Dynamic learning method and system of intelligent controller

Abstract

The invention discloses a dynamic learning method and system for an intelligent controller, and the method specifically comprises the steps: building a system dynamic characteristic model for the intelligent controller through employing a dynamic model building method based on the real-time data of an industrial process; based on the system dynamic characteristic model, a parameter optimization objective function is set for the intelligent controller, and the parameter optimization objective function is used for adjusting the balance among control precision, strategy flexibility and real-time indexes; performing global optimization on the parameter optimization objective function by adopting a particle swarm optimization algorithm, and dynamically adjusting the particle speed and position in the optimization process through a dynamic learning mechanism to obtain an optimization result; and dynamically adjusting the control parameters of the intelligent controller based on the optimization result. According to the invention, the adaptability and flexibility of the intelligent controller are improved, rapid response and accurate control are realized, and the performance of the intelligent controller in complex industrial process control is improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent control technology, and in particular to a dynamic learning method and system for an intelligent controller. Background Art

[0002] In the field of complex industrial process control, the performance of intelligent controllers is directly related to production efficiency and safety. However, due to the complexity of industrial processes, intelligent controllers face numerous challenges. First, the rapid changes in operating conditions require control systems to be highly adaptable and flexible. Traditional intelligent controllers are often designed based on static models, which fail to accurately reflect the dynamic characteristics of the system during actual operation. This leads to slow response to unexpected situations and poor control effectiveness.

[0003] Secondly, the nonlinear dynamic characteristics of industrial processes are also a significant challenge for intelligent controllers. These nonlinear dynamics cause the system's operating state and parameters to change over time, making traditional static models unable to accurately predict and control these changes. This makes it difficult for intelligent controllers to find the global optimal solution during parameter optimization, limiting control accuracy and stability.

[0004] Furthermore, the time-varying nature of system parameters further complicates intelligent controller design. In fields such as petrochemicals, electric power, and metallurgy, system parameters often fluctuate significantly over time. While traditional intelligent controller parameter optimization methods, such as genetic algorithms and particle swarm optimization, have improved controller performance to some extent, they often struggle to meet the demands for rapid response and precise control in industrial control systems with demanding real-time requirements. Especially in situations where system dynamics are significant, these methods can easily become trapped in local optimal solutions, resulting in suboptimal controller performance.

[0005] Most existing intelligent controller design methods ignore the impact of system dynamics and rely on static models for parameter optimization. However, in real industrial processes, the dynamic characteristics of a system significantly impact controller performance. Therefore, how to dynamically learn and optimize intelligent controller parameters based on the system's dynamic characteristics has become a pressing issue.

[0006] Furthermore, traditional parameter optimization methods suffer from significant limitations, including high computational complexity, slow convergence, and a tendency to fall into local optima. In industrial control systems with demanding real-time performance, these methods struggle to meet the demands for rapid response and precise control. Parameter changes during system operation, in particular, often require the controller to rapidly adjust to maintain system stability and efficiency. However, existing intelligent controllers often struggle to adjust parameters in a timely manner in these complex situations, resulting in insufficient control accuracy and potentially even causing control system instability. Summary of the Invention

[0007] The purpose of the present invention is to provide a dynamic learning method and system for an intelligent controller, which improves the adaptability and flexibility of the intelligent controller, realizes rapid response and precise control, and enhances the performance of the intelligent controller in complex industrial process control, so as to solve at least one of the above-mentioned prior art problems.

[0008] In a first aspect, the present invention provides a dynamic learning method for an intelligent controller, the method specifically comprising:

[0009] Based on the real-time data of industrial processes, a system dynamic characteristic model is established for the intelligent controller using a dynamic modeling method. The real-time data of industrial processes includes the trend of operating condition changes, nonlinear characteristics of the system, and time-varying parameter information.

[0010] Based on the system dynamic characteristic model, setting a parameter optimization objective function for the intelligent controller, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility and real-time performance indicators;

[0011] A particle swarm optimization algorithm is used to globally optimize the parameter optimization objective function, and the particle speed and position are dynamically adjusted during the optimization process through a dynamic learning mechanism to obtain the optimization result;

[0012] Based on the optimization result, dynamically adjust the control parameters of the intelligent controller, wherein the control parameters include control gain, integral time and differential time;

[0013] During the optimization process, if it is detected that the diversity of the particle swarm decreases or the objective function value does not improve for a long time, a global search strategy is used to jump out of the local optimum.

[0014] In a second aspect, the present invention provides a dynamic learning system for an intelligent controller, the system specifically comprising:

[0015] A first dynamic learning module is configured to establish a system dynamic characteristic model for the intelligent controller using a dynamic modeling method based on real-time industrial process data, wherein the real-time industrial process data includes operating condition change trends, system nonlinear characteristics, and time-varying parameter information;

[0016] A second dynamic learning module is used to set a parameter optimization objective function for the intelligent controller based on the system dynamic characteristic model, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility and real-time performance indicators;

[0017] The third dynamic learning module is used to use the particle swarm optimization algorithm to perform global optimization on the parameter optimization objective function, and dynamically adjust the particle speed and position during the optimization process through the dynamic learning mechanism to obtain the optimization result;

[0018] A fourth dynamic learning module, configured to dynamically adjust control parameters of the intelligent controller based on the optimization result, the control parameters including control gain, integral time, and differential time;

[0019] The fifth dynamic learning module is used to use a global search strategy to jump out of the local optimum if it is detected that the diversity of the particle swarm is reduced or the objective function value has not improved for a long time during the optimization process.

[0020] In a third aspect, the present invention provides a computer device comprising: a memory and a processor and a computer program stored in the memory, wherein when the computer program is executed on the processor, a dynamic learning method of an intelligent controller as described in any one of the above methods is implemented.

[0021] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the dynamic learning method of the intelligent controller as described in any one of the above methods is implemented.

[0022] Compared with the prior art, the present invention has at least one of the following technical effects:

[0023] 1. The present invention improves the adaptability and flexibility of the intelligent controller, achieves rapid response and precise control, and enhances the performance of the intelligent controller in complex industrial process control.

[0024] 2. The dynamic learning method of the intelligent controller of the present invention establishes a system dynamic characteristic model through real-time data, and sets parameters to optimize the objective function according to this model, thereby achieving a balanced adjustment of control accuracy, strategy flexibility and real-time indicators, and improving the adaptability and control performance of the intelligent controller to complex industrial processes.

[0025] 3. The present invention adopts technologies such as sliding window, operating condition feature extraction, time-varying parameter identification and deep feature network, which can accurately extract operating condition change trends, system nonlinear characteristics and time-varying parameter information from real-time industrial process data, and establish an accurate system dynamic characteristic model for the intelligent controller, thereby improving the controller's prediction and control capabilities.

[0026] 4. The present invention comprehensively evaluates the control accuracy, strategy flexibility and real-time indicators through the time domain integral formula, parameter change sensitivity formula and time penalty function, provides a clear optimization direction for the particle swarm optimization algorithm, and ensures that the intelligent controller has good control performance in complex industrial processes.

[0027] 5. The present invention can quickly find the global optimal solution of the parameter optimization objective function through the particle swarm optimization algorithm. The nonlinear attenuation strategy enables the algorithm to smoothly transition between global optimization and local optimization, avoiding premature convergence to the local optimal solution, and improving the convergence speed and stability of the algorithm.

[0028] 6. This invention enhances the diversity of the particle swarm by dynamically adjusting particle speed and position, combined with strategies such as Gaussian perturbation terms, and prevents the algorithm from falling into a local optimum. Furthermore, the learning factor is dynamically adjusted based on parameters such as the standard deviation of the particle swarm distribution, improving the algorithm's adaptability and convergence performance.

