Intelligent measurement and control optimization method and device for dynamic parameter adaptive calibration, equipment and medium

Abstract

The invention relates to an intelligent measurement and control optimization method and device for dynamic parameter adaptive calibration, equipment and a medium. The method comprises the following steps: executing invariance causal test based on an external environment context to obtain a cross-environment invariant explanatory variable subset, and performing anti-fact simulation on each parameter in the cross-environment invariant explanatory variable subset to generate a causal attribution report; based on an under-excitation parameter subset and a cross-environment invariant explanatory variable subset in the baseline recognizable atlas, performing safe active excitation planning on the parameters to obtain a safe micro-perturbation excitation plan; and based on the residual error, the uncertainty estimation, the delay cross-correlation feature, the hysteresis loop area feature, the causal contribution score and the excitation-response fragment, performing calibration estimation on the target parameter by using a hierarchical estimator, and generating a calibration packet according to a calibration estimation result. By adopting the method, self-adaptive calibration of intelligent measurement and control dynamic parameters can be realized through residual attribution and cross-environment invariance test in combination with safe perturbation excitation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent measurement and control, and in particular relates to an intelligent measurement and control optimization method, device, equipment and medium for dynamic parameter adaptive calibration. Background Technology

[0002] With the rapid development of industrial automation, energy systems, and advanced manufacturing, intelligent measurement and control technologies based on the integration of model-driven and data-driven approaches have emerged. These technologies combine physical mechanism models with real-time acquired data, using model prediction, parameter estimation, and control optimization to achieve high-precision monitoring and dynamic control of complex systems. Their key features include rapid response in changing environments and a degree of self-adaptability to cope with performance fluctuations caused by external disturbances and system aging.

[0003] In traditional techniques, dynamic parameter calibration primarily relies on periodic offline calibration or simple online recursive least squares estimation methods. Offline calibration is typically performed with equipment shut down or in a laboratory environment, using a standard input signal to excite the system and collect the response to calculate and update parameters. This method offers high calibration accuracy but cannot reflect the dynamic changes of the system in actual operating environments, and incurs high downtime costs. Online recursive least squares estimation methods continuously update parameters during system operation, avoiding downtime, but often rely on continuous excitation signals and assume that system noise follows a simple statistical distribution, lacking consideration for changes in the external environment and complex nonlinear factors.

[0004] Under certain operating conditions, some key parameters lack sufficient identification information from natural operating data, leading to unstable parameter estimation results. To improve identification accuracy, traditional methods often apply strong external stimuli, but this can easily affect normal system operation and even introduce safety risks. Existing parameter update methods are typically based on correlation analysis, which cannot effectively distinguish causal relationships, thus causing misjudgments in multivariate coupled systems and reducing the reliability of parameter updates. Traditional algorithms lack robustness to changes in the external environment, causing parameters to quickly become inaccurate in new environments, requiring recalibration. Existing methods often fail to assess the system's residual control capability and safety boundaries during parameter updates, posing potential risks. Summary of the Invention

[0005] Therefore, it is necessary to provide an intelligent measurement and control optimization method, device, equipment, and medium that enables more refined and safer adaptive calibration of dynamic parameters to address the aforementioned technical problems.

[0006] Firstly, this application provides an intelligent measurement and control optimization method for dynamic parameter adaptive calibration, including:

[0007] Acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object, distributed over time.

[0008] Using a differentiable twin model, the current state of the controlled object and the sensor output are predicted based on the control input and sensor output. Based on the prediction results, feature analysis is performed to obtain residuals, uncertainty estimates, delay cross-correlation characteristics, and hysteresis area characteristics. The delay cross-correlation characteristics include time delay estimation and excitation-response coupling strength.

[0009] Based on the external environment context, an invariance causality test is performed to obtain a subset of explanatory variables that are invariant across the environment. Counterfactual simulations are then performed on each parameter in the subset of explanatory variables that are invariant across the environment to generate a causal attribution report. The causal attribution report includes the causal attribution contribution score of each parameter.

[0010] Based on the subset of underexcited parameters and the subset of explanatory variables that are invariant across environments in the baseline identifiable spectrum, a safe active excitation plan is performed on the parameters to obtain a safe micro-perturbation excitation plan. The safe micro-perturbation excitation plan includes the excitation order, perturbation sequence, perturbation time, perturbation frequency band, perturbation amplitude, and minimum safety margin.

[0011] Based on the safety margin constraint strategy, a safety micro-perturbation excitation plan is executed on the target parameter by superimposing the perturbation sequence with the original control input. The sensor output and external environmental context corresponding to the execution of the safety micro-perturbation excitation plan are collected to obtain the excitation-response segment. The excitation-response segment includes the excitation signal, the response signal and the external environmental context during execution.

[0012] Based on residuals, uncertainty estimation, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and excitation-response segments, a hierarchical estimator is used to calibrate and estimate the target parameters, and a calibration package is generated based on the calibration estimation results. The calibration package includes a set of parameter calibration values, posterior covariance, excitation-response segment parameters, timestamps, and version information. The calibration package is used to calibrate and update the dynamic parameters of the intelligent measurement and control system.

[0013] In one embodiment, before acquiring observation data collected at fixed intervals, the method further includes:

[0014] Based on the structural variance model corresponding to the intelligent measurement and control system, a differentiable twin model is obtained by fitting and estimating representative working condition data from historical observation data.

[0015] The numerical expression corresponding to the differentiable twin model can be obtained through the following formula:

[0016]

[0017] Where, xt The state of the controlled object at time t; u t For control input; For sensor output; The dynamic parameters of the sensor to be calibrated; The dynamic parameters of the actuator to be calibrated; These are the dynamic parameters to be calibrated in the process. The control input is the result of nonlinear / hysteresis / time-delay mapping of the actuator; ω t For process noise; v t For measuring noise;

[0018] Based on the twin linearization model, the Jacobian matrix is ​​obtained by linearizing the representative operating points corresponding to the representative operating conditions data; the twin linearization model is obtained by performing first-order linearization on the model function in the differentiable twin model;

[0019] Based on the preset control input sequence, the sensitivity sequence of the sensor output to each parameter is calculated according to the measurement noise variance and Jacobian matrix of historical observation data, and the Fischer information matrix is ​​obtained by discrete summation of the sensitivity sequence.

[0020] The identifiability index of each parameter is determined based on the Fischer information matrix, and then visualized by combining the classic excitation frequency bands of each parameter to obtain the baseline identifiability map.

[0021] Based on a pre-defined underexcitation judgment strategy, underexcitation parameters in the baseline identifiable spectrum are labeled to obtain a subset of underexcitation parameters.

[0022] In one embodiment, a differentiable twin model is used to predict the current state of the controlled object and the sensor output based on the control input and sensor output. Feature analysis is then performed based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation characteristics, and hysteresis area characteristics, including:

[0023] The control input and the state of the controlled object at the previous moment are input into the differentiable twin model to obtain the predicted state of the controlled object at the current moment; the previous moment is a fixed period based on the past of the current moment;

[0024] The predicted state of the controlled object is input into the differentiable twin model to obtain the predicted sensor output at the current moment;

[0025] The predicted sensor output is compared with the actual sensor output at the current moment to obtain the residual;

[0026] By calculating the cross-correlation between the change in control input and the change in sensor output under different delays, a delay cross-correlation characteristic including time delay estimation and excitation-response coupling strength is obtained; the change in control input is obtained from the difference between the control inputs of two adjacent fixed periods; the change in sensor output is obtained from the difference between the sensor outputs of two adjacent fixed periods.

[0027] Based on a sliding time window of preset length, the loop area formed by the control input and sensor output is calculated to obtain the hysteresis loop area characteristics.

[0028] In one embodiment, an invariance causality test is performed based on the external environment context to obtain a subset of explanatory variables that are invariant across the environment. Counterfactual simulations are then performed on each parameter in this subset to generate a causal attribution report, including:

[0029] The external environment context is divided into multiple environment subsets based on the environmental scenario, and an error generation model is fitted for each environment subset.

[0030] Based on the error generation model, the cross-environment stability index of each parameter is calculated, and the parameters are screened according to the cross-environment stability index to obtain a subset of explanatory variables that are invariant across the environment.

[0031] Based on the differentiable twin model, each parameter of the subset of explanatory variables that are invariant across environments is slightly intervened, and the control input at the current moment and the external environmental context are replayed to obtain the counterfactual sensor output.

[0032] The causal contribution score for each parameter is obtained by calculating the difference between the counterfactual sensor output, the predicted sensor output, and the actual sensor output at the current moment.

[0033] The causal contribution score can be obtained using the following formula:

[0034]

[0035] Where, α i The causal contribution score for the i-th parameter; y t This represents the actual sensor output at the current moment. Output of the counterfactual sensor; To predict the sensor output; δ is the disturbance value corresponding to a small intervention; ||·||2 is the L2 norm;

[0036] A causal attribution report is generated based on the causal contribution score.

[0037] In one embodiment, based on a subset of underexcited parameters in the baseline identifiable map and a subset of explanatory variables that are invariant across environments, a safe active excitation plan is performed on the parameters to obtain a safe micro-perturbation excitation plan, including:

[0038] Within a sliding window of a preset step size, the differentiable twin model is linearized using twins based on the current operating point to obtain the current Jacobian matrix;

[0039] The current sensitivity sequence of each parameter is calculated based on the current Jacobian matrix, and the current sensitivity sequence is discretized and summed to obtain the current Fisher information matrix.

[0040] Based on the parameters in the underexcited parameter subset and the cross-environment invariant explanatory variable subset, the target parameter submatrix is ​​extracted from the current Fisher information matrix, and the instantaneous identifiability corresponding to each parameter is calculated based on the target parameter submatrix.

