Internet of things sensing and data monitoring integrated system

Abstract

The invention relates to the technical field of industrial big data, in particular to an Internet of Things sensing and data monitoring integrated system, which comprises an Internet of Things sensing module used for acquiring a sensing parameter set of a distributed simulation unit of an intelligent power plant and a historical performance evolution trend reflecting a dynamic evolution rule; the twin mapping module is used for realizing signal space-time synchronization by utilizing a time service protocol, calling a multi-physics field evolution model to calculate thermal stress offset so as to execute numerical compensation, and generating a high-fidelity physical feature vector; the collaborative decision-making module shares observation information by using a multi-agent distributed collaborative architecture, identifies an electromagnetic-dynamic coupling response mode through a joint reward function and outputs a collaborative optimization strategy; and the execution module is used for generating an evaluation result by calculating the numerical deviation between the physical feature vector and the design envelope, and adjusting the excitation control variable according to a strategy until the system converges to a target equilibrium value. According to the method, the steady state monitoring of the power system is realized through twin mapping and collaborative decision.

Description

Technical Field

[0001] This invention relates to the field of industrial big data technology, specifically to an integrated system for IoT sensing and data monitoring. Background Technology

[0002] With the advancement of digital transformation in the energy industry, distributed simulation and real-time operation status monitoring technologies for smart power plants have become core means to ensure the safe and stable operation of complex power systems. During production operations, distributed simulation units need to continuously acquire multimodal characteristic indicators reflecting system operating efficiency and structural response. Through real-time monitoring and analysis of these indicators, an in-depth assessment of equipment operating health can be achieved.

[0003] Due to the extreme complexity of the internal logical topology of smart power plants and the high-frequency dynamic interaction of multiple physical fields involved in the operation of large units, the data generated by the monitoring system exhibits significant characteristics of industrial big data: on the one hand, the sensing signals are massive in scale, have extremely high sampling frequency, and exhibit significant temporal heterogeneity; on the other hand, during the dynamic evolution of the system, there are high cross-domain correlations and nonlinear interferences between the state variables of different functional units within the system, resulting in extremely complex transient evolution characteristics in the data distribution.

[0004] In existing technological applications, monitoring systems typically utilize IoT sensing nodes to collect underlying signals and transmit them to a monitoring platform for processing. However, under the multi-source boundary constraints of actual unit operation, how to ensure that these massive, heterogeneous, and interconnected sensing signals can achieve high-precision spatiotemporal alignment and logical mapping with the nonlinear evolution process of physical entities, so as to accurately characterize the transient response and long-term evolution trend of physical entities in the digital space, is a key technical problem currently faced in the field of integrated sensing and monitoring of smart power plants.

[0005] To address this, an integrated system for IoT sensing and data monitoring is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide an integrated IoT sensing and data monitoring system that achieves steady-state monitoring of a power system through twin mapping and collaborative decision-making. It includes an IoT sensing module for acquiring the set of sensing parameters from a distributed simulation unit of a smart power plant and historical performance evolution trends reflecting dynamic evolution laws; a twin mapping module for achieving spatiotemporal synchronization of signals using a timing protocol and calling a multi-physics evolution model to calculate thermal stress offsets for numerical compensation, generating high-fidelity physical feature vectors; a collaborative decision-making module for sharing observation information using a multi-agent distributed collaborative architecture, identifying electromagnetic-dynamic coupling response modes through a joint reward function, and outputting collaborative optimization strategies; and an execution module for generating evaluation results by calculating the numerical deviation between the physical feature vectors and the design envelope, and adjusting incentive control variables according to the strategy until the system converges to the target equilibrium value.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An integrated IoT sensing and data monitoring system includes: IoT sensing module: acquires the state parameters of the distributed simulation unit of the smart power plant, including initial characteristic frequency, real-time temperature, real-time power factor, real-time speed, rated load demand, identification code and historical performance evolution trend; Twin mapping module: Maps state parameters to simulation nodes in the digital twin space based on the identification code; calls the evolution model to calculate thermal stress offset and performs numerical compensation to generate physical feature vectors; Collaborative decision-making module: Input the physical feature vector, real-time speed, real-time power factor and historical performance evolution trend into the multi-agent model, define the distributed simulation unit as a dynamic simulation agent, identify the electromagnetic-dynamic coupling response mode through the sharing of observation information among agents and the constraint of joint reward function, and output the collaborative optimization strategy; Execution module: Calculates the numerical deviation between the physical feature vector and the design envelope limit to generate monitoring and evaluation results. When the evaluation results reach the preset steady-state threshold, it extracts the real-time power factor based on the control parameters, issues instructions to the actuator through the supervisory control logic to adjust the excitation control variables, adjust the torque output, until the real-time speed and operating parameters reach the target equilibrium value corresponding to the collaborative optimization strategy.

[0008] Preferably, the process of obtaining the state parameters of the distributed simulation unit of the smart power plant includes: using sensors deployed at the monitoring points of the physical equipment corresponding to the distributed simulation unit to capture the underlying physical signals corresponding to the initial characteristic frequency, real-time temperature, real-time speed, and real-time power factor; retrieving the logical topology identification code bound to the physical equipment from the system configuration library, and retrieving the historical operating data sequence of the corresponding period from the time series database based on the logical topology identification code; performing data standardization processing on the historical operating data sequence, and using a time sliding window to extract the discrete deviation distribution of each state parameter relative to the preset design benchmark value; analyzing the slope of the numerical evolution of the discrete deviation distribution within the time sliding window, and extracting the stable component of the numerical evolution slope as the historical performance evolution trend reflecting the dynamic evolution law of the unit.

[0009] Preferably, the process of mapping to the simulation node in the digital twin space includes: calling the built-in time protocol of the supervision, control and data acquisition system to align the asynchronous sensing signals obtained through the IoT data acquisition interface to a unified system master clock frequency; encapsulating timestamps for each aligned asynchronous sensing signal; using the timestamps as index keys to aggregate the discrete sampling points in the state parameters into a correlated monitoring dataset that is synchronously associated in the time dimension; extracting the logical topology identification code and retrieving the preset virtual model index in the digital twin space to establish a ternary mapping binding relationship consisting of timestamps, state parameters and virtual model indexes; and pushing the state parameters to the corresponding simulation node in the digital twin platform in real time according to the ternary mapping binding relationship to realize the spatiotemporal synchronous display of the real-time operating conditions of the physical equipment in the three-dimensional virtual space.

[0010] Preferably, the evolution model includes: a multi-source boundary integration layer: acquiring real-time temperature, rated load demand, centrifugal load generated by real-time rotation speed and fluid pressure components as composite boundary constraints, and combining them with the pre-set structural material parameters of the simulation unit to construct a multi-physics simulation environment in the digital twin space; Thermo-mechanical coupling stress evolution layer: Identify the temperature field gradient and stress field distribution of the simulation unit in the multiphysics simulation environment, and calculate the structural thermal stress offset induced by temperature rise, load and pressure based on the thermal expansion properties of materials and elasticity mechanism. Nonlinear stiffness correction layer: Analyzes the influence of structural thermal stress offset on the equivalent stiffness and dynamic damping characteristics of the unit physical structure, and calculates the numerical drift of the initial characteristic frequency based on the nonlinear change of the equivalent stiffness, and performs deviation compensation for the initial characteristic frequency. High-fidelity feature synthesis layer: The compensated feature frequencies are fused with the historical performance evolution trend reflecting the service status of the material to output the physical feature vector.

[0011] Preferably, the step of inputting the physical feature vector, real-time rotational speed, real-time power factor, and historical performance evolution trend into the multi-agent model includes: concatenating the physical feature vector, real-time rotational speed, real-time power factor, and historical performance evolution trend according to a preset time sequence to construct a multi-dimensional state tensor corresponding to each simulation unit; using the communication bus of the supervision control and data acquisition system, synchronously sharing the multi-dimensional state tensor among the controllers of each simulation unit through register address mapping; each simulation unit extracts the correlation feature terms reflecting the mechanical vibration and electrical fluctuations of adjacent units based on the received multi-dimensional state tensors of adjacent units, and merges them with its own multi-dimensional state tensor to generate a joint input matrix for strategy generation by the dynamic simulation agent.