[0029] 7. The control gain of the present invention is dynamically adjusted according to the deviation between the system output and the set value, and the deviation change is smoothed by strategies such as exponential decay terms to improve the response speed and stability of the controller.

[0030] 8. The integral time of the present invention is dynamically adjusted according to the rate of change of the deviation, and the deviation information is accumulated through the integral term, thereby enhancing the controller's ability to correct the system steady-state error.

[0031] 9. The differential time of the present invention is dynamically adjusted according to the acceleration of the deviation change, and the control amount is prevented from being too large through strategies such as saturation function, thereby improving the sensitivity and robustness of the controller to the dynamic characteristics of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0033] Figure 1 This is a flow chart of a dynamic learning method for an intelligent controller provided by one embodiment of the present invention;

[0034] Figure 2 1 is a structural diagram of a dynamic learning system of an intelligent controller provided by one embodiment of the present invention;

[0035] Figure 3 It is a structural diagram of a computer device provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0036] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0037] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0038] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0039] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0040] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0041] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0042] In the embodiments of the present application, the execution subject of the process includes a terminal device, which includes but is not limited to: a server, a computer, a smart phone, a tablet computer, and other devices capable of executing the method disclosed in the present application. Figure 1 A flow chart of a dynamic learning method for an intelligent controller according to an embodiment of the present invention is shown, and is described in detail as follows:

[0043] S101 , based on real-time industrial process data, a system dynamic characteristic model is established for the intelligent controller using a dynamic modeling method, wherein the real-time industrial process data includes operating condition change trends, system nonlinear characteristics, and time-varying parameter information.

[0044] In this embodiment, a real-time data stream is acquired from an industrial process sensor network. The data stream contains information about operating condition variations, nonlinearity, and parameter variations. In the data preprocessing module, a sliding window method is used to segment the data stream to obtain data segments. Principal component analysis is used to extract eigenvalues ​​for each data segment. These eigenvalues ​​contain information about operating condition variations, nonlinearity, and parameter variations. Based on the eigenvalues, a support vector machine is used to classify the data and determine the operating mode to which the data belongs. If the data belongs to a known operating mode, the corresponding dynamic model is retrieved from a pre-established model library. If the data belongs to an unknown operating mode, a recursive neural network is used to establish a new dynamic model that covers the operating condition variations, nonlinearity, and parameter variations. In the model training module, gradient descent is used to optimize the dynamic model parameters, with the optimization objective being to minimize prediction error. Cross-validation is used to determine the model's generalization capability. If the model's generalization capability is insufficient, the model is rebuilt by returning to the feature extraction stage. The optimized dynamic model is stored in a model library, which uses a hash table structure with the hash key being the operating mode identifier. In the model prediction module, the corresponding dynamic model is retrieved from the model library based on the current operating mode identifier for real-time prediction. The model prediction results are corrected using the Kalman filter method, taking into account system noise and measurement noise. The corrected prediction results are fed back to the industrial process control system to achieve closed-loop control.

[0045] For example, during the production process at a chemical plant, precise control of the reactor temperature and pressure is required to ensure product quality and production safety. Sensors installed on the reactor collect real-time data on key parameters such as temperature, pressure, and flow. Simultaneously, they record operating condition trends, such as raw material input and reaction rate. The collected data is cleaned and filtered to remove outliers and noise. Time-varying parameter information is smoothed to reduce the impact of fluctuations on the model. Based on the nonlinear characteristics of the system and the time-varying parameter information, an appropriate dynamic model structure is selected, such as a nonlinear autoregressive moving average model (NARMAX) or a state-space model. Model parameters are estimated using system identification methods, such as least squares or Kalman filtering. Parameters are dynamically adjusted based on operating condition trends. Model output is compared with actual data to verify model accuracy and reliability. Model performance is continuously optimized by adjusting the model structure and parameters. Based on the established dynamic model, an intelligent controller is designed. Appropriate control algorithms, such as fuzzy control, neural network control, or adaptive control, are selected to achieve precise temperature and pressure control. The intelligent controller is embedded in the plant's automated control system for real-time online control. Through the monitoring interface, you can view the control effect and system status in real time.

[0046] In this embodiment, a dynamic modeling approach is employed to accurately describe the system's dynamic characteristics, thereby improving the control accuracy of the intelligent controller. Because the model takes into account operating condition trends and time-varying parameter information, the intelligent controller can adapt to control requirements under varying operating conditions, enhancing the system's robustness and adaptability. When raw material input or reaction rate changes, the intelligent controller can rapidly adjust the control strategy to maintain stable system operation. Furthermore, the intelligent controller can dynamically adjust the control strategy based on operating condition trends, further optimizing energy efficiency.

[0047] S102 , based on the system dynamic characteristic model, setting a parameter optimization objective function for the intelligent controller, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility and real-time performance indicators.

[0048] In this embodiment, the system state variables and control input variables of the dynamic model are obtained to establish a state-space equation. A gradient descent algorithm is used to weight the control accuracy, strategy flexibility, and real-time performance indicators in the objective function. Based on the weight distribution results of the objective function, the parameter optimization target value of the intelligent controller is calculated. The optimal controller parameter combination is searched using the particle swarm optimization algorithm. Based on the optimal parameter combination, the control strategy of the intelligent controller is updated. System output data is collected in real time to determine whether the current control accuracy has reached a preset threshold. If the control accuracy does not reach the preset threshold, the objective function weight distribution is readjusted and iterative optimization is performed.

[0049] For example, control accuracy is measured by the root mean square error (RMSE) or maximum absolute error (MaxAE). Strategy flexibility is measured by the adaptability and adjustment speed of the controller under different operating conditions, which can be evaluated by calculating the rate of change of controller parameters under different operating conditions. Real-time performance is measured by the time difference between the controller receiving data and responding. The objective function can be designed as: J = α·RMSE + β·Flexibility_Index + γ·Latency, where α, β, and γ are weight coefficients used to adjust the relative importance of each indicator in the objective function. RMSE is a quantitative indicator of control accuracy, Flexibility_Index is a quantitative indicator of strategy flexibility, and Latency is a quantitative indicator of real-time performance. The weight coefficients are adjusted through expert experience, simulation experiments, or actual operating data to ensure that the objective function reflects the actual needs of the control system. Multi-objective optimization methods, such as the weight sum method and goal programming method, can be used to determine the optimal combination of weight coefficients.

[0050] In this embodiment, by optimizing the control accuracy term in the objective function, the error of the control system can be significantly reduced, and product quality and production stability can be improved. The optimized intelligent controller can better adapt to the control requirements under different working conditions and improve the robustness and adaptability of the system. When the working conditions change, the controller can quickly adjust the control strategy to maintain the stable operation of the system. By optimizing the real-time term in the objective function, the time difference between the controller receiving data and responding can be shortened, and the response speed and real-time performance of the control system can be improved. This is especially important for production processes that require fast response. By balancing control accuracy, strategy flexibility and real-time indicators, the optimized intelligent controller has been significantly improved in comprehensive performance. This can not only improve production efficiency, but also reduce energy consumption and production costs, bringing greater economic benefits to the enterprise.

[0051] S103, using a particle swarm optimization algorithm to perform global optimization on the parameter optimization objective function, and dynamically adjusting the particle speed and position during the optimization process through a dynamic learning mechanism to obtain an optimization result.