[0041] The priority score of each parameter is calculated based on the immediate identifiability and causal contribution score, and the parameters are arranged in order of priority score from highest to lowest to obtain the incentive order.

[0042] Based on the information gain maximization objective and constraints, a safety micro-perturbation incentive plan is obtained by superimposing small-amplitude safety incentives through Bayesian optimization and combining the incentive sequence.

[0043] In one embodiment, based on the information gain maximization objective and constraints, a safety micro-perturbation incentive plan is obtained by solving a Bayesian optimization solution with superimposed small-amplitude safety incentives and combining the incentive sequence, including:

[0044] The predicted state of the controlled object is filtered and corrected by adaptive Kalman filtering to obtain the corrected predicted state of the controlled object and the uncertainty estimate.

[0045] The state of the controlled object is corrected and input into the differentiable twin model to obtain the state of the controlled object at future time, and the corresponding prediction Jacobian matrix is ​​calculated based on the state of the controlled object at future time.

[0046] The predicted sensitivity sequence of each target parameter is calculated based on the predicted Jacobian matrix, and the predicted Fisher information matrix is ​​obtained by discrete summation based on the predicted sensitivity sequence.

[0047] If the instantaneous identifiability of the parameter is less than the preset threshold, a small safety excitation is superimposed on the information obtained by Bayesian optimization until the gain maximization objective and constraints are met, and a perturbation sequence is generated. The constraints include safety function constraints, performance deviation constraints, and frequency band constraints. The safety function constraints include safety sets, control barrier functions, and linear constraints. The performance deviation constraints include limiting sensor output deviation and limiting perturbation energy.

[0048] The minimum safety margin corresponding to the current moment is calculated based on the safety function constraints, and the safety micro-perturbation excitation plan is obtained by combining the perturbation sequence and its corresponding superimposed small-amplitude safety excitation parameters and excitation order.

[0049] In one embodiment, the hierarchical estimator consists of a fast layer and a slow layer connected together; the fast layer uses Kalman filtering to model the sensor dynamic parameters to be calibrated and the fast drift term as a random walk; the slow layer uses a moving horizon of a preset length to solve the constraint optimization, which includes error, state transition error and the deviation of the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated from the prior value, under the condition of satisfying the constraints.

[0050] Based on residuals, uncertainty estimates, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and stimulus-response segments, a hierarchical estimator is used to perform calibration estimation of the target parameters, and a calibration package is generated based on the calibration estimation results, including:

[0051] Based on the residual and uncertainty estimates, the dynamic parameters of the sensor to be calibrated are iteratively calculated until the accuracy converges, and the posterior distribution of the dynamic parameters of the sensor to be calibrated is obtained; the posterior distribution includes the mean and covariance of the dynamic parameters of the sensor to be calibrated; the mean is the parameter calibration value of the dynamic parameters of the sensor to be calibrated.

[0052] Based on a preset moving time window, the parameter calibration values ​​of the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated are obtained through constraint optimization solution state and data-guided joint estimation. Data-guided corresponding time delay estimation serves as the prior center of the time delay parameters in the process parameters, the hysteresis loop area feature serves as the regularization guide for the actuator hysteresis model, the causal contribution score above the preset attribution threshold in the causal attribution report is used to relax the uncertainty variance of the prior estimation, the objective function information is weighted according to the instantaneous identifiability, and the convergence of the parameters is improved by setting the weight of the excitation-response segment.

[0053] A calibration package is obtained based on the calibration values ​​of the dynamic parameters to be calibrated for the sensors, the dynamic parameters to be calibrated for the actuators, and the dynamic parameters to be calibrated for the processes.

[0054] Secondly, this application also provides an intelligent measurement and control optimization device for dynamic parameter adaptive calibration, comprising:

[0055] The measurement and control observation module is used to acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object, distributed over time.

[0056] The feature analysis module is used to predict the state of the controlled object and the sensor output at the current moment using a differentiable twin model based on the control input and sensor output, and to perform feature analysis based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation features, and hysteresis area features; the delay cross-correlation features include time delay estimation and excitation-response coupling strength;

[0057] The invariant causality module is used to perform invariant causality tests based on the external environment context, obtain a subset of explanatory variables that are invariant across the environment, and perform counterfactual simulations on each parameter in the subset of explanatory variables to generate a causal attribution report; the causal attribution report includes the causal attribution contribution score of each parameter;

[0058] The excitation module is used to perform safe active excitation planning on parameters based on a subset of underexcited parameters and a subset of explanatory variables that are invariant across environments in the baseline identifiable map, so as to obtain a safe micro-perturbation excitation plan;

[0059] The response module is used to execute a safety micro-perturbation excitation plan on the target parameters by superimposing the perturbation sequence with the original control input based on the safety margin constraint strategy, and to collect the corresponding sensor output and external environmental context when executing the safety micro-perturbation excitation plan to obtain the excitation-response segment;

[0060] The calibration module is used to perform calibration estimation of target parameters using a hierarchical estimator based on residuals, uncertainty estimates, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and excitation-response segments, and to generate a calibration package based on the calibration estimation results.

[0061] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the intelligent measurement and control optimization method for adaptive calibration of any of the above-mentioned dynamic parameters.

[0062] Fourthly, this application also provides a computer-readable storage medium storing a computer program thereon, wherein the computer program, when executed by a processor, implements the steps of the intelligent measurement and control optimization method for adaptive calibration of any of the above-mentioned dynamic parameters.

[0063] The aforementioned intelligent measurement and control optimization method, device, equipment, and medium for dynamic parameter adaptive calibration, through the combination of a differentiable twin model and real-time observation data, can not only simulate system dynamics but also update model parameters in real time, balancing accuracy and adaptability. Introducing an invariant causal test eliminates spurious correlation features dependent on the environment, ensuring consistency and reliability of calibration results across different operating environments and adapting to complex and changing actual working conditions. Through a safe micro-perturbation excitation design, the identifiability of under-excited parameters is significantly improved without affecting system stability, solving the problem of difficult active excitation in continuously operating systems using traditional methods. Utilizing multiple types of information such as residuals, uncertainties, delay cross-correlation, hysteresis characteristics, and causal contributions, a hierarchical estimator comprehensively analyzes the data to achieve higher accuracy and lower uncertainty in dynamic parameter calibration. Standardized calibration packages enable automated parameter updates and version management, allowing for rapid deployment of updates and rollback to historical versions when necessary, reducing operational risks. It supports real-time data acquisition, excitation execution, and calibration updates during system operation, avoiding downtime and improving the availability and economy of the measurement and control system. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0065] Figure 1 This is a flowchart illustrating the intelligent measurement and control optimization method for dynamic parameter adaptive calibration according to the present invention.

[0066] Figure 2 This is a flowchart illustrating the steps of step S103.

[0067] Figure 3 This is a flowchart illustrating the steps of step S104.

[0068] Figure 4 This is a structural diagram of the intelligent measurement and control optimization device for dynamic parameter adaptive calibration according to the present invention. Detailed Implementation

[0069] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0070] In one embodiment, such as Figure 1As shown, an intelligent measurement and control optimization method with dynamic parameter adaptive calibration is provided. This embodiment illustrates the application of this method to a terminal. It can be understood that this method can also be applied to a server, and to a system including both a terminal and a server, and implemented through the interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0071] S101. Acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object, distributed over time.

[0072] As an illustration, during the operation of an intelligent measurement and control system, observation data needs to be continuously collected at fixed intervals. The observation data specifically includes the control input u distributed according to a time series. t Sensor output y t and external environment context c t The system includes parameters such as temperature and load, as well as the state of the controlled object. Control inputs are the command signals sent by the system to the actuators, directly driving their actions. Sensor outputs are the sensor measurements of the controlled object's state, reflecting the system's actual operating status. The external environment context encompasses environmental factors affecting system operation, which may cause parameter drift. The controlled object's state is the system's core dynamic variable, such as the equipment's internal temperature and the robotic arm's position, and is a key indicator that needs to be tracked during calibration.

[0073] Furthermore, using fixed-period data acquisition aligns with the dynamic characteristics of intelligent measurement and control systems. The fixed period ensures the temporal consistency of the data, facilitating subsequent analysis of parameter changes over time.

[0074] S102. Using a differentiable twin model, predict the state of the controlled object and the sensor output at the current moment based on the control input and sensor output, and perform feature analysis based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation characteristics, and hysteresis area characteristics; delay cross-correlation characteristics include time delay estimation and excitation-response coupling strength.

[0075]

[0076] In a schematic representation, after acquiring observational data, a differentiable twin model is used to predict the current state of the controlled object and the sensor output. A differentiable twin model is a digital mapping of a physical system, encompassing physical mechanisms such as heat conduction equations and laws of mechanical motion, while also integrating data-driven models. It possesses differentiability, enabling the calculation of parameter sensitivity to output. Specifically, the twin model is used to predict the state of the controlled object at the previous moment based on its position. Control input after actuator parameter mapping and process parameters Predict the state at the current moment Then, based on the predicted state and sensor parameters Obtain the predicted output Furthermore, calculate the actual output y of the sensor. t With predicted output The difference between them, i.e., the residual r t .

[0077] Furthermore, the predicted values ​​are compared with the actual sensor outputs to obtain residuals. These residuals reflect the deviation between the model predictions and the actual system. Larger residuals indicate a lower degree of matching between the model parameters and the actual system, necessitating greater calibration. Optionally, the dynamic characteristics of the system can be captured by extracting delay cross-correlation features and hysteresis loop area features. Specifically, delay cross-correlation features are used to determine the time difference between input and output responses (i.e., hysteresis estimation) and the degree of input influence on output (i.e., excitation-response coupling strength) by calculating the correlation between changes in control input and sensor output at different time delays. For example, in a flow control system, there may be a 0.5-second time delay between pump voltage changes and flow rate changes; this time delay needs to be reflected in the parameters. Hysteresis loop area features are used to quantify the nonlinear hysteresis effect of the actuator or sensor by calculating the area of ​​the closed loop formed by the input-output trajectories; the larger the area, the more significant the hysteresis.

