A digital human twin edge AI control terminal for a supercritical CO2 power generation system

Abstract

The application discloses a supercritical CO2 power generation system digital human twin edge AI management and control terminal and belongs to the technical field of industrial automatic control. In view of the problems of operation control lag in the near critical region, strong coupling of parameters, and dispersed traditional terminal hardware and low-efficiency interaction in the prior art, the application realizes active prediction and real-time dynamic optimization of complex working conditions by deeply integrating perception and execution through an integrated edge AI architecture, decoupling strong coupling parameters and predicting states by relying on an autonomous decision module. Meanwhile, the application provides operation pre-rehearsal and intuitive guidance in combination with a digital human twin multi-modal interaction module, builds a safety barrier against misoperation, and cooperates with a hierarchical fault self-healing and closed-loop self-calibration mechanism, thereby greatly reducing the operation and maintenance threshold and comprehensively guaranteeing the stability, safety and comprehensive power generation efficiency of the system in the whole life cycle.

Description

Technical Field

[0001] This invention relates to the field of industrial automatic control technology, specifically to a digital human twin edge AI control terminal for a supercritical CO2 power generation system. Background Technology

[0002] Supercritical CO2 power generation systems have broad application prospects in the field of new energy power generation due to their high thermal efficiency and compact design. However, the physical state of the system during operation exhibits inherent characteristics such as strong nonlinearity, large time lag, and strong coupling of multiple parameters. Furthermore, it operates under high temperature and high pressure conditions under normal circumstances, which places extremely high demands on the real-time performance, adaptive capabilities, safety response speed, and human-machine interaction efficiency of the terminal.

[0003] Most existing commercial supercritical CO2 control equipment uses traditional industrial control architectures, primarily relying on PLCs (Programmable Logic Controllers) combined with PID control to achieve setpoint control of system operating parameters, and conventional industrial touchscreens for two-dimensional data display at the human-machine interface. Furthermore, although emerging technologies such as digital twins, virtual digital humans, and edge computing have begun to be applied in some industrial monitoring fields, when comprehensively applied to the control of supercritical CO2 power generation systems, existing technologies still have the following overall technical shortcomings:

[0004] 1) Passive control and limited computing power: Traditional control architecture can only achieve passive setpoint adjustment and passive alarm after the fact, without active early warning and predictive control capabilities. It cannot solve the problem of strong coupling of multiple parameters, resulting in low power generation efficiency and large fluctuations in operating status. At the same time, existing terminals lack local edge AI computing power units. If they rely on cloud AI computing, they will face high transmission latency, network failure and network security risks, and cannot meet the requirements of millisecond-level real-time control and closed-loop handling of the entire process in industrial sites.

[0005] 2) Inefficient human-computer interaction mode: The interaction mode remains in the traditional numerical / curve visualization interface, without humanized and visualized guided interaction, resulting in high learning costs for operation and maintenance personnel, slow response to fault handling, and prominent risk of on-site misoperation; Although digital twin technology is applied to industrial equipment monitoring, it has not achieved integration with digital human multimodal interaction, nor does it have a state mapping mechanism for the physical properties of supercritical CO2.

[0006] 3) Dispersed hardware with poor adaptability: The existing terminal's monitoring, control, diagnostic, interaction, and security protection modules are independent of each other, usually requiring multiple devices such as industrial control computers, PLCs, displays, and communication gateways. This dispersed hardware architecture leads to complex field wiring, poor compatibility between devices, high maintenance costs, and extremely low overall integration, making it difficult to adapt to the harsh installation environment of industrial sites. Summary of the Invention

[0007] To address the aforementioned issues, the purpose of this invention is to propose a digital human twin edge AI control terminal for supercritical CO2 power generation systems, achieving high hardware integration, autonomous edge AI decision-making, and multimodal interaction of the digital human, thereby resolving the technical shortcomings of existing systems such as passive control, inefficient interaction, and fragmented architecture.

[0008] This was achieved through the following technical solutions:

[0009] A digital human twin edge AI control terminal for a supercritical CO2 power generation system, characterized in that it includes:

[0010] Edge AI computing chips are used to carry out control logic operations and real-time rendering of digital human twins, and output control commands;

[0011] An isolated interface circuit, including a multi-source data acquisition unit and an isolated control circuit unit, is used to receive field operating parameter signals input from the multi-source sensor array and action feedback signals from the actuator, convert control commands into electrical signals that the actuator can recognize, and push them to the actuator.

[0012] The internal communication bus connects to the edge AI computing chip and the isolated interface circuit, and is used to transmit field operation parameter signals, actuator action feedback signals and control commands between the edge AI computing chip and the isolated interface circuit.