[0052] In this embodiment, the objective function is obtained, the particle swarm position and velocity are initialized, and initial parameters are set. The current fitness value of the particle swarm is calculated based on historical information, and the optimal solution is recorded. If the current fitness value is better than the historical optimal solution, the historical optimal solution is updated, and the particle velocity and position are adjusted. A dynamic learning mechanism is used to recalculate the particle velocity and position based on historical information. If the number of iterations reaches a preset threshold, the optimization process is terminated and the global optimal solution is output. Based on the convergence of the optimization process, the learning factor and inertia weight are dynamically adjusted. The final optimization result is achieved by reducing the computational effort and optimizing the particle update strategy.

[0053] For example, to globally optimize the objective function, a particle swarm optimization (PSO) algorithm is employed. A dynamic learning mechanism dynamically adjusts particle speeds and positions during the optimization process to obtain optimal control parameters. First, a certain number of particles are randomly generated, each representing a set of candidate parameters for the intelligent controller. Each particle's position and speed, as well as its individual historical optimal position (pBest) and global historical optimal position (gBest), are initialized. Then, each particle's current position (i.e., candidate parameters) is evaluated to calculate the objective function value. Each particle's pBest is updated if the objective function value at the current position is better than the previous pBest. The global gBest is updated if a particle's pBest is better than the current gBest. The dynamic learning mechanism dynamically adjusts particle speeds based on the particle's current speed, the distance between its individual optimal position and the current position, the distance between its global optimal position and the current position, and a random acceleration factor. The particle's position is then dynamically adjusted based on the updated speed. Based on the convergence of the optimization process, parameters such as the inertia weight and learning factor are dynamically adjusted to balance global and local search capabilities and avoid premature convergence. Repeat the steps of velocity updating, position updating, and evaluating the objective function until a preset number of iterations is reached or the objective function value meets the convergence condition.

[0054] In this embodiment, the particle swarm optimization algorithm can effectively search for the global optimal solution in a complex multi-dimensional space by simulating the foraging behavior of a flock of birds. In the parameter optimization of the intelligent controller, the algorithm can find the optimal parameter combination that minimizes the objective function value. Through the dynamic learning mechanism, the algorithm can dynamically adjust the speed and position of the particles, as well as the optimization parameters, according to the convergence of the optimization process, thereby avoiding premature convergence and improving the search efficiency and the quality of the optimization results. Since the particle swarm optimization algorithm has the characteristics of parallel search and can gradually approach the global optimal solution through multiple iterations, the optimization results have good stability and reliability. After optimizing the parameters of the intelligent controller using the particle swarm optimization algorithm, the control accuracy, strategy flexibility and real-time indicators of the control system are significantly improved.

[0055] S104: Dynamically adjust control parameters of the intelligent controller based on the optimization result, where the control parameters include control gain, integral time, and differential time.

[0056] In this embodiment, after obtaining the optimization results, the parameters corresponding to the particle with the minimum objective function value are used as the initial parameters of the intelligent controller. Based on these initial parameters, the adjustment range of the control gain, integral time, and differential time is determined. Within this adjustment range, the control gain, integral time, and differential time are fine-tuned using the gradient descent method to ensure a further decrease in the objective function value. The adjusted parameters are then re-entered into the particle swarm optimization algorithm, and optimization continues until the objective function value reaches the preset accuracy requirement.

[0057] For example, the optimal control gain, integral time, and differential time are read from the output of the particle swarm optimization algorithm. If the parameters output by the optimization algorithm do not match the parameter range or format actually used by the intelligent controller, appropriate mapping or conversion is required. For example, if the parameters output by the optimization algorithm are normalized values, they need to be converted back to the actual parameter range. The optimal control gain, integral time, and differential time are set as the corresponding parameters of the intelligent controller respectively. If the control system is running in real time, these parameters can be dynamically updated in each iteration or at a specific time interval to adapt to changes in operating conditions. A parameter update mechanism, such as a smooth transition strategy, can be designed to avoid system instability caused by sudden changes in parameters. The performance indicators of the control system, such as control accuracy, overshoot, stabilization time, etc., are monitored in real time. If the performance indicators do not meet the preset standards, a re-optimization process can be triggered, or the weight coefficient can be adjusted to optimize specific indicators.

[0058] In this embodiment, by dynamically adjusting the control gain, integral time, and differential time, the intelligent controller can track the set value more accurately, reduce the steady-state error, and improve the control accuracy. The optimized control parameters can reduce the overshoot and oscillation of the system and improve the stability and robustness of the system. This is particularly important for processing nonlinear, time-varying, and highly uncertain industrial processes. By updating the control parameters in real time or periodically, the intelligent controller can better adapt to changes in operating conditions and maintain the optimal control performance of the system. The optimized control system can respond to changes in operating conditions more quickly, reduce fluctuations and downtime in the production process, and thus improve production efficiency. Based on the optimization results of the particle swarm optimization algorithm, dynamically adjusting the control gain, integral time, and differential time of the intelligent controller can significantly improve the performance and stability of the control system, bring significant economic benefits to the factory's production process control, and provide an effective parameter optimization and adjustment strategy for industrial process control.

[0059] S105, during the optimization process, if it is detected that the diversity of the particle swarm decreases or the objective function value does not improve for a long time, a global search strategy is used to jump out of the local optimum.

[0060] In this embodiment, during the optimization process, the diversity index of the particle swarm is calculated in real time. If the diversity falls below a preset threshold, a global search is performed by randomly initializing the positions and velocities of some particles. Based on the improvement in the objective function value, if the objective function value does not decrease after ten consecutive iterations, the optimization process is determined to have fallen into a local optimum, and the position and velocity of the particle swarm are reinitialized. A threshold is set based on historical data, and statistical hypothesis testing is used to determine whether operating conditions have changed or whether parameters have time-varying. If it is determined that operating conditions have changed or parameters have time-varying, the least squares method is triggered for dynamic learning. The optimized parameters are obtained based on the dynamic learning process, and the parameters of the proportional-integral-derivative controller are updated. Fuzzy logic is used to adjust the control rules and weights based on the updated proportional-integral-derivative controller parameters. A new control strategy is generated using the adjusted control rules and weights. Based on the new control strategy, adaptive control of the proportional-integral-derivative controller is executed. The adaptive control of the proportional-integral-derivative controller updates the system operating status.

[0061] In the application of the Particle Swarm Optimization (PSO) algorithm, two common problems are reduced particle swarm diversity and prolonged stagnation of objective function values. Both of these issues can cause the algorithm to become stuck in a local optimum, preventing it from finding a global optimal solution. To overcome this challenge, this embodiment proposes a global search strategy to escape the local optimum when these issues are detected.

[0062] For example, during the algorithm iteration process, a particle swarm diversity index is regularly calculated. Common diversity indexes include the standard deviation of inter-particle distances and the standard deviation of particle velocities. When the diversity index falls below a preset threshold, the particle swarm diversity is considered to have decreased, and a global search strategy is required. The objective function value after each iteration is recorded, and its rate of change is calculated. If the objective function value does not significantly improve after multiple consecutive iterations (i.e., the rate of change is less than a preset threshold), the algorithm is considered to have fallen into a local optimum, and a global search strategy is also required. When a decrease in particle swarm diversity is detected or the objective function value has not improved for a long time, a global search strategy is introduced. The global search strategy may include reinitializing some particles, increasing the randomness of the particle swarm, introducing new search directions, etc. For example, a portion of particles can be randomly selected and their positions and velocities reinitialized to introduce a new search space. Alternatively, parameters of the particle swarm optimization algorithm, such as the inertia weight and learning factor, can be adjusted to change the search behavior of the particles. After introducing the global search strategy, the algorithm iteration process continues. The objective function value is regularly evaluated to determine whether the algorithm has converged to the global optimal solution. If the algorithm converges or reaches the preset number of iterations, the iteration stops and the optimal solution is output.