[0078] S103. Perform invariance causality tests based on the external environment context to obtain a subset of explanatory variables that are invariant across the environment. Then, perform counterfactual simulations on each parameter in the subset of explanatory variables that are invariant across the environment to generate a causal attribution report. The causal attribution report includes the causal attribution contribution score of each parameter.

[0079] Indicatively, changes in the external environmental context can alter the dynamic characteristics of the system. The parameters truly requiring calibration should maintain a stable impact on the residuals under different environments. Therefore, an invariant causality test (ICP) is performed under various external environmental contexts, such as load, temperature, and mode. The ICP compares the correlation between residuals and parameters under different environments to identify a subset of parameters with stable residual effects. Specifically, these different environmental contexts are first divided into multiple environments. A simplified error generation model is fitted to each environment, containing a design matrix column consisting of twin sensitivity, features, and activation indicators for the corresponding parameter families. Through testing, a subset of explanatory variables that maintain a stable relationship with the sensor output across environments—the invariant explanatory subset—is identified as the root cause parameter family causing the current residuals. Further, counterfactual simulations are performed. Specifically, in a differentiable twin, parameters in the invariant explanatory subset are slightly modified one by one while keeping other parameters and inputs constant. The simulation is then replayed using historical or current inputs to obtain the counterfactual output. Based on the differences between the counterfactual output and the actual and predicted outputs, parameter attribution scores are calculated to quantify the causal contribution of each parameter to the residuals. A higher score indicates a greater impact of the parameter on the residuals. The causal attribution report integrates the causal contribution scores of each parameter.

[0080] S104. Based on the subset of underexcited parameters and the subset of explanatory variables that are invariant across environments in the baseline identifiable spectrum, perform safe active excitation planning on the parameters to obtain a safe micro-perturbation excitation plan; the safe micro-perturbation excitation plan includes excitation order, perturbation sequence, perturbation time, perturbation frequency band, perturbation amplitude and minimum safety margin.

[0081] This paper illustrates how, based on a subset of under-excited parameters and a subset of environment-invariant explanatory variables in the baseline identifiable graph, proactive excitation planning for safety is performed to generate a safety micro-perturbation excitation plan. The aim is to improve parameter identifiability without compromising system safety. Specifically, the baseline identifiable graph, generated from historical data during system initialization, marks the subset of under-excited parameters that are difficult to identify under normal operating conditions due to insufficient excitation. For example, the time constant in process parameters hardly changes during steady-state operation, masking its impact on the output; these are considered under-excited parameters. Combining this with the subset of environment-invariant explanatory variables, the set of parameters requiring focused excitation can be determined.

[0082] The core of incentive planning is to design a sequence of micro-perturbations that maximizes information gain under the safety constraints defined by the control barrier function (CBF). The control barrier function ensures that the incentives do not cause the system to go out of bounds by quantifying the distance between the system state and the safety boundary; the information gain is measured by the Fisher information matrix, reflecting the amount of information that the incentive signal can provide for parameter estimation.

[0083] The excitation order is sorted according to the causal contribution score, and parameters with greater impact are excited first; the disturbance sequence can be a sine wave or a pseudo-random signal; the disturbance time should avoid the critical task period of the system; the disturbance frequency band should match the sensitive frequency band of the parameter, such as hysteresis parameters being more sensitive to low-frequency signals; the disturbance amplitude should be small and within the safe range to avoid affecting system performance; the minimum safety margin ensures the minimum distance between the system state and the safety boundary during the excitation process.

[0084] S105. Based on the safety margin constraint strategy, a safety micro-perturbation excitation plan is executed on the target parameter by superimposing the perturbation sequence with the original control input, and the corresponding sensor output and external environment context are collected when the safety micro-perturbation excitation plan is executed to obtain the excitation-response segment; the excitation-response segment includes the excitation signal, the response signal and the external environment context during execution.

[0085] When executing a safety-based micro-disturbance excitation plan, a safety margin constraint strategy must be implemented. The disturbance sequence is superimposed on the original control input and applied to the system, while excitation-response segments are acquired simultaneously. The safety margin constraint strategy ensures that the system state remains within the safety set by real-time monitoring of the control barrier function value. If the safety margin approaches a threshold, the disturbance amplitude is automatically reduced or the excitation is paused to avoid safety risks. Optionally, passive calibration can be performed. The disturbance sequence is superimposed on the original control input. Where, δu t It is a perturbation sequence, u t The original control input is superimposed onto the original control input. This superimposed input retains the core functionality of the original command while introducing targeted disturbances, enabling both normal system operation and the stimulation of characteristic parameter responses. For example, in flow control, the original control input is the valve opening command to maintain stable flow. The superimposed disturbance sequence consists of minute opening fluctuations that do not affect overall flow stability and allow for the inference of valve dead-zone parameters through subtle changes in flow. The acquired excitation-response segment includes the excitation signal (the superimposed control input), the response signal (changes in sensor output), and the external environmental context during execution.

[0086] S106. Based on residuals, uncertainty estimation, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and excitation-response segments, a hierarchical estimator is used to calibrate and estimate the target parameters, and a calibration package is generated based on the calibration estimation results. The calibration package includes a set of parameter calibration values, posterior covariance, excitation-response segment parameters, timestamps, and version information. The calibration package is used to calibrate and update the dynamic parameters of the intelligent measurement and control system.

[0087] Based on previously acquired residuals, uncertainty estimates, various features, causal contribution scores, and stimulus-response segments, a hierarchical estimator is used to calibrate and estimate the target parameters, generating a calibration package for system updates. The hierarchical estimator is designed based on the dynamic characteristics of the parameters: for rapidly drifting parameters such as sensor bias and scale, an adaptive Kalman filter (AKF) is used to quickly adjust parameters by tracking residual changes in real time. For example, sensor bias may change by 0.1 units per hour due to temperature fluctuations; AKF can offset this drift through high-frequency updates. For slowly varying parameters such as actuator dead zone and process time constant, a short-window moving horizon estimation (MHE) is used, combining historical data and stimulus-response segments for constraint optimization. For example, the change period of the process time constant may be as long as several days; MHE improves estimation accuracy through accumulated data within a sliding window.

[0088] In the estimation process, information from multiple sources needs to be integrated. Specifically, the residual provides real-time deviation feedback, the uncertainty estimate is used to adjust the estimation weight, the time delay estimate in the delay cross-correlation feature serves as the prior value of the process time delay parameter, the hysteresis loop area guides the regularization of the actuator hysteresis parameter, the causal contribution score determines the update priority of the parameter, and the excitation-response segment is weighted and included in the calculation due to its rich information.

[0089] The final calibration package contains a set of parameter calibration values, posterior covariance, excitation-response segment parameters, timestamps, and version information. After the calibration package is distributed, the system automatically updates the sensor compensation table, actuator inverse model, and controller parameters, achieving closed-loop calibration of dynamic parameters.

[0090] In the aforementioned intelligent measurement and control optimization method for dynamic parameter adaptive calibration, fixed-period data acquisition ensures the continuity and uniformity of data over time, avoiding feature extraction distortion caused by irregular sampling. Introducing external environmental context enables the measurement and control system to maintain robust adaptability under dynamically changing external environments, providing more comprehensive data support for subsequent causal analysis and parameter calibration. Differentiable twin models retain gradient information during model inference, allowing parameter optimization and feature extraction to be performed end-to-end, improving the consistency between prediction and calibration. Residual and uncertainty estimation allows for real-time identification of prediction bias and signal reliability, providing a reliable basis for subsequent active excitation and calibration. Extraction of delay cross-correlation features and hysteresis loop area features helps to quantitatively describe the system's dynamic response characteristics and nonlinear hysteresis effects, thereby more accurately identifying key factors leading to system performance degradation. Invariant causal testing based on external environmental context and counterfactual simulation generate causal attribution reports. Invariant causal testing eliminates spurious correlation features that only hold true in specific environments, extracting cross-environmentally stable explanatory variables and improving the generalization ability of parameter identification and adjustment. Counterfactual simulation can not only quantify the causal contribution scores of each explanatory variable, but also predict the adjustment effect without interfering with the real system, reducing experimental risks and costs. Targeted excitation design for underexcited parameters avoids repeated perturbations to well-identified parameters, improving calibration efficiency. Constraints from a safe micro-perturbation excitation scheme can effectively improve parameter identifiability without causing system instability or performance degradation. The superposition of excitation signals and control inputs based on a safety margin constraint strategy ensures that the impact of the excitation signal on the system is controllable and does not exceed safety boundaries, making it suitable for critical measurement and control scenarios with continuous operation. Synchronous acquisition of excitation and response can directly form a high-value excitation-response dataset, providing targeted, high-quality samples for parameter calibration. A hierarchical estimator combines multiple features and causal contribution scores for parameter calibration estimation. Hierarchical estimation can integrate residuals, uncertainties, dynamic features, and causal attribution information to achieve comprehensive correction of target parameters from multiple sources, improving parameter estimation accuracy and stability. The output of posterior covariance can quantify the uncertainty after calibration, facilitating dynamic risk management in the subsequent operation of the measurement and control system. By generating calibration packages and dynamically updating parameters, and using structured information such as calibration value sets, version information, and timestamps, the measurement and control system can be made traceable, rollbackable, and versioned, thus improving the manageability of system maintenance and upgrades.