[0013] The interactive display unit is used to display the digital twin interactive interface and collect touch commands.

[0014] Optionally, the edge AI computing chip is configured to run the following modules:

[0015] Multi-source sensing and acquisition module: collects multi-source data in real time, removes abnormal data, and classifies and stores valid data in the terminal's local cache library;

[0016] Data preprocessing module: retrieves multi-source data from the local cache library and preprocesses it, then constructs a feature dataset based on the preprocessed multi-source data;

[0017] Efficiency self-optimization module: Based on the feature dataset, a full life cycle efficiency prediction and optimization model is built to calculate the real-time degradation trend of the supercritical CO2 power generation system and generate preventive maintenance suggestions and efficiency optimization data;

[0018] Autonomous decision-making and control module: Based on the feature dataset, it performs strong coupling and decoupling calculations on the parameters of the feature dataset to predict the drift trend of the supercritical CO2 power generation system's operating state; and generates control commands and constraints based on the prediction results.

[0019] Isolated execution control module: Receives control instructions and constraints from the autonomous decision control module, sends control instructions to the execution mechanism, and performs real-time feedback on the status of the execution mechanism;

[0020] Multimodal interaction module: Constructing a digital human twin that is synchronized in real time with the supercritical CO2 power generation system, and controlling the entire process of multimodal feedback interaction;

[0021] Fault self-healing and safety management module: Real-time monitoring of data throughout the entire process, determination of fault level, and implementation of corresponding self-healing strategies based on the fault level;

[0022] Closed-loop self-calibration and iteration module: used to analyze the causes of deviations in the end-to-end control effect of the terminal and to calibrate the parameters of other modules.

[0023] Optionally, the multi-source data includes the operating parameters of the supercritical CO2 power generation system, the working status of the actuators, and the status parameters of the terminal hardware; among which, the operating parameters of the supercritical CO2 power generation system include pressure, temperature, flow rate, density, turbine speed, compressor load, and heat exchanger heat exchange efficiency; the terminal hardware status parameters include the load of the edge AI computing chip, the accuracy of the acquisition module, and the communication status of the control module.

[0024] Optionally, outlier data removal includes: cross-validating multi-source data based on the supercritical CO2 property equation, identifying outlier data, marking and removing outlier data, and storing valid data in the terminal's local cache after timestamping and format standardization. Outlier data includes drift, jumps, and no response. By removing outlier data, the authenticity and consistency of the data stored in the terminal's local cache are effectively ensured, avoiding misjudgments caused by outlier values ​​interfering with subsequent control decisions.

[0025] Optionally, the data preprocessing module retrieves standardized data from the local cache library, performs preprocessing operations on the multi-source data using a supercritical CO2 property model, extracts key feature values ​​from the preprocessed multi-source data, and constructs a feature dataset. The preprocessing operations include completing and correcting data with missing or drifting features; normalizing and compressing the dimensionality of the preprocessed multi-source data; the key feature values ​​of the multi-source data include near-critical CO2 property mutation characteristics, actuator action response characteristics, system load change characteristics, and fault precursor characteristics.

[0026] Optionally, the autonomous decision control module inputs the feature dataset into a multivariate time-series decoupling network, allocates decoupling weights for each parameter through an attention mechanism, and performs independent decoupling calculations on strongly coupled parameters; the decoupled data is then input into a time-series prediction model based on a Long Short-Term Memory (LSTM) network to determine whether the system deviates from the critical operating range, predict the drift trend of the supercritical CO2 power generation system's operating state, and output the prediction results.

[0027] Optionally, based on the efficiency data of the supercritical CO2 power generation system during long-term operation, the efficiency self-optimization module calculates the physical property parameters of supercritical CO2 in real time. Using a pre-calculated CO2 physical property database, the calculated values ​​of physical property parameters in the near-critical region are corrected through an interpolation algorithm. Combined with the load changes of the supercritical CO2 power generation system, a full life cycle efficiency prediction and optimization model is constructed to generate efficiency optimization data for compressor pressure ratio, turbine inlet temperature, and heat exchanger medium flow rate, so as to improve the heat exchange efficiency of the system and achieve optimal matching of the circulation pressure ratio.

[0028] Optionally, the control command includes multi-dimensional attributes, which are divided into an execution target layer, a timing attribute layer, a safety constraint layer, and a verification feedback layer. The isolated execution control module generates electrical signals by parsing the execution target layer, executes actions based on the priority of the timing attribute layer, performs fault tolerance judgment based on the safety constraint layer, and uses the verification feedback layer for status back-checking.