[0063] In this embodiment, by introducing a global search strategy, the algorithm can escape from local optimal solutions when the diversity of the particle swarm decreases or the objective function value does not improve for a long time, thereby increasing the possibility of finding the global optimal solution. The global search strategy can guide particles to explore a wider search space, preventing particles from lingering near the local optimal solution, thereby improving search efficiency. Faced with complex optimization problems, the algorithm can more flexibly adapt to different search environments, improving the robustness and versatility of the algorithm. By escaping local optimal solutions and improving search efficiency, the algorithm can find higher-quality optimal solutions, thus meeting the needs of practical applications.

[0064] In some embodiments, in step S101, the process of establishing a system dynamic characteristic model for the intelligent controller using a dynamic modeling method based on real-time industrial process data specifically includes:

[0065] Acquire real-time industrial process data, and segment the real-time industrial process data using a sliding window method to form standard data;

[0066] Extracting working condition characteristics from the standard data using a working condition characteristic extraction formula to generate a working condition change trend indicator;

[0067] Using a time-varying parameter identification formula to perform time-varying parameter identification on the standard data to generate a time-varying parameter vector;

[0068] Using a deep feature network to extract nonlinear features from the standard data to generate system nonlinear features;

[0069] A system dynamic characteristic model is formed according to the operating condition change trend index, the time-varying parameter vector and the system nonlinear characteristics.

[0070] In this example, assume that a chemical production process requires real-time monitoring of system dynamics to enable timely adjustment of production parameters and ensure product quality and efficiency. The system is equipped with various sensors and monitoring equipment capable of collecting real-time parameters such as temperature, pressure, and flow rate during the production process. Through these sensors, monitoring equipment, and other real-time monitoring systems, various real-time data from the chemical production process are collected. This data includes parameters such as temperature, pressure, flow rate, humidity, and vibration, which reflect the real-time status of the production process. A sliding window of fixed size is set, such as a window containing the most recent 100 data points. As new data arrives, the sliding window continuously moves forward, incorporating the latest data points with each shift and removing the oldest data points. The data within the sliding window serves as standard data for subsequent feature extraction and parameter identification. Based on the actual conditions of chemical production, a process feature extraction formula is designed. This formula is applied to the standard data to extract indicators reflecting process condition trends, such as temperature trends and pressure fluctuations. Based on the mathematical model of chemical production and the theory of time-varying parameter identification, a time-varying parameter identification formula is designed. This formula is applied to the standard data to identify time-varying parameters in the production process, such as reaction rate and mass transfer coefficient. These time-varying parameters are organized into vectors for subsequent modeling of the system's dynamic characteristics. A deep feature network, such as a convolutional neural network (CNN) or deep neural network (DNN), is constructed. Standard data is input into the deep feature network for nonlinear feature extraction. Through network training and learning, features reflecting the system's nonlinear characteristics are extracted. The operating condition trend indicators, time-varying parameter vectors, and the system's nonlinear characteristics are integrated. Using machine learning or data mining techniques, a system dynamic characteristics model is constructed. This model can reflect the real-time status of the production process, operating condition trends, and the impact of time-varying parameters, providing support for production optimization and fault warning.

[0071] In this embodiment, real-time data segmentation is achieved through a sliding window approach, ensuring data timeliness and accuracy. The application of operating condition feature extraction formulas and time-varying parameter identification formulas enables accurate extraction of indicators and parameters reflecting the characteristics of the production process. By integrating operating condition trend indicators, time-varying parameter vectors, and system nonlinear characteristics, the constructed system dynamic characteristics model possesses enhanced predictive capabilities. This model can monitor dynamic changes in the production process in real time, provide early warning of potential failures, and provide decision support for production optimization.

[0072] In some embodiments, the use of a sliding window method to segment the industrial process real-time data to form standard data specifically includes:

[0073] The real-time data stream of the industrial process is segmented into sliding windows by a preset window length and a preset sliding step size to form a continuous data window sequence;

[0074] Perform wavelet threshold denoising on each data window in the data window sequence to obtain preprocessed data;

[0075] Based on the preprocessed data, the mean and standard deviation in each data window are calculated, and Z-score normalization is performed to generate standard data.

[0076] In some embodiments, the step of using a time-varying parameter identification formula to perform time-varying parameter identification on the standard data to generate a time-varying parameter vector specifically includes:

[0077] Extract regression vectors and observation outputs from standard data to build a dynamic model;

[0078] Based on the dynamic model, the Kalman gain and covariance matrix are updated by the recursive least squares algorithm with forgetting factor to calculate the time-varying parameter estimates;

[0079] The covariance matrix is ​​compressed and the time-varying parameter estimates are smoothed using the exponential weighted averaging method to generate the target time-varying parameter vector.

[0080] In some embodiments, the use of a deep feature network to extract nonlinear features from the standard data to generate system nonlinear features specifically includes:

[0081] Input the standard data into the deep feature network and generate hidden features through layer-by-layer nonlinear transformation;

[0082] Perform multi-scale one-dimensional convolution operations on hidden features to extract local time domain patterns, and then concatenate the hidden features and local time domain patterns to generate fused features;

[0083] The fusion features are subjected to principal component analysis and dimensionality reduction to generate the nonlinear characteristics of the system.

[0084] Furthermore, the sliding window method satisfies Among them, represents each sampling window, N represents the number of data points in the sampling window, represents the process variable timestamp of the i-th data point, represents the control signal timestamp of the i-th data point;

[0085] The working condition feature extraction formula satisfies in, represents the working condition change trend index, n represents the total number of control variables, ω j represents the weight of the j-th control variable, Δx j represents the rate of change of the j-th control variable, Δt represents the unit time, λ represents the time attenuation factor, Indicates the time of the most recent working mode switching;

[0086] The time-varying parameter identification formula satisfies y(t)=f(x(t-1),...,x(tn y ),u(t-1),...,u(tn u ))+∈(t), Among them, θ(t) represents the current time-varying parameter vector, θ(t-1) represents the previous time-varying parameter vector, K(t) represents the Kalman gain, and y(t) represents the current system output. represents the previous system output, x represents the state vector, x(t-1),...,x(tn y ) indicates the output from the t-1th to the tnth y Multiple state vectors output at the same time, u represents the control input vector, u(t-1),...,u(tn u ) indicates the t-1th input to the tnth u The input control input vector, ∈(t) represents the modeling error, P(t-1) represents the covariance matrix, represents the regression vector, λ0 represents the forgetting factor, represents the transposed regression vector;

[0087] The deep feature network satisfies h(t)=σ(W h ·[x;u]+b h ), where h(t) represents the nonlinear characteristics of the system, W h and b h represents the learnable parameter matrix, σ(·) represents the activation function, and [;] represents the vector concatenation operation.

[0088] In this example, the sampling window size N is set to, for example, N = 100, indicating that each window contains the most recent 100 data points. For each newly arrived data point, the sliding window is updated, removing the oldest data point and adding the newest data point. Data points include the timestamps and corresponding values ​​of process variables (such as temperature and pressure) and control signals (such as motor speed and valve opening).

[0089] Set the weights and time decay factors for the control variables and calculate the rate of change of each control variable, which is the difference between adjacent data points divided by the unit time. Calculate the operating condition trend indicator based on the operating condition feature extraction formula, taking into account the impact of the most recent operating condition switch time on the indicator.