[0012] Preferably, the process of the output collaborative optimization strategy includes: setting the calculation components of the joint reward function, including: using the deviation value of the output active power relative to the rated load as the reward weight, and using the value of the vibration amplitude in the physical feature vector exceeding a preset safety threshold as the penalty weight; by comparing the phase fluctuation curve of the real-time power factor with the frequency offset feature curve in the physical feature vector, and calculating the cross-correlation coefficient, if the cross-correlation coefficient exceeds a preset threshold, it is identified as the existence of an electromagnetic-dynamic coupling response mode that causes speed fluctuation, and a sensitivity mapping relationship between electromagnetic torque and mechanical speed is established; according to the highest score guidance of the joint reward function, the excitation current correction increment step size of the excitation regulator controller is calculated, and the correction increment step size is sent to the control register of the excitation regulator controller as the collaborative optimization strategy.

[0013] Preferably, the process of generating monitoring and evaluation results includes: retrieving a multidimensional design envelope preset in the digital twin space, and calculating the Euclidean distance of the physical feature vector relative to the boundary of the design envelope at the current moment; defining the Euclidean distance as the numerical deviation, and calculating the variance volatility of the numerical deviation within a preset statistical time window; if the numerical deviation exceeds a preset safety alarm limit and / or the variance volatility is less than a preset convergence constant, then the evaluation result is determined to have reached a preset steady-state threshold, and the supervisory control logic is triggered.

[0014] Preferably, the process of issuing instructions to the actuator through the supervisory control logic includes: retrieving the corresponding excitation correction coefficient from a preset adjustment characteristic lookup table based on the real-time power factor, and calculating the target adjustment increment of the excitation control variable; using the communication protocol of the supervisory control and data acquisition system, sending a control message containing the target adjustment increment to the instruction register of the actuator to drive the excitation source to change the input energy intensity of the distributed simulation unit; monitoring the rotational speed change of the distributed simulation unit in real time, and using the target equilibrium value corresponding to the collaborative optimization strategy as the feedback target, adjusting the excitation control variable in a step-by-step cyclic manner to optimize the electromechanical energy conversion characteristics, so that the dynamic response of the distributed simulation unit converges to the collaborative optimization strategy, until the composite residual of the real-time rotational speed, real-time power factor and their respective target equilibrium values ​​enters a preset allowable error range.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By constructing a multiphysics simulation environment and performing nonlinear stiffness correction, the evolution of physical characteristics of the simulation unit under varying working conditions can be accurately captured. The thermal-mechanical coupling analysis is used to identify the thermal stress shift of the structure, and numerical deviation compensation is performed for the characteristic frequency. This effectively improves the matching accuracy between the physical feature vector and the actual operating state of the equipment, significantly enhances the ability of the digital twin space to characterize the service state of materials under complex working conditions, and improves the modeling accuracy of the monitoring system when dealing with nonlinear dynamic drift.

[0016] 2. By introducing a multi-agent distributed collaborative architecture and joint reward function constraints, the deep integration of the generator set's electromagnetic torque and dynamic response is achieved. The electromagnetic-dynamic coupling response mode is identified by calculating the cross-correlation coefficient, which can sensitively detect micro-speed fluctuation causes that are difficult to detect by traditional monitoring methods. Based on this, the optimal excitation current correction increment is generated. While ensuring the stability of the output active power, the mechanical vibration safety threshold is also taken into account. Through the dynamic weight balancing of multiple objectives, the overall operational stability and collaborative operation efficiency of the unit in the complex power grid fluctuation environment are optimized.

[0017] 3. By applying a precise time protocol and a three-dimensional mapping binding mechanism, deep synchronization of multiple asynchronous sensing signals in the time and spatial dimensions is achieved, aligning each discrete sampling point and establishing an association mapping between timestamps, state parameters, and virtual model indexes. This ensures the spatiotemporal consistency of physical real-time operating conditions in the digital twin platform, effectively mitigating the impact of latency jitter during cross-system data acquisition. It provides a precise, correlated, and time-series-logical structured monitoring dataset for real-time operating condition monitoring, multi-dimensional data tracing, and three-dimensional visualization of smart power plants. Attached Figure Description

[0018] Figure 1This is a schematic diagram of the structure of an integrated IoT sensing and data monitoring system according to the present invention; Figure 2 This is a schematic diagram of the process for generating the joint input matrix according to the present invention; Figure 3 This is a flowchart illustrating the output collaborative optimization strategy of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1 to 3 This invention provides an integrated system for IoT sensing and data monitoring, the technical solution of which is as follows: Example 1: An integrated IoT sensing and data monitoring system, with the following specific structure: Figure 1 As shown, it includes: IoT sensing module: acquires the state parameters of the distributed simulation unit of the smart power plant, including initial characteristic frequency, real-time temperature, real-time power factor, real-time speed, rated load demand, identification code and historical performance evolution trend; Twin mapping module: Maps state parameters to simulation nodes in the digital twin space based on the identification code; calls the evolution model to calculate thermal stress offset and performs numerical compensation to generate physical feature vectors; Collaborative decision-making module: Input the physical feature vector, real-time speed, real-time power factor and historical performance evolution trend into the multi-agent model, define the distributed simulation unit as a dynamic simulation agent, identify the electromagnetic-dynamic coupling response mode through the sharing of observation information among agents and the constraint of joint reward function, and output the collaborative optimization strategy; Execution module: Calculates the numerical deviation between the physical feature vector and the design envelope limit to generate monitoring and evaluation results. When the evaluation results reach the preset steady-state threshold, it extracts the real-time power factor based on the control parameters, issues instructions to the actuator through the supervisory control logic to adjust the excitation control variables, adjust the torque output, until the real-time speed and operating parameters reach the target equilibrium value corresponding to the collaborative optimization strategy.

[0021] Furthermore, the process of acquiring the state parameters of the distributed simulation unit of the smart power plant includes: using sensors deployed at the monitoring points of the physical equipment corresponding to the distributed simulation unit to capture the underlying physical signals corresponding to the initial characteristic frequency, real-time temperature, real-time speed, and real-time power factor; retrieving the logical topology identification code bound to the physical equipment from the system configuration library, and retrieving the historical operating data sequence of the corresponding period from the time series database based on the logical topology identification code; performing data standardization processing on the historical operating data sequence, and extracting the discrete deviation distribution of each state parameter relative to the preset design benchmark value using a time sliding window; analyzing the slope of the numerical evolution of the discrete deviation distribution within the time sliding window, and extracting the stable component of the numerical evolution slope as the historical performance evolution trend reflecting the dynamic evolution law of the unit.

[0022] Specifically, sensor arrays deployed at key locations of the rotating power conversion unit capture underlying physical signals in real time. Piezoelectric vibration sensors are installed at the bearing support locations of the unit to record the mechanical vibration characteristics of the shaft system at a sampling frequency of 20,000 times per second. By performing discrete transformation analysis on the raw vibration data, the reference vibration frequency of the unit under no-load operation is determined as the initial characteristic frequency. Simultaneously, the real-time temperature of the core components of the unit is obtained through thermistor temperature sensors embedded in the stator housing and cooling circuit. At the end of the rotating spindle, a laser pulse sensor measures the angular velocity of the spindle and converts it into real-time rotational velocity. Furthermore, by connecting to the unit's power output bus, the phase difference between current and voltage is collected using digital power monitoring instruments, and the real-time power factor is calculated to characterize the unit's load characteristics during the electromagnetic energy conversion process. The rated load demand is the target output power command parameter corresponding to the distributed simulation unit. Preferably, it is retrieved by the edge control unit from the preset peak-shaving operation plan table in the plant-level dispatching system or the unit control system as the target load reference value within the time sliding window. When not connected to an external dispatching system, the rated load demand can be replaced by the rated power on the unit nameplate recorded in the system configuration library.