[0091] In one embodiment, before acquiring observation data collected at fixed intervals, the method further includes:

[0092] S11. Based on the structural variance model corresponding to the intelligent measurement and control system, a differentiable twin model is obtained by fitting and estimating representative working condition data from historical observation data.

[0093] The numerical expression corresponding to the differentiable twin model can be obtained through the following formula:

[0094]

[0095] Where, x t The state of the controlled object at time t; u t For control input; For sensor output; The dynamic parameters of the sensor to be calibrated; The dynamic parameters of the actuator to be calibrated; These are the dynamic parameters to be calibrated in the process. The control input is the result of nonlinear / hysteresis / time-delay mapping of the actuator; ω t For process noise; v t For measuring noise.

[0096] A differentiable twin model is a differentiable system model, including physical, gray-box, or black-box models, used for forward simulation and sensitivity / Jacobi calculations. It reflects the dynamic behavior of the system, obtaining information on the sensitivity of the system state and output to parameters by differentiating the functions (f, g, h) in the model, providing a foundation for subsequent analysis and calculations. For example, a differentiable twin model can be constructed in two ways: one is by directly using an existing physical or mechanistic model in the field; if no suitable model is available in the field, a system identification method, such as NARX (Nonlinear Autoregressive Moving Average) or a state-space black-box model, is used to fit a differentiable model. In representative operating condition data D... hist ={u t ,y t ,c t On the}, the selected or fitted model is estimated to obtain a differentiable twin model (f, g, h).

[0097] S12. Based on the twin linearization model, the Jacobian matrix is ​​obtained by linearizing the representative operating points corresponding to the representative operating conditions data. The twin linearization model is obtained by performing first-order linearization on the model function in the differentiable twin model.

[0098] The twin linearization model is based on the differentiable twin model, at the current operating point. The model obtained by performing a first-order linearization on the functions (f, g, h) in the differentiable twin model yields the partial derivative (Jacobi) matrix. The linearized model is simpler in form and is a linear approximation of the original differentiable twin model near the local operating point.

[0099] Indicatively, based on a twin linearization model, for representative operating points corresponding to representative operating conditions, the model function in the differentiable twin model is linearized to order one, yielding the Jacobian matrix. Specifically, this is calculated using automatic differentiation or symbolic differentiation. The partial derivative matrix, where x is the state of the controlled object, u is the control output, and θ is the partial derivative matrix. S θ A and θ P The distribution consists of dynamic parameters to be calibrated from sensors, actuators, and processes. The Jacobian matrix reflects the local rate of change of the model function at the current operating point and is an important basis for subsequent calculations such as the Fischer information matrix.

[0100] S13. Based on the preset control input sequence, calculate the sensor output sensitivity sequence to each parameter according to the measurement noise variance and Jacobian matrix of historical observation data, and obtain the Fischer information matrix by discrete summation of the sensitivity sequence.

[0101] Based on the preset control input sequence, and combined with the measurement noise variance and Jacobian matrix obtained from historical observation data, the sensitivity sequence of the sensor output to each parameter is calculated. Wherein, CΦ t,0 G represents the influence of process parameters, S represents the influence of sensor parameters, and the parts related to actuator parameters, etc., where Φ t,τ The state transition matrix is ​​obtained recursively from A. The Fischer information matrix is ​​obtained by discretizing and summing the sensitivity sequences. Where R is the measurement noise variance.

[0102] S14. Determine the identifiability index of each parameter based on the Fischer information matrix, and visualize it in combination with the classic excitation frequency band of each parameter to obtain the baseline identifiability map.

[0103] Schematic illustration: Determining the discriminability index of each parameter based on the Fischer information matrix. Where ∈ is a very small positive number to avoid matrix distortion, and I is an identity matrix with the same dimension as the Fisher information matrix. It also preserves the marginal information of each parameter component, such as the diagonal, condition number, and eigenvalue. Each parameter θ... i Information intensity, such as Or an index based on A-opt / D-opt, and its composition The typical excitation frequency bands obtained from the spectrum are visualized to obtain a baseline identifiable spectrum. This spectrum intuitively shows the information intensity of each parameter and the characteristics of the excitation frequency band, providing a clear reference for judging the identifiability of the parameters.

[0104] S15. Based on the preset underexcitation judgment strategy, the underexcitation parameters in the baseline identifiable spectrum are marked to obtain a subset of underexcitation parameters.

[0105] For example, let the threshold τ be... or θ i The corresponding sensitivity frequency band almost does not overlap with the energy frequency band of the current operating condition, and is therefore marked as underexcited. Marking the underexcited parameter subset is to clarify the objects that need to be focused on for subsequent excitation, and to solve the problem of parameters being unidentifiable due to insufficient excitation under normal operating conditions.

[0106] In one embodiment, a differentiable twin model is used to predict the current state of the controlled object and the sensor output based on the control input and sensor output. Feature analysis is then performed based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation characteristics, and hysteresis area characteristics, including:

[0107] S21. Input the control input and the state of the controlled object from the previous moment into the differentiable twin model to obtain the predicted state of the controlled object at the current moment; the previous moment is a fixed period based on the past of the current moment.

[0108] Indicatively, the control input and the controlled object state from the previous moment are input into a differentiable twin model to obtain the predicted controlled object state at the current moment. The previous moment is a fixed period prior to the current moment, ensuring the continuity of the time sequence and making the prediction based on the recent operating state of the system, which conforms to the evolution law of dynamic systems.

[0109] S22. Input the predicted state of the controlled object into the differentiable twin model to obtain the predicted sensor output at the current moment.

[0110] The predicted state of the controlled object is input into the differentiable twin model to obtain the predictive sensor output at the current moment. Sensor output prediction is a crucial link connecting system state with actual measurement, and its accuracy directly affects the effectiveness of subsequent residual calculation.

[0111] S23. Compare the predicted sensor output with the actual sensor output at the current moment to obtain the residual.

[0112] The residual is obtained by comparing the predicted sensor output with the actual sensor output at the current moment.

[0113] S24. By calculating the cross-correlation between the change in control input and the change in sensor output under different delays, the delay cross-correlation characteristics including time delay estimation and excitation-response coupling strength are obtained; the change in control input is obtained from the difference between the control inputs of two adjacent fixed periods; the change in sensor output is obtained from the difference between the sensor outputs of two adjacent fixed periods.

[0114] The delay cross-correlation characteristics are obtained by calculating the cross-correlation between changes in control input and sensor output at different delays. The change in control input Δu is obtained from the difference between control inputs in two adjacent fixed periods, while the change in sensor output Δy is obtained from the difference between sensor outputs in adjacent periods. This differential processing highlights the dynamic components of input and output changes and eliminates the influence of static bias. Cross-correlation calculation ρ uy (l)=xcorr(Δu t-l:t ,Δy t-l:t ), l∈[-LL], taking the peak position l* as the time delay estimate, and the peak height reflects the excitation-response coupling strength. It is used to measure the correlation between input changes and output changes under different delays l, where the peak position l* is regarded as the time delay estimate, representing the time difference that the input has a significant impact on the output, and the peak height reflects the excitation-response coupling strength, that is, the magnitude of the impact of the input change on the output change.

[0115] S25. Based on a sliding time window of a preset length, calculate the loop area formed by the control input and the sensor output to obtain the hysteresis loop area characteristics.

[0116] Based on a sliding time window of preset length, the loop area formed by the control input and sensor output is calculated to obtain the hysteresis area characteristics. Optionally, (u,y) or The loop area and hysteresis area of ​​the trajectory are calculated by approximate integration. hyst ≈∮ydu, which is positively correlated with the nonlinear characteristics of actuators or sensors, such as hysteresis and dead zone. For example, in valve control, if there is hysteresis in the relationship between valve opening and flow rate, the trajectories of the two will form a closed loop. The larger the loop area, the more significant the hysteresis effect.

[0117] In one embodiment, such as Figure 2 As shown, an invariance causality test is performed based on the external environment context to obtain a subset of explanatory variables that are invariant across environments. Counterfactual simulations are then performed on each parameter in this subset to generate a causal attribution report, including:

[0118] S201. Divide the external environment context into multiple environment subsets according to the environmental scenario, and fit an error generation model for each environment subset.

[0119] Based on the environmental scenario, the external environmental context is divided into multiple environmental subsets e, and an error generation model is fitted to each environmental subset. Where r t For the residual, z i,t It is composed of twin sensitivity and eigenvector φ t =[ρ uy (l),l*,A hyst Explanatory variables consisting of ,Δu,Δy, etc. and For model parameters, ∈ t This represents the error term. The external environmental context encompasses factors such as load, temperature, and operating mode that may affect system dynamics; different environmental scenarios lead to different system characteristics. Dividing these contexts into multiple environmental subsets, such as high-temperature environments, low-temperature environments, light-load conditions, and heavy-load conditions, allows for targeted analysis of the impact of parameters in different environments. For each environmental subset, an error generation model is fitted to describe the relationship between residuals and parameter-related variables.

[0120] S202. Calculate the cross-environmental stability index of each parameter based on the error generation model, and filter the parameters according to the cross-environmental stability index to obtain a subset of explanatory variables that are invariant across environments.

[0121] Based on the error generation model, cross-environment stability indices are calculated for each parameter, and a subset of explanatory variables that remain unchanged across environments is selected accordingly. These cross-environment stability indices are measured using statistical tests, such as statistical tests for parameter drift and likelihood ratio tests for pooled regressions, to assess the correlation between the model coefficients and the parameters. Consistency across different environmental subsets e. If the coefficient of a certain parameter is statistically invariant across all environmental subsets—that is, the regression residual distribution is consistent, the parameter is stable, and it has strong explanatory power for the residuals—then this parameter is included in the cross-environment invariant explanatory variable subset S. This subset represents the core parameters that have a stable impact on the residuals under different environments, and are the key potential factors leading to systematic bias, pointing the way for subsequent accurate calibration.