[0029] Optionally, the entire process of multimodal feedback control interaction includes: driving the digital human twin to provide visual feedback on its status and actions based on real-time synchronization parameters; receiving external operation commands and converting them to the autonomous decision-making control module, and performing animation pre-playing through the digital human twin before the command is executed.

[0030] Optionally, a fault classification model is trained based on fault cases and operating data of supercritical CO2 power generation systems. The model uses a dual criterion of parameter deviation and rate of change to determine the fault level and executes a preset self-healing strategy based on the fault level. The fault levels are divided into Level 1 warning, Level 2 self-healing, and Level 3 protection.

[0031] Optionally, the causes of deviations in control are analyzed, including: comparing control objectives with actual implementation results, using the backpropagation algorithm to analyze deviation values, and identifying the causes of deviations; the causes of deviations include: data collection errors, decision-making errors, and execution deviations.

[0032] The beneficial effects of this invention compared to the prior art are:

[0033] 1) Enhance dynamic control and self-healing capabilities: Leveraging edge AI chips to achieve local millisecond-level lightweight inference, eliminating reliance on the cloud. Through multi-parameter decoupling and advanced prediction, the system can dynamically track the optimal operating point, overcoming the challenge of drastic property changes in the near-critical region of supercritical CO2, and effectively improving the power generation efficiency of the Brayton cycle; at the same time, it constructs an active safety defense system to achieve autonomous closed-loop handling of most common faults, significantly extending the system's fault-free operation time.

[0034] 2) Improved on-site interaction and troubleshooting efficiency: Digital human twin multimodal interaction is applied to this field, replacing traditional fragmented two-dimensional numerical charts. Through millisecond-level state synchronization, multimodal interaction, and fault visualization guidance, the underlying thermodynamic state is intuitively mapped into a human-like expression. This significantly reduces the learning cost for operators, substantially shortens fault handling response time, and fundamentally reduces the on-site error rate.

[0035] 3) High hardware integration and strong reliability: The integrated architecture deeply integrates core units such as control, communication, and edge computing, replacing traditionally distributed PLCs, industrial PCs, and communication gateways. Combined with industrial-grade wide-temperature and anti-interference design, it significantly reduces the number of on-site hardware devices, effectively lowers procurement and maintenance costs, simplifies deployment in harsh industrial environments, and comprehensively enhances the system's overall resilience. Attached Figure Description

[0036] Figure 1 This is an overall architecture diagram of a digital human twin edge AI control terminal for a supercritical CO2 power generation system.

[0037] Figure 2 A flowchart illustrating the autonomous decision-making and execution control process of a digital human twin edge AI control terminal for a supercritical CO2 power generation system;

[0038] Figure 3 This is a fault prediction and multimodal mapping data flow diagram for a digital human twin edge AI control terminal of a supercritical CO2 power generation system. Detailed Implementation

[0039] The technical solutions in the embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0040] Figure 1 This is an overall architecture diagram of a digital human twin edge AI control terminal for a supercritical CO2 power generation system, showcasing the overall architecture of the terminal modules. The terminal adopts a hardware-integrated architecture, and at the physical level, the terminal includes:

[0041] Edge AI computing chips are used to carry out control logic operations and real-time rendering of digital human twins, and generate control commands;

[0042] The isolated interface circuit includes a multi-source data acquisition unit and an isolated control circuit unit. The multi-source data acquisition unit is used to receive field operating parameter signals input from the multi-source sensor array and action feedback signals from the actuator. The isolated control circuit unit is responsible for converting the control commands generated by the edge AI computing chip into electrical signals that the actuator can recognize and pushing them to the actuator.

[0043] The internal communication bus connects to the edge AI computing chip and the isolated interface circuit, and is used to transmit field operation parameter signals, actuator action feedback signals and control commands between the edge AI computing chip and the isolated interface circuit.

[0044] The interactive display unit is used to display the digital twin interactive interface and collect touch commands.

[0045] The edge AI computing chip is configured to run a series of functional modules. These modules form a logical closed loop through a multi-source sensing and acquisition module, a data preprocessing module, an autonomous decision-making and control module, an isolated execution and control module, a multimodal interaction module, a fault self-healing and security management module, and a closed-loop self-calibration and iteration module. The business flow relationships between these functional modules are clearly illustrated. A detailed introduction to each functional module follows.

[0046] Multi-source sensing and acquisition module: This module collects multi-source data in real time and removes abnormal data, classifying and storing valid data in the terminal's local cache. Specifically, the multi-source sensing and acquisition module calls the multi-source data acquisition unit in the isolated interface circuit to collect multi-source data at a sampling rate of no less than 100Hz. The multi-source data includes the operating parameters of the supercritical CO2 power generation system, the working status of the actuators, and the terminal hardware status parameters. Among them, the supercritical CO2 power generation system operating parameters include pressure, temperature, flow rate, density, turbine speed, compressor load, and heat exchanger heat exchange efficiency; the terminal hardware status parameters include the edge AI computing chip load, the accuracy of the acquisition module, and the communication status of the control module.