[0023] After completing the acquisition of the underlying signals, the logic identification and historical data association logic is initiated. The edge control unit first retrieves the logical topology identification code bound to the current physical device hardware address from the locally preset device configuration library. This identification code is not only the unique identifier of the unit in the smart power plant production network, but also defines its hierarchical relationship in the dynamic simulation topology. The system uses this identification code as the query key to automatically access the time series database in the plant area and retrieve the historical operating data sequence of the specific unit for 168 consecutive hours before the current sampling point. This sequence covers the unit's speed, temperature and energy output trajectory during the complete peak shaving cycle, providing a time domain reference for subsequent performance degradation analysis.

[0024] To eliminate numerical differences between different physical dimensions such as speed, temperature, and power factor, the system employs a standard deviation scaling algorithm to uniformly map all parameters to a standard interval with a mean of 0 and a variance of 1, generating a dimensionless feature flow. Subsequently, the system initiates a time sliding window program, setting the window length to 600 seconds and the sliding step size to 10 seconds. Within each sliding window period, the system compares the real-time sensor observations with the design reference values ​​(i.e., the theoretical ideal parameters of the unit under rated full-load conditions) stored internally. By calculating the degree of deviation of the real-time values ​​from the design reference values, a discrete deviation distribution reflecting the deviation characteristics of the unit's operating conditions is formed within the window.

[0025] Finally, the least squares linear fitting method is used to calculate the slope of the deviation value change over time within the sliding window. In order to eliminate random noise caused by instantaneous load fluctuations or measurement environment interference, the system performs first-order hysteresis filtering on the fitted slope sequence and extracts the numerical component with continuous monotonicity as the stable component. This stable component is defined as the historical performance evolution trend, and its magnitude and direction intuitively reflect the physical degradation rate of the rotating power conversion unit during long-term service. This trend parameter is then encapsulated and transmitted to the digital twin platform as the core weight for correcting material fatigue and structural stiffness in the simulation node.

[0026] By capturing signals at the bottom layer and mapping logical recognition, the association between physical units and digital nodes is realized. By using data standardization and trend extraction logic, the evolution law of the equipment relative to the benchmark value is quantified, providing objective data with performance degradation characteristics, improving the data consistency of the monitoring process, and providing reliable support for the dynamic evaluation of the unit status.

[0027] Furthermore, the process of mapping to the simulation node in the digital twin space includes: invoking the built-in time protocol of the supervision, control, and data acquisition system to align the asynchronous sensing signals obtained through the IoT data acquisition interface to a unified system master clock frequency; encapsulating timestamps for each aligned asynchronous sensing signal; using the timestamps as index keys to aggregate the discrete sampling points in the state parameters into a synchronously associated monitoring dataset in the time dimension; extracting the logical topology identification code and retrieving the preset virtual model index in the digital twin space to establish a ternary mapping binding relationship consisting of timestamps, state parameters, and virtual model indexes; and pushing the state parameters to the corresponding simulation node in the digital twin platform in real time according to the ternary mapping binding relationship to achieve spatiotemporal synchronous display of the real-time operating conditions of the physical equipment in the three-dimensional virtual space.

[0028] During the signal time alignment phase, the system first invokes the precise time protocol built into the supervisory control and data acquisition system. This protocol provides a time reference for the distributed simulation unit through the fiber optic network. Since the vibration, temperature, speed and electrical signal acquisition frequencies on site are different and there is an asynchronous phenomenon, the system uses the unified master clock frequency established by this protocol to perform interpolation or downsampling processing on each asynchronous sensing signal. Through this operation, the acquisition signals that were originally out of phase are aligned to the same time reference, ensuring the logical consistency of multi-source monitoring data under the same instantaneous physical conditions.

[0029] During the data aggregation phase, the system encapsulates timestamps for each aligned sensing signal. Using these timestamps as globally unique index keys, the system aggregates discrete sampling points distributed across different acquisition channels. For example, it packages the rotational speed pulse, bearing temperature, and power factor corresponding to a certain moment into a data tuple, thereby forming a correlated monitoring dataset that is synchronously associated in the time dimension. This time-linked aggregation method solves the signal delay jitter problem commonly found in large-scale distributed monitoring and provides high-precision data packets for subsequent simulation and deduction.

[0030] In the ternary mapping binding stage, the logical topology identification code obtained in the previous steps is extracted and retrieved into the virtual model database of the digital twin space. The system automatically matches the virtual model index corresponding to the identification code. This index points to a specific geometric entity node in the three-dimensional space. Subsequently, the system establishes a ternary mapping binding relationship consisting of a timestamp, real-time status parameters, and a virtual model index. This relationship is stored in the memory database in the form of structured data, realizing the logical coupling of the physical entity's operating state, time dimension, and virtual three-dimensional coordinates.

[0031] During the real-time synchronous display phase, the system pushes the associated monitoring dataset to the digital twin platform through the industrial real-time data bus according to the preset three-dimensional mapping binding relationship. After receiving the data, the simulation nodes in the platform immediately update the corresponding three-dimensional model attributes. In this embodiment, the real-time temperature is mapped to the color gradient change of the geometric surface, and the real-time rotation speed is mapped to the rotational angular velocity of the three-dimensional entity. This process realizes the spatiotemporal synchronous display of the real-time operating conditions of the physical equipment in the three-dimensional virtual space, enabling managers to intuitively observe the operation and evolution of the power conversion unit.

[0032] By introducing a precise time protocol and a ternary mapping mechanism, this module solves the problem of time-domain inaccuracy in the acquisition of multi-source heterogeneous signals, realizes data alignment and aggregation, and ensures a high degree of consistency between the physical entity's operating conditions and the virtual simulation nodes in the time dimension by establishing a binding relationship between time, parameters and models. This improves the real-time feedback performance of the digital twin system and provides reliable technical support for the remote monitoring and spatiotemporal simulation of complex power units.

[0033] Furthermore, the evolution model includes: a multi-source boundary integration layer: acquiring real-time temperature, rated load requirements, centrifugal load generated by real-time rotation speed, and fluid pressure components as composite boundary constraints, and combining them with the pre-set structural and material parameters of the simulation unit to construct a multi-physics simulation environment in the digital twin space; Thermo-mechanical coupling stress evolution layer: Identifies the temperature field gradient and stress field distribution of the simulation unit in the multiphysics simulation environment, and calculates the thermal stress offset of the structure induced by temperature rise, load and pressure based on the thermal expansion properties of the material and the elasticity mechanism. Nonlinear stiffness correction layer: Analyzes the influence of structural thermal stress offset on the equivalent stiffness and dynamic damping characteristics of the unit physical structure, and calculates the numerical drift of the initial characteristic frequency based on the nonlinear change of the equivalent stiffness, and performs deviation compensation for the initial characteristic frequency. High-fidelity feature synthesis layer: The compensated feature frequencies are fused with the historical performance evolution trend reflecting the service status of the material to output the physical feature vector.

[0034] Specifically, the system first launches the dynamic simulation engine in the digital twin platform, retrieving the preset structural material parameters of the simulation unit, including the material's elastic modulus, Poisson's ratio, coefficient of thermal expansion, and mass density. Then, the system acquires real-time temperature, real-time rotational speed, and fluid pressure of the cooling medium through the IoT sensing module. The system divides the geometric model of the power unit into discrete finite element meshes and applies the acquired real-time temperature as a thermal load boundary condition to the mesh nodes. Simultaneously, based on the real-time rotational speed and the mass distribution of the rotating components, the system calculates the angular velocity vector of each mesh element and converts it into a centrifugal inertial force load acting at the center of mass. Combining the torque boundary generated by the rated load demand and the fluid pressure distribution, the system establishes a multiphysics simulation environment with composite boundary constraints on the three-dimensional mesh model, providing a numerical benchmark for subsequent stress solutions.

