Robust adaptive control method for industrial control system and related device

Abstract

The application belongs to the field of industrial control, and discloses a robust adaptive control method for an industrial control system and a related device. Firstly, real-time and historical working condition data are standardized and disturbance detection classification is performed, disturbance features are extracted, and a pre-trained large model is used to predict disturbance trends. Then, customized control parameters are matched or generated through a control strategy library, and after multi-working condition simulation verification, disturbance feedforward compensation and dynamic weight synthesis are combined to obtain the final control output. The method significantly improves the control performance and adaptability in complex industrial scenarios, realizes active disturbance perception, autonomous control strategy generation and dynamic optimization, and overcomes the core defects of the prior art, such as dependence on manual operation, strong conservatism and lack of prediction ability.

Description

Technical Field

[0001] This invention belongs to the field of industrial control technology, and in particular relates to a robust adaptive control method and related device for industrial control systems. Background Technology

[0002] Industrial control systems are a core component of industrial production systems, and their control performance directly affects production efficiency, product quality, and operational safety. In actual industrial operations, control systems are generally affected by multiple uncertainties, such as process parameter deviations caused by equipment aging, process disturbances caused by raw material composition fluctuations, and external interference caused by changes in environmental conditions. These uncertainties can easily lead to decreased system control accuracy and reduced operational stability. Robust adaptive control, as a core control technology for addressing system uncertainties, aims to ensure that the system maintains stable operation and expected control performance under various uncertain conditions through controller design and optimization. It is currently a key research direction in the field of industrial control. With the rapid development of large model technology, its capabilities in time series prediction, state perception, and decision generation are gradually emerging, providing a new technical path for the technological upgrading and performance breakthroughs of robust adaptive control.

[0003] Currently, mainstream robust adaptive dependent control techniques can be broadly categorized into three types: First, classical robust control, typically represented by H∞ control and sliding mode control. This type of technology covers the range of uncertain fluctuations by designing a controller with a fixed structure and possesses a mature theoretical foundation. However, the controller design is conservative, making it difficult to balance control performance and stability under complex operating conditions. Second, adaptive control, typically represented by model reference adaptive control. This type of technology adjusts controller parameters in real time through online parameter estimation, adapting to dynamic changes in system parameters. However, the convergence of the parameter estimation process and the analysis of system closed-loop stability are challenging. Third, predictive control, typically represented by model predictive control. This type of technology solves for the optimal control sequence through rolling optimization, effectively handling multiple constraints of the system. However, online optimization computation is large, requiring high hardware computing power. In addition, existing control methods are generally unable to effectively adapt to strongly coupled and highly nonlinear complex industrial systems, and they generally lack the ability to proactively predict process disturbances. The adjustment rules and adaptive mechanisms of control strategies still heavily rely on manual design and cannot autonomously cope with unknown types of uncertain disturbances.

[0004] It is evident that existing robust adaptive control technologies suffer from core defects such as insufficient adaptability to complex nonlinear systems, lack of proactive disturbance prediction capabilities, and reliance on manual design for control strategy adjustments. Consequently, they fail to meet the high-performance, highly adaptable, and stable control requirements of complex industrial scenarios. Summary of the Invention

[0005] This invention provides a robust adaptive control method and related apparatus for industrial control systems. This method can effectively solve the problems of insufficient adaptability to complex nonlinear systems, lack of active disturbance prediction capability, and reliance on manual design for control strategy adjustment in existing robust adaptive control technologies. It can meet the high-performance, highly adaptable, and stable control requirements in complex industrial scenarios.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A robust adaptive control method for industrial control systems includes: Outlier detection and time-series standardization are performed on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. Based on a standardized time-series operating condition dataset, operating condition disturbance detection is performed. When an operating condition disturbance is determined to exist, the disturbance type is classified and identified. At the same time, multi-dimensional disturbance features are extracted to generate a disturbance feature vector. Historical operating data, disturbance classification results, and disturbance feature vectors are combined into a prediction input sequence, which is then input into a pre-trained large model to infer disturbance trends. The output includes disturbance amplitude range, direction of change, expected duration, and prediction confidence. Extract the current system operating state features, construct the retrieval feature vector by combining the disturbance prediction results, call the preset control strategy library to perform similarity matching on the retrieval feature vector, and obtain the control strategy to be verified if the matching result meets the preset threshold; if the matching result does not meet the preset threshold, combine the system dynamic model and constraints to call the large model to generate customized control strategy parameters, which are used as the control strategy to be verified. The control strategy to be verified is subjected to multi-condition closed-loop simulation test, the comprehensive performance score is calculated, and the strategy is determined to be an executable control strategy after the performance constraints are met. The disturbance feedforward compensation is calculated based on the disturbance prediction results. The weight coefficients of the feedforward and feedback control are dynamically adjusted according to the prediction confidence. The feedforward and feedback control are then weighted and synthesized with the feedback control quantity corresponding to the executable control strategy to obtain the final control output quantity.

[0007] Furthermore, the real-time process operation data includes process variable measurements, equipment operating status signals, and control loop signals; When classifying and identifying disturbance types, a distinction is made between measurable disturbances that can be directly obtained through sensing measurements and unmeasurable disturbances that cannot be directly measured; the disturbance feature vector includes at least four dimensions: disturbance amplitude, direction of change, duration, and rate of change.

[0008] Furthermore, when the prediction confidence of the disturbance prediction result is lower than the preset prediction confidence threshold, the calculation of the disturbance feedforward compensation amount is stopped, and the weight coefficient of the feedforward control is set to zero, wherein the final control output amount is generated based on the feedback control amount.

[0009] Furthermore, before performing multi-condition closed-loop simulation testing on the control strategy to be verified, the following steps are also included: A preset system reduced-order equivalent model is used to perform closed-loop stability verification of the control strategy to be verified. If the verification passes, multi-condition closed-loop simulation test is performed. If the verification fails, the control strategy matching or generation steps are re-executed.

[0010] Furthermore, the multi-condition closed-loop simulation test includes at least the rated operating condition, step disturbance condition, slope disturbance condition, and noise disturbance condition. The comprehensive performance score is calculated by weighting three performance indicators: overshoot, settling time, and steady-state error. The performance constraints also include controlling energy consumption to not exceed a preset proportion of the energy consumption of the benchmark control strategy.

[0011] Furthermore, the disturbance feedforward compensation amount is calculated based on the dynamic transmission relationship from the disturbance channel to the controlled output; The weighting coefficient of the feedforward control increases synchronously with the increase of the disturbance prediction confidence and the increase of the disturbance amplitude, and the weighting coefficient of the feedforward control is set with a preset upper limit.

[0012] Furthermore, after obtaining the final control output, the method further includes: The final control output is encapsulated into a standard control command and sent to the field controller for execution. Process variable response data and control signal data after control execution are collected to form an execution feedback dataset and stored in historical operating condition data. When the preset model update trigger conditions are met, the newly added execution feedback data is used to perform incremental fine-tuning and updating of the large model.

[0013] A robust adaptive control system for industrial control systems, comprising: The standardization module is used to perform outlier detection and time-series standardization on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. The identification module is used to detect operating condition disturbances based on a standardized time-series operating condition dataset, classify and identify the type of disturbance when an operating condition disturbance is determined, and extract multi-dimensional features of the disturbance to generate a disturbance feature vector. The inference module is used to combine historical operating data, disturbance classification results and disturbance feature vectors into a prediction input sequence, input it into a pre-trained large model to perform disturbance trend inference, and output disturbance prediction results including disturbance amplitude range, change direction, expected duration and prediction confidence. The matching module is used to extract the current system operating state features, construct a retrieval feature vector by combining the disturbance prediction results, and call the preset control strategy library to perform similarity matching on the retrieval feature vector. If the matching result meets the preset threshold, the control strategy to be verified is obtained; if the matching result does not meet the preset threshold, the large model is called to generate customized control strategy parameters by combining the system dynamic model and constraints, which are used as the control strategy to be verified. The calculation module is used to perform multi-condition closed-loop simulation tests on the control strategy to be verified, calculate the comprehensive performance score, and determine the executable control strategy after the performance constraints are met. The synthesis module is used to calculate the disturbance feedforward compensation amount based on the disturbance prediction results, dynamically adjust the weight coefficients of the feedforward and feedback control according to the prediction confidence, and synthesize the weighted control amount with the feedback control amount corresponding to the executable control strategy to obtain the final control output amount.

[0014] A robust adaptive control device for industrial control systems, comprising: Memory, used to store computer programs; A processor is used to implement the robust adaptive control method of the industrial control system described above when executing the computer program.