[0090] Initialize the Kalman gain, covariance matrix, and forgetting factor. For each sampling window, calculate the state vector and regression vector based on the system output and control input vectors. Apply the Kalman filter formula to update the time-varying parameter vector and covariance matrix. Consider the impact of the forgetting factor on the covariance matrix to adapt to system dynamics.

[0091] Design a deep feature network architecture consisting of an input layer, hidden layers, and an output layer. The input layer receives standard data (data points within a sliding window). The hidden layer contains multiple neurons, using learnable parameter matrices W and b and activation functions (such as ReLU) to perform nonlinear transformations. The output layer generates nonlinear features of the system and combines the outputs of multiple hidden layers through vector concatenation.

[0092] The system integrates operating condition trend indicators, time-varying parameter vectors, and nonlinear characteristics. A system dynamics model is constructed using machine learning algorithms (such as support vector machines and neural networks). This model can reflect the real-time state of the production process, operating condition trends, and the impact of time-varying parameters.

[0093] In this embodiment, the sliding window method ensures data timeliness and accuracy, enabling timely reflection of the latest status of the production process. The operating condition feature extraction formula and the time-varying parameter identification formula accurately extract key information, providing strong support for modeling the system's dynamic characteristics. The integration of operating condition trend indicators, time-varying parameter vectors, and system nonlinear characteristics improves the model's predictive accuracy and robustness. The model can monitor dynamic changes in the production process in real time, provide early warning of potential failures, and provide decision support for production optimization.

[0094] In some embodiments, the parameter optimization objective function includes a time domain integral formula for control accuracy quantification, a parameter change sensitivity formula for evaluating strategy flexibility, and a time penalty function for real-time constraints;

[0095] The time domain integration formula satisfies e(t)=r(t)-y(t), Δu(t)=u(t)-u(t-Δt), where, f acc represents the time domain integral index, e(t) represents the deviation between the system output and the set value, Δu(t) represents the rate of change of the control quantity, ω e represents the bias weight, ω u represents the weight of the control variable change rate, r(t) represents the set value, y(t) represents the system output, u(t) represents the current control variable, u(t-Δt) represents the control variable at the previous moment, Δt represents the time interval, and t represents time;

[0096] The parameter change sensitivity formula satisfies Among them, f flex Indicates the sensitivity of parameter changes, n0 indicates the number of historical working condition samples, J k represents the performance index under the kth working condition, Indicates the intensity of the effect of parameter changes on performance indicators, represents the gradient operator, ∈0=1e-6 and is a zero-proof constant;

[0097] The time penalty function satisfies Among them, f rt represents the time penalty function, λ0 represents the penalty coefficient, t calc It represents the actual calculation time taken by the intelligent controller to complete a parameter optimization or control quantity calculation, t max Indicates the maximum allowable time required to complete a parameter optimization or control variable calculation.

[0098] In this embodiment, a deviation weight and a control quantity change rate weight are set. These two weights reflect the importance the system attaches to control accuracy and stability. At each sampling time t, the deviation e(t) = r(t) - y(t) between the system output y(t) and the set value r(t) is calculated. At the same time, the control quantity change rate Δu(t) = u(t) - u(t-Δt) is calculated, where u(t) and u(t-Δt) are the control quantities at the current and previous moments, respectively. According to the time domain integral formula, the time domain integral index is calculated. This index combines the effects of the deviation and the control quantity change rate and is used to quantify the control accuracy.

[0099] Collect historical working condition samples, each sample contains a set of parameters and their corresponding performance indicators. For each working condition sample, calculate the impact of parameter changes on performance indicators, which can be achieved through the gradient operator. To achieve this, a small constant ∈ 0 is added to prevent division by zero. According to the parameter change sensitivity formula, the parameter change sensitivity is calculated. This indicator reflects the flexibility and robustness of the strategy to parameter changes.

[0100] Set the penalty coefficient and maximum allowable time. During each parameter optimization or control variable calculation, record the actual calculation time. Based on the time penalty function, calculate a time penalty proportional to the difference between the actual calculation time and the maximum allowable time.

[0101] The objective function for parameter optimization is constructed by combining the time-domain integral metric, parameter change sensitivity, and time penalty. Appropriate optimization algorithms (such as gradient descent and genetic algorithms) are employed to minimize the objective function. During the optimization process, system parameters are continuously adjusted until convergence conditions are met or a predetermined number of optimizations are reached.

[0102] In this embodiment, the application of the time-domain integral formula quantifies the deviation between the system output and the setpoint, thereby improving control accuracy. The introduction of the control variable change rate helps avoid drastic fluctuations in the control variable while ensuring control accuracy, thereby improving system stability. The calculation of parameter change sensitivity reflects the strategy's sensitivity to parameter changes, helping to consider the system's robustness and adaptability when designing the control strategy. By optimizing the parameter change sensitivity term in the objective function, the system can more quickly adjust the control strategy in the face of parameter changes, maintaining stable system performance. The introduction of a time penalty function ensures that the system considers the impact of computational time when optimizing parameters, thereby avoiding system performance degradation caused by computational delays. By setting a maximum allowable time and a penalty coefficient, computational time can be effectively controlled, ensuring that the system can complete parameter optimization or control variable calculation within the specified time. The construction of the parameter optimization objective function comprehensively considers multiple aspects, including control accuracy, strategy flexibility, and real-time constraints, resulting in comprehensive improvements in system performance. The application of the optimization algorithm enables precise adjustment of system parameters, thereby improving the overall performance and stability of the system.

[0103] In some embodiments, in the above step S103, the particle swarm optimization algorithm is used to perform global optimization on the parameter optimization objective function, specifically including:

[0104] Setting initial parameters of the particle swarm, calculating the current fitness value of the parameter optimization objective function using a particle swarm optimization algorithm based on the initial parameters, and recording the optimal solution;

[0105] The nonlinear attenuation strategy formula is used to make the particle swarm optimization algorithm smoothly transition between global optimization and local optimization. The nonlinear attenuation strategy formula satisfies Among them, ω(t0) represents the current inertia weight and is used to control the tendency of the particle to maintain the current speed during the optimization process, ω max Indicates the maximum value of inertia weight, ω min Indicates the minimum value of the inertia weight, t0 indicates the current number of iterations, T max Represents the maximum number of iterations, γ represents the decay exponent and is used to control the nonlinear degree of inertia weight decay.

[0106] Furthermore, the setting of initial parameters of the particle swarm, calculating the current fitness value of the parameter optimization objective function using a particle swarm optimization algorithm based on the initial parameters, and recording the optimal solution specifically includes:

[0107] Determine the size of the particle swarm, randomly generate the initial position and initial velocity for each particle in the particle swarm, set the maximum and minimum values ​​of the inertia weight, the decay exponent, and the maximum number of iterations;

[0108] For each particle, its position is substituted into the parameter optimization objective function to calculate the fitness value;

[0109] Based on the fitness value, record the global optimal solution and the corresponding position, as well as the individual optimal solution and the corresponding position of each particle;

[0110] According to the velocity update formula of the PSO algorithm, the new velocity of each particle is calculated and the position of the particle is updated according to the new velocity.

[0111] Furthermore, the nonlinear attenuation strategy formula is used to enable the particle swarm optimization algorithm to smoothly transition between global optimization and local optimization, specifically including:

[0112] In each iteration of the particle swarm optimization process, the current inertia weight is calculated using the nonlinear decay strategy formula according to the current number of iterations and the maximum number of iterations;

[0113] Use the current inertia weight in the velocity update formula for the next iteration;

[0114] In each iteration, the global optimal solution and the individual optimal solution are updated;

[0115] After the iteration is completed, the global optimal solution and the corresponding position are output.