[0035] In the simulation environment, the system first uses the heat conduction balance equation to solve for the temperature field gradient distribution inside the dynamic unit. Based on the thermal expansion properties of the material, the system calculates the free thermal strain generated by temperature rise in each grid unit. Due to the mechanical constraints on the unit structure, this thermal strain cannot be completely released. Based on the generalized Hooke's law in elasticity, the system transforms the confined thermal strain into an equivalent thermal stress component. Simultaneously, the system calculates the mechanical stress tensor generated by centrifugal load and fluid pressure. Through a numerical iterative algorithm, the system vectorically superimposes the thermal stress component and the mechanical stress tensor to calculate the structural thermal stress offset. This offset quantitatively characterizes the internal potential energy reorganization caused by temperature rise and load fluctuations, and serves as the fundamental data for subsequent correction of the structural dynamic characteristics.

[0036] The system further analyzes the nonlinear influence of structural thermal stress offset on the equivalent stiffness of the unit physical structure. In this embodiment, the system extracts the stress tensor and constructs a geometric stiffness matrix related to the current stress state. The system then performs matrix summation on this geometric stiffness matrix and the linear elastic stiffness matrix under the initial state to obtain the corrected total structural stiffness matrix. Subsequently, the system uses a modal analysis algorithm to solve for the eigenvalues ​​of the corrected total stiffness matrix and extracts its frequency values ​​of a specific order. The system performs a difference operation between this value and the initial characteristic frequency under no-load conditions to obtain the numerical drift of the initial characteristic frequency induced by stress hardening or softening effects. Finally, the system performs deviation compensation by superimposing this drift on the initial characteristic frequency to generate a dynamically compensated characteristic frequency reflecting the real-time stress state.

[0037] After obtaining the compensated characteristic frequency, the system enters the feature synthesis stage, which aims to transform physical parameters with different dimensions into a unified physical feature vector. The system first receives three core inputs: the compensated characteristic frequency, real-time stress distribution parameters, and historical performance evolution trends reflecting long-term degradation. Since the numerical units of these parameters (such as Hertz, Megapascals, and evolution slope) are not uniform, the system adopts deviation standardization processing. Specifically, the system subtracts the preset health benchmark value within the system from the real-time observed value of each parameter, and then divides it by the maximum fluctuation range recorded in the historical database for that parameter. Through this operation, all input parameters are transformed into dimensionless values ​​between 0 and 1, ensuring that physical quantities of different properties are compared in the same dimension. In this embodiment, the physical feature vector includes at least the compensated characteristic frequency deviation sensitivity component, the typical stress distribution statistical component, and the standardized deviation component of the historical performance evolution trend. The vector dimension can be configured between 3 and several dimensions according to monitoring needs, but all can be directly determined by those skilled in the art based on the above processing flow.

[0038] The system determines the deviation sensitivity of each parameter by comparing the difference between the standardized parameters and the ideal state value (i.e., the value 0). This sensitivity quantitatively represents the degree to which a specific physical index deviates from the normal operating trajectory. For example, when thermal stress concentration occurs in the unit structure, causing a sharp drift in the compensated characteristic frequency, the corresponding deviation sensitivity will increase significantly. The system uses this sensitivity value as an input vector to drive the dynamic inference of the subsequent weight prediction model.

[0039] The system pre-configures a weight prediction model based on a multilayer perceptron structure to output the optimal weight allocation scheme for the current operating condition. The model's input layer contains three nodes, corresponding to the aforementioned standardized features; the hidden layer contains eight neurons with non-linear activation functions; and the output layer contains three nodes, corresponding to the dynamic weight coefficients of the three types of features. To ensure the model has accurate weight allocation capabilities, this embodiment discloses the model's training process: First, the system retrieves 50,000 sample sequences covering four typical operating conditions—"steady-state operation," "peak-shaving load variation," "material fatigue," and "typical structural instability"—from the historical operating database as a training set. Second, a reverse evaluation method is used for sample labeling. That is, in historical samples where performance evolution is known, sensitivity analysis is used to determine which type of parameter first captures the abnormal precursor, and the ideal weight of that parameter at the current moment is marked as a high value (e.g., above 0.7), thereby generating labeled training pairs. Finally, the model weights are iteratively optimized using the backpropagation algorithm. When the model's prediction accuracy on the validation set reaches over 95% and the loss function tends to stabilize, the model is considered to have converged and training is complete.

[0040] During real-time monitoring, the system inputs the calculated deviation sensitivity into the trained weight prediction model. The model calculates through forward propagation and outputs three dynamic weight coefficients for the current instantaneous working condition in real time, ensuring that the sum of the three is always equal to 1. For example, when the model identifies that the current frequency deviation sensitivity is high and conforms to the stress-induced frequency drift mode, it will automatically increase the dynamic weight coefficient of the characteristic frequency to 0.6, while proportionally reducing the weights of other parameters.

[0041] After determining the dynamic weight coefficients, the system performs the final synthesis operation. Each standardized parameter obtained is multiplied by its corresponding dynamic weight coefficient derived in real-time, and the three products are arranged in order to construct a multi-dimensional numerical matrix, i.e., a high-fidelity physical feature vector. This synthesis method ensures that when a power unit exhibits abnormal tendencies, the corresponding abnormal features will occupy a larger numerical proportion in the feature vector, thereby significantly enhancing the subsequent collaborative decision-making module's ability to identify equipment fault characteristics. This vector is then encapsulated into a standardized data package and output to the digital twin platform.

[0042] The high-fidelity feature synthesis layer also includes parameter self-calibration logic: extracting the real state sequence of the physical entity after the actuator's action, and calculating the transient response residual between it and the physical feature vector; using variational inference algorithm to analyze the deviation characteristics of the transient response residual in the probability distribution dimension, and identifying the drift increment of material physical parameters or boundary constraints in the multiphysics evolution model; using the drift increment as feedback bias to correct the operator weights in the nonlinear stiffness correction layer in real time, thereby realizing the dynamic tracking of the degradation of physical properties in the digital space.

[0043] When the execution module, based on the strategy output by the collaborative decision, issues commands to the actuator (such as the excitation regulator) through the supervisory control logic and adjusts the excitation control variables (such as increasing the excitation current step size) to change the energy input intensity of the distributed simulation unit, the system uses the physical disturbance generated by this action as an active excitation signal to start the calibration procedure. The system utilizes bearing vibration sensors and speed encoders deployed at key locations on the generator set to capture in real-time the actual state sequence of the physical entity within a preset sampling window after the execution action (including transient speed fluctuations and shaft vibration response). Subsequently, the system synchronously extracts the theoretical physical feature vector output by the twin mapping module through the multiphysics evolution model within the same time step. Since this feature vector represents the model's predicted response under the current parameters, the system performs point-by-point comparison calculations between the measured sequence and the predicted curve to obtain the transient response residual. This residual intuitively quantifies the simulation deviation of the digital twin space for the transient response of the physical entity in both the time and frequency domains.

[0044] Considering the complex random measurement noise and transient load interference in industrial environments, the system introduces a variational inference algorithm to deeply decouple the aforementioned residuals. The algorithm approximates the true posterior probability distribution of the unit's implicit physical variables (such as equivalent stiffness coefficients and structural damping factors) by constructing an analytical variational distribution (e.g., a mean field distribution). During processing, the system analyzes the offset characteristics of the residual sequence in the probability distribution dimension. By minimizing the difference between the variational distribution and the true feedback (i.e., maximizing the lower bound of evidence), it effectively distinguishes which deviations belong to random fluctuations of the sensors and which belong to persistent physical degradation. Through this probabilistic inference, the system can accurately identify the drift increments of material physical parameters (e.g., the decrease in equivalent stiffness caused by rotor material fatigue) or boundary constraints (e.g., the deviation in stress distribution caused by changes in cooling fluid pressure) in the multiphysics evolution model.

[0045] The system uses the identified physical property drift increments as feedback biases, transmitting them in real-time to the nonlinear stiffness correction layer in the twin mapping module. The system has a pre-set set of sensitivity mapping relationships to quantify the impact of different physical parameter variations on the overall structural stiffness matrix. Based on the drift increment values, the system automatically recalibrates the operator weights in the nonlinear stiffness correction layer. Specifically, the system adjusts the balancing coefficients used to balance linear deformation and nonlinear response in the overall structural stiffness matrix calculation logic, ensuring that the compensation calculation for the initial characteristic frequency automatically includes the current physical property degradation values. Through the corrected operator weights, the physical feature vectors output by the high-fidelity feature synthesis layer can be highly aligned with the actual operating conditions of the physical entity. This process enables real-time dynamic tracking of physical property degradation in digital space, ensuring the objectivity and accuracy of monitoring and evaluation results throughout the unit's entire lifecycle.