[0015] A computer-readable storage medium storing a computer program, which, when executed by a processor, is used to implement the robust adaptive control method of the above-described industrial control system.

[0016] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a robust adaptive control method for industrial control systems. First, real-time and historical operating condition data are standardized and disturbance detection and classification are performed. Disturbance features are extracted, and a pre-trained large model is used to predict disturbance trends. Then, customized control parameters are generated through control strategy library matching or large model generation. After verification through multi-condition simulation, the final control output is synthesized by combining disturbance feedforward compensation and dynamic weights. This method achieves proactive prediction of unknown disturbances through the temporal reasoning capability of the large model, replacing traditional manual rule design. Feature matching and dynamic model generation reduce the controller's conservatism and improve its adaptability to strongly coupled nonlinear systems. Adaptive adjustment of feedforward and feedback weights enhances the system's robustness to uncertain disturbances. This method significantly improves control performance and adaptability in complex industrial scenarios, achieving proactive disturbance perception, autonomous control strategy generation, and dynamic optimization, overcoming the core shortcomings of existing technologies such as reliance on manual intervention, strong conservatism, and lack of predictive capability. Attached Figure Description

[0017] Figure 1 This is an overall architecture diagram of the robust adaptive control system provided in an embodiment of the present invention; Figure 2This is a flowchart of disturbance prediction and strategy generation provided in an embodiment of the present invention; Figure 3 A flowchart of the two-stage robustness verification and control execution closed loop provided in this embodiment of the invention; Figure 4 A flowchart of a robust adaptive control method for an industrial control system provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of a robust adaptive control system for an industrial control system provided in an embodiment of the present invention. Detailed Implementation

[0018] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0019] As described in the background section, traditional control methods struggle to handle complex nonlinear systems, lack the ability to proactively predict disturbances, and rely on manual design for adjusting control strategies.

[0020] To address the aforementioned issues, this embodiment provides a robust adaptive control method for industrial control systems. The method aims to: achieve proactive prediction and feedforward compensation of disturbance trends through in-depth analysis of historical operating data using a large model, shifting the control paradigm from passive feedback to active feedforward; utilize a control strategy knowledge base and the large model for intelligent matching or dynamic generation of customized control parameters to achieve integrated adaptive control encompassing "perception-reasoning-decision-making"; add a two-stage verification process of stability checking and performance evaluation before strategy execution to eliminate the "illusion" risk of the large model and ensure safety; reduce controller conservatism and improve response speed and accuracy through auxiliary parameter optimization using the large model; integrate feedforward, feedback, and predictive control modes for coordinated operation, leveraging their respective advantages; simultaneously construct a hierarchical computing architecture to ensure real-time performance and achieve continuous self-optimization throughout the entire lifecycle through effect feedback; ultimately achieving the goals of reducing manual intervention, improving intelligence, and increasing operational efficiency, meeting the high-performance, highly adaptable, and stable control requirements of complex industrial scenarios.

[0021] For example, such as Figure 4 As shown, this embodiment provides a robust adaptive control method for an industrial control system, including: Outlier detection and time-series standardization are performed on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. Based on a standardized time-series operating condition dataset, operating condition disturbance detection is performed. When an operating condition disturbance is determined to exist, the disturbance type is classified and identified. At the same time, multi-dimensional disturbance features are extracted to generate a disturbance feature vector. Historical operating data, disturbance classification results, and disturbance feature vectors are combined into a prediction input sequence, which is then input into a pre-trained large model to infer disturbance trends. The output includes disturbance amplitude range, direction of change, expected duration, and prediction confidence. Extract the current system operating state features, construct the retrieval feature vector by combining the disturbance prediction results, call the preset control strategy library to perform similarity matching on the retrieval feature vector, and obtain the control strategy to be verified if the matching result meets the preset threshold; if the matching result does not meet the preset threshold, combine the system dynamic model and constraints to call the large model to generate customized control strategy parameters, which are used as the control strategy to be verified. The control strategy to be verified is subjected to multi-condition closed-loop simulation test, the comprehensive performance score is calculated, and the strategy is determined to be an executable control strategy after the performance constraints are met. The disturbance feedforward compensation is calculated based on the disturbance prediction results. The weight coefficients of the feedforward and feedback control are dynamically adjusted according to the prediction confidence. The feedforward and feedback control are then weighted and synthesized with the feedback control quantity corresponding to the executable control strategy to obtain the final control output quantity.

[0022] The robust adaptive control method for industrial control systems provided in this embodiment will be further explained below with reference to the accompanying drawings: For example, this embodiment provides a robust adaptive control method for an industrial control system. The overall technical approach is as follows: First, a real-time operating condition data acquisition module acquires sensor measurements and equipment operating status data with a sampling period of no more than 1 second, ensuring data integrity of no less than 95%. Then, a disturbance analysis module classifies the data, distinguishing between measurable and unmeasurable disturbances. Measurable disturbances are directly acquired by sensors, while unmeasurable disturbances enter the large-scale model prediction process. Next, the large-scale model disturbance prediction module uses 1800 sampling points from the past 30 minutes as input, and outputs a quantitative prediction and confidence level of the disturbance change within the next 15 minutes under a temperature parameter of 0.3. When the confidence level is below 70%, a conservative control mode is triggered. Subsequently… The control strategy matching module retrieves the optimal matching strategy from a strategy library containing at least 15 strategies based on the current system state and disturbance prediction results. When the highest matching degree is lower than 0.75, a large model generative strategy suggestion is activated, and the temperature parameter is adjusted to 0.5 with a parameter tuning range. Subsequently, the two-stage robustness verification module performs stability verification (phase margin ≥ 30°, gain margin ≥ 6dB, Lyapunov function derivative < 0) and performance evaluation (overshoot ≤ 15%, settling time ≤ 90 seconds, steady-state error ≤ ±1%, comprehensive performance score ≤ 1.0) on the candidate strategies. Finally, the verified control strategy is sent to the controller for execution, and the execution effect data is fed back to the data acquisition stage to form a closed-loop optimization. The specific steps are as follows: Step S1: Real-time operating condition data acquisition: Specific operations: Process data is acquired in real time from the fieldbus, sensor network, and equipment status interface of the industrial control system. The acquired data includes: (a) Sensor measurements: key process variables such as temperature, pressure, flow rate, liquid level, and valve opening; (b) Equipment operating status: pump start / stop status, valve on / off status, and start / stop signals; (c) Control signals: setpoint (SP), process variable (PV), and operation output (MV); (d) Historical operating condition records: retrieving the operating condition data sequence from the past 30 minutes from the historical database. Outlier detection is performed simultaneously with data acquisition, using the 3σ criterion to identify and mark abrupt data points.

[0023] Key parameters: Sampling period T_s≤1 second; Data channels cover all key process variables (temperature, pressure, flow rate, liquid level, etc., ≥4 variables); Data delay Δt≤3 seconds; Outlier detection adopts the 3σ criterion; Data storage format is standardized time-series data (timestamp + variable name + value).

[0024] Judgment criteria: Data integrity ≥ 95% (percentage of valid data points per period); outlier detection single point marking does not exceed 5% of the total data points; data timestamps are continuous without breaks (maximum allowable interruption ≤ 2 sampling periods); outliers do not exceed 2 consecutive periods.

[0025] Step S2: Disturbance type classification and measurable disturbance acquisition: Specific operation: The current disturbance is classified based on an expert rule base and a machine learning classifier (random forest or support vector machine). First, the deviation of the process variable from the set value is detected. When the deviation exceeds ±5% of the set value, it is determined that a disturbance exists. Then, the classifier distinguishes between measurable disturbances (such as load changes and changes in raw material composition, which can be directly measured by dedicated sensors) and unmeasurable disturbances (such as catalyst activity decline, heat exchanger fouling, and unknown external interference). At the same time, the duration and rate of change of the disturbance are statistically analyzed to generate a disturbance feature vector (containing four dimensions: disturbance amplitude, direction, duration, and rate of change).

[0026] Key parameters: Measurable disturbance detection threshold ±5% of set value change; machine learning classifier accuracy ≥90%; disturbance duration statistical window 60 seconds; disturbance feature vector dimension 4; classifier output confidence threshold ≥85%.

[0027] Judgment conditions: Measurable disturbances can be directly acquired by the sensor and enter the feedforward channel in step S3; the proportion of unmeasurable disturbances does not exceed 30% of the total number of disturbances; the confidence level of the disturbance type determination result is ≥85%; if the confidence level is <85%, the human-computer interaction confirmation process is triggered (step S11 human-computer interaction confirmation).