[0116] In this embodiment, the Particle Swarm Optimization (PSO) algorithm is a swarm intelligence-based optimization method that simulates the foraging behavior of bird flocks to search for the optimal solution to a problem. This embodiment aims to use the PSO algorithm to solve a parameter optimization objective function and improve the algorithm's performance by introducing a nonlinear decay strategy formula to balance global and local optimization capabilities.

[0117] By introducing a nonlinear decay strategy formula, the particle swarm optimization algorithm (PSO) algorithm uses a larger inertia weight in the early stages of the iteration, facilitating global search and preventing premature convergence. Later in the iteration, the inertia weight gradually decreases, facilitating local search and improving the algorithm's convergence speed and accuracy. The nonlinear decay strategy allows the algorithm to smoothly transition between global and local optimization, avoiding the "oscillation" or "stagnation" that can occur during the iterations of traditional PSO algorithms. This improves the algorithm's performance, enabling it to more quickly find optimal or near-optimal solutions. By adjusting the parameters in the nonlinear decay strategy formula, the algorithm can adapt to optimization problems of varying scale and complexity.

[0118] Furthermore, the dynamic learning mechanism satisfies

[0119]

[0120] in, represents the velocity vector of the lth particle at the t0+1th iteration, represents the velocity vector of the particle at the t0th iteration, represents the position vector of the lth particle at the t0th iteration, represents the individual optimal position found by the lth particle in the historical iteration, G best represents the global optimal position found by the entire particle swarm in historical iterations, r1 and r2 represent uniformly distributed random vectors, c1(t0) represents the dynamically adjusted individual learning factor, c2(t0) represents the dynamically adjusted social learning factor, λ1 represents the perturbation intensity coefficient, N(0,∑) represents a Gaussian distributed random vector with mean 0 and covariance matrix ∑, λ1N(0,∑) represents the Gaussian perturbation term, and represents the basic learning factor, σ(t0) represents the standard deviation of the particle swarm distribution, σ max represents the preset maximum standard deviation threshold, and k0 represents the adjustment gain.

[0121] In this embodiment, in the particle swarm optimization (PSO) algorithm, the particle velocity and position update formula is the core of the algorithm. This embodiment aims to propose an improved PSO algorithm, in which the particle velocity update formula not only considers the influence of the current velocity, individual optimal position, and global optimal position, but also introduces a dynamically adjusted individual learning factor, social learning factor, and Gaussian perturbation term to enhance the algorithm's search capability and avoid premature convergence.

[0122] By dynamically adjusting individual and social learning factors, the algorithm balances global and local search capabilities during iteration, avoiding premature convergence. The introduction of a Gaussian perturbation term increases the algorithm's randomness and diversity, helping it escape local optima and explore a wider search space. Dynamically adjusting the standard deviation ensures that the Gaussian perturbation term has a greater impact in the early stages of the iteration, facilitating rapid exploration of the search space. Its influence gradually diminishes in the later stages of the iteration, facilitating refined search and convergence. The improved PSO algorithm maintains the concise mathematical expression and easily understood physical meaning of the original algorithm, making it easier to implement and expand.

[0123] Specifically, the velocity vector determines the direction and speed of a particle's movement in the search space. Each particle has a velocity vector, which is updated based on its current position, its individual optimal position, and its global optimal position. By adjusting the velocity vector, particles can efficiently explore new areas in the search space while maintaining a convergence trend toward the optimal solution.

[0124] The position vector represents the particle's current position in the search space. The particle's position is updated based on the velocity vector, continuously approaching the optimal solution. This updating of the position vector allows the particle to gradually converge to the optimal solution while maintaining a certain level of diversity and avoiding premature convergence.

[0125] The individual optimal position records the best position a particle has found in its previous iterations. Each particle has an individual optimal position, which guides its movement. By comparing its current position with the individual optimal position, the particle can determine whether its movement direction is correct, thereby adjusting its velocity vector and accelerating convergence to the optimal solution.

[0126] The global optimal position records the best position found by the entire particle swarm in previous iterations. All particles share this global optimal position, which guides the movement of the entire particle swarm. The global optimal position provides a common goal for the particle swarm, enabling them to work together and accelerate convergence to the global optimal solution.

[0127] Uniformly distributed random vectors are used to increase the randomness of particle motion and prevent particles from becoming trapped in local optima. These random vectors, multiplied by the individual learning factor and the social learning factor, collectively determine the particle's velocity update. By introducing random vectors, particles maintain a certain level of exploration within the search space, preventing premature convergence and improving the robustness of the algorithm.

[0128] The individual learning factor determines the degree to which a particle learns to its optimal position. By dynamically adjusting the individual learning factor, we can balance the particle's local and global search capabilities. This dynamically adjusted individual learning factor enables particles to have strong global search capabilities in the early stages of the search, while gradually shifting to local search in the later stages, improving the algorithm's convergence speed and accuracy.

[0129] The social learning factor determines the degree to which particles learn to move toward the global optimal position. By dynamically adjusting the social learning factor, we can coordinate the cooperation and competition between particles. This dynamically adjusted social learning factor allows the particle swarm to maintain a certain level of diversity during the search process while accelerating convergence to the global optimal solution.

[0130] The perturbation intensity coefficient determines the strength of the Gaussian perturbation term. By introducing the Gaussian perturbation term, we can increase the randomness of particle motion and improve the algorithm's search capabilities. The Gaussian perturbation term allows particles to maintain a certain level of exploration in the search space, avoiding being trapped in local optimal solutions, thereby improving the algorithm's robustness and global search capabilities.

[0131] A Gaussian distributed random vector is used to generate a Gaussian perturbation term. This random vector, multiplied by the perturbation intensity coefficient, determines the magnitude and direction of the Gaussian perturbation term. By introducing a Gaussian distributed random vector, particles are subjected to a certain amount of random perturbation during the search process, increasing the algorithm's search power and diversity.

[0132] The basic learning factor is the base value for dynamically adjusting the individual learning factor and the social learning factor. These base values ​​determine the basic learning behavior of particles during the search process. By setting a reasonable basic learning factor, we can balance the local and global search capabilities of particles, improving the convergence speed and accuracy of the algorithm.

[0133] The particle swarm distribution standard deviation reflects the distribution of the particle swarm in the search space. By calculating the particle swarm's position standard deviation, we can assess the diversity of the particle swarm. The particle swarm distribution standard deviation is used to monitor the diversity of the particle swarm. A small particle swarm distribution standard deviation indicates that the particle swarm may be trapped in a local optimal solution or premature convergence. A large particle swarm distribution standard deviation indicates that the particle swarm has good diversity and can continue to explore new search areas.

[0134] The preset maximum standard deviation threshold is used to determine whether the particle swarm maintains sufficient diversity. When the standard deviation of the particle swarm distribution is less than this threshold, measures may need to be taken to increase the diversity of the particle swarm. By setting the preset maximum standard deviation threshold, it is possible to promptly identify local optimal solutions or premature convergence problems that the particle swarm may fall into, and take appropriate measures to intervene, thereby improving the robustness and global search capabilities of the algorithm.

[0135] The tuning gain adjusts the coefficient in the speed update formula to balance the particle's exploration and exploitation capabilities. Adjusting the tuning gain influences the speed and direction of particle movement during the search process. This allows the algorithm to have strong exploration capabilities in the early stages of the search while gradually shifting towards exploitation capabilities in the later stages, improving its convergence speed and accuracy. Furthermore, by dynamically adjusting the tuning gain, the algorithm can be adaptively adjusted based on actual conditions during the search, enhancing its robustness and flexibility.