[0046] By constructing a multiphysics simulation environment and introducing nonlinear stiffness correction logic, the impact of temperature rise and load fluctuation on structural dynamics was effectively quantified. Dynamic compensation of physical properties was achieved through the superposition of geometric stiffness matrices and eigenvalue solving. Feature fusion using a deep learning-based weighted prediction model improved the system's accuracy and discriminative power in identifying early signs of equipment anomalies, providing high-fidelity physical data support for subsequent precise decision-making.

[0047] Further, the step of inputting the physical feature vector, real-time rotational speed, real-time power factor, and historical performance evolution trend into the multi-agent model includes: concatenating the physical feature vector, real-time rotational speed, real-time power factor, and historical performance evolution trend according to a preset time sequence to construct a multi-dimensional state tensor corresponding to each simulation unit; using the communication bus of the supervision control and data acquisition system, synchronously sharing the multi-dimensional state tensor among the controllers of each simulation unit through register address mapping; each simulation unit extracts the correlation feature terms reflecting the mechanical vibration and electrical fluctuations of adjacent units based on the received multi-dimensional state tensors of adjacent units, and merges them with its own multi-dimensional state tensor to generate a joint input matrix for strategy generation by the dynamic simulation agent. The specific process is as follows: Figure 2 As shown.

[0048] Specifically, in the process of constructing the multidimensional state tensor, the system performs data standardization and structured reconstruction through the main control CPU of the local controller. In order to eliminate the influence of different physical dimensions on model training, the system performs normalization processing on the input parameters: the original values ​​such as real-time speed, real-time power factor and characteristic frequency are linearly mapped to the numerical range of 0 to 1 based on the maximum and minimum values ​​in the historical sequence of the past 168 hours. The system presets a sliding time window with a length of 50 sampling periods, and concatenates the normalized physical feature vector (length D), real-time speed, real-time power factor and historical performance evolution trend in the time dimension according to the sampling order to construct a data matrix with a dimension of 50 rows * D + 3 columns. In order to adapt to the deep reinforcement learning algorithm, the system further adds the simulation unit ID as the index dimension, and finally generates a three-dimensional state tensor with a dimension of 1 * 50 * total number of features. This tensor not only records the current instantaneous physical state, but also implies the dynamic trajectory of state evolution through the stacking of features in the time dimension, providing a complete data foundation for the model to identify transient impacts.

[0049] During the distributed data sharing phase, considering the bandwidth limitations of the industrial SCADA system communication bus, the system adopts a register mapping mechanism based on quantization compression. Due to the large amount of original tensor data, the system converts it into a 16-bit short integer sequence before sending to reduce the communication load. Each controller (such as a PLC or DCS) is allocated a continuous range of 256 register addresses in the communication memory. The system uses read and write instructions of Profinet or Modbus TCP protocol to write the multidimensional state tensor summary (i.e., key feature points) of the unit to the input register range of the adjacent unit in real time in a peer-to-peer manner. In order to solve the communication latency problem between heterogeneous systems, the system sets up a ping-pong buffer with a depth of 3 frames inside the controller to ensure that the time deviation between the data of the adjacent unit and the data of the unit is controlled within 10 milliseconds during collaborative simulation, thereby ensuring the real-time performance of electromagnetic-dynamic collaborative control at the hardware level.

[0050] During the generation of the joint input matrix, the agent of each unit processes the received data from neighboring units through the built-in sliding window cross-correlation operator. Specifically, the operator calculates the Pearson correlation coefficient between the unit's speed fluctuation sequence and the power factor fluctuation sequence of neighboring units, thereby quantifying the intensity of electromagnetic interference generated by neighboring units on the unit. The system then performs a matrix merging operation: taking the 50*feature dimension matrix of the unit as the benchmark, it performs a fast Fourier transform on the received multidimensional state tensors of neighboring units to analyze the energy distribution of each order of harmonic components to extract the current harmonic ratio, and uses the sliding window standard deviation algorithm to calculate the dispersion of the real-time vibration sequence relative to the design benchmark value to extract the vibration amplitude deviation. Then, these correlation feature terms, which have undergone frequency domain processing and statistical quantization, are concatenated column by column after the benchmark matrix to generate the dimension-expanded joint input matrix. This process ensures that the input data not only includes the operating conditions of a single unit, but also couples the topological environmental features of the entire power grid, enabling the agent to perceive nonlinear interference signals transmitted by shared buses or physical bases.

[0051] Regarding the internal logic implementation of the multi-agent model, this embodiment adopts a deep reinforcement learning architecture with centralized training and distributed execution. During the offline training phase, the system pre-sets a joint reward function, which is mathematically defined as follows: taking the absolute value of the deviation between the output active power and the rated load as the independent variable, constructing a reward weight term based on a negative exponential decay function as the positive reward component, specifically, the reward value smoothly converges to the maximum value of 1 as the deviation decreases. By introducing an exponential mapping, the numerical singularity and model gradient explosion risk caused by directly taking the reciprocal when the deviation tends to zero are avoided. The value of the vibration amplitude in the physical feature vector that exceeds the safety threshold is used as the negative penalty component. The weight ratio of reward to penalty is set at 0.6 to 0.4 to ensure that the system prioritizes mechanical safety while pursuing economic operation. The policy network in the model adopts a three-layer convolutional neural network structure, which extracts spatial features from the joint input matrix through convolution kernels. During training, the system sets the learning rate between 0.0001 and 0.001 and uses an experience playback mechanism to store 100,000 sets of historical working condition samples until the variance volatility of the reward function is less than the preset convergence constant (e.g., 0.01) and no longer decreases for 100 consecutive iterations. The model is then judged to have reached the optimal convergence state. During the field execution phase, each unit controller only needs to run the lightweight policy network inference part to output the corresponding excitation current correction increment within 50 milliseconds based on the real-time joint input matrix, thereby realizing the autonomous collaborative optimization of distributed units.

[0052] By constructing a multidimensional state tensor and using register address mapping to achieve data sharing, real-time synchronization and interoperability of key operating indicators among distributed units are ensured. By merging the mechanical vibration and electrical fluctuation characteristics of adjacent units to generate a joint input matrix, the dynamic simulation agent can effectively identify cross-unit nonlinear interference transmitted by shared bus or physical base. This provides globally correlated data support for the generation of collaborative strategies for multi-agent models in complex topological environments, and enhances the monitoring system's perception depth of group control conditions.

[0053] Further, the process of the output collaborative optimization strategy includes: setting the calculation components of the joint reward function, including: using the deviation of the output active power relative to the rated load as the reward weight, and using the value of the vibration amplitude in the physical feature vector exceeding a preset safety threshold as the penalty weight; calculating the cross-correlation coefficient between the phase fluctuation curve of the real-time power factor and the frequency offset feature curve in the physical feature vector by comparing the two; if the cross-correlation coefficient exceeds a preset threshold, it is identified as the existence of an electromagnetic-dynamic coupling response mode that causes speed fluctuation, and a sensitivity mapping relationship between electromagnetic torque and mechanical speed is established; based on the highest score guidance of the joint reward function, the excitation current correction increment step size of the excitation regulator is calculated, and the correction increment step size is used as the collaborative optimization strategy to be sent to the control register of the excitation regulator. The specific process is as follows: Figure 3 As shown.

[0054] In setting the components of the joint reward function, the system aims to achieve optimal operation of the generator unit through the weight balancing of multiple objectives. Specifically, the system uses the absolute value of the deviation of the output active power relative to the rated load as the input independent variable of the reward function, and calculates the reward weight using a negative exponential decay algorithm. The value of this term approaches 1 when the deviation is zero and decays rapidly as the deviation increases, ensuring the accuracy of the unit's frequency regulation response. At the same time, the system extracts the vibration deviation component obtained by the standard deviation algorithm from the high-fidelity physical feature vector in real time. If the physical displacement corresponding to this component exceeds the preset 50-micrometer safety threshold, the excess value is multiplied by a preset amplification factor (1.5 in this embodiment) as a penalty weight. By linearly weighting and summing the reward weight and the penalty weight according to a dynamic weight of 0.6:0.4, a joint reward function that can take into account both power generation efficiency and structural safety is constructed, providing a clear evolutionary guide for the reinforcement learning of the multi-agent model.