[0047] The multi-source sensing and acquisition module performs cross-validation on multi-source data based on the supercritical CO2 property equation, identifies abnormal data with drift and jump characteristics, marks and removes abnormal data, and stores the valid multi-source data in the terminal local cache after timestamp synchronization and format standardization.

[0048] It should be noted that cross-validation refers to verifying the acquired multi-source data by constraining it with the physical property model of supercritical CO2, and based on the redundancy consistency of multiple sensors, the rate of change of time, and the detection of no response. A comprehensive confidence score is output, and the data is divided into three categories: normal, suspicious, and abnormal based on the comprehensive confidence score. Suspicious data is filtered or interpolated for correction, while abnormal data is directly marked and removed, triggering a hardware self-diagnostic mechanism. The hardware self-diagnostic mechanism comprehensively detects the sensor power supply status, signal output characteristics, communication link status, acquisition module operating status, and multi-source data consistency to identify abnormal situations. Abnormal situations include constant signal, exceeding the measurement range, no response, communication timeout, or multi-channel anomalies.

[0049] Based on the identified anomalies, determine whether the source of the anomaly is a sensor failure, communication failure, acquisition module failure, or system-level anomaly, and output the anomaly type, location, and severity level for the fault self-healing and safety management module to use.

[0050] The data preprocessing module preprocesses multi-source data from the local cache and constructs a feature dataset based on the preprocessed multi-source data. Specifically, the data preprocessing module retrieves standardized data from the local cache, performs preprocessing operations on the multi-source data using a supercritical CO2 property model, and constructs a feature dataset based on the preprocessed multi-source data. The preprocessing operations include completing and correcting data with missing or drifting features, performing high-frequency time-series alignment on the multi-source data, and reconstructing multi-dimensional features.

[0051] The feature dataset includes a real-time running dataset, a historical comparison dataset, and a fault feature dataset. Specifically, the real-time running dataset stores real-time running data; the historical comparison dataset stores historical operating parameter data under historical synchronization periods or standard stable operating conditions, serving as a dynamic benchmark for comparative analysis and status health assessment; and the fault feature dataset stores typical cases, serving as the judgment benchmark for fault self-healing and safety management modules.

[0052] The real-time running dataset and historical comparison dataset are pushed to the autonomous decision control module and the efficiency self-optimization module; the fault feature dataset is pushed to the efficiency self-optimization module and the fault self-healing and safety management module, while it is stored in the local cache database for model iteration training.

[0053] Efficiency self-optimization module: Based on the feature dataset, a full life cycle efficiency prediction and optimization model is built to calculate the real-time degradation trend of the supercritical CO2 power generation system and generate preventive maintenance suggestions and efficiency optimization data.

[0054] Specifically, the efficiency self-optimization module calls the feature dataset to extract the physical property parameters of supercritical CO2 in real time, including density, specific heat capacity, and thermal conductivity; using a pre-calculated CO2 physical property database, it corrects the calculated values ​​of physical property parameters in the near-critical region through an interpolation algorithm; combining the load changes of the supercritical CO2 power generation system, it constructs a full life cycle efficiency prediction and optimization model, inputs the corrected parameters into the full life cycle efficiency prediction and optimization model for thermodynamic cycle optimization, and calculates the optimal operating point that maximizes the system's power generation efficiency by adjusting the compressor pressure ratio, turbine inlet temperature, and heat exchanger medium flow rate.

[0055] The full life-cycle efficiency prediction and optimization model compares physical property parameters with the optimal operating point over a long period to calculate the real-time degradation trend of the supercritical CO2 power generation system. When the real-time degradation trend exceeds a set threshold, corresponding preventive maintenance suggestions are generated. Simultaneously, based on the optimal operating point, efficiency optimization data for compressor pressure ratio, turbine inlet temperature, and heat exchanger medium flow rate are generated, and the preventive maintenance suggestions and efficiency optimization data are pushed to the autonomous decision-making and control module in real time. This optimizes heat exchange efficiency and circulation pressure ratio, effectively overcoming the impact of physical property fluctuations in supercritical CO2 in the near-critical region and improving power generation efficiency.

