a multi-fidelity simulation coordination unit, which automatically selects the simulation accuracy based on the real-time load and operating conditions through a model switching function, so as to achieve dynamic coordination between the high-fidelity CFD mechanism model and the low-fidelity AI approximation model and perform parameter recalibration.

8. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 7, characterized in that, The system support layer includes: The power management unit is used to provide stable power to all electrical components; Data bus and communication gateway are used to build a system intranet using industry-standard protocols for communication. The safety rollback controller is used to switch to a preset PID control strategy to maintain basic system operation when an edge computing unit inference timeout, communication link interruption, or main controller failure is detected.

9. The wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence according to claim 8, characterized in that, The core processing and execution layer also includes: The adsorbent regeneration unit, connected to the parallel adsorption reactor, regenerates the saturated adsorption packing material after receiving the control strategy to restore its adsorption capacity. The adsorbent regeneration unit supports at least one of the following regeneration modes: electrochemical regeneration, chemical elution regeneration, low-temperature thermal regeneration, vacuum desorption regeneration, ultrasonic regeneration, and microwave regeneration. The selection of the regeneration mode is intelligently decided by the final control strategy or a preset PID control strategy. The circulating pumps and valves, controlled by the main controller, are used to switch between adsorption reactors, reflux, and regulate the flow rate within the system. The AI-GenMOF design unit, based on a generative artificial intelligence model, generates MOF candidate structures according to the coordination characteristics of the target heavy metal ions. Reinforcement learning is then used to optimize the specific surface area, pore size distribution, and charge density distribution of the MOF candidate structures. Finally, density functional theory (DFT) calculations are used to screen and obtain the optimal MOF structure. The predictive maintenance and fault-tolerant coordination unit, located between the main controller and the safety rollback controller, is used to monitor equipment degradation trends, generate proactive protection strategies through the predictive maintenance model, and actively coordinate the rollback controller to enter steady-state mode when the risk exceeds the threshold, thereby achieving proactive fault tolerance and continuous operation assurance.

10. A method for heavy metal recovery from wastewater based on generative artificial intelligence and edge intelligence, characterized in that, The method is used to implement the wastewater heavy metal recovery system based on generative artificial intelligence and edge intelligence as described in any one of claims 1 to 9, and the method includes the following steps: At least two adsorption reactors are arranged in parallel and / or in series, and the adsorption reactors are filled with highly selective metal-organic framework adsorption materials designed and screened using generative artificial intelligence based on the target heavy metal ions. The wastewater containing heavy metals to be treated is introduced, and effluent data and system status data are collected in real time. Real-time state prediction is performed based on current effluent data, current material characteristic data of adsorption packing, and current system status data to obtain future state data; Based on future state data, corresponding control strategies are generated with the goals of maximizing adsorption efficiency, minimizing energy consumption, and minimizing regeneration costs. The control strategy is simulated and evaluated, and the final control strategy is obtained based on the evaluation results; A preset PID control strategy is used, and the final control strategy or the preset PID control strategy is selected according to the operating status to control the flow rate of the wastewater containing heavy metals to be treated and the adsorption and regeneration operating parameters of the adsorption reactor, so as to continuously adsorb and regenerate the wastewater containing heavy metals to be treated, thereby realizing the recovery of heavy metals from wastewater.

Images