[0136] Furthermore, the control gain is Where ΔK p represents the control gain, η p represents the learning rate of the proportional gain, e(t) represents the deviation between the system output and the set value, sign(e(t)) represents the sign function of the error, e max Indicates the maximum allowable deviation between the system output and the set value, min(|e(t)|,e max ) represents the cutoff value of the deviation, σ e represents the deviation attenuation coefficient, An exponential decay term representing the bias.

[0137] In this embodiment, the control gain determines the control system's response to deviations. When the system output deviates from the setpoint, the control gain determines the magnitude of the adjustment, thus affecting the system's convergence speed and stability. By adjusting the control gain, the system's sensitivity to deviations can be controlled, thereby affecting the system's response speed and stability. A larger control gain can speed up the system's response but may cause system oscillations; a smaller control gain can improve system stability but may slow down the response.

[0138] The learning rate of the proportional gain determines how quickly the control gain adjusts to the deviation. When the deviation is large, a system with a higher learning rate can adjust the control gain more quickly, thereby reducing the deviation more quickly. By adjusting the learning rate, you can control how quickly the system adapts to changes in the deviation. A higher learning rate can speed up the system's adaptation process but may cause system instability; a lower learning rate can improve system stability but may slow down the adaptation process.

[0139] The maximum allowable deviation limits the range of control gain growth, preventing excessive increases in control gain when the deviation is too large, which could lead to system instability. By setting the maximum allowable deviation, you can protect the system from the impact of excessive deviations and improve system robustness.

[0140] The deviation cutoff value is used to limit the influence of the deviation on the control gain, preventing the control gain from increasing abnormally when the deviation is too large. By setting the deviation cutoff value, the growth of the control gain can be further controlled, improving the stability and robustness of the system.

[0141] The deviation attenuation coefficient determines how quickly the deviation's effect on the control gain decreases. When the deviation is small, a system with a larger attenuation coefficient can reduce the control gain more quickly, thus preventing over-adjustment. By adjusting the attenuation coefficient, you can control how quickly the system adjusts the control gain when the deviation is small, improving system stability and accuracy.

[0142] The exponential decay term of the deviation utilizes the decay characteristics of the exponential function to smooth the deviation and prevent excessive fluctuations in the control gain when the deviation is small. By introducing the exponential decay term, the change in the control gain can be further smoothed, improving the stability and smoothness of the system.

[0143] Furthermore, the integration time is Where ΔT i represents the integration time, η i represents the learning rate of the integration time, represents the integration time window, represents the rate of change of the deviation, Represents the integral term of the deviation.

[0144] In this embodiment, the integral time determines the influence of the deviation integral term on the control output. The longer the integral time, the greater the influence of the integral term on the control output and the smaller the system's steady-state error. By adjusting the integral time, the system's sensitivity to the deviation integral term can be controlled, thereby affecting the system's steady-state error and stability.

[0145] The learning rate of the integration time determines how quickly the integration time adjusts to the rate of change of the deviation. When the rate of change of the deviation is large, a system with a larger learning rate can adjust the integration time more quickly, thereby reducing the steady-state error more quickly. By adjusting the learning rate, you can control how quickly the system adapts to the rate of change of the deviation, improving the system's adaptive capabilities and the speed of convergence of the steady-state error.

[0146] The integral time window defines the time range for the integral term calculation. The longer the integral time window, the more historical deviation information the integral term contains, and the stronger its compensation for the system's steady-state error. By setting the integral time window, you can control the degree to which the integral term compensates for the system's steady-state error, thereby improving the system's steady-state performance.

[0147] Furthermore, the differential time is Where ΔT d represents the differential time, η d represents the learning rate of the differentiation time, sat(·) represents the saturation function, It represents the acceleration of the deviation change, and ∈1 represents the zero-proof constant.

[0148] In this embodiment, the derivative time determines the degree to which the acceleration of the deviation change affects the control output. The longer the derivative time, the greater the influence of the derivative term on the control output, and the better the system's dynamic performance. By adjusting the derivative time, the system's sensitivity to the acceleration of the deviation change can be controlled, thereby affecting the system's dynamic performance and stability.

[0149] The learning rate of the derivative time determines how quickly the derivative time adjusts to the acceleration of the deviation change. When the deviation changes rapidly, a system with a higher learning rate can adjust the derivative time more quickly, thereby improving the system's dynamic performance more quickly. By adjusting the learning rate, you can control how quickly the system adapts to the acceleration of the deviation change, improving the system's adaptive capabilities and dynamic performance.

[0150] The saturation function is used to limit the output range of the differential term, preventing it from growing abnormally when the deviation acceleration is too large, which could lead to system instability. By introducing the saturation function, the system can be protected from the impact of excessive deviation acceleration and improve system robustness.

[0151] The zero division prevention constant is used to prevent the differential term from dividing by zero when the deviation acceleration is zero. By setting a non-zero constant, you can ensure that the differential term still has a small output value when the deviation acceleration is zero. By setting the zero division prevention constant, you can avoid zero division errors in the differential term when the deviation acceleration is zero, thereby improving system reliability and stability.

[0152] Reference Figure 2 An embodiment of the present invention provides a dynamic learning system 2 for an intelligent controller, wherein the system 2 specifically includes:

[0153] A first dynamic learning module 201 is configured to establish a system dynamic characteristic model for the intelligent controller using a dynamic modeling method based on real-time industrial process data, wherein the real-time industrial process data includes operating condition change trends, system nonlinear characteristics, and time-varying parameter information;

[0154] A second dynamic learning module 202 is configured to set a parameter optimization objective function for the intelligent controller based on the system dynamic characteristic model, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility, and real-time performance indicators;

[0155] The third dynamic learning module 203 is used to use the particle swarm optimization algorithm to perform global optimization on the parameter optimization objective function, and dynamically adjust the particle speed and position during the optimization process through a dynamic learning mechanism to obtain an optimization result;

[0156] A fourth dynamic learning module 204 is configured to dynamically adjust control parameters of the intelligent controller based on the optimization result, wherein the control parameters include control gain, integral time, and differential time;

[0157] The fifth dynamic learning module 205 is used to adopt a global search strategy to jump out of the local optimum if it is detected that the diversity of the particle swarm decreases or the objective function value does not improve for a long time during the optimization process.

[0158] It is understandable that if Figure 1 The contents of the embodiment of the dynamic learning method of the intelligent controller shown in FIG. 1 are applicable to the embodiment of the dynamic learning system of the intelligent controller. The functions specifically implemented by the embodiment of the dynamic learning system of the intelligent controller are similar to those in FIG. Figure 1 The embodiment of the dynamic learning method of the intelligent controller shown is the same as that of Figure 1 The beneficial effects achieved by the embodiment of the dynamic learning method of the intelligent controller shown are also the same.

[0159] It should be noted that the information interaction, execution process, etc. between the above-mentioned systems are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0160] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0161] Reference Figure 3 An embodiment of the present invention further provides a computer device 3, comprising: a memory 302, a processor 301, and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, a dynamic learning method of an intelligent controller as described in any one of the above methods is implemented.

[0162] The computer device 3 may be a desktop computer, a notebook computer, a PDA, a cloud server or other computing devices. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that Figure 3 This is merely an example of the computer device 3 and does not constitute a limitation on the computer device 3. The computer device 3 may include more or fewer components than shown in the figure, or a combination of certain components, or different components. For example, it may also include input and output devices, network access devices, etc.