[0028] Step S3: Large Model Perturbation Trend Prediction: Specific operation: Combine historical operating condition data sequences (time series of 1800 sampling points over the past 30 minutes) and disturbance classification results (output of step S2) to construct a Prompt, which is then input into the large model. Based on semantic understanding of the industrial process mechanism and pattern recognition of historical data, the large model outputs a quantitative prediction of disturbance changes within the next 15 minutes, including: disturbance amplitude range (minimum and maximum values), direction of change (increasing, decreasing, or fluctuating), expected duration (minutes), and probability distribution (high, medium, or low confidence levels). If the prediction confidence level is <70%, the system automatically switches to conservative control mode (only feedback control is enabled, without applying feedforward compensation).

[0029] Key parameters: Prediction time window T_pred = 15 minutes; Input sequence length L = 1800 sampling points (30 minutes × 60 samples / minute); Prediction step size Δt_pred = 1 minute; Large model temperature parameter temp = 0.3 (retrieval mode); Maximum number of output tokens 512; top_p = 0.9.

[0030] Judgment criteria: The prediction result must include the amplitude range [lower limit, upper limit] and probability distribution; when the confidence level is ≥70%, the prediction result is adopted and proceeds to step S4; when the confidence level is <70%, the conservative control mode is triggered (step S8 feedforward compensation is skipped, and only feedback control is performed); the response time of the large model is ≤3 seconds.

[0031] Step S4: Control policy library matching: Specific steps: Based on the current system state feature vector (current values ​​and trends of key process variables extracted from step S1 over the past 10 minutes) and the perturbation prediction results from step S3, construct a retrieval query vector. Call the control strategy library retrieval algorithm to calculate the cosine similarity between the query vector and the feature vector of each strategy in the library, and return the Top-3 strategies and their matching scores in descending order of similarity. Simultaneously calculate the strategy recommendation confidence (based on a weighted average of matching score and prediction confidence). If the highest matching score SIM ≥ 0.75 and the recommendation confidence ≥ 65%, select the strategy with the highest matching score and proceed to step S6; if SIM < 0.75 or confidence < 65%, trigger step S5, large-model generative strategy suggestion.

[0032] Key parameters: The control strategy inventory contains at least 15 strategies (≥5 for normal operating conditions, ≥5 for disturbance operating conditions, and ≥5 for abnormal operating conditions); the matching similarity threshold SIM ≥ 0.75; the strategy recommendation confidence threshold ≥ 65%; the number of Top-3 returns = 3; the retrieval algorithm uses cosine similarity calculation.

[0033] Judgment criteria: The strategy library retrieves the top-3 strategies with the highest matching degree and their scores; if the highest matching degree SIM≥0.75 and the confidence level≥65%, select the top-1 strategy and proceed to step S6; if any condition is not met, trigger step S5 generative suggestion.

[0034] Step S5: Generative Strategy Recommendation for Large Models (Conditional Triggering): Specific operation: When the matching degree (SIM < 0.75) or confidence degree is insufficient (< 65%) in step S4, the large model generative capability is activated. The system dynamic model description (expressed in ODE or PDE equation form), the current disturbance prediction result (output of step S3), and system constraints (control variable upper and lower limits, process variable constraints) are combined to construct a generative prompt, which directly generates customized control law parameters by calling the large model. The generated content includes: PID parameters (proportional coefficient Kp, integral coefficient Ki, derivative coefficient Kd) or MPC weight matrix (control weight R, terminal weight P), along with the parameter tuning range [min, max] and suggested effective time. After generation, additional stability verification is required in step S6.

[0035] Key parameters: The dimension of the generated strategy matches the dimension of the controlled system (e.g., a 2×2 weight matrix is ​​generated for a dual-input dual-output system); the generation temperature temp=0.5 (higher than step S3 to increase diversity); the generation suggestion includes the parameter tuning range [min, max]; the maximum number of generated parameters is 20; the response time of large models is ≤5 seconds.

[0036] Judgment conditions: The policy dimension of the generative suggestion is consistent with the controller interface (dimension verification passed); the policy parameters are within the physically achievable range (upper and lower limit verification passed); the triggering condition is that the matching degree SIM < 0.75 or the confidence degree < 65% in step S4; the generative suggestion must undergo the additional stability verification in step S6.

[0037] Step S6: Robust stability check: Specific operations: Use a simplified reduced-order linear model (order ≤ 3) or a piecewise linear model to perform closed-loop stability verification on the matching strategy selected in step S4 or the strategy generated in step S5. The verification method adopts the Lyapunov direct method and / or frequency domain method: (a) Lyapunov direct method: Construct a Lyapunov candidate function V(x) = x T Px (where P is a symmetric positive definite matrix, obtained by solving linear matrix inequalities), verify that the closed-loop system state trajectory makes the derivative of V(x) dV / dt < 0; (b) Frequency domain method: plot the open-loop Nyquist curve or Bode plot, and calculate the phase margin PM and gain margin GM. The stability judgment must simultaneously satisfy: dV / dt < 0 (Lyapunov method) or PM ≥ 30° and GM ≥ 6dB (frequency domain method).

[0038] Key parameters: Simplified model order ≤ 3; Lyapunov candidate function is quadratic V(x) = x T Px; stability margin PM≥30°, GM≥6dB; linear matrix inequality (LMI) solver tolerance 1e-6; upper limit of verification computation time ≤5 seconds; the dimension of the candidate positive definite matrix P is consistent with the state dimension of the simplified model.

[0039] Judgment conditions: If the stability verification is passed (Lyapunov function derivative dV / dt<0 and or phase margin PM≥30°, gain margin GM≥6dB), step S7 can be entered; if the verification fails, the current strategy is rejected and an alarm signal is output (sent to step S11), and the system automatically returns to step S4 to rematch (maximum of 2 retries, if more than that, the effective strategy of the previous cycle is maintained).

[0040] Step S7: Control performance evaluation: Specific operations: In a closed-loop simulation environment (based on a high-precision process simulation model, the object model can be a nonlinear differential-algebraic equations model), Monte Carlo simulation evaluation is performed on the candidate control strategy that passes the stability verification in step S6. The simulation scenarios include four types: rated operating condition, step disturbance condition (load ±20% variation), ramp disturbance condition (gradual change in raw material quality), and noise disturbance condition (measurement noise standard deviation 5%). Key performance indicators (KPIs) are calculated for each scenario: overshoot σ (the percentage difference between the maximum step response value and the steady-state value divided by the steady-state value), settling time t_s (the shortest time required for the response to enter and remain within the ±2% error band), steady-state error e_ss (the deviation of the process variable from the setpoint at steady state), and control energy consumption. Take a weighted average of the KPIs from multiple scenarios (with equal weight for each scenario), and then compare it with the performance threshold.

[0041] Key parameters: Simulation duration T_sim = 3 times the settling time (maximum 300 seconds); Simulation step size 0.1 seconds; Performance index thresholds: σ≤15%, t_s≤90 seconds, e_ss≤±1%; Weighting coefficient overshoot: settling time: steady-state error = 0.4:0.3:0.3; Control energy consumption benchmark is the consumption value of the current operating strategy; Number of scenarios = 4 (rated / step / ramp / noise).

[0042] Judgment conditions: The overall performance score W = 0.4·σ_norm + 0.3·t_s_norm + 0.3·e_ss_norm ≤ 1.0 (normalized value, normalized according to the threshold); the energy consumption is controlled to not exceed 120% of the baseline strategy; if the above conditions are met, the strategy passes the evaluation and proceeds to step S8; if any condition is not met, return to step S4 for rematching (maximum of 2 retries).

[0043] Step S8: Feedforward-feedback synthesis and compensation calculation: Specific operation: Based on the disturbance prediction result in step S3, calculate the feedforward compensation amount and synthesize it with the feedback control amount. Based on the transfer function model G_d(s) from the disturbance channel to the controlled output, the feedforward compensation amount u_ff(t) = G_d(s)·d(t), where d(t) is the predicted disturbance amount. Simultaneously, dynamically adjust the feedforward-feedback weight allocation coefficient according to the prediction confidence and disturbance amplitude: when the confidence level ≥ 70%, the feedforward weight is 0.3 to 0.5 (adjusted according to the disturbance amplitude; the larger the disturbance amplitude, the higher the feedforward weight); when the confidence level < 70% (conservative mode), the feedforward weight is 0. The final control amount u(t) = (1-λ_ff)·u_fb(t) + λ_ff·u_ff(t), where λ_ff is the feedforward weight (0≤λ_ff≤0.5), and u_fb(t) is the feedback control amount.