[0122] S203. Based on the differentiable twin model, each parameter of the subset of explanatory variables that are invariant across environments is slightly intervened, and the control input at the current moment and the external environmental context are replayed to obtain the counterfactual sensor output.

[0123] Based on a differentiable twin model, each parameter in the subset of explanatory variables that are invariant across environments is slightly intervened, and the results are replayed using the current control input and the external environmental context to obtain the counterfactual sensor output. Specifically, in the differentiable twin model, for a certain parameter θ... iBy applying a small perturbation while keeping other parameters and inputs constant, the sensor output after the parameter change is simulated by replaying the current control input and external environmental context. This process leverages the simulation capabilities of differentiable twin models to construct a counterfactual scenario of how the output would be affected by parameter changes, enabling the individual quantification of the impact of each parameter on the output.

[0124] Optionally, different historical input segments or consecutive current inputs can be replayed multiple times to comprehensively evaluate the impact of parameter changes on the model output. To analyze the impact of sensor parameter changes, the model can be run before and after the sensor parameters are intervened using the same historical input, and the output differences can be compared. In real-time monitoring scenarios, the current input can also be replayed to promptly detect the immediate impact of parameter changes on the system.

[0125] S204. Calculate the difference between the counterfactual sensor output, the predicted sensor output, and the actual sensor output at the current moment to obtain the causal contribution score corresponding to each parameter.

[0126] The causal contribution score can be obtained using the following formula:

[0127]

[0128] Where, α i The causal contribution score for the i-th parameter; y t This represents the actual sensor output at the current moment. Output of the counterfactual sensor; δ is the value of the disturbance corresponding to a small intervention; ||·||2 is the L2 norm.

[0129] S205. Generate a causal attribution report based on the causal contribution score.

[0130] A causal attribution report is generated based on the causal contribution score. The report includes a list of suspicious parameters and their confidence intervals, sorted by causal contribution score.

[0131] In one embodiment, such as Figure 3 As shown, based on the subset of underexcited parameters and the subset of explanatory variables that are invariant across environments in the baseline identifiable spectrum, a safe active excitation program is developed for the parameters, resulting in a safe micro-perturbation excitation scheme, including:

[0132] S301. Within a sliding window with a preset step size, perform twin linearization on the differentiable twin model based on the current working point to obtain the current Jacobian matrix.

[0133] Within a sliding window of preset step size t:tW, the differentiable twin model is linearized using twinning based on the current operating point, yielding the current Jacobian matrix. The sliding window is used to focus on the system's recent dynamic characteristics, ensuring the linearization result reflects the current operating conditions. The current operating point, determined by the system's real-time state, control inputs, and parameters, serves as the benchmark for linearization. First-order linearization is performed on the state evolution function, sensor output function, and actuator mapping function in the differentiable twin model using automatic differentiation or symbolic differentiation, resulting in Jacobian matrices reflecting the sensitivity of parameters to the output, such as partial derivative matrices related to process parameters and sensor parameters.

[0134] S302. Calculate the current sensitivity sequence of each parameter based on the current Jacobian matrix, and perform discrete summation based on the current sensitivity sequence to obtain the current Fisher information matrix.

[0135] Calculate the current sensitivity sequence of each parameter based on the current Jacobian matrix. The current Fisher information matrix is ​​obtained by discrete summation of the sensitivity sequence. Sensitivity sequences describe the degree to which a sensor output is sensitive to various parameters. Their calculation integrates information from the Jacobian matrix and system state transitions. For example, the sensitivity of process parameters needs to be derived using the state transition matrix, while the sensitivity of actuator parameters is related to the mapping relationship with the control input. Based on this, the inverse matrix of the measurement noise variance is used... The current Fisher information matrix is ​​obtained by weighted summation of the sensitivity sequences. The larger the element value of this matrix, the richer the identifiable information of the corresponding parameter, which is a key indicator for measuring the current identifiability of the parameter.

[0136] S303. Based on the parameters in the underexcited parameter subset and the cross-environment invariant explanatory variable subset, extract the target parameter submatrix from the current Fisher information matrix, and calculate the instantaneous identifiability corresponding to each parameter based on the target parameter submatrix.

[0137] For parameters within the underexcited parameter subset and the cross-environment invariant explanatory variable subset, extract the target parameter submatrix from the current Fisher information matrix. Based on this, the instantaneous identifiability of each parameter is calculated. The under-stimulated parameter subset consists of parameters identified in the baseline analysis that lack sufficient identifiability under normal operating conditions. The cross-environment invariant explanatory variable subset represents the core parameters with a stable impact on the residuals. The intersection or union of these two subsets constitutes the target parameters requiring focused stimulation. By extracting the sub-matrices corresponding to the target parameters and calculating the logarithm of their determinants (i.e., the instantaneous identifiability index), the current identifiability level of these parameters can be quantified. If the index is below a threshold, it indicates that active stimulation is needed to enhance their information content.

[0138] S304. Calculate the priority score of each parameter based on the immediate identifiability and causal contribution score, and arrange the parameters in order of priority score from highest to lowest to obtain the excitation order.

[0139] Priority scores P for each parameter are calculated based on immediate identifiability and causal contribution score. i =w1·(1-J i (t))+w2·α i Here, w1 and w2 are weighting coefficients used to balance the degree of information insufficiency with the degree of contribution to the residuals, and are sorted by score to obtain the incentive order. The lower the immediate identifiability, the more information the parameter currently needs to be supplemented; the higher the causal contribution score, the more significant the parameter's impact on the residuals. The priority score obtained by combining the two can comprehensively reflect the necessity of parameter incentive. Parameters with higher priority should receive incentive resources first to ensure that limited incentive energy can maximize the overall calibration effect of the system.

[0140] S305. Based on the information gain maximization objective and constraints, a safety micro-perturbation incentive plan is obtained by superimposing small-amplitude safety incentives through Bayesian optimization and combining the incentive sequence.

[0141] Based on the information gain maximization objective and constraints, Bayesian optimization is used to solve for superimposed small-amplitude safety stimuli, and a safe micro-perturbation stimulus plan is generated by combining the stimulus sequence. The information gain maximization objective aims to design a small-amplitude perturbation sequence that maximizes the increment of the determinant of the Fisher information matrix of the target parameters, i.e., maximizing its identifiability. Constraints include a safety set defined by the control barrier function, ensuring the system state does not exceed limits, an upper limit on performance deviation to avoid stimuli affecting normal system operation, and limits on perturbation energy and frequency band to ensure small and targeted stimuli. Bayesian optimization can efficiently search for the optimal perturbation sequence while satisfying the constraints. Combined with the stimulus sequence, it determines the timing of stimulus, perturbation amplitude, frequency band, and other details for each parameter. The resulting safe micro-perturbation stimulus plan ensures both the safety and effectiveness of the stimulus and provides clear guidance for subsequent execution and data acquisition.

[0142] In one embodiment, based on the information gain maximization objective and constraints, a safety micro-perturbation incentive plan is obtained by solving a Bayesian optimization solution with superimposed small-amplitude safety incentives and combining the incentive sequence, including:

[0143] S31. The predicted state of the controlled object is filtered and corrected by adaptive Kalman filtering to obtain the corrected predicted state of the controlled object and the uncertainty estimate.

[0144] Optionally, an uncertainty estimate can be calculated using a Kalman filter algorithm to quantify the reliability of the predicted and measured values; the smaller the uncertainty, the higher the data reliability. Adaptive Kalman filtering (KF) or moving horizon estimation (MHE) can then be used for filtering correction to obtain the corrected state. Uncertainty estimation of state And the output uncertainty estimate. Adaptive Kalman filtering can continuously adjust the filter gain using real-time observation data, thereby correcting the prediction bias of the differentiable twin model and making the corrected state of the controlled object closer to the actual system state. Simultaneously, this process also outputs an uncertainty estimate, quantifying the reliability of the corrected state and providing an accuracy reference for subsequent state prediction and constraint verification.

[0145] S32. Input the corrected state of the controlled object into the differentiable twin model to obtain the state of the controlled object at future time, and calculate the corresponding prediction Jacobian matrix based on the state of the controlled object at future time.

[0146] The corrected state of the controlled object is input into a differentiable twin model to obtain the future state of the controlled object, and the corresponding prediction Jacobian matrix is ​​calculated accordingly. The differentiable twin model, through a state evolution function, infers the future state changes based on the corrected current state and a preset control input trend. This process considers the dynamic characteristics of the system and the influence of parameters. Based on this, the model function corresponding to the future state is linearized to obtain the prediction Jacobian matrix. This matrix reflects the sensitivity of the parameters to the output at future times and is crucial for calculating the future information gain.

[0147] S33. Calculate the predicted sensitivity sequence of each target parameter based on the predicted Jacobian matrix, and perform discrete summation based on the predicted sensitivity sequence to obtain the predicted Fisher information matrix.

[0148] The predicted sensitivity sequence for each target parameter is calculated based on the predicted Jacobian matrix, and the predicted Fisher information matrix is ​​obtained through discrete summation. The predicted sensitivity sequence integrates the influence of future parameters on the sensor output, and its calculation combines the predicted Jacobian matrix and state transition information. Discrete summation of this sequence, combined with weighting using the inverse matrix of the measurement noise variance, yields the predicted Fisher information matrix, which quantifies the potential for identifiable information about future parameters, providing a basis for evaluating the information gain brought about by the stimulus.