[0055] In the process of identifying electromagnetic-dynamic coupling response modes, the system utilizes the collaborative observation capability of a multi-agent model to extract the phase variation trajectory of the real-time power factor and compares it with the frequency drift component in the physical feature vector after FFT transformation in time series. The system uses a cross-correlation algorithm to calculate the correlation coefficient between the two curves within a sliding window. If the coefficient exceeds a preset threshold (e.g., 0.85), it is determined that the unit is currently in a coupling state where abnormal fluctuations in electromagnetic torque induce mechanical shaft resonance. At this time, the model automatically calls the sensitivity calculation operator to calculate the partial derivative of the rate of change of electromagnetic torque with respect to the mechanical speed deviation, thereby establishing a sensitivity mapping relationship between electromagnetic torque and mechanical speed. This relationship can quantify in real time the effect of fine adjustments to the excitation current on suppressing speed fluctuations, providing a physical basis for the subsequent generation of precise strategies.

[0056] In the process of outputting the collaborative optimization strategy, the policy network in the multi-agent model takes maximizing the joint reward function as the optimization objective. Combining the aforementioned sensitivity mapping relationship, it deduces the optimal excitation current regulation scheme under the current operating condition. Specifically, the model outputs an excitation current correction increment step size that has been limited (for example, the single adjustment step size is limited to within ±0.01A) to prevent the system from oscillating due to excessively rapid adjustment. This correction increment step size is then encapsulated into a 16-bit data message conforming to the industrial control protocol and accurately written into the control register of the excitation regulator controller through the communication bus of the SCADA system. The actuator adjusts the current output of the excitation system in real time according to the received instructions, thereby changing the electromagnetic torque, offsetting the speed deviation on the mechanical side, until the real-time speed and power factor are restored to the target equilibrium value set by the collaborative optimization strategy.

[0057] Through a multi-agent distributed cooperative architecture and joint reward function constraints, a deep integration of electromagnetic torque and dynamic response is achieved. By using cross-correlation coefficients to identify electromagnetic-dynamic coupling modes, the microscopic causes of speed fluctuations can be sensitively detected and the excitation current correction increment can be calculated. While maintaining stable output active power, the safety threshold of mechanical vibration is also taken into account. Through dynamic weight balancing of multiple objectives, the overall operational stability of the unit in complex grid fluctuation environments is optimized.

[0058] Furthermore, the process of generating monitoring and evaluation results includes: retrieving a multidimensional design envelope preset in the digital twin space, and calculating the Euclidean distance of the physical feature vector relative to the boundary of the design envelope at the current moment; defining the Euclidean distance as the numerical deviation, and calculating the variance volatility of the numerical deviation within a preset statistical time window; if the numerical deviation exceeds a preset safety alarm limit and / or the variance volatility is less than a preset convergence constant, then the evaluation result is determined to have reached a preset steady-state threshold, and the supervisory control logic is triggered.

[0059] In the process of generating monitoring and evaluation results, the system first retrieves the multidimensional design envelope corresponding to the current distributed simulation unit from the storage module of the digital twin platform. This envelope is a hyperdimensional boundary pre-trained and deployed in the virtual space based on the equipment's full life cycle health operation data. It defines a set of safety thresholds for multiple physical dimensions, including characteristic frequency, stress distribution, rotational speed, and temperature. Subsequently, the system acquires the high-fidelity physical feature vector synthesized in the aforementioned steps in real time and uses a spatial geometric analysis algorithm to calculate the shortest Euclidean distance between the coordinate point of this vector at the current moment and the hypersurface boundary of the design envelope. The system defines this distance value as the numerical deviation, which is used to intuitively quantify the degree to which the physical entity deviates from the ideal design conditions.

[0060] In constructing the multidimensional design envelope, the system first extracts a health condition dataset of the distributed simulation unit throughout its entire lifecycle from the historical operation database. This dataset covers stable operation records under rated load, peak load variations, and different ambient temperatures. The system employs a support vector machine algorithm, using the characteristic frequencies, stress distribution, rotational speed, and temperature from high-fidelity physical feature vectors as input dimensions, to find the minimum hypersphere or hypersurface that can enclose the vast majority of healthy sample points in the multidimensional feature space. By setting kernel function parameters (such as the Gaussian width of the radial basis kernel function) and anomaly point proportion thresholds, the system automatically defines the evolutionary boundary of the physical entity under healthy service conditions and maps it to a set of high-dimensional geometric vector benchmarks, serving as the reference plane for subsequent Euclidean distance calculations. This unsupervised learning-based modeling approach ensures that the envelope can adaptively cover the nonlinear safety range under complex operating conditions, providing a physically and statistically significant benchmark for real-time risk assessment.

[0061] To further evaluate the stability of system evolution, the system opens a preset statistical time window of 300 seconds and performs variance analysis on the continuously generated numerical deviation sequence within this window. Specifically, the system calculates the variance volatility of the numerical deviation sequence in the time dimension to characterize whether the current operating condition deviation is in a state of violent oscillation or has entered a stable abnormal trend. In the evaluation logic, the system adopts a dual trigger mechanism: on the one hand, if the numerical deviation at the current moment directly exceeds the preset safety alarm limit (generally selected within the range of 1.1 to 1.5 times the design benchmark value, and 1.2 times in this embodiment), it is determined to be an instantaneous risk of exceeding the limit; on the other hand, if the variance volatility of the numerical deviation is less than the preset convergence constant (e.g., 0.005), it indicates that the system has entered a stable deviation state and reached the preset steady-state threshold.

[0062] Once any of the above triggering conditions are met, the system immediately determines that the monitoring and evaluation results have reached the intervention requirements and automatically triggers the back-end supervisory control logic. Based on the severity of the evaluation results, this logic calls the optimization strategy output by the aforementioned collaborative decision-making module. By adjusting the incentive control variables, it guides the dynamic response of the distributed simulation unit to converge into the design envelope, thereby realizing an integrated closed loop from IoT sensing, digital mapping, state evaluation to closed-loop control.

[0063] By calculating the Euclidean distance between the physical feature vector and the multidimensional design envelope, a precise quantitative characterization of the degree to which the unit deviates from the ideal design operating conditions is achieved. Combined with the variance volatility analysis of the deviation sequence, a dual discrimination mechanism of instantaneous over-limit and steady-state trend is constructed, which can effectively identify abnormal drift of equipment and determine steady-state threshold. This not only improves the objectivity and multidimensionality of monitoring and evaluation, but also provides a robust quantitative basis for the automated triggering of back-end monitoring and control logic, ensuring the convergence of dynamic response and operational safety under complex operating conditions.

[0064] Furthermore, the process of issuing instructions to the actuator through the supervisory control logic includes: retrieving the corresponding excitation correction coefficient from a preset adjustment characteristic lookup table based on the real-time power factor, and calculating the target adjustment increment of the excitation control variable; using the communication protocol of the supervisory control and data acquisition system, sending a control message containing the target adjustment increment to the instruction register of the actuator to drive the excitation source to change the input energy intensity of the distributed simulation unit; monitoring the rotational speed change of the distributed simulation unit in real time, and using the target equilibrium value corresponding to the collaborative optimization strategy as the feedback target, adjusting the excitation control variable in a step-by-step cyclic manner to optimize the electromechanical energy conversion characteristics, so that the dynamic response of the distributed simulation unit converges to the collaborative optimization strategy until the composite residual of the real-time rotational speed, real-time power factor and their respective target equilibrium values ​​enters a preset allowable error range.