[0056] The autonomous decision-making control module receives real-time operating datasets and historical comparison datasets, performs strong coupling and decoupling calculations on the parameters of the feature datasets, and predicts the drift trend of the supercritical CO2 power generation system's operating state. Based on the prediction results, it generates control commands and constraints. Specifically, the autonomous decision-making control module incorporates an edge AI control model based on deep learning and reinforcement learning. It inputs the feature datasets into a multivariate temporal decoupling network, allocates decoupling weights for each parameter through an attention mechanism, and achieves independent decoupling calculations for strongly coupled parameters. The decoupled data is then input into a temporal prediction model based on a Long Short-Term Memory (LSTM) network to determine whether the system deviates from the critical operating range, predict the drift trend of the supercritical CO2 power generation system's operating state, and outputs the prediction results.

[0057] Furthermore, such as Figure 2 As shown, based on the prediction results of the autonomous decision-making control module, the autonomous decision-making control module performs multi-objective optimization decisions. Specifically, the multi-objective optimization decision-making steps include:

[0058] S201: Invoke the edge AI control model to construct a multi-objective reward function. The edge AI control model constructs a multi-objective reward function with the objectives of maximizing power generation efficiency, operational stability, and low operating energy consumption. It combines the efficiency optimization data output by the efficiency self-optimization module as an alignment benchmark to generate control commands and constraints. Control commands include adjusting compressor load, regulating valve opening, and heat exchanger medium flow rate; constraints are the parameter safety ranges of the actuators.

[0059] The specific calculation method for the multi-objective reward function is as follows:

[0060]

[0061] in, for The overall reward value at any given moment. , as well as The preset reward function weight coefficients, represent The real-time power generation efficiency bonus is negatively correlated with the real-time deviation of compressor load and heat exchanger efficiency in the on-site operating parameters. represent The operational stability reward item at any given time shows a negative correlation between operational stability and the variance of the pressure and temperature fluctuation curves. represent Energy-saving reward item for real-time operation (this item is positive when the actual energy consumption of the system is lower than the preset benchmark; it is negative or zero when it is higher than the benchmark). represent Penalties for exceeding safety constraints at any time.

[0062] S202: Verify the feasibility and prioritize control commands. Upon receiving a control command, the isolated execution control module first reads the safety constraint layer within the command. It determines whether the command is within the actuator's range and whether it will cause sudden changes in system parameters. If the verification is successful, the control command is prioritized and pushed to the isolated execution control module. Control commands are prioritized into primary core parameter commands and secondary auxiliary parameter commands. Primary core parameter commands include pressure and temperature, while secondary auxiliary parameter commands include flow rate and load. Control commands contain multi-dimensional attributes, categorized into execution target layer, timing attribute layer, safety constraint layer, and verification feedback layer. The isolated execution control module generates electrical signals by parsing the execution target layer, executes actions based on the priority of the timing attribute layer, performs fault tolerance judgment based on the safety constraint layer, and uses the verification feedback layer for status re-checking.

[0063] S203: Push the control instructions to the isolated execution control module. Specifically, the autonomous decision-making control module sends the generated control instructions to the isolated execution control module via the internal communication bus, and simultaneously sends them to the multimodal interaction module to trigger the digital human twin's action rehearsal before instruction execution. At the same time, the autonomous decision-making control module stores the generated control instructions, predicted states, and corresponding feature datasets in a local cache for archiving, allowing the closed-loop self-calibration and iteration module to analyze and clarify the causes of deviations by comparing actual execution effects.

[0064] The isolated execution control module receives control commands from the autonomous decision-making control module, sends these commands to the actuator, and performs real-time feedback on the actuator's status. Specifically, the isolated execution control module receives control commands from the autonomous decision-making control module via an internal communication bus, performs safety verification on the commands, and checks whether the commands exceed the actuator's preset safety boundaries. If the limits are exceeded, safety restrictions are applied. After the safety verification passes, the module generates the corresponding drive electrical signal for the actuator based on the execution target layer parsed from the control commands. The digital control target is then converted into a voltage or current output that the actuator can recognize through an opto-isolation circuit.

[0065] When executing parallel or multi-step tasks, the module executes actions based on priority tags defined in the timing attribute layer, prioritizing the execution timing of first-level core parameter (including pressure and temperature) commands, and then executing second-level auxiliary parameter (including flow and load) commands. Finally, after the electrical signal is output, the module uses the verification feedback layer to perform status feedback on the actuator's status feedback signal, compares the actual execution quantity with the target layer to see if they are consistent, and sends the result back, thereby ensuring a tight connection between the multi-dimensional attribute layer and the underlying control hardware.

[0066] Multimodal Interaction Module: Constructing a digital twin that is synchronized in real time with the supercritical CO2 power generation system to achieve multimodal interactive feedback control throughout the entire process. Specifically, constructing the digital twin includes: building a digital twin model of the supercritical CO2 power generation system, digitally mapping information such as the system's equipment layout, operating parameters, and fault status, providing an underlying industrial data environment for the construction of the digital twin; constructing a three-dimensional geometric model, behavioral logic model, and voice semantic model of the digital human entity, and establishing targeted action libraries, dialogue libraries, and interaction scenario libraries to support the operation of the digital twin.