[0044] Key parameters: Feedforward compensation calculation is based on the transfer function G_d(s); the feedforward weight λ_ff ranges from 0 to 0.5 (0.3≤λ_ff≤0.5 when confidence level ≥70%, λ_ff=0 when confidence level <70%); the upper and lower limits of the compensation amount are ±20% of the controller output range; the synthesis formula is u(t)=(1-λ_ff)·u_fb+λ_ff·u_ff.

[0045] Judgment conditions: The feedforward compensation amount does not exceed ±20% of the controller output range; the synthesized control amount does not exceed the upper and lower limits of the actuator; when the confidence level is <70%, λ_ff is forced to 0 (no feedforward is applied); the compensation amount calculation time is ≤0.5 seconds.

[0046] Step S9: Control strategy issuance and execution: Specific operation: The final control strategy parameters (PID parameters or MPC weight matrix and feedforward-feedback synthesis coefficients) synthesized in steps S7 and S8 are encapsulated into a control command frame and sent to the controller for execution via an industrial communication protocol (such as Modbus, OPCUA, or PROFINET). The control command issuance cycle T_ctrl is synchronized with the sampling cycle T_s (both ≤ 1 second), and the execution delay tolerance is ≤ 2 sampling cycles (i.e., the delay from strategy determination to execution effectiveness is ≤ 2 seconds). Simultaneously, the actual system response data (actual values ​​of process variables and actual values ​​of control signals) are recorded. The execution result data is sent to step S10 for effect feedback and simultaneously sent to step S1 for data acquisition to form a closed loop.

[0047] Key parameters: Control command issuance cycle T_ctrl=T_s≤1 second; Execution delay tolerance≤2 sampling cycles; Communication protocol supports OPCUA / ModbusTCP / PROFINET; Command frame includes strategy type (PID or MPC), parameter array, and timestamp; Feedback data recording cycle is synchronized with sampling.

[0048] Judgment conditions: Control command is successfully sent to the controller (acknowledgment signal returned); execution delay does not exceed 2 sampling cycles; no error code is returned by the controller during execution; safety protection is triggered in case of abnormality (switch to manual mode or backup controller).

[0049] Step S10: Feedback on execution results and model update: Specific operation: The execution effect data (actual overshoot, actual settling time, actual steady-state error, and measured control energy consumption) recorded in step S9 are compared with the predicted values ​​or thresholds to evaluate the actual effectiveness of the strategy. Simultaneously, the execution effect data is stored in the historical database (the historical database called in step S1) for subsequent periodic retraining or online fine-tuning of the large model perturbation prediction model. The model update trigger condition is: periodic triggering (interval ≥ 7 days) or a sudden drop in performance score (overall performance score W decreases by ≥ 20% relative to the previous period). The model update adopts an incremental learning method, using newly accumulated feedback data to fine-tune the parameters using a small number of epochs based on the original model parameters.

[0050] Key parameters: Model update trigger condition 1: periodic (interval ≥ 7 days); trigger condition 2: performance score drops sharply by ≥ 20%; incremental learning fine-tunes the number of epochs by 3 to 5; historical data retention period ≥ 180 days; feedback data is used to update the parameters of the three major model perturbation prediction models in the module.

[0051] Judgment criteria: The actual overshoot does not exceed 130% of the predicted value; the steady-state error is within ±1.5%; the performance score recovers to W≤1.0 after the model update; the model version number is automatically incremented and recorded after the update.

[0052] As another preferred embodiment, the method further includes: Step S11: Human-computer interaction confirmation and alarm handling: Specific operation: When the system encounters the following situations, an alarm message will be pushed to the operator via the human-machine interface and confirmation will be requested: (a) Disturbance classification confidence level <85% in step S2; (b) Policy matching fails twice consecutively in step S4; (c) Stability verification rejects the policy in step S6; (d) Performance evaluation fails in step S7; (e) The controller returns an exception code during the execution of step S9. Alarms are classified according to severity: Red alarm (requires immediate handling, such as stability verification rejection), Yellow alarm (requires prompt handling, such as policy matching failing twice), Blue alarm (information prompt, such as model update completed). Operators can confirm acceptance of the system's recommended policy through the interface or manually select other policies. Manual intervention commands have higher priority than automatic system decisions.

[0053] Key parameters: Alarm classification thresholds: Red (stability rejection or execution anomaly), Yellow (policy matching failure 2 times or performance degradation ≥20%), Blue (information prompt); Alarm response time limit: Red ≤1 minute, Yellow ≤10 minutes; Manual confirmation timeout defaults to accepting the system recommendation (timeout 10 minutes); Manual intervention has higher priority than automatic decision-making.

[0054] Judgment criteria: Red alarm confirmation rate 100% (must be handled before continuing); Yellow alarm confirmation rate ≥90%; After manual intervention, the system executes according to the strategy selected by the operator; Alarm records are completely saved to the system log.

[0055] For example, in order to implement the robust adaptive control method for industrial control systems provided in the above embodiments, a hardware structure for a robust adaptive control system for industrial control systems is also provided, including a data acquisition layer, an analysis and inference layer, a strategy generation layer, a verification and decision layer, and an execution and control layer, such as... Figure 1As shown, the data acquisition layer consists of a Real-Time Operating Data Acquisition Module (RTDAM), which provides data input to the Large Model Disturbance Trend Prediction Module (LDPM) and Disturbance Type Classification Module (DCM) of the analysis and inference layer. The analysis and inference layer outputs disturbance prediction results to the strategy generation layer, which includes a Control Strategy Library Management Module (CSLMM) and a Large Model Generative Strategy Recommendation Module (LGSAM), determining whether to enable retrieval or generation mode based on the matching degree. The verification decision layer includes a Robust Stability Verification Module (RSVM) and a Control Performance Evaluation Module (CPAM), forming a two-stage verification guarantee. The execution control layer includes a Control Command Issuance and Execution Module (CCDEM) and a Feedforward Compensation Calculation Module (FCCM), which simultaneously receives disturbance prediction results from the analysis and inference layer. The Human-Machine Interaction and Alarm Management Module (HMIAMM) interacts with all modules, and the Historical Database and Model Update Module (HDMUM) provides historical data support and continuous model optimization capabilities for the system. The specific components are as follows: Module 1: Real-time Operating Condition Data Acquisition Module (RTDAM): Name: Real-Time Data Acquisition Module (RTDAM).

[0056] Function: Collect process variable data in real time from various sensors and equipment status interfaces of industrial control systems at a fixed sampling period, including key process variables such as temperature, pressure, flow rate, and liquid level, as well as equipment operating status data such as pumps, valves, and start-up / shutdown status, and complete data cleaning, outlier detection, and timestamp alignment.

[0057] Input: Industrial control system fieldbus or sensor signals (4-20mA analog or digital switch signals).

[0058] Output: A standardized time-series data stream containing an array of process variables, a device state vector, and a timestamp sequence.

[0059] Interface relationships: It provides data input to the disturbance analysis module; it interacts with the historical database module to obtain historical operating condition records; and it sets the sampling period and control command issuance period synchronously.

[0060] Module 2: Disturbance Type Classification Module (DCM): Name: Disturbance Classification Module (DCM).

[0061] Function: Based on an expert rule base and machine learning classifier, the current disturbance is classified into measurable and unmeasurable disturbances. The disturbance type label and its confidence level are output, and the duration and trend of the disturbance are statistically analyzed.

[0062] Input: Real-time operating data (from module 1), disturbance detection threshold parameters (±5% change of set value).

[0063] Output: Perturbation type classification results (measurable or unmeasurable), confidence score (≥85%), and perturbation statistical feature vector.

[0064] Interface relationship: Receives the output of module 1 and provides classification results input for module 3 and the large model perturbation prediction module.

[0065] like Figure 2 As shown, real-time operating data, after being classified into measurable and unmeasurable disturbances by the disturbance type classification module, is combined with the historical data sequence of the past 30 minutes and input into the large model. The large model outputs the disturbance prediction result and confidence level. When the confidence level is ≥70%, the policy library matching process is initiated, calculating the cosine similarity between the current state and each policy in the policy library. If the highest matching degree SIM ≥0.75 and the confidence level ≥65%, the optimal policy is selected; if SIM <0.75 or the confidence level <65%, the large model's generative policy suggestion is triggered, generating customized control parameters based on the system dynamic model. All candidate policies must undergo two-stage verification: the first stage is Lyapunov stability verification or frequency domain margin verification, and the second stage is four-scenario simulation performance evaluation. After all stages pass, the policy is combined with the feedforward compensation amount and then executed.

[0066] Module 3: Large Model Perturbation Trend Prediction Module (LDPM): Name: Large Model-based Disturbance Prediction Module (LDPM).