[0065] Specifically, during the execution of the command issuance process, the supervisory control logic first searches the adjustment characteristic lookup table preset by the edge control unit based on the real-time collected power factor. This table pre-stores the mapping relationship between the power factor and the excitation correction coefficient under different load conditions. The system matches the current correction coefficient accordingly and, in combination with the parameters output in the decision-making stage, calculates the specific target adjustment increment of the excitation control variable.

[0066] During the data communication and drive phase, the system calls the communication protocol of the supervisory control and data acquisition system, encapsulates the target adjustment increment into a control message of a specific format, and sends the message precisely to the instruction register of the actuator through the industrial Ethernet bus, directly driving the power adjustment unit to change the output intensity of the excitation source, thereby adjusting the energy input of the distributed simulation unit and changing the torque output state from the source.

[0067] During the feedback monitoring and convergence control phase, the system enters a dynamic adjustment loop, tracking the speed fluctuations of the distributed simulation unit in real time. The system uses the target equilibrium value set in the collaborative optimization strategy as the endpoint, and through step-by-step cyclic adjustment logic, it makes small, multiple continuous corrections to the excitation control variables to optimize the electromechanical energy conversion process.

[0068] In the adjustment termination judgment logic, the system calculates the deviation of the current speed and power factor from their respective target equilibrium values ​​in real time and synthesizes them into a composite residual. The system guides the dynamic response to converge continuously through step adjustment until the composite residual enters the preset allowable error range (for example, the error ratio drops to the preset threshold of 10,000). At this time, the adjustment process ends and the system enters the steady-state monitoring state of the next cycle.

[0069] The execution module further includes: extracting the response residual sequence during the adjustment of the excitation control variables, wherein the response residual sequence is the numerical difference between the actual physical state of the physical entity after the actuator's action and the physical feature vector output by the twin mapping module; using a variational inference algorithm to analyze the deviation characteristics of the response residual sequence in the probability distribution dimension, identifying the drift increment of material physical parameters or boundary constraints in the multiphysics evolution model; and feeding the drift increment back to the twin mapping module to achieve online parameter calibration of structural parameters in the evolution model, reducing the representation distortion between physical space and digital space.

[0070] In the operation and control of steam turbine generator units in smart power plants, when the execution module issues instructions according to the collaborative optimization strategy and adjusts the excitation control variables (such as adjusting the excitation current increment) through the supervisory control logic, the system uses the physical disturbance generated by the execution action as the excitation source to start the self-calibration program. After the execution mechanism moves, the IoT sensing module captures the real state sequence of the physical entity in real time through vibration sensors and speed encoders deployed at the bearing support position and the end of the main shaft. Subsequently, the system synchronously extracts the physical feature vector output by the twin mapping module within the same time step. This vector represents the theoretical response value predicted by the current digital twin model based on the existing material parameters. The system performs point-by-point numerical difference operation between the measured real state value and the model prediction value to generate a response residual sequence. This residual sequence intuitively reflects the simulation distortion caused by the degradation of physical properties in the digital twin space in the time domain dimension.

[0071] The system utilizes a variational inference algorithm to perform a deep deviation feature analysis on the above response residual sequence in the probability distribution dimension. In the processing flow, the algorithm approximates the true posterior distribution of material physical parameters (such as equivalent stiffness coefficient and damping attenuation factor) by constructing an analytical variational distribution (such as mean field distribution), thereby decoupling the implicit physical property variations from the residual signal containing random measurement noise and transient load interference. By minimizing the distribution difference between the variational distribution and the true physical feedback (i.e., maximizing the lower bound of evidence), the system can accurately identify which residual components originate from random fluctuations of the sensor and which belong to the drift increments of material physical parameters (such as stiffness reduction caused by spindle material fatigue) or boundary constraints (such as stress distribution changes caused by cooling circuit pressure deviation) due to long-term system service.

[0072] The system uses the identified drift increment as a feedback bias, transmitting it in real time to the twin mapping module to achieve online calibration of structural parameters in the evolution model. Specifically, this feedback bias directly affects the operator weights in the nonlinear stiffness correction layer. Based on the drift increment value, the system automatically adjusts the balancing coefficients used to balance elastic deformation and nonlinear response in the overall structural stiffness matrix calculation formula. This ensures that the compensation calculation for the initial characteristic frequency automatically includes the current physical property degradation value. This closed-loop feedback mechanism effectively reduces the representational distortion between the physical and digital spaces, ensuring that the high-fidelity physical feature vectors are highly aligned with the actual operating conditions of the physical entity. It achieves real-time dynamic tracking of physical property degradation throughout the generator set's entire lifecycle, ensuring the accuracy of subsequent monitoring and evaluation results and collaborative decision-making instructions.

[0073] By combining real-time power factor with a preset regulation characteristic lookup table, accurate prediction and incremental calculation of excitation control variables are achieved. By using register-level instruction issuance and step-by-step regulation logic based on composite residuals, the dynamic response can be smoothly converged to the optimization target under complex operating conditions, significantly improving the control accuracy of electromechanical energy conversion and effectively solving the response hysteresis and regulation overshoot problems of smart power plants in multi-field coupling environments.

[0074] By integrating IoT sensing, multi-physics digital twin, and multi-agent collaborative decision-making technologies, the system achieves integrated monitoring of the smart power plant's power system from data acquisition to closed-loop execution. It utilizes a thermo-mechanical coupling evolution model to accurately compensate for characteristic frequency drift and identifies complex electromagnetic-dynamic coupling modes through a multi-agent architecture to generate optimal control strategies. Combined with a step-by-step adjustment logic based on composite residuals, it significantly improves the accuracy of equipment condition characterization and operational stability under varying operating conditions, effectively solving the collaborative control challenges of large-scale distributed units under spatiotemporal synchronization and nonlinear disturbances.

[0075] Example 2: A 300MW steam turbine generator unit in a smart power plant was used as the monitoring object. The system provided by this invention integrates IoT sensing and data monitoring. The system first captures the initial characteristic frequency by using a piezoelectric vibration sensor deployed on the bearing support. It obtains the real-time temperature using a thermal resistance sensor embedded in the stator housing and simultaneously collects underlying physical signals such as real-time speed and real-time power factor. In order to quantify the degradation law of the equipment, the system retrieves the historical operating data of the unit over the past 168 hours from the time series database. It uses a standard deviation scaling algorithm to map the heterogeneous data to a unified interval and uses a time sliding window to extract the slope of numerical evolution. Its stable component is defined as the historical performance evolution trend that reflects the dynamic evolution law of the unit.

[0076] The system calls the SCADA system's built-in precise time protocol to align the acquired asynchronous sensing signals to a unified master clock frequency and encapsulates the timestamp as an index key. The system retrieves the virtual model index through the logical topology identification code and establishes a three-dimensional mapping binding relationship consisting of timestamp, status parameters, and the virtual model. In the digital twin platform, the system maps the color gradient of the geometric surface according to the real-time temperature and drives the rotation of the three-dimensional entity according to the real-time rotation speed, thereby realizing the spatiotemporal synchronous display of the real-time operating conditions of the physical equipment in the three-dimensional virtual space.

[0077] The system constructs a multiphysics simulation environment, applying real-time temperature, centrifugal load, and fluid pressure as composite boundary constraints to the finite element mesh nodes. The thermo-mechanical coupling evolution layer calculates the structural thermal stress offset based on the material's thermal expansion properties, and uses a nonlinear stiffness correction layer to analyze the impact of this offset on the equivalent stiffness. The system calculates the numerical drift of the initial characteristic frequency through eigenvalue solving, performs deviation compensation, and then fuses the compensated characteristic frequency with the historical performance evolution trend in a multidimensional manner to output a high-fidelity physical feature vector.

[0078] The system concatenates physical feature vectors, real-time rotational speed, real-time power factor, and historical performance evolution trends according to a preset time sequence to construct a three-dimensional state tensor with dimensions of 1*50*total number of features. Through register address mapping, each unit controller synchronously shares this tensor on the communication bus. Each simulation unit extracts associated feature terms (such as vibration amplitude deviation and current harmonic ratio) from adjacent units using built-in operators and generates a joint input matrix. The multi-agent model, based on a centralized training and distributed execution architecture, identifies electromagnetic-dynamic coupling response modes and outputs collaborative optimization strategies.