[0067] Furthermore, the multimodal interaction module drives the digital human twin and the physical system to achieve state synchronization of less than or equal to 10ms through the edge AI computing chip, and uses the UDP real-time transmission protocol to achieve high-speed transmission of operating parameters. It establishes a 1:1 mapping relationship between the action library of the digital human twin and the device actions of the physical system to realize the visual feedback of the state and actions of the digital human twin.

[0068] When maintenance personnel issue operation commands through the interactive display unit, the multimodal interaction module captures the personnel's operational intentions, converts the commands into standardized data, and pushes it to the autonomous decision-making and control module, achieving seamless integration of manual intervention, AI decision-making, and execution by the implementing agency. Simultaneously, before the actual execution of the command, the digital twin performs an animated rehearsal, intuitively demonstrating the expected results of the operation to avoid misoperation. When a system anomaly occurs, the digital twin immediately locates the abnormal area, using human-like gesture guidance and equipment animation to achieve intuitive visualization of fault deduction.

[0069] When a supercritical CO2 power generation system experiences abnormal parameter changes or abnormal actuator operation, the digital twin, based on the hardware self-diagnosis mechanism of the multi-source sensing and acquisition module, immediately locates the abnormal part, marks the abnormal point in a visual way, and uses a visualization modeling algorithm based on fault tree analysis to transform the cause of the fault, the scope of impact, and the handling steps into gesture guidance and equipment animation of the digital twin, thereby realizing the intuitiveness of fault deduction.

[0070] The digital twin provides real-time feedback on the entire control process through a multimodal approach combining voice broadcasting, gesture guidance, and animation demonstrations. This includes the control objectives, core parameter control values, actuator actions, and system parameter change trends, thus enabling visualization of the control process.

[0071] The fault self-healing and safety management module monitors all data in real time, determines the fault level, and implements corresponding self-healing strategies based on the fault level. Specifically, this module retrieves all data from other modules in real time, monitors the system's operating status, terminal hardware status, and inter-module coordination, and constructs a multi-dimensional fault feature identification system. A fault classification and early warning model is trained based on fault cases and operating data from a supercritical CO2 power generation system. The model uses a dual criterion of parameter deviation and rate of change to determine the fault level and executes preset self-healing strategies based on the fault level.

[0072] like Figure 3 As shown, the fault self-healing and safety management module classifies fault levels based on the dual standards of parameter deviation and rate of change. The fault levels include: Level 1 warning fault: parameter deviation ≤ C1, rate of change ≤ R1. At this time, the parameters are slightly deviated and there is no risk of fault. For Level 1 warning fault, the digital human twin promptly broadcasts the warning information, and the edge AI autonomous decision control module automatically fine-tunes the control instructions without human intervention. The efficiency self-optimization module operates normally.

[0073] Level 2 self-healing fault: C1 < parameter deviation ≤ C2, R1 < rate of change ≤ R2. At this time, there is a minor fault, and the system can self-heal. For Level 2 self-healing faults, the self-healing strategy is triggered. The isolated execution control module automatically switches redundant execution channels or adjusts the control scheme. The digital twin visualizes the self-healing process. The efficiency self-optimization module is paused to prioritize system stability.

[0074] Level 3 protection fault: When parameter deviation > C2 or change rate > R2, or when core equipment malfunctions, a major fault exists and emergency handling is required. For Level 3 protection faults, the terminal immediately disconnects the dangerous loop of the isolated execution control module. The emergency shutdown procedure is initiated, and the digital twin quickly locates the fault point and demonstrates the emergency response steps. Simultaneously, the fault information is uploaded to the host computer, and the system enters safety protection mode.

[0075] It should be noted that, in this embodiment, parameter deviation refers to the percentage by which the actual value of the physical quantity collected in real time deviates from the rated design value of the system; the rate of change refers to the percentage by which the fluctuation amplitude of the physical quantity per second is relative to the rated design value; C1 and C2 are preset parameter deviation thresholds, and R1 and R2 are preset rates of change. In this embodiment, C1 is 5%, C2 is 15%; R1 is 0.5% / s, and R2 is 2% / s.

[0076] After the differentiated handling strategy is completed, information such as the time of the fault occurrence, the basis for judgment, the handling process, the handling result, and the system recovery status are archived, saved to the fault case library, and pushed to the fault feature dataset. This provides samples for the iterative training of subsequent fault early warning models and decision-making models, thereby improving the fault identification accuracy of the models.