[0067] Function: The large model takes historical operating condition data series (1800 sampling points in the past 30 minutes) and disturbance classification results as context input. The large model outputs a quantitative prediction of disturbance changes in the next 15 minutes, including the disturbance amplitude range, direction of change, duration and probability distribution, as well as the prediction confidence level. The prediction results are adopted when the confidence level is ≥70%, and the conservative control mode is triggered when the confidence level is <70%.

[0068] Input: Historical operating condition sequence (30 minutes × 60 samples / minute = 1800 sampling points), disturbance classification results (from module 2), large model temperature parameters (temp = 0.3), prediction time window (T_pred = 15 minutes).

[0069] Output: Perturbation prediction result structure (amplitude range, direction, duration, probability distribution), prediction confidence score.

[0070] Interface relationship: Receives the outputs of module 1 and module 2, and provides prediction input for the control strategy library matching module of module 4; the prediction results are simultaneously sent to the feedforward compensation unit of module 8.

[0071] Module 4: Control Policy Library Management Module (CSLMM): Name: Control Strategy Library Management Module (CSLMM).

[0072] Function: Manage and maintain a multi-condition control strategy knowledge base, which stores no less than 15 sets of standard control strategies, covering three types of conditions: normal conditions (≥5 sets), disturbance conditions (≥5 sets), and abnormal conditions (≥5 sets). It supports cosine similarity matching search based on system status and disturbance prediction results, and returns the top-3 matching strategies and their matching scores.

[0073] Input: Current system state feature vector, perturbation prediction result (from module 3), matching similarity threshold (SIM≥0.75), recommendation confidence threshold (≥65%).

[0074] Output: A list of Top-3 recommendation strategies, their matching scores, and recommendation confidence.

[0075] Interface relationship: Receives the output of module 3; triggers generative suggestions from module 5 when the highest matching degree is <0.75; the output is sent to the stability verification module of module 6.

[0076] Module 5: Large Model Generative Policy Proposal Module (LGSAM): Name: Large Model Generative Strategy Advisory Module (LGSAM).

[0077] Function: When the matching degree of module 4 is lower than the 0.75 threshold, the large model generative capability is enabled. Based on the system dynamic model description (ordinary differential equation or partial differential equation form) and the current disturbance prediction results, customized control law parameters (such as PID parameter proportional-integral-derivative coefficients, predictive controller weight matrix, gain scheduling parameters) are directly generated, along with the parameter tuning range and effective time.

[0078] Input: System dynamic model description (ODE or PDE equation), disturbance prediction results (from module 3), generated temperature parameter (temp=0.5), and strategy dimension information.

[0079] Output: Generative policy parameter structure (including specific parameter values, tuning range [min, max], and effective time), policy dimension verification results.

[0080] Interface relationship: The call is triggered by module 4 when the matching degree is insufficient; the output is sent to module 6 for additional verification.

[0081] Module 6: Robust Stability Verification Module (RSVM): Name: Robust Stability Verification Module (RSVM).

[0082] Function: Perform closed-loop stability verification of candidate control strategies using a reduced-order linear system model (order ≤ 3) or a piecewise linear model. Employ either the Lyapunov direct method (verifying that the derivative of the Lyapunov candidate function V(x) = x^TPx is less than zero) or a frequency domain method (Nyquist criterion), outputting the stability determination result, phase margin (PM), and gain margin (GM). Requirement: phase margin ≥ 30°, gain margin ≥ 6dB, verification calculation time ≤ 5 seconds.

[0083] Input: Candidate control policy parameters (from module 4 or module 5), simplified system model (≤3rd order reduced model), margin threshold (PM≥30°, GM≥6dB).

[0084] Output: Stability determination result (pass or reject), phase margin value, gain margin value, and verification calculation time.

[0085] Interface relationship: Receives output from module 4 or module 5; outputs to module 7 performance evaluation module when successful; outputs an alarm signal and triggers module 4 to re-match when rejected.

[0086] Module 7: Control Performance Evaluation Module (CPAM): Name: Control Performance Assessment Module (CPAM).

[0087] Function: To perform performance simulation evaluation on candidate control strategies that have passed stability verification in a closed-loop simulation environment, and calculate key performance indicators (KPIs): overshoot (σ), settling time (t_s, ±2% error band), steady-state error (e_ss), and control energy consumption (∫u²dt). The comprehensive performance score is W=0.4·σ_norm+0.3·t_s_norm+0.3·e_ss_norm, with requirements that W≤1.0 and control energy consumption not exceed 120% of the baseline strategy.

[0088] Input: Candidate control strategy parameters (from module 6), simulation model, weighting coefficients (overshoot: settling time: steady-state error = 0.4:0.3:0.3), performance thresholds (σ≤15%, t_s≤90 seconds, e_ss≤±1%).

[0089] Output: KPI calculation result structure (σ, t_s, e_ss, control energy), comprehensive performance score W, evaluation judgment result.

[0090] Interface relationship: Receive module 6 stability pass signal; if pass, output to module 9 strategy execution module; if fail, return to module 4 for rematch.

[0091] Module 8: Feedforward Compensation Calculation Module (FCCM) Name: Feedforward Compensation Calculation Module (FCCM).

[0092] Function: Based on the disturbance prediction results output by module 3 and the system dynamic model, calculate the feedforward compensation amount and superimpose it onto the feedback control output. The compensation amount calculation is based on the transfer function model from the disturbance to the controlled output, and the compensation strategy adaptively adjusts the compensation intensity coefficient according to the disturbance type and amplitude.

[0093] Inputs: Disturbance prediction results (amplitude range, direction, duration, from module 3), system transfer function model, compensation intensity coefficient (range from 0 to 1), and feedforward-feedback weight allocation coefficients.

[0094] Outputs: feedforward compensation amount u_ff, control quantity correction value, feedforward-feedback combined proportional coefficient.

[0095] Interface relationship: Receives the output of module 3; the output is superimposed with the feedback control quantity of module 9 and then sent to the actuator.

[0096] Module 9: Control Command Issuance and Execution Module (CCDEM): Name: Control Command Dispatch and Execution Module (CCDEM).

[0097] Function: Sends the final control strategy parameters and control commands to the industrial controller (such as DCS, PLC or dedicated controller) for execution. The command sending cycle is synchronized with the sampling cycle (T_ctrl=T_s≤1 second), and the execution delay tolerance is ≤2 sampling cycles. At the same time, it monitors the actuator feedback signal and triggers the safety protection mechanism (switching to manual mode or backup controller) in case of abnormality.

[0098] Input: final control strategy parameters (from the combined results of modules 7 and 8), control command issuance cycle (T_ctrl≤1 second), and safety protection threshold.

[0099] Outputs: Control command frame (sent to the controller), execution confirmation signal, and abnormal alarm signal.

[0100] Interface relationship: Receives outputs from module 7 and module 8; outputs are sent to the field controller and actuator; and the execution result data is fed back to the data acquisition module 1.

[0101] Module 10: Historical Database and Model Update Module (HDMUM): Name: Historical Database and Model Update Module (HDMUM).

[0102] Functions: Stores and manages historical operating condition data, control strategy execution records, performance evaluation results, etc.; Based on execution feedback data, it periodically retrains or fine-tunes the large-scale disturbance prediction model online to achieve continuous self-optimization of the control system; The database adopts time-series database storage with a retention period of no less than 180 days.

[0103] Input: Execution effect feedback data (actual values ​​of process variables, actual values ​​of control signals, and actual values ​​of performance indicators, from module 9), and model update trigger conditions (cycle ≥ 7 days or performance score W drops by ≥ 20%).

[0104] Output: Updated large model prediction model, historical working condition query results, model version number.

[0105] Interface relationship: Receives feedback data from module 9; outputs model parameters to update module 3; provides historical data support for module 3 through the historical query service.

[0106] Module 11: Human-Computer Interaction and Alarm Management Module (HMIAMM): Name: Human-Machine Interface and Alarm Management Module (HMIAMM).

[0107] Functions: Provides a visual human-computer interaction interface to display system operating status, disturbance prediction results, policy matching status, performance indicator trends, etc.; centrally manages all alarm information, including policy rejection alarms, data anomaly alarms, performance degradation alarms, and equipment failure alarms, and displays and records them according to severity level (red / yellow / blue); supports manual intervention by operators and policy confirmation operations.

[0108] Inputs: System operating status data (from various modules), alarm trigger signals (from module 6, etc.), and operator instructions.

[0109] Output: Visualized interface data stream, alarm logs and confirmation status, and manual intervention commands.

[0110] Interface Relationship: Interacts with all modules, receives status and alarm data from each module; can send manual intervention commands to modules 4 and 9.

