Intelligent industrial control parameter self-tuning method and equipment

Abstract

The invention relates to an intelligent industrial control parameter self-tuning method and equipment. The method comprises the following steps: performing piecewise linearization processing on an actuator input and output characteristic curve in the industrial control system, constructing an upper-layer agent model and a lower-layer agent model, and combining to obtain a system dynamic characteristic prediction model; performing state observer design and gain matrix optimization on the industrial control system based on the system dynamic characteristic prediction model to obtain a system state observation vector and an interference compensation vector; inputting the system state observation vector and the interference compensation vector into a fixed time convergence controller for control parameter calculation to obtain a system control input matrix; and performing adaptive gradient descent operation and momentum item optimization on the system control input matrix, and outputting an optimal control parameter matrix in combination with the system dynamic characteristic prediction model. According to the invention, real-time automatic adjustment of the control parameters is realized, the steady-state error of the system is effectively reduced, and the control performance of the control system under different working conditions is ensured.

Description

Technical Field

[0001] The present invention relates to the field of intelligent industrial control technology, and in particular to an intelligent industrial control parameter self-tuning method and device. Background Art

[0002] The parameter setting of traditional industrial control systems usually adopts empirical debugging or fixed parameter calculation methods, which leads to large steady-state errors and dynamic response deviations during the operation of the system. In actual industrial production, due to various external interferences, system nonlinear characteristics and parameter uncertainties in the industrial process, it is difficult for traditional parameter setting methods to meet the requirements of high-precision control. In addition, in the industrial process, the nonlinear characteristics of the actuator and the time-varying dynamic characteristics of the system also bring great challenges to parameter setting.

[0003] Faced with complex and ever-changing industrial environments, traditional parameter tuning methods such as PID control lack adaptive and intelligent learning capabilities and cannot accurately capture the dynamic characteristics of the system. At the same time, due to the various uncertain factors in the industrial process, such as equipment aging, environmental changes, and load disturbances, it is difficult for fixed parameter control strategies to ensure the stability and robustness of the control system. Summary of the invention

[0004] The main purpose of the present invention is to provide an intelligent industrial control parameter self-tuning method and device, which realizes real-time automatic adjustment of control parameters, effectively reduces the system steady-state error, and ensures the control performance of the control system under different working conditions.

[0005] To achieve the above object, the present invention provides a method for self-tuning intelligent industrial control parameters, comprising the following steps: The input-output characteristic curve of the actuator in the industrial control system is processed piecewise linearly, and the upper-layer agent model and the lower-layer agent model are constructed to obtain the system dynamic characteristic prediction model; Based on the system dynamic characteristic prediction model, the state observer design and gain matrix optimization are performed on the industrial control system to obtain a system state observation vector and a disturbance compensation vector; Inputting the system state observation vector and the disturbance compensation vector into a fixed time convergence controller to calculate control parameters, thereby obtaining a system control input matrix; An adaptive gradient descent operation and momentum term optimization are performed on the system control input matrix, and an optimal control parameter matrix is ​​output in combination with the system dynamic characteristic prediction model.

[0006] The present invention also provides an intelligent industrial control parameter self-tuning device, comprising: A construction module is used to perform piecewise linearization processing on the input-output characteristic curve of the actuator in the industrial control system, and to construct an upper-layer proxy model and a lower-layer proxy model, and to combine them to obtain a system dynamic characteristic prediction model; An optimization module, used for designing a state observer and optimizing a gain matrix of the industrial control system based on the system dynamic characteristic prediction model, to obtain a system state observation vector and a disturbance compensation vector; A calculation module, used for inputting the system state observation vector and the interference compensation vector into a fixed time convergence controller to calculate control parameters and obtain a system control input matrix; The output module is used to perform adaptive gradient descent operation and momentum term optimization on the system control input matrix, and output the optimal control parameter matrix in combination with the system dynamic characteristic prediction model.

[0007] In summary, the technical solution provided by the present invention combines the nonlinear mapping ability of deep neural networks with the fast response characteristics of local linear models by constructing a two-layer proxy model structure, thereby significantly improving the prediction accuracy and computational efficiency of the system dynamic characteristics; adopts a fixed-time convergence controller design method to ensure that the system state converges to the equilibrium point within a finite time, overcoming the problem of slow convergence speed of traditional asymptotic convergence methods; introduces adaptive gradient descent and momentum term optimization algorithms to realize real-time automatic adjustment of control parameters, effectively reducing the system steady-state error; designs intelligent observers and interference compensation mechanisms to enhance the system's anti-interference ability and robustness; and through multi-objective optimization and online parameter adjustment strategies, ensures the control performance of the control system under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] Figure 1 It is a schematic diagram of the steps of a method for self-tuning intelligent industrial control parameters in one embodiment of the present invention; Figure 2 It is a structural block diagram of an intelligent industrial control parameter self-tuning device in one embodiment of the present invention.