[0079] The system retrieves the preset multidimensional design envelope and calculates the shortest Euclidean distance of the physical feature vector relative to the boundary of the envelope, which is defined as the numerical deviation. Within a statistical time window of 300 seconds, the system continuously calculates the variance volatility of this deviation. If the numerical deviation exceeds 1.2 times the design benchmark value, or the variance volatility is less than the preset convergence constant (0.005), the evaluation result is determined to have reached the preset steady-state threshold, and the supervisory control logic is immediately triggered to call the collaborative optimization strategy for intervention.

[0080] The supervisory control logic retrieves the excitation correction coefficient from the regulation characteristic lookup table based on the real-time power factor and calculates the target regulation increment. The system sends the control message to the instruction register of the actuator through the industrial communication protocol, driving the excitation source to change the input energy intensity. The system uses the target equilibrium value as the feedback target and adjusts the excitation control variables in a step-by-step cyclic manner until the composite residual of the real-time speed, real-time power factor and their respective target equilibrium values ​​is reduced to within the allowable error range (e.g., 0.1%), thus achieving smooth convergence of the dynamic response.

[0081] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An integrated system of internet of things, perception and data monitoring, characterized in that, include: IoT sensing module: acquires the state parameters of the distributed simulation unit of the smart power plant, including initial characteristic frequency, real-time temperature, real-time power factor, real-time speed, rated load demand, identification code and historical performance evolution trend; Twin mapping module: Maps state parameters to simulation nodes in the digital twin space based on the identification code; calls the evolution model to calculate thermal stress offset and performs numerical compensation to generate physical feature vectors; Collaborative decision-making module: Input the physical feature vector, real-time rotational speed, real-time power factor and historical performance evolution trend into the multi-agent model, define the distributed simulation unit as a dynamic simulation agent, identify the electromagnetic-dynamic coupling response mode through the sharing of observation information among agents and the constraint of joint reward function, and output the collaborative optimization strategy; Execution module: Calculates the numerical deviation between the physical feature vector and the design envelope limit to generate monitoring and evaluation results. When the evaluation results reach the preset steady-state threshold, it extracts the real-time power factor based on the control parameters, issues instructions to the actuator through the supervisory control logic to adjust the excitation control variables, adjust the torque output, until the real-time speed and operating parameters reach the target equilibrium value corresponding to the collaborative optimization strategy.

2. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The process of acquiring the state parameters of the distributed simulation unit of the smart power plant includes: using sensors deployed at the monitoring points of the physical equipment corresponding to the distributed simulation unit to capture the underlying physical signals corresponding to the initial characteristic frequency, real-time temperature, real-time speed, and real-time power factor; retrieving the logical topology identification code bound to the physical equipment from the system configuration library, and retrieving the historical operating data sequence of the corresponding period from the time series database based on the logical topology identification code; performing data standardization processing on the historical operating data sequence, and using a time sliding window to extract the discrete deviation distribution of each state parameter relative to the preset design benchmark value; analyzing the slope of the numerical evolution of the discrete deviation distribution within the time sliding window, and extracting the stable component of the numerical evolution slope as the historical performance evolution trend reflecting the dynamic evolution law of the unit.

3. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The process of mapping to the simulation node in the digital twin space includes: calling the built-in time protocol of the supervision, control and data acquisition system to align the asynchronous sensing signals obtained through the IoT data acquisition interface to a unified system master clock frequency; encapsulating timestamps for each aligned asynchronous sensing signal; using the timestamps as index keys to aggregate the discrete sampling points in the state parameters into a synchronously associated monitoring dataset in the time dimension; extracting the logical topology identification code and retrieving the pre-set virtual model index in the digital twin space to establish a ternary mapping binding relationship consisting of timestamps, state parameters and virtual model indexes; and pushing the state parameters to the corresponding simulation node in the digital twin platform in real time according to the ternary mapping binding relationship to achieve spatiotemporal synchronous display of the real-time operating conditions of the physical equipment in the three-dimensional virtual space.

4. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The evolution model includes: a multi-source boundary integration layer: acquiring real-time temperature, rated load requirements, centrifugal load generated by real-time rotation speed, and fluid pressure components as composite boundary constraints, and combining them with the pre-set structural and material parameters of the simulation unit to construct a multi-physics simulation environment in the digital twin space; Thermo-mechanical coupling stress evolution layer: Identify the temperature field gradient and stress field distribution of the simulation unit in the multiphysics simulation environment, and calculate the structural thermal stress offset induced by temperature rise, load and pressure based on the thermal expansion properties of materials and elasticity mechanism. Nonlinear stiffness correction layer: Analyzes the influence of structural thermal stress offset on the equivalent stiffness and dynamic damping characteristics of the unit physical structure, and calculates the numerical drift of the initial characteristic frequency based on the nonlinear change of the equivalent stiffness, and performs deviation compensation for the initial characteristic frequency. High-fidelity feature synthesis layer: The compensated feature frequencies are fused with the historical performance evolution trend reflecting the service status of the material to output the physical feature vector.

5. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The step of inputting physical feature vectors, real-time rotational speed, real-time power factor, and historical performance evolution trends into the multi-agent model includes: concatenating the physical feature vectors, real-time rotational speed, real-time power factor, and historical performance evolution trends according to a preset time sequence to construct a multi-dimensional state tensor corresponding to each simulation unit; using the communication bus of the supervisory control and data acquisition system, synchronously sharing the multi-dimensional state tensor among the controllers of each simulation unit through register address mapping; each simulation unit extracts correlation feature terms reflecting the mechanical vibration and electrical fluctuations of adjacent units based on the received multi-dimensional state tensors of adjacent units, and merges them with its own multi-dimensional state tensor to generate a joint input matrix for strategy generation by the dynamic simulation agent.

6. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The process of the output collaborative optimization strategy includes: setting the calculation components of the joint reward function, including: using the deviation value of the output active power relative to the rated load as the reward weight, and using the value of the vibration amplitude in the physical feature vector exceeding the preset safety threshold as the penalty weight; by comparing the phase fluctuation curve of the real-time power factor with the frequency offset feature curve in the physical feature vector, and calculating the cross-correlation coefficient, if the cross-correlation coefficient exceeds the preset threshold, it is identified as the existence of an electromagnetic-dynamic coupling response mode that causes speed fluctuation, and a sensitivity mapping relationship between electromagnetic torque and mechanical speed is established; according to the highest score guide of the joint reward function, the excitation current correction increment step size of the excitation regulator controller is calculated, and the correction increment step size is sent to the control register of the excitation regulator controller as the collaborative optimization strategy.

7. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The process of generating monitoring and evaluation results includes: retrieving a multidimensional design envelope preset in the digital twin space, and calculating the Euclidean distance of the physical feature vector relative to the boundary of the design envelope at the current moment; defining the Euclidean distance as the numerical deviation, and calculating the variance volatility of the numerical deviation within a preset statistical time window; if the numerical deviation exceeds a preset safety alarm limit and / or the variance volatility is less than a preset convergence constant, then the evaluation result is determined to have reached a preset steady-state threshold, and the supervisory control logic is triggered.

8. The integrated IoT sensing and data monitoring system according to claim 1, characterized in that, The process of issuing instructions to the actuator through the supervisory control logic includes: retrieving the corresponding excitation correction coefficient from a preset adjustment characteristic lookup table based on the real-time power factor, and calculating the target adjustment increment of the excitation control variable; using the communication protocol of the supervisory control and data acquisition system, sending a control message containing the target adjustment increment to the instruction register of the actuator to drive the excitation source to change the input energy intensity of the distributed simulation unit; monitoring the rotational speed change of the distributed simulation unit in real time, and using the target equilibrium value corresponding to the collaborative optimization strategy as the feedback target, adjusting the excitation control variable in a step-by-step cyclic manner to optimize the electromechanical energy conversion characteristics, so that the dynamic response of the distributed simulation unit converges to the collaborative optimization strategy until the composite residual of the real-time rotational speed, real-time power factor and their respective target equilibrium values ​​enters a preset allowable error range.

Images