[0149] S34. If the instantaneous identifiability corresponding to the parameter is less than the preset threshold, then a small safety excitation is superimposed on the information obtained by Bayesian optimization until the gain maximization objective and constraint conditions are met, and a disturbance sequence is generated. The constraint conditions include safety function constraints, performance deviation constraints and frequency band constraints. The safety function constraints include safety set, control barrier function and linear constraints. The performance deviation constraints include limiting sensor output deviation and limiting disturbance energy.

[0150] If the instantaneous identifiability of a parameter is less than a preset threshold, it indicates that the current information for that parameter is insufficient, and Bayesian optimization is needed to solve for the superimposed small-amplitude safety stimulus. Bayesian optimization aims to maximize information gain, specifically by maximizing the difference in determinant between the predicted Fisher information matrix and the current Fisher information matrix. This objective ensures that the stimulus effectively improves parameter identifiability. The small perturbation input sequence δu to be superimposed is then selected. t:t+K-1 As a design variable, it can be restricted to a sine wave, pseudo-random binary sequence (PRBS), or narrowband swept frequency form at several frequency points for ease of calculation and implementation, enabling convex approximation and simplifying the subsequent optimization process. The objective is to maximize the predicted Fisher information matrix (FIM) gain, approximating the D-opt criterion. This is achieved through calculation... To achieve this, in which The result is obtained from the predicted Jacobian. The objective is to find the excitation signal that maximizes the improvement of the parameter discriminability index, allowing the system to significantly enhance the discriminability of its parameters after acquiring new information. Let the safety function be h(x), such as the boundary constraint function of physical quantities like temperature, position, and pressure. Define the safety set S = {x: h(x) ≥ 0}, and use the control barrier function (CBF) inequality to constrain the input. Transform into linear constraint A CBF (x)u≤b CBF (x) ensures that the system state remains within a safe range under the excitation, preventing abnormal system operation due to the excitation. Limits output deviation to ensure... Ensure that the excitation does not cause excessive deviation between the system output and the expected value, maintaining normal system performance. Simultaneously, limit the disturbance energy ||δu||²≤η to prevent excessive excitation energy from adversely affecting the system. The spectrum of δu should avoid the main task frequency band to prevent conflicts between the excitation signal and the system's main task frequency band, ensuring the normal operation of the system's main functions.

[0151] The optimal solution is searched within the constraint space using Bayesian optimization, ultimately generating a perturbation sequence that meets the conditions.

[0152] S35. Calculate the minimum safety margin corresponding to the current moment based on the safety function constraints, and obtain the safety micro-perturbation excitation plan by combining the perturbation sequence and its corresponding superimposed small-amplitude safety excitation parameters and excitation order.

[0153] Calculate the minimum safety margin m at the current moment based on the safety function constraints. t =min τ∈[t-W,t]This refers to the minimum value of the safety function within the sliding window, which intuitively reflects the minimum margin between the system and the safety boundary. Combining the disturbance sequence, including the amplitude and timing of the excitation signal, with the addition of small-amplitude safety excitation parameters such as energy upper limit, frequency band range, and excitation order, a safety micro-disturbance excitation plan is formed. This plan clarifies the excitation details of each parameter and ensures the safety of the excitation process through a minimum safety margin, providing complete guidance for subsequent execution.

[0154] In one embodiment, the hierarchical estimator consists of a fast layer and a slow layer connected together; the fast layer uses Kalman filtering to model the sensor dynamic parameters to be calibrated and the fast drift term as a random walk; the slow layer uses a moving horizon of a preset length to solve the constraint optimization, which includes errors, state transition errors and deviations between the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated and the prior values, under the condition of satisfying constraints.

[0155] A hierarchical estimator structure is used for parameter estimation. The fast layer employs either an adaptive noise Kalman filter (AKF) or an unscented Kalman filter (UKF) for the sensor parameter θ. S (e.g., zero bias, scale, noise) and fast drift terms are modeled as random walks, and a large process noise is set. To achieve fast response, the slow layer employs short-window moving horizon estimation (MHE) based on actuator parameter θ. A and process parameter θ P Solve a small-scale constrained optimization problem within a moving time window of length H. The objective function includes output error, state transition error, and deviation of parameters from prior values, while also considering constraints such as physical boundaries and hysteresis / delay parameter intervals.

[0156] Based on residuals, uncertainty estimates, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and stimulus-response segments, a hierarchical estimator is used to perform calibration estimation of the target parameters, and a calibration package is generated based on the calibration estimation results, including:

[0157] S41. Based on the residual and uncertainty estimates, iteratively calculate the dynamic parameters of the sensor to be calibrated until the accuracy converges, and obtain the posterior distribution of the dynamic parameters of the sensor to be calibrated; the posterior distribution includes the mean and covariance of the dynamic parameters of the sensor to be calibrated; the mean is the parameter calibration value of the dynamic parameters of the sensor to be calibrated.

[0158] Sensor parameters often exhibit rapid drift characteristics. The fast layer employs Kalman filtering to model these parameters and the rapid drift term as a random walk process. This assumes that the parameters change slowly and randomly over time, and by setting a large process noise level, the parameters are allowed to adjust rapidly to adapt to the drift. During iteration, the residuals provide real-time feedback on parameter deviations, while the uncertainty estimate is used to adjust the filter gain to ensure the stability of the estimate. When the change in the parameter estimate is less than a set threshold, the accuracy is considered to have converged. The resulting posterior distribution includes the mean and covariance. The mean can be directly used for zero-bias compensation, scale correction, and other operations at the sensor front end to assess the sensor's zero-bias, scale error, and noise characteristics.

[0159] S42. Based on a preset length of moving time window, the parameter calibration values ​​of the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated are obtained through constraint optimization solution state and data-guided joint estimation. The data-guided corresponding time delay estimation is used as the prior center of the time delay parameters in the process parameters, the hysteresis loop area feature is used as the regularization guide for the actuator hysteresis model, the causal contribution score in the causal attribution report that is higher than the preset attribution threshold is used to relax the uncertainty variance of the prior estimation, the objective function information is weighted according to the instantaneous identifiability, and the convergence of the parameters is improved by setting the weight of the excitation-response segment.

[0160] For actuator dynamic parameters to be calibrated, such as dead zone and hysteresis, and process dynamic parameters to be calibrated, such as time delay and gain, the corresponding parameter calibration values ​​are obtained through constraint optimization and data-guided joint estimation based on a preset-length moving-horizon window H. Both actuator and process dynamic parameters are slowly varying and require constraints based on historical data; therefore, a preset-length moving-horizon estimation (short window MHE) is used for the slow-moving layer. The moving-horizon window focuses on recent historical data, and an objective function is constructed that includes sensor output error, state transition error, and parameter deviations from prior values. Where Q is the variance of the disturbance noise, and the constrained optimization problem is solved to achieve joint estimation of state and parameters, including the dead zone and hysteresis of the actuator, and the time delay, gain, time constant and other slowly varying parameters of the process.

[0161] Specifically, constraints include ensuring parameters satisfy physical boundaries and the monotonicity of hysteresis characteristics to guarantee estimation results conform to actual physical laws. The time delay estimation and hysteresis loop area characteristics within the delay cross-correlation features are crucial for the slow layer. Time delay estimation, as the prior center for time delay parameters in the process parameters, narrows the estimation range, allowing the slow layer to focus more on reasonable time delay values ​​during estimation. Hysteresis loop area characteristics, as a regularization guide for actuator hysteresis models such as Bouc-Wen parameters, help the slow layer better capture the hysteresis characteristics of the actuator, more accurately reflecting the actual system situation when estimating actuator and process parameters. High and low attribution parameters identified in the causal attribution report influence the slow layer's estimation. For high attribution parameters, the slow layer relaxes the prior θ. prior The variance of these parameters makes the estimator more inclined to update them, as they have a significant impact on system bias. For low-attribution parameters, the prior is tightened to avoid unnecessary adjustments, thereby improving the accuracy and efficiency of the estimation. Immediate identifiability guides the estimation of the slow layer. When immediate identifiability is low, it indicates insufficient information about certain parameters. The slow layer adds information weighting to the objective function, prioritizing the use of stimulus-response segments and giving them higher weights. This is because stimulus-response segments are specifically designed for those difficult-to-identify parameters, have high information content, and can significantly improve the convergence speed and confidence of these parameters. The stimulus-response segments are concatenated with regular data, and the stimulus-response segments are multiplied by information weights in the cost function. In this way, the slow layer fully utilizes the high information content in the stimulus-response segments, more accurately reflecting the true parameter situation of the system when estimating actuator and process parameters.

[0162] The slow layer employs a moving horizon of length H to solve a small-scale constrained optimization problem that includes output error, state transition error, and parameter deviations from prior values, considering constraints such as physical boundaries, hysteresis parameter ranges, and monotonicity. Under these constraints, the posterior distributions of the actuator and process parameters are obtained, including their mean and covariance. The mean of the actuator parameters to be calibrated is used in the actuator inverse model / feedforward, while the mean of the process parameters to be calibrated is synchronized to the controller and updated via sigma. Together, they constitute the calibration values ​​of the actuator and process parameters in the calibration vector.

[0163] S43. Based on the parameter calibration values ​​of the dynamic parameters to be calibrated of the sensor, the dynamic parameters to be calibrated of the actuator, and the dynamic parameters to be calibrated of the process, obtain the calibration package.

[0164] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0165] Based on the same inventive concept, this application also provides an intelligent measurement and control optimization device for dynamic parameter adaptive calibration, used to implement the intelligent measurement and control optimization method for dynamic parameter adaptive calibration described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the intelligent measurement and control optimization device for dynamic parameter adaptive calibration provided below can be found in the limitations of the intelligent measurement and control optimization method for dynamic parameter adaptive calibration described above, and will not be repeated here.