[0077] After the fault handling is completed and the supercritical CO2 power generation system returns to normal, all suspended modules will be automatically restarted, the closed-loop management and normal operation status will be restored, and the recovery information will be fed back to the operation and maintenance personnel through the digital human twin.

[0078] The closed-loop self-calibration and iteration module analyzes the causes of deviations in terminal control effectiveness and calibrates parameters of other modules. Specifically, it retrieves control instructions and feature datasets from the local cache, compares the control targets with the actual execution results, and uses an error backpropagation algorithm to analyze deviation values ​​and identify their causes. These causes include: acquisition errors (originating from the multi-source sensing acquisition module), decision-making errors (originating from the autonomous decision-making control module's algorithm model), and execution deviations (originating from the isolated execution control module). Targeted parameter correction rules are then applied to address different causes of deviation.

[0079] Regarding acquisition errors, when the raw voltage / current sensor signals acquired by the multi-source data acquisition unit are input to the supercritical CO2 property equation for cross-validation, if the validation result shows a constant electrical drift residual, an acquisition error is determined to exist. At this point, the closed-loop self-calibration and iteration module generates a corresponding correction value based on this residual and feeds it back to the multi-source sensing acquisition module to dynamically correct its existing calibration coefficients. Through these calibration coefficients, the raw voltage / current sensor signals are directly and accurately converted into standardized parameter values ​​of actual physical quantities that can be directly validated using the property equation during subsequent acquisitions, achieving the technical effect of eliminating electrical drift in the field acquisition link and improving acquisition accuracy.

[0080] Regarding decision bias, when a strong nonlinear and drastic change in the system state causes a deviation between the control strategy output by the autonomous decision-making control module and the actual system evolution trend, the closed-loop self-calibration and iteration module uses this deviation as input to locally trigger the backpropagation algorithm. Through backpropagation calculation, the connection strength parameters of neurons in the deep learning network (including LSTM layers and attention mechanism layers) built into the autonomous decision-making control module are dynamically fine-tuned, optimizing the influence weights of different parameters on the system state in the multivariate temporal decoupling network, thereby eliminating decision bias and optimizing the decision logic.

[0081] Regarding execution deviation, an execution deviation is determined when there is a response residual between the AI ​​command output by the decision and the actual physical execution quantity fed back by the actuator. At this point, the closed-loop self-calibration and iteration module calculates the conversion ratio of the AI ​​command to the electrical signal, generates a corresponding gain compensation coefficient, and feeds it back to the isolated execution control module to correct its internal command coefficients. This command coefficient determines the proportional relationship between the control output and the physical execution quantity, ensuring that the actuator achieves a response consistent with the preset target value under different loads, reducing execution deviation and improving the accuracy of command execution.

[0082] After completing the parameter calibration, a new round of closed-loop management and control process is initiated, and the calibrated parameters are used as the benchmark for the new round of management and control.

[0083] In summary, this invention constructs a digital human twin edge AI control terminal for supercritical CO2 power generation systems. Addressing the problems of lag in control and strong parameter coupling in the near-critical region of supercritical CO2 power generation systems, as well as the dispersed hardware and inefficient interaction of traditional terminals, this invention achieves proactive prediction and real-time dynamic optimization under complex operating conditions through a deeply integrated edge AI architecture that deeply fuses perception and execution. Relying on an autonomous decision-making module to decouple strongly coupled parameters and predict states, it provides operational pre-visualization and intuitive guidance, constructing a safety barrier against misoperation. Furthermore, it coordinates a graded fault self-healing and closed-loop self-calibration mechanism, significantly reducing the operation and maintenance threshold and comprehensively ensuring the stability, safety, and overall power generation efficiency of the system throughout its entire lifecycle, demonstrating significant progress.

[0084] The above embodiments are merely illustrative of the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of this invention.