[0009] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0010] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0011] Reference Figure 1 , this embodiment provides a method for self-tuning intelligent industrial control parameters, comprising the following steps: S1, perform piecewise linearization on the input-output characteristic curve of the actuator in the industrial control system, and construct an upper-layer agent model and a lower-layer agent model to combine and obtain a system dynamic characteristic prediction model; Among them, for the input-output characteristic curve of the actuator in the industrial control system, by analyzing the change trend of the curve, the slope mutation point detection algorithm is applied to identify the key change points of the curve. These mutation points reflect the turning areas of different dynamic characteristics in the characteristic curve. By detecting the significant change of the slope, the curve is effectively divided into several intervals with different characteristics, and each interval is approximately linear within a certain range. Linear regression analysis is performed on each data segment in the linearized sub-interval sequence, and the data in the interval is fitted to obtain a set of regression model parameters describing the linear relationship. At the same time, the input-output relationship in the interval is modeled as the transfer function form of the dynamic system using the transfer function fitting technology, and the parameter set of the piecewise linearization model is obtained, which includes the key dynamic characteristic parameters in each interval, such as gain, time constant and inertia term. According to the piecewise linearization model parameter set, the dynamic equation of the system is established to describe the dynamic characteristics of the entire system, which includes the combination of multiple linear sub-models and the connection relationship between them. After the dynamic equation is established, the statistical characteristics of the system noise and measurement noise are analyzed. By analyzing the noise distribution and correlation of the system operation data, the influence of noise on the dynamic characteristics of the system is determined. Based on these analysis results, the state space basic equations of the system are derived. These equations formally include state equations and observation equations, which describe the dynamic behavior and measurement behavior of the system respectively. On the basis of the state space basic equations, the system state matrix, control input matrix and output observation matrix are constructed to form a state space model. At the same time, the performance indicators such as steady-state deviation, dynamic overshoot and adjustment time of the industrial control system operation data are introduced into the multi-objective optimizer. By weighting these performance indicators, a comprehensive performance indicator function is generated. Using the state space model and performance indicator function, an agent model that adapts to complex industrial control scenarios is constructed. The deep neural network structure is used as the upper agent model to capture the global dynamic characteristics; at the same time, combined with the combination of local linear models, it is used as the lower agent model to describe the local characteristics of each linearized subinterval. The collaborative mode of the upper and lower models can effectively balance the modeling requirements between the global dynamic characteristics and the local fine characteristics, and form a system dynamic characteristics prediction model.

[0012] The system state vector, control parameter vector and external disturbance vector in the state space model are input into the deep neural network for feature extraction. Through its multi-layer structure, the deep neural network extracts complex features from the input data layer by layer, and gradually optimizes the weight parameters of each layer during the network training process to form an upper-level proxy model that can highly summarize the dynamic characteristics of the system. The core of this network structure lies in its deep learning ability, which can capture the potential nonlinear relationship and dynamic change trend in the state space model. At the same time, the working range of the industrial control system is reasonably divided into multiple working sub-intervals with relatively independent characteristics. For each sub-interval, a separate mathematical model is established using the local linear modeling method. These models can effectively reflect the linear dynamic characteristics within the sub-interval and form the basic model group of the lower-level proxy model. Each basic model obtains parameters through techniques such as linear regression or transfer function fitting to characterize the dynamic characteristics within a specific working range. In the process of building the lower-level proxy model, the output results of all basic models are weighted fused to obtain a comprehensive output result. The weighted fusion method is dynamically adjusted according to the applicable scope and accuracy of each basic model to ensure that the comprehensive output of the lower-level proxy model can fully reflect the characteristics of the system in different working intervals. The fusion process uses the optimal weight distribution algorithm to achieve automatic weight adjustment, making the output of the lower-level proxy model more adaptable. The network structure of the upper-level proxy model is trained through back propagation, and the loss value is calculated by introducing the performance indicator function, and the parameters of the deep neural network are updated based on the loss value. The performance indicator function is a comprehensive evaluation standard that combines multiple key performance indicators such as system steady-state error, dynamic overshoot, and adjustment time, and quantifies the deviation of system performance from the expected target in a weighted manner. During the back propagation training process, the neural network continuously adjusts the weight and bias parameters through the gradient descent algorithm, so that the loss value gradually decreases, thereby optimizing the network performance and obtaining the trained upper-level proxy model. The output results of the trained upper-level proxy model are combined with the comprehensive output results of the lower-level proxy model to generate the system state prediction value. The combined operation process adopts a fusion strategy to coordinately integrate the prediction of the global dynamic characteristics of the upper-level proxy model with the detailed description of the local characteristics of the lower-level proxy model, thereby improving the accuracy and reliability of the system state prediction. According to the deviation between the system state prediction value and the actual system state, the parameters of the upper-level proxy model and the lower-level proxy model are adjusted online. The online learning algorithm is used to update the model parameters according to the real-time feedback data, so that it can quickly adapt to the changes in the system operating environment. In this way, the upper and lower agent models gradually approach the real dynamic characteristics of the system in the continuous adjustment and optimization, and obtain the system dynamic characteristics prediction model.