[0163] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. A general-purpose processor may be a microprocessor or any conventional processor.

[0164] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. equipped on the computer device 3. Furthermore, the memory 302 may include both an internal storage unit of the computer device 3 and an external storage device. The memory 302 is used to store an operating system, application programs, a boot loader, data, and other programs, such as the program code of the computer program. The memory 302 may also be used to temporarily store data that has been output or is about to be output.

[0165] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the dynamic learning method of the intelligent controller as described in any one of the above methods is implemented.

[0166] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the process of the above-mentioned method embodiment by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of each of the above-mentioned method embodiments. The computer program includes computer program code, which can be in source code form, object code form, executable file, or some intermediate form. The computer-readable medium can at least include: any entity or device capable of carrying computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal, and software distribution medium. For example, a USB flash drive, mobile hard drive, magnetic disk, or optical disk. In some jurisdictions, based on legislation and patent practice, computer-readable media cannot be electric carrier signals or telecommunication signals.

[0167] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0168] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0169] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely schematic. For example, the division of the modules or units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0170] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A dynamic learning method for an intelligent controller, characterized in that: The method specifically includes: Based on the real-time data of industrial processes, a system dynamic characteristic model is established for the intelligent controller using a dynamic modeling method. The real-time data of industrial processes includes the trend of operating condition changes, nonlinear characteristics of the system, and time-varying parameter information. Based on the system dynamic characteristic model, setting a parameter optimization objective function for the intelligent controller, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility and real-time performance indicators; A particle swarm optimization algorithm is used to globally optimize the parameter optimization objective function, and the particle speed and position are dynamically adjusted during the optimization process through a dynamic learning mechanism to obtain the optimization result; Based on the optimization result, dynamically adjust the control parameters of the intelligent controller, wherein the control parameters include control gain, integral time and differential time; During the optimization process, if it is detected that the diversity of the particle swarm decreases or the objective function value does not improve for a long time, a global search strategy is used to jump out of the local optimum.

2. The method according to claim 1, characterized in that The method of using a dynamic model to build a system dynamic characteristic model for the intelligent controller based on real-time industrial process data specifically includes: Acquire real-time industrial process data, and segment the real-time industrial process data using a sliding window method to form standard data; Extracting working condition characteristics from the standard data using a working condition characteristic extraction formula to generate a working condition change trend indicator; Using a time-varying parameter identification formula to perform time-varying parameter identification on the standard data to generate a time-varying parameter vector; Using a deep feature network to extract nonlinear features from the standard data to generate system nonlinear features; A system dynamic characteristic model is formed according to the operating condition change trend index, the time-varying parameter vector and the system nonlinear characteristics.

3. The method according to claim 2, characterized in that The use of the sliding window method to segment the industrial process real-time data to form standard data specifically includes: The real-time data stream of the industrial process is segmented into sliding windows by a preset window length and a preset sliding step size to form a continuous data window sequence; Perform wavelet threshold denoising on each data window in the data window sequence to obtain preprocessed data; Based on the preprocessed data, the mean and standard deviation in each data window are calculated, and Z-score normalization is performed to generate standard data.

4. The method according to claim 3, characterized in that The operating condition feature extraction formula is used to extract the operating condition features of the standard data to generate an operating condition change trend indicator, specifically including: The window mean, linear slope and variance of each data window are extracted to form time domain features; Based on the time domain characteristics, the time domain trend index is generated through the weighted fusion formula; Perform fast Fourier transform on the standard data to divide it into low-frequency band and high-frequency band, calculate the ratio of high-frequency energy to low-frequency energy, and obtain the frequency domain trend index; The time domain trend index and the frequency domain trend index are logarithmically weighted and fused to generate the working condition change trend index.

5. The method according to claim 2, characterized in that The time-varying parameter identification formula is used to identify the time-varying parameters of the standard data to generate a time-varying parameter vector, specifically including: Extract regression vectors and observation outputs from standard data to build a dynamic model; Based on the dynamic model, the Kalman gain and covariance matrix are updated by the recursive least squares algorithm with forgetting factor to calculate the time-varying parameter estimates; The covariance matrix is ​​compressed and the time-varying parameter estimates are smoothed using the exponential weighted averaging method to generate the target time-varying parameter vector.

6. The method according to claim 2, characterized in that The use of a deep feature network to extract nonlinear features from the standard data to generate system nonlinear features specifically includes: Input the standard data into the deep feature network and generate hidden features through layer-by-layer nonlinear transformation; Perform multi-scale one-dimensional convolution operations on hidden features to extract local time domain patterns, and then concatenate the hidden features and local time domain patterns to generate fused features; The fusion features are subjected to principal component analysis and dimensionality reduction to generate the nonlinear characteristics of the system.

7. The method according to claim 1, characterized in that The particle swarm optimization algorithm is used to perform global optimization on the parameter optimization objective function, specifically including: Setting initial parameters of the particle swarm, calculating the current fitness value of the parameter optimization objective function using a particle swarm optimization algorithm based on the initial parameters, and recording the optimal solution; The nonlinear attenuation strategy formula is used to make the particle swarm optimization algorithm transition smoothly between global optimization and local optimization.

8. The method according to claim 7, characterized in that The setting of initial parameters of the particle swarm, calculating the current fitness value of the parameter optimization objective function using a particle swarm optimization algorithm based on the initial parameters, and recording the optimal solution specifically includes: Determine the size of the particle swarm, randomly generate the initial position and initial velocity for each particle in the particle swarm, set the maximum and minimum values ​​of the inertia weight, the decay exponent, and the maximum number of iterations; For each particle, its position is substituted into the parameter optimization objective function to calculate the fitness value; Based on the fitness value, record the global optimal solution and the corresponding position, as well as the individual optimal solution and the corresponding position of each particle; According to the velocity update formula of the PSO algorithm, the new velocity of each particle is calculated and the position of the particle is updated according to the new velocity.

9. The method according to claim 8, characterized in that The nonlinear attenuation strategy formula is used to make the particle swarm optimization algorithm smoothly transition between global optimization and local optimization, specifically including: In each iteration of the particle swarm optimization process, the current inertia weight is calculated using the nonlinear decay strategy formula according to the current number of iterations and the maximum number of iterations; Use the current inertia weight in the velocity update formula for the next iteration; In each iteration, the global optimal solution and the individual optimal solution are updated; After the iteration is completed, the global optimal solution and the corresponding position are output.

10. A dynamic learning system for an intelligent controller, characterized in that: The system specifically includes: A first dynamic learning module is configured to establish a system dynamic characteristic model for the intelligent controller using a dynamic modeling method based on real-time industrial process data, wherein the real-time industrial process data includes operating condition change trends, system nonlinear characteristics, and time-varying parameter information; A second dynamic learning module is used to set a parameter optimization objective function for the intelligent controller based on the system dynamic characteristic model, wherein the parameter optimization objective function is used to adjust the balance between control accuracy, strategy flexibility and real-time performance indicators; The third dynamic learning module is used to use the particle swarm optimization algorithm to perform global optimization on the parameter optimization objective function, and dynamically adjust the particle speed and position during the optimization process through the dynamic learning mechanism to obtain the optimization result; A fourth dynamic learning module, configured to dynamically adjust control parameters of the intelligent controller based on the optimization result, the control parameters including control gain, integral time, and differential time; The fifth dynamic learning module is used to use a global search strategy to jump out of the local optimum if it is detected that the diversity of the particle swarm is reduced or the objective function value has not improved for a long time during the optimization process.

Images