Claims

1. A digital human twin edge AI control terminal for a supercritical CO2 power generation system, characterized in that, include: Edge AI computing chips are used to carry out control logic operations and real-time rendering of digital human twins, and output control commands; The isolated interface circuit includes a multi-source data acquisition unit and an isolated control circuit unit, which are used to receive field operating parameter signals input from the multi-source sensor array and action feedback signals from the actuator, convert control commands into electrical signals that the actuator can recognize, and push them to the actuator. The internal communication bus connects to the edge AI computing chip and the isolated interface circuit, and is used to transmit field operation parameter signals, actuator action feedback signals and control commands between the edge AI computing chip and the isolated interface circuit. An interactive display unit is used to display the digital human twin interactive interface and collect touch commands; Among them, edge AI computing chips include: Multi-source sensing and acquisition module: Real-time acquisition of multi-source data and removal of abnormal data, and classification and storage of valid data in the terminal local cache; multi-source data includes operating parameters of the supercritical CO2 power generation system, working status of actuators, and terminal hardware status parameters; among which, the operating parameters of the supercritical CO2 power generation system include pressure, temperature, flow rate, density, turbine speed, compressor load, and heat exchanger heat exchange efficiency; terminal hardware status parameters include edge AI computing chip load, acquisition module accuracy, and control module communication status; Data preprocessing module: retrieves multi-source data from the local cache library and preprocesses it, then constructs a feature dataset based on the preprocessed multi-source data; Efficiency self-optimization module: Based on the feature dataset, a full life cycle efficiency prediction and optimization model is built to calculate the real-time degradation trend of the supercritical CO2 power generation system and generate preventive maintenance suggestions and efficiency optimization data; Autonomous Decision Control Module: Based on the feature dataset, it performs strong coupling and decoupling calculations on the parameters of the feature dataset to predict the drift trend of the supercritical CO2 power generation system's operating state; it generates control commands and constraints based on the prediction results; the autonomous decision control module inputs the feature dataset into a multivariate time-series decoupling network, allocates decoupling weights for each parameter through an attention mechanism, and performs independent decoupling calculations on the strongly coupled parameters; it inputs the decoupled data into a time-series prediction model based on a Long Short-Term Memory (LSTM) network to determine whether the system deviates from the critical operating range, predicts the drift trend of the supercritical CO2 power generation system's operating state, and outputs the prediction results; Isolated execution control module: Receives control instructions and constraints from the autonomous decision control module, sends control instructions to the execution mechanism, and performs real-time feedback on the status of the execution mechanism; Multimodal interaction module: Constructing a digital human twin that is synchronized in real time with the supercritical CO2 power generation system, and controlling the entire process of multimodal feedback interaction; Fault self-healing and safety management module: Real-time monitoring of data throughout the entire process, determination of fault level, and implementation of corresponding self-healing strategies based on the fault level; Closed-loop self-calibration and iteration module: used to analyze the causes of deviations in the end-to-end control effect of the terminal and to calibrate the parameters of other modules.

2. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, Abnormal data removal includes: cross-validating multi-source data based on the supercritical CO2 property equation, identifying abnormal data, marking and removing abnormal data, and storing valid data in the terminal's local cache after timestamping and format standardization; abnormal data includes drift, jump, and no response.

3. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, The data preprocessing module retrieves standardized data from the local cache library, performs preprocessing operations on the multi-source data using the supercritical CO2 property model, extracts the key feature values ​​of the preprocessed multi-source data, and constructs a feature dataset. The preprocessing operations include completing and correcting data with missing or drifting features; normalizing and compressing the dimensions of the processed multi-source data; and the key feature values ​​of the multi-source data include the abrupt change characteristics of CO2 properties near the critical region, the action response characteristics of actuators, the characteristics of system load changes, and the characteristics of fault precursors.

4. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, Based on the efficiency data from the long-term operation of the supercritical CO2 power generation system, the efficiency self-optimization module calculates the physical property parameters of supercritical CO2 in real time. Using a pre-calculated CO2 physical property database, the calculated values ​​of physical property parameters in the near-critical region are corrected through interpolation algorithms. Combined with the load changes of the supercritical CO2 power generation system, a full life cycle efficiency prediction and optimization model is constructed to generate efficiency optimization data for compressor pressure ratio, turbine inlet temperature, and heat exchanger medium flow rate, thereby optimizing heat exchange efficiency and circulation pressure ratio.

5. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, The control command contains multi-dimensional attributes, which are divided into an execution target layer, a timing attribute layer, a safety constraint layer, and a verification feedback layer. The isolated execution control module generates electrical signals by parsing the execution target layer, executes actions based on the priority of the timing attribute layer, performs fault tolerance judgment based on the safety constraint layer, and uses the verification feedback layer for status back-checking.

6. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, The entire process of multimodal feedback control interaction includes: driving the digital human twin to provide visual feedback on its status and actions based on real-time synchronization parameters; receiving external operation commands and converting them to the autonomous decision-making control module, and performing animation pre-playing through the digital human twin before the command is executed.

7. The digital human twin edge AI control terminal for a supercritical CO2 power generation system according to claim 1, characterized in that, Determining the fault level includes: training a fault classification model based on fault cases and operating data of supercritical CO2 power generation systems; the model uses a dual criterion of parameter deviation and rate of change to determine the fault level; and executing a preset self-healing strategy based on the fault level; among which, the fault levels are divided into Level 1 warning, Level 2 self-healing, and Level 3 protection.

Images