[0111] Module 12: System Configuration and Management Module (SCAMM): Name: System Configuration and Management Module (SCAMM) Functions: Manage system configuration parameters, including working parameters of each module, large model call parameters (temperature, maximum number of tokens, etc.), control threshold parameters, performance benchmark parameters, etc.; support adding, deleting, modifying and querying configuration information; manage the operation status monitoring and fault diagnosis of each module; record system operation logs.

[0112] Input: Configuration parameters and system operation status monitoring data.

[0113] Outputs: Configuration parameter distribution, system health status report, and operation log recording.

[0114] Interface relationships: Connects with all modules, providing a unified configuration management service; configuration files can be imported and exported.

[0115] In summary, this embodiment provides a robust adaptive control method for industrial control systems, with the following specific technical advantages: First, active perturbation prediction and feedforward compensation techniques based on large models: This method is the first to apply large model technology to disturbance prediction in industrial control systems, constructing an active control framework of "historical operating conditions → large model semantic understanding → disturbance trend → feedforward compensation". By using time-series data from 1800 sampling points over the past 30 minutes as context input, the large model can identify complex spatiotemporal patterns in industrial processes and predict the direction, amplitude range, and duration of disturbances within the next 15 minutes, achieving a fundamental shift in control paradigm from "passive feedback" to "active feedforward". The feedforward compensation is calculated in real time based on the disturbance channel transfer function and is activated when the prediction confidence level is ≥70%. When the confidence level is insufficient, it automatically degrades to pure feedback control to ensure safety.

[0116] This method produces the following technical effects: main steam temperature fluctuations are reduced by 60% (temperature standard deviation reduced from ±2.5℃ to ±1.0℃); overshoot is reduced by more than 45% (from 28% to below 15%); feedforward compensation takes effect approximately 30 seconds earlier (relative to the actual time of disturbance occurrence); and disturbance prediction accuracy reaches over 85% in samples with confidence levels ≥70%.

[0117] It is evident that this method overcomes the fundamental shortcomings of traditional robust control and adaptive control, which can only passively respond to disturbances that have already occurred. It provides a brand-new active anti-interference technology path for industrial control systems, and is particularly suitable for continuous production processes such as thermal power and chemical industries where load changes are predictable.

[0118] Second, the intelligent matching and generation mechanism of adaptive control strategies driven by large models: This method establishes a multi-condition control strategy knowledge base (containing no fewer than 15 strategies across three categories: normal operating conditions, disturbance operating conditions, and abnormal operating conditions), and innovatively uses a large model as the intelligent hub for strategy retrieval and generation. Based on real-time system status and disturbance prediction results, the large model intelligently matches the optimal strategy from the strategy base (matching degree threshold 0.75) through cosine similarity calculation. When there is no matching strategy in the base (matching degree < 0.75), it automatically switches to generative capabilities, directly generating customized PID parameters or MPC weight matrices based on the system's ODE or PDE dynamic model description, thus achieving integrated "perception-reasoning-decision" control.

[0119] This method produces the following technical effects: the policy matching success rate reaches over 85%; the adaptive adjustment time of large model generative policies is shortened to less than 10 seconds (compared to the convergence time of 3 to 5 minutes for traditional MRAC); the control success rate of nonlinear time-varying systems is improved by 50% (from 60% to over 90%); and the policy recommendation confidence reaches over 65%.

[0120] As can be seen, this method breaks through the bottleneck of slow convergence of traditional adaptive control relying on online parameter estimation. Large models can be directly inferred to generate control strategies, which greatly shortens the adaptive adjustment time and reduces the dependence on accurate mathematical models, enabling the system to quickly cope with nonlinear, time-varying complex industrial environments.

[0121] Third, a two-stage robust verification and assurance system: This method employs a two-stage verification mechanism: stability verification and performance evaluation. Stage 1 (Stability Verification): Using a reduced-order system model (≤3rd order), closed-loop stability is determined using either the Lyapunov direct method (verifying dV / dt < 0) or a frequency domain method (Nyquist criterion, requiring phase margin ≥ 30° and gain margin ≥ 6dB). Unstable strategies are rejected and an alarm is triggered. Stage 2 (Performance Evaluation): Key performance indicators such as overshoot σ, settling time t_s, and steady-state error e_ss are evaluated in a closed-loop simulation environment encompassing four scenarios: rated, step, ramp, and noise. The overall performance score W = 0.4·σ_norm + 0.3·t_s_norm + 0.3·e_ss_norm is calculated. Execution is only possible after both stages are passed; failure in either stage triggers strategy re-matching or an alarm. Figure 3As shown, the candidate strategy first enters Phase 1 (stability verification), where a simplified system model of no more than order 3 is used to determine closed-loop stability. The Lyapunov direct method is used to verify dV / dt < 0, or the frequency domain method requires a phase margin ≥ 30° and a gain margin ≥ 6dB. After passing stability verification, Phase 2 (performance evaluation) is entered. In four simulation scenarios—rated, step disturbance, ramp disturbance, and noise disturbance—overshoot σ, settling time t_s, steady-state error e_ss, and control energy consumption are calculated. The comprehensive performance score W = 0.4·σ_norm + 0.3·t_s_norm + 0.3·e_ss_norm is calculated. After both phases pass, the feedforward compensation (dynamically calculated based on the disturbance transfer function and prediction confidence) and the feedback control are combined according to weights. The final control is then sent to the industrial controller for execution via OPC UA or Modbus TCP protocol. The execution effect data is simultaneously fed back to the historical database and the model update module, forming a complete closed-loop optimization.

[0122] This method produces the following technical effects: two-stage verification reduces the probability of unsafe execution of the control strategy to below 5%; stability verification calculation time is ≤5 seconds, meeting the real-time requirements of industry; performance evaluation improves the overall performance of the control system by about 30%; and convergence to an executable strategy is achieved in no more than 2 rematches after strategy rejection.

[0123] It is evident that this method, while fully leveraging the powerful reasoning capabilities of large models, safeguards the bottom line of safety through a rigorous verification mechanism. It solves the core safety problem that the "illusion" of large models may lead to dangerous control decisions, and provides a feasible safety assurance solution for the application of large models in safety-critical systems, enabling large model-assisted control to move from theoretical possibility to engineering practice.

[0124] Fourth, an integrated intelligent control framework that combines feedforward, feedback, and prediction: This method proposes an integrated control framework that combines feedforward control (active compensation based on disturbance prediction), feedback control (closed-loop correction based on real-time error), and predictive control (multi-step control sequence based on rolling optimization). A large model serves as the core scheduling unit, coordinating the weight switching and collaborative operation of the three control modes based on disturbance prediction confidence and system operating state: when the disturbance is predictable and the confidence level is high (feedforward weight 0.3 to 0.5), feedforward dominates; when the disturbance is unpredictable or the confidence level is low, the system switches to feedback-dominated mode; and MPC rolling optimization is enabled in setpoint tracking scenarios. The framework provides a fusion view of the system state through a unified state estimator, enabling smooth switching and complementary advantages of the three control modes.

[0125] This method produces the following technical effects: the integrated framework improves control performance by more than 40% under complex disturbance conditions compared to a single control mode; the system's overall anti-interference capability (IAE index) is reduced by more than 50%; the feedforward-feedback switching time is ≤0.5 seconds; there is no overshoot or oscillation during control mode switching; and the system can maintain stable operation for more than 90 days without manual intervention under slow-changing uncertainties such as equipment aging and raw material changes.

[0126] As can be seen, this method provides a complete high-performance intelligent control architecture for industrial control systems, organically combining the cognitive intelligence of large models with the stability guarantee of classical control theory, thus opening up a new direction for the integrated development of "AI + control". The versatility of the framework makes it applicable to various continuous industrial processes such as thermal power, chemical industry, and metallurgy, and it has broad industrial application and promotion value.

[0127] Fifth, a continuous self-optimization operation mechanism throughout the entire lifecycle: This method establishes a closed-loop feedback self-optimization mechanism for the control system. By continuously monitoring and analyzing the actual effect data after strategy execution, it automatically determines whether the large-scale perturbation prediction model needs to be updated. Model updates employ incremental learning methods (lightweight fine-tuning based on LoRA or adapters), fine-tuning the original model parameters over 3 to 5 epochs using newly accumulated feedback data, avoiding the computational overhead of full retraining. Simultaneously, the system automatically performs shadow testing (comparing the old and new models in an offline environment) after model updates. Only if performance does not degrade is the new model allowed to go live, ensuring the safety of the update.

[0128] This method produces the following technical effects: the system can achieve autonomous and stable operation for more than 90 days without human intervention (tolerance for slow-changing uncertainty); the model update cycle is ≥7 days, avoiding instability caused by frequent updates; each incremental fine-tuning uses ≥1000 newly collected feedback data samples; the performance score W after the model update is no higher than before the update (performance degradation is not allowed); the model version is traceable and supports rollback to historical stable versions.