[0166] In one exemplary embodiment, such as Figure 4 As shown, an intelligent measurement and control optimization device with dynamic parameter adaptive calibration is provided, comprising:

[0167] The measurement and control observation module 401 is used to acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object distributed over time.

[0168] The feature analysis module 402 is used to predict the state of the controlled object and the sensor output at the current moment using a differentiable twin model based on the control input and sensor output, and to perform feature analysis based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation features and hysteresis area features; the delay cross-correlation features include time delay estimation and excitation-response coupling strength;

[0169] The invariant causality module 403 is used to perform invariant causality tests based on the external environment context, obtain a subset of explanatory variables that are invariant across the environment, and perform counterfactual simulations on each parameter in the subset of explanatory variables that are invariant across the environment to generate a causal attribution report; the causal attribution report includes the causal attribution contribution score of each parameter;

[0170] The excitation module 404 is used to perform safe active excitation planning on the parameters based on the subset of underexcited parameters and the subset of explanatory variables that are invariant across the environment in the baseline identifiable map, so as to obtain a safe micro-perturbation excitation plan.

[0171] The response module 405 is used to execute a safety micro-perturbation excitation plan on the target parameter by superimposing the perturbation sequence with the original control input based on the safety margin constraint strategy, and to collect the sensor output and external environment context corresponding to the execution of the safety micro-perturbation excitation plan to obtain the excitation-response segment;

[0172] The calibration module 406 is used to perform calibration estimation of the target parameters using a hierarchical estimator based on residuals, uncertainty estimates, delay cross-correlation characteristics, hysteresis area characteristics, causal contribution fractions, and excitation-response segments, and to generate a calibration package based on the calibration estimation results.

[0173] In one embodiment, it further includes:

[0174] The model building module is used to obtain a differentiable twin model by fitting and estimating representative working condition data from historical observation data based on the structural variance model corresponding to the intelligent measurement and control system.

[0175] The linearization module is used to perform linearization processing based on the twin linearization model and the representative operating points corresponding to the representative operating conditions data to obtain the Jacobian matrix.

[0176] The information matrix module is used to calculate the sensor output sensitivity sequence to each parameter based on the preset control input sequence, the measurement noise variance and Jacobian matrix of historical observation data, and obtain the Fischer information matrix by discrete summation of the sensitivity sequence.

[0177] The spectral module is used to determine the identifiability index of each parameter based on the Fischer information matrix, and to visualize it in combination with the classic excitation frequency band of each parameter to obtain the baseline identifiability spectral.

[0178] The underexcitation labeling module is used to label the underexcitation parameters in the baseline identifiable spectrum based on a preset underexcitation judgment strategy, thereby obtaining a subset of underexcitation parameters.

[0179] In one embodiment, the feature analysis module 402 is further configured to:

[0180] The control input and the state of the controlled object at the previous moment are input into the differentiable twin model to obtain the predicted state of the controlled object at the current moment; the previous moment is a fixed period based on the past of the current moment;

[0181] The predicted state of the controlled object is input into the differentiable twin model to obtain the predicted sensor output at the current moment;

[0182] The predicted sensor output is compared with the actual sensor output at the current moment to obtain the residual;

[0183] By calculating the cross-correlation between the change in control input and the change in sensor output under different delays, the delay cross-correlation characteristics, including time delay estimation and excitation-response coupling strength, are obtained.

[0184] Based on a sliding time window of preset length, the loop area formed by the control input and sensor output is calculated to obtain the hysteresis loop area characteristics.

[0185] In one embodiment, the invariant causality module 403 is further configured to:

[0186] The external environment context is divided into multiple environment subsets based on the environmental scenario, and an error generation model is fitted for each environment subset.

[0187] Based on the error generation model, the cross-environment stability index of each parameter is calculated, and the parameters are screened according to the cross-environment stability index to obtain a subset of explanatory variables that are invariant across the environment.

[0188] Based on the differentiable twin model, each parameter of the subset of explanatory variables that are invariant across environments is slightly intervened, and the control input at the current moment and the external environmental context are replayed to obtain the counterfactual sensor output.

[0189] The causal contribution score for each parameter is obtained by calculating the difference between the counterfactual sensor output, the predicted sensor output, and the actual sensor output at the current moment.

[0190] A causal attribution report is generated based on the causal contribution score.

[0191] In one embodiment, the excitation module 404 is further configured to:

[0192] Within a sliding window of a preset step size, the differentiable twin model is linearized using twins based on the current operating point to obtain the current Jacobian matrix;

[0193] The current sensitivity sequence of each parameter is calculated based on the current Jacobian matrix, and the current sensitivity sequence is discretized and summed to obtain the current Fisher information matrix.

[0194] Based on the parameters in the underexcited parameter subset and the cross-environment invariant explanatory variable subset, the target parameter submatrix is ​​extracted from the current Fisher information matrix, and the instantaneous identifiability corresponding to each parameter is calculated based on the target parameter submatrix.

[0195] The priority score of each parameter is calculated based on the immediate identifiability and causal contribution score, and the parameters are arranged in order of priority score from highest to lowest to obtain the incentive order.

[0196] Based on the information gain maximization objective and constraints, a safety micro-perturbation incentive plan is obtained by superimposing small-amplitude safety incentives through Bayesian optimization and combining the incentive sequence.

[0197] In one embodiment, an optimization solution module is also included, for:

[0198] The predicted state of the controlled object is filtered and corrected by adaptive Kalman filtering to obtain the corrected predicted state of the controlled object and the uncertainty estimate.

[0199] The state of the controlled object is corrected and input into the differentiable twin model to obtain the state of the controlled object at future time, and the corresponding prediction Jacobian matrix is ​​calculated based on the state of the controlled object at future time.

[0200] The predicted sensitivity sequence of each target parameter is calculated based on the predicted Jacobian matrix, and the predicted Fisher information matrix is ​​obtained by discrete summation based on the predicted sensitivity sequence.

[0201] If the instantaneous identifiability of the parameter is less than the preset threshold, a small safety excitation is superimposed on the information obtained by Bayesian optimization until the gain maximization objective and constraints are met, and a perturbation sequence is generated. The constraints include safety function constraints, performance deviation constraints, and frequency band constraints. The safety function constraints include safety sets, control barrier functions, and linear constraints. The performance deviation constraints include limiting sensor output deviation and limiting perturbation energy.

[0202] The minimum safety margin corresponding to the current moment is calculated based on the safety function constraints, and the safety micro-perturbation excitation plan is obtained by combining the perturbation sequence and its corresponding superimposed small-amplitude safety excitation parameters and excitation order.

[0203] In one embodiment, the calibration module 406 is further configured to:

[0204] Based on the residual and uncertainty estimates, the dynamic parameters of the sensor to be calibrated are iteratively calculated until the accuracy converges, and the posterior distribution of the dynamic parameters of the sensor to be calibrated is obtained; the posterior distribution includes the mean and covariance of the dynamic parameters of the sensor to be calibrated; the mean is the parameter calibration value of the dynamic parameters of the sensor to be calibrated.

[0205] Based on a preset moving time window, the parameter calibration values ​​of the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated are obtained through constraint optimization solution state and data-guided joint estimation. Data-guided corresponding time delay estimation serves as the prior center of the time delay parameters in the process parameters, the hysteresis loop area feature serves as the regularization guide for the actuator hysteresis model, the causal contribution score above the preset attribution threshold in the causal attribution report is used to relax the uncertainty variance of the prior estimation, the objective function information is weighted according to the instantaneous identifiability, and the convergence of the parameters is improved by setting the weight of the excitation-response segment.

[0206] A calibration package is obtained based on the calibration values ​​of the dynamic parameters to be calibrated for the sensors, the dynamic parameters to be calibrated for the actuators, and the dynamic parameters to be calibrated for the processes.

[0207] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps in the above method embodiments.

[0208] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0209] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0210] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. A smart measurement and control optimization method with dynamic parameter adaptive calibration, characterized in that, The method includes: Acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object, distributed over time. Using a differentiable twin model, the current state of the controlled object and the sensor output are predicted based on the control input and the sensor output. Based on the prediction results, feature analysis is performed to obtain residuals, uncertainty estimates, delay cross-correlation features, and hysteresis area features. The delay cross-correlation features include time delay estimation and excitation-response coupling strength. Based on the external environment context, an invariance causality test is performed to obtain a subset of explanatory variables that are invariant across the environment. Counterfactual simulations are then performed on each parameter in the subset of explanatory variables to generate a causal attribution report. The causal attribution report includes the causal attribution contribution score of each parameter. Based on the subset of underexcited parameters in the baseline identifiable map and the subset of explanatory variables that are invariant across environments, a safe active excitation plan is performed on the parameters to obtain a safe micro-perturbation excitation plan; the safe micro-perturbation excitation plan includes excitation order, perturbation sequence, perturbation time, perturbation frequency band, perturbation amplitude and minimum safety margin; Based on the safety margin constraint strategy, the safety micro-perturbation excitation plan is executed on the target parameter by superimposing the perturbation sequence with the original control input, and the sensor output and the external environment context corresponding to the execution of the safety micro-perturbation excitation plan are collected to obtain an excitation-response segment; the excitation-response segment includes an excitation signal, a response signal and the external environment context at the time of execution; Based on the residuals, the uncertainty estimate, the delay cross-correlation characteristics, the hysteresis area characteristics, the causal contribution score, and the excitation-response segment, a hierarchical estimator is used to calibrate and estimate the target parameters, and a calibration package is generated based on the calibration estimation results. The calibration package includes a set of parameter calibration values, posterior covariance, excitation-response segment parameters, timestamps, and version information. The calibration package is used to calibrate and update the dynamic parameters of the intelligent measurement and control system.

2. The method according to claim 1, characterized in that, Before acquiring the observation data collected at fixed intervals, the process also includes: Based on the structural variance model corresponding to the intelligent measurement and control system, the differentiable twin model is obtained by fitting and estimating representative working condition data from historical observation data. The numerical expression corresponding to the differentiable twin model can be obtained through the following formula: Where, x t The state of the controlled object at time t; u t For control input; For sensor output; The dynamic parameters of the sensor to be calibrated; The dynamic parameters of the actuator to be calibrated; These are the dynamic parameters to be calibrated in the process. The control input is the result of nonlinear / hysteresis / time-delay mapping of the actuator; ω t For process noise; v t For measuring noise; Based on the twin linearization model, the Jacobian matrix is ​​obtained by linearizing the representative operating points corresponding to the representative operating conditions data; the twin linearization model is obtained by performing first-order linearization on the model function in the differentiable twin model. Based on the preset control input sequence, the sensor output sensitivity sequence to each parameter is calculated according to the measurement noise variance of the historical observation data and the Jacobian matrix, and the Fischer information matrix is ​​obtained by discrete summation of the sensitivity sequence. The identifiability index of each parameter is determined based on the Fischer information matrix, and the baseline identifiability map is obtained by combining the classical excitation frequency band of each parameter with the visualization. Based on a preset underexcitation judgment strategy, the underexcitation parameters in the baseline identifiable spectrum are marked to obtain the subset of underexcitation parameters.

3. The method according to claim 1, characterized in that, The method of using a differentiable twin model to predict the current state of the controlled object and the sensor output based on the control input and the sensor output, and performing feature analysis based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation characteristics, and hysteresis area characteristics, includes: The control input and the state of the controlled object at the previous moment are input into the differentiable twin model to obtain the predicted state of the controlled object at the current moment; the previous moment is based on a fixed period past the current moment; The predicted state of the controlled object is input into the differentiable twin model to obtain the predicted sensor output at the current moment; The residual is obtained by comparing the predicted sensor output with the actual sensor output at the current moment. By calculating the cross-correlation between the change in control input and the change in sensor output under different delays, the delay cross-correlation feature, which includes the time delay estimation and the excitation-response coupling strength, is obtained; the change in control input is obtained by the difference between the control inputs of two adjacent fixed periods; the change in sensor output is obtained by the difference between the sensor outputs of two adjacent fixed periods. Based on a sliding time window of preset length, the loop area formed by the control input and the sensor output is calculated to obtain the hysteresis loop area characteristics.

4. The method according to claim 3, characterized in that, The process of performing an invariant causal test based on the external environment context to obtain a subset of explanatory variables that are invariant across the environment, and then performing counterfactual simulations on each parameter in the subset of explanatory variables to generate a causal attribution report, includes: The external environment context is divided into multiple environment subsets according to the environmental scenario, and an error generation model is fitted for each environment subset; Based on the error generation model, the cross-environment stability index of each parameter is calculated, and the parameters are screened according to the cross-environment stability index to obtain the subset of explanatory variables that are invariant across the environment. Based on the differentiable twin model, each parameter of the cross-environment invariant subset of explanatory variables is slightly intervened, and the control input and the external environmental context at the current moment are used to replay the data to obtain the counterfactual sensor output. The causal contribution score corresponding to each parameter is obtained by calculating the difference between the counterfactual sensor output, the predictive sensor output, and the actual sensor output at the current moment. The causal contribution score is obtained using the following formula: Where, α i The causal contribution score for the i-th parameter; y t This represents the actual sensor output at the current moment. Output of the counterfactual sensor; To predict the sensor output; δ is the disturbance value corresponding to a small intervention; ||·||2 is the L2 norm; The causal attribution report is generated based on the causal contribution score.

5. The method according to claim 3, characterized in that, Based on the subset of underexcited parameters in the baseline identifiable map and the subset of explanatory variables that are invariant across environments, a safe active excitation planning is performed on the parameters to obtain a safe micro-perturbation excitation plan, including: Within a sliding window of a preset step size, the differentiable twin model is linearized according to the current working point to obtain the current Jacobian matrix; The current sensitivity sequence of each parameter is calculated based on the current Jacobian matrix, and the current sensitivity sequence is discretely summed to obtain the current Fisher information matrix. Based on the parameters in the underexcited parameter subset and the cross-environment invariant explanatory variable subset, a target parameter submatrix is ​​extracted from the current Fisher information matrix, and the instantaneous identifiability corresponding to each parameter is calculated based on the target parameter submatrix. The priority score of each parameter is calculated based on the instantaneous identifiability and the causal contribution score, and the parameters are arranged in order of priority score from highest to lowest to obtain the excitation order. Based on the information gain maximization objective and constraints, the safety micro-perturbation incentive plan is obtained by superimposing small-amplitude safety incentives through Bayesian optimization and combining the incentive sequence.

6. The method according to claim 5, characterized in that, The method for obtaining the safety micro-perturbation incentive plan, based on the information gain maximization objective and constraints, through Bayesian optimization and superimposed small-amplitude safety incentives, and combined with the incentive sequence, includes: The predicted state of the controlled object is filtered and corrected by adaptive Kalman filtering to obtain the corrected predicted state of the controlled object and the uncertainty estimate. The corrected state of the controlled object is input into the differentiable twin model to obtain the state of the controlled object at future time, and the corresponding prediction Jacobian matrix is ​​calculated based on the state of the controlled object at future time. The predicted sensitivity sequence of each target parameter is calculated based on the predicted Jacobian matrix, and the predicted sensitivity sequence is discretely summed to obtain the predicted Fisher information matrix. If the instantaneous identifiability corresponding to the parameter is less than a preset threshold, then a small-amplitude safety excitation is superimposed on the information obtained through Bayesian optimization until the gain maximization objective and constraints are met, thereby generating the perturbation sequence. The constraints include safety function constraints, performance deviation constraints, and frequency band constraints. The safety function constraints include safety sets, control barrier functions, and linear constraints. The performance deviation constraints include limiting sensor output deviation and limiting perturbation energy. The minimum safety margin corresponding to the current moment is calculated based on the safety function constraint, and the safety micro-perturbation excitation plan is obtained by combining the perturbation sequence and its corresponding superimposed small-amplitude safety excitation parameters and the excitation order.

7. The method according to claim 5, characterized in that, The hierarchical estimator consists of a fast layer and a slow layer connected together. The fast layer uses Kalman filtering to model the sensor dynamic parameters to be calibrated and the fast drift term as a random walk. The slow layer uses a moving horizon of a preset length to solve the constraint optimization, which includes errors, state transition errors, and deviations between the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated and the prior values, under the condition of satisfying constraints. The method involves using a hierarchical estimator to perform calibration estimation on the target parameter based on the residual, the uncertainty estimate, the delay cross-correlation characteristic, the hysteresis area characteristic, the causal contribution fraction, and the excitation-response segment, and generating a calibration package based on the calibration estimation results, including: Based on the residual and the uncertainty estimate, the dynamic parameters of the sensor to be calibrated are iteratively calculated until the accuracy converges, and the posterior distribution of the dynamic parameters of the sensor to be calibrated is obtained; the posterior distribution includes the mean and covariance of the dynamic parameters of the sensor to be calibrated; the mean is the parameter calibration value of the dynamic parameters of the sensor to be calibrated. Based on a preset length of moving time window, the parameter calibration values ​​of the actuator dynamic parameters to be calibrated and the process dynamic parameters to be calibrated are obtained through joint estimation of the constraint optimization solution state and data guidance. The data guidance corresponds to the time delay estimation as the prior center of the time delay parameters in the process parameters, the hysteresis loop area feature as the regularization guide of the actuator hysteresis model, the causal contribution score in the causal attribution report that is higher than the preset attribution threshold for relaxing the uncertainty variance of the prior estimation, the objective function information weighted judgment based on the instantaneous identifiability, and the setting of the weight of the excitation-response segment to improve the convergence of the parameters. The calibration package is obtained based on the calibration values ​​of the sensor dynamic parameters to be calibrated, the actuator dynamic parameters to be calibrated, and the process dynamic parameters to be calibrated.

8. An intelligent measurement and control optimization device with dynamic parameter adaptive calibration, characterized in that, The device includes: The measurement and control observation module is used to acquire observation data collected at fixed intervals; the observation data includes control inputs, sensor outputs, external environmental context, and the state of the controlled object, distributed over time. The feature analysis module is used to predict the state of the controlled object and the sensor output at the current moment using a differentiable twin model based on the control input and the sensor output, and to perform feature analysis based on the prediction results to obtain residuals, uncertainty estimates, delay cross-correlation features, and hysteresis area features; the delay cross-correlation features include time delay estimation and excitation-response coupling strength; The invariant causality module is used to perform invariant causality tests based on the external environment context, obtain a subset of explanatory variables that are invariant across the environment, and perform counterfactual simulations on each parameter in the subset of explanatory variables to generate a causal attribution report; the causal attribution report includes the causal attribution contribution score of each parameter; The excitation module is used to perform safe active excitation planning on the parameters based on the underexcited parameter subset in the baseline identifiable map and the cross-environment invariant explanatory variable subset, to obtain a safe micro-perturbation excitation plan; The response module is used to execute the safety micro-perturbation excitation plan on the target parameters by superimposing the perturbation sequence with the original control input based on the safety margin constraint strategy, and to collect the sensor output and the external environment context corresponding to the execution of the safety micro-perturbation excitation plan to obtain the excitation-response segment; The calibration module is used to perform calibration estimation on the target parameter using a hierarchical estimator based on the residual, the uncertainty estimate, the delay cross-correlation feature, the hysteresis area feature, the causal contribution fraction, and the excitation-response segment, and to generate a calibration package based on the calibration estimation result.

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

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

Images