[0129] As can be seen, this method overcomes the fundamental drawback of traditional control systems that require manual readjustment after equipment aging or operating condition drift, achieving the maintenance of the control system's adaptive capability throughout its entire lifecycle and significantly reducing maintenance costs. Simultaneously, incremental learning and shadow testing mechanisms ensure that the model update process itself does not introduce new risks.

[0130] The robust adaptive control method for industrial control systems provided in this embodiment will be further implemented below: The specific hardware and software configuration parameters are shown in Tables 1 and 2: Table 1 shows the hardware configuration parameters.

[0131] Table 2 shows the software configuration parameters.

[0132] Applying this control method to the main steam temperature control of thermal power units, under a 300MW rated load, the standard deviation of main steam temperature fluctuation decreased from ±3.2℃ to ±1.0℃ (a reduction of 68.8%), the overshoot decreased from 28% to 8.2% (a reduction of 70.7%), and the settling time was shortened from 180 seconds to 62 seconds (a reduction of 65.6%). In a ±20% load step change scenario, the peak temperature fluctuation decreased from ±12.3℃ to ±5.1℃. Feedforward compensation began adjusting fuel and feedwater flow approximately 30 seconds before the disturbance occurred, effectively suppressing the upward trend of the main steam temperature. Performance comparisons in main steam temperature control of thermal power units, performance comparisons under load change scenarios, and the accuracy evaluation of large-scale model disturbance prediction are shown in Tables 3, 4, and 5.

[0133] Table 3 compares the performance of different control methods in the main steam temperature control of thermal power units.

[0134] Table 4 compares the main steam temperature control performance under load variation scenarios (load ±20% step variation).

[0135] Table 5 shows the test data for advance detection of operating condition switching.

[0136] For example, the control method provided in this embodiment is also applied to another application, namely, robust temperature control in the stripping section of a distillation column. The specific hardware and software configuration parameters are shown in Tables 6 and 7. Table 6 shows the hardware configuration parameters.

[0137] Table 7 shows the software configuration parameters.

[0138] As can be seen, when the control method provided in this embodiment is applied to the temperature control of the stripping section of the distillation column, in the face of combined disturbances such as changes in feed composition and fluctuations in heating steam pressure, the large model predicts the disturbance trend 5 to 10 minutes in advance and applies feedforward compensation, reducing product purity fluctuations from ±1.5% to ±0.4%, and achieving a feed disturbance suppression rate of 88% (a 95.6% improvement compared to the 45% of traditional PID), effectively reducing the output of defective products. The specific performance comparison of the temperature control in the stripping section of the distillation column is shown in Table 8.

[0139] Table 8 compares the temperature control performance of the stripping section of the distillation column.

[0140] For example, the control method provided in this embodiment is further applied to another application, namely, robust control of dissolved oxygen concentration in a wastewater treatment process. The specific hardware and software configuration parameters are shown in Tables 9 and 10: Table 9 shows the hardware configuration parameters.

[0141] Table 10 shows the software configuration parameters.

[0142] As can be seen, applying the control method provided in this embodiment to the dissolved oxygen concentration control in the aeration tank of a wastewater treatment process, and predicting the changing trend of influent organic load through a large model, allows for advance adjustment of the aeration rate, achieving feedforward-feedback coordinated control. The dissolved oxygen concentration overshoot decreased from 35% to 9.5% (a reduction of 72.9%), the adjustment time was shortened from 240 seconds to 75 seconds (a reduction of 68.8%), and the aeration energy consumption decreased from 1850 kWh / d to 1180 kWh / d (a reduction of 36.2%). Simultaneously, the ammonia nitrogen removal rate increased from 78.5% to 91.2%. This embodiment demonstrates that the method of the present invention also possesses significant technical advantages and economic benefits in environmental protection fields such as wastewater treatment.

[0143] For example, such as Figure 5 As shown, this embodiment also provides a robust adaptive control system for industrial control systems, including: The standardization module is used to perform outlier detection and time-series standardization on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. The identification module is used to detect operating condition disturbances based on a standardized time-series operating condition dataset, classify and identify the type of disturbance when an operating condition disturbance is determined, and extract multi-dimensional features of the disturbance to generate a disturbance feature vector. The inference module is used to combine historical operating data, disturbance classification results and disturbance feature vectors into a prediction input sequence, input it into a pre-trained large model to perform disturbance trend inference, and output disturbance prediction results including disturbance amplitude range, change direction, expected duration and prediction confidence. The matching module is used to extract the current system operating state features, construct a retrieval feature vector by combining the disturbance prediction results, and call the preset control strategy library to perform similarity matching on the retrieval feature vector. If the matching result meets the preset threshold, the control strategy to be verified is obtained; if the matching result does not meet the preset threshold, the large model is called to generate customized control strategy parameters by combining the system dynamic model and constraints, which are used as the control strategy to be verified. The calculation module is used to perform multi-condition closed-loop simulation tests on the control strategy to be verified, calculate the comprehensive performance score, and determine the executable control strategy after the performance constraints are met. The synthesis module is used to calculate the disturbance feedforward compensation amount based on the disturbance prediction results, dynamically adjust the weight coefficients of the feedforward and feedback control according to the prediction confidence, and synthesize the weighted control amount with the feedback control amount corresponding to the executable control strategy to obtain the final control output amount.

[0144] The present invention also provides a robust adaptive control device for an industrial control system, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of the robust adaptive control method for an industrial control system.

[0145] The present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the robust adaptive control method for the industrial control system.

[0146] When the processor executes the computer program, it implements the steps of the robust adaptive control of the industrial control system described above, such as: performing outlier detection and time-series standardization on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. Based on a standardized time-series operating condition dataset, operating condition disturbance detection is performed. When an operating condition disturbance is determined to exist, the disturbance type is classified and identified. At the same time, multi-dimensional disturbance features are extracted to generate a disturbance feature vector. Historical operating data, disturbance classification results, and disturbance feature vectors are combined into a prediction input sequence, which is then input into a pre-trained large model to infer disturbance trends. The output includes disturbance amplitude range, direction of change, expected duration, and prediction confidence. Extract the current system operating state features, construct the retrieval feature vector by combining the disturbance prediction results, call the preset control strategy library to perform similarity matching on the retrieval feature vector, and obtain the control strategy to be verified if the matching result meets the preset threshold; if the matching result does not meet the preset threshold, combine the system dynamic model and constraints to call the large model to generate customized control strategy parameters, which are used as the control strategy to be verified. The control strategy to be verified is subjected to multi-condition closed-loop simulation test, the comprehensive performance score is calculated, and the strategy is determined to be an executable control strategy after the performance constraints are met. The disturbance feedforward compensation is calculated based on the disturbance prediction results. The weight coefficients of the feedforward and feedback control are dynamically adjusted according to the prediction confidence. The feedforward and feedback control are then weighted and synthesized with the feedback control quantity corresponding to the executable control strategy to obtain the final control output quantity.

[0147] For example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules / units can be a series of computer program instruction segments capable of performing preset functions, the instruction segments describing the execution process of the computer program in the robust adaptive control device of the industrial control system. For example, the computer program can be divided into a standardization module, an identification module, an inference module, a matching module, a calculation module, and a synthesis module; the specific functions are as follows: the standardization module is used to perform outlier detection and time-series standardization processing on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset; the identification module is used to perform operating condition disturbance detection based on the standardized time-series operating condition dataset, classify and identify the disturbance type when an operating condition disturbance exists, and extract multi-dimensional disturbance features to generate a disturbance feature vector; the inference module is used to combine historical operating condition data, disturbance classification results, and disturbance feature vectors into a prediction input sequence, input it into a pre-trained large model for disturbance trend inference, and output a disturbance prediction result including the disturbance amplitude range, direction of change, expected duration, and prediction confidence. The matching module extracts the current system operating state features, constructs a retrieval feature vector based on the disturbance prediction results, and calls a preset control strategy library to perform similarity matching on the retrieval feature vector. If the matching result meets a preset threshold, the control strategy to be verified is obtained; if the matching result does not meet the preset threshold, a large model is called to generate customized control strategy parameters based on the system dynamic model and constraints, which is then used as the control strategy to be verified. The calculation module performs multi-condition closed-loop simulation tests on the control strategy to be verified, calculates the comprehensive performance score, and determines the executable control strategy after meeting the performance constraints. The synthesis module calculates the disturbance feedforward compensation amount based on the disturbance prediction results, dynamically adjusts the weight coefficients of the feedforward and feedback control according to the prediction confidence, and weights and synthesizes them with the feedback control amount corresponding to the executable control strategy to obtain the final control output amount.

[0148] The robust adaptive control device for an industrial control system can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. This device may include, but is not limited to, processors and memory. Those skilled in the art will understand that the above are examples of robust adaptive control devices for industrial control systems and do not constitute a limitation on such devices. The device may include more components than described above, or combine certain components, or use different components. For example, the robust adaptive control device may also include input / output devices, network access devices, buses, etc.

[0149] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or any conventional processor. This processor is the control center of the robust adaptive control system in the industrial control system, connecting various parts of the robust adaptive control equipment of the entire industrial control system via various interfaces and lines.

[0150] The memory can be used to store the computer program and / or modules. The processor implements various functions of the robust adaptive control device of the industrial control system by running or executing the computer program and / or modules stored in the memory and calling the data stored in the memory.

[0151] The memory may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a function (such as sound playback, image playback, etc.). The data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). Furthermore, the memory may include high-speed random access memory and non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital cards (SD cards), flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0152] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the robust adaptive control method for an industrial control system.

[0153] If the modules / units integrated in the robust adaptive control system of the industrial control system are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.

[0154] Based on this understanding, the present invention can implement all or part of the processes in the robust adaptive control method for industrial control systems described above, or it can be accomplished by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the robust adaptive control method for industrial control systems described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or preset intermediate forms, etc.

[0155] The computer-readable storage medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0156] It should be noted that the content contained in the computer-readable storage medium may be appropriately increased or decreased according to the requirements of legislation and methodological practice in the jurisdiction. For example, in some jurisdictions, according to legislation and methodological practice, the computer-readable storage medium may not include electrical carrier signals and telecommunication signals.

[0157] Compared with existing control methods, this invention provides a robust adaptive control method and related device for industrial control systems, which has the following advantages: First, this invention enables active prediction and feedforward compensation of various disturbances in industrial control systems, reducing main steam temperature fluctuations by 60% and overshoot by more than 45%, thus transforming the control paradigm from "passive feedback" to "active feedforward". Secondly, this invention achieves adaptive intelligent matching and dynamic generation of control strategies, with a strategy matching success rate of over 85%, and the control success rate of nonlinear time-varying systems is increased from 60% to over 90%, significantly reducing the reliance on manual design. Third, this invention reduces the probability of unsafe execution of the control strategy to below 5% through a two-stage robustness verification and assurance system, ensuring the robust stability of the system under large model-assisted control. Fourth, while ensuring robust stability, this invention reduces the conservatism of the control system, improves overall performance by approximately 30%, and keeps control energy consumption below 120% of the baseline strategy. Fifth, this invention automates all aspects of disturbance prediction, strategy matching, stability verification, and performance evaluation, reducing manual intervention and improving the intelligence level and operating efficiency of the control system. Sixth, this invention integrates three modes: feedforward control, feedback control, and predictive control, achieving complementary advantages and reducing the overall anti-interference capability (IAE index) of the system by more than 50%. Seventh, through a continuous self-optimization mechanism, the system can maintain autonomous and stable operation for more than 90 days without human intervention.

[0158] The above embodiments are merely one of the implementation methods for achieving the technical solution of the present invention. The scope of protection claimed by the present invention is not limited to this embodiment, but also includes any variations, substitutions and other implementation methods that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention.

[0159] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A robust adaptive control method for an industrial control system, characterized in that, include: Outlier detection and time-series standardization are performed on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. Based on a standardized time-series operating condition dataset, operating condition disturbance detection is performed. When an operating condition disturbance is determined to exist, the disturbance type is classified and identified. At the same time, multi-dimensional disturbance features are extracted to generate a disturbance feature vector. Historical operating data, disturbance classification results, and disturbance feature vectors are combined into a prediction input sequence, which is then input into a pre-trained large model to infer disturbance trends. The output includes disturbance amplitude range, direction of change, expected duration, and prediction confidence. Extract the current system operating state features, construct the retrieval feature vector by combining the disturbance prediction results, call the preset control strategy library to perform similarity matching on the retrieval feature vector, and obtain the control strategy to be verified if the matching result meets the preset threshold; if the matching result does not meet the preset threshold, combine the system dynamic model and constraints to call the large model to generate customized control strategy parameters, which are used as the control strategy to be verified. The control strategy to be verified is subjected to multi-condition closed-loop simulation test, the comprehensive performance score is calculated, and the strategy is determined to be an executable control strategy after the performance constraints are met. The disturbance feedforward compensation is calculated based on the disturbance prediction results. The weight coefficients of the feedforward and feedback control are dynamically adjusted according to the prediction confidence. The feedforward and feedback control are then weighted and synthesized with the feedback control quantity corresponding to the executable control strategy to obtain the final control output quantity.

2. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, The real-time process operation data includes process variable measurements, equipment operating status signals, and control loop signals; When classifying and identifying disturbance types, a distinction is made between measurable disturbances that can be directly obtained through sensing measurements and unmeasurable disturbances that cannot be directly measured; the disturbance feature vector includes at least four dimensions: disturbance amplitude, direction of change, duration, and rate of change.

3. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, When the prediction confidence of the disturbance prediction result is lower than the preset prediction confidence threshold, the calculation of the disturbance feedforward compensation amount is stopped, and the weight coefficient of the feedforward control is set to zero. The final control output amount is generated based on the feedback control amount.

4. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, Before performing multi-condition closed-loop simulation testing on the control strategy to be verified, the following steps are also included: A preset system reduced-order equivalent model is used to perform closed-loop stability verification of the control strategy to be verified. If the verification passes, multi-condition closed-loop simulation test is performed. If the verification fails, the control strategy matching or generation steps are re-executed.

5. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, The multi-condition closed-loop simulation test includes at least the rated operating condition, step disturbance condition, slope disturbance condition and noise disturbance condition. The comprehensive performance score is calculated by weighting three performance indicators: overshoot, settling time, and steady-state error. The performance constraints also include controlling energy consumption to not exceed a preset proportion of the energy consumption of the benchmark control strategy.

6. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, The disturbance feedforward compensation amount is calculated based on the dynamic transmission relationship from the disturbance channel to the controlled output; The weighting coefficient of the feedforward control increases synchronously with the increase of the disturbance prediction confidence and the increase of the disturbance amplitude, and the weighting coefficient of the feedforward control is set with a preset upper limit.

7. The robust adaptive control method for an industrial control system according to claim 1, characterized in that, After obtaining the final control output, the process also includes: The final control output is encapsulated into a standard control command and sent to the field controller for execution. Process variable response data and control signal data after control execution are collected to form an execution feedback dataset and stored in historical operating condition data. When the preset model update trigger conditions are met, the newly added execution feedback data is used to perform incremental fine-tuning and updating of the large model.

8. A robust adaptive control system for industrial control systems, characterized in that, include: The standardization module is used to perform outlier detection and time-series standardization on the collected real-time process operation data and historical operating condition data to obtain a standardized time-series operating condition dataset. The identification module is used to detect operating condition disturbances based on a standardized time-series operating condition dataset, classify and identify the type of disturbance when an operating condition disturbance is determined, and extract multi-dimensional features of the disturbance to generate a disturbance feature vector. The inference module is used to combine historical operating data, disturbance classification results and disturbance feature vectors into a prediction input sequence, input it into a pre-trained large model to perform disturbance trend inference, and output disturbance prediction results including disturbance amplitude range, change direction, expected duration and prediction confidence. The matching module is used to extract the current system operating state features, construct a retrieval feature vector by combining the disturbance prediction results, and call the preset control strategy library to perform similarity matching on the retrieval feature vector. If the matching result meets the preset threshold, the control strategy to be verified is obtained; if the matching result does not meet the preset threshold, the large model is called to generate customized control strategy parameters by combining the system dynamic model and constraints, which are used as the control strategy to be verified. The calculation module is used to perform multi-condition closed-loop simulation tests on the control strategy to be verified, calculate the comprehensive performance score, and determine the executable control strategy after the performance constraints are met. The synthesis module is used to calculate the disturbance feedforward compensation amount based on the disturbance prediction results, dynamically adjust the weight coefficients of the feedforward and feedback control according to the prediction confidence, and synthesize the weighted control amount with the feedback control amount corresponding to the executable control strategy to obtain the final control output amount.

9. A robust adaptive control device for an industrial control system, characterized in that, include: Memory, used to store computer programs; A processor, configured to implement the robust adaptive control method for an industrial control system according to any one of claims 1-7 when executing the computer program.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it is used to implement the robust adaptive control method for industrial control systems according to any one of claims 1-7.

Images