[0013] The operation data of the industrial control system is clustered according to the distribution of the working points to identify the target working point set within the working range. By collecting the input and output data of the industrial control system under different operating conditions, cluster analysis is performed using the distribution characteristics of the working points to divide the working range of the system. Reasonable characteristic variables, such as the input signal amplitude, the output response amplitude, and the dynamic characteristic parameters of the system, are selected to reflect the operating state of the system. The goal of cluster analysis is to extract a set of target working points that can represent different dynamic characteristic regions. The extracted target working point set is input into the K-means clustering algorithm for automatic partition calculation. The K-means algorithm iteratively optimizes the position of the cluster center and divides the entire working range into m working sub-intervals according to the principle of minimizing the distance from the target point to the cluster center. Each sub-interval represents a working area with relatively consistent dynamic characteristics. After completing the division of the working sub-intervals, a corresponding dynamic model is established for each sub-interval to accurately describe its inherent operating characteristics. By analyzing the system operation law in the sub-interval, the state equation and output equation are constructed respectively to obtain the dynamic model matrix of each working sub-interval. These dynamic model matrices contain key characteristic parameters of the system, such as state transfer matrix, control input matrix and output matrix, which reflect the state change law and input-output relationship in the sub-interval. In the process of constructing the dynamic model matrix, physical laws, experimental data fitting and mathematical optimization methods are introduced to ensure the accuracy and stability of the model. In order to ensure the continuity of each sub-interval model at the boundary, the boundary condition constraint calculation is performed on the dynamic model matrix. It is ensured that the transition of state variables and output variables is smooth and without obvious jump when switching sub-intervals. Through the continuity constraint calculation, the model parameters near the sub-interval boundary are adjusted so that the state transfer matrix meets the continuity requirements in the entire working range. This processing method can improve the overall consistency of the model and effectively avoid the control instability caused by boundary jump. By performing zero-pole analysis on the state transfer matrix, the dynamic characteristics of the system, such as natural frequency, damping ratio and stability boundary, are obtained. The results of zero-pole analysis help determine whether each working sub-interval is in a stable state and whether its dynamic response meets expectations. If the stability of some sub-interval models is poor, their parameters need to be further adjusted to ensure that all sub-interval models can maintain stable operation within the effective working range. Based on the stability analysis results, the effective working range of each sub-interval is determined, and the parameters of the local linear model are optimized. After the parameter optimization is completed, the local linear models of all working sub-intervals are fused to form the basic model group of the lower-level proxy model.

[0014] S2, based on the system dynamic characteristics prediction model, the state observer design and gain matrix optimization of the industrial control system are carried out to obtain the system state observation vector and the interference compensation vector; Specifically, the state prediction data generated by the system dynamic characteristics prediction model and the actual state data of the industrial control system are used. By inputting these two types of data into the observer design unit, the basic structure model of the observer is established. The basic structure model is used to combine the system dynamic characteristics prediction model and the feedback of the actual system state to provide a high-precision estimate of the unmeasured system state. In the observer design process, the basic model needs to accurately reflect the dynamic behavior of the system and the coupling relationship between the states. The state error in the basic structure model is processed. Through the weighted least squares operation, the sensitivity of the observer state error to external interference is effectively reduced, and the observer error weight is dynamically optimized during the model parameter adjustment process. Through the numerical optimization method, the deviation between the observer estimate and the actual system state is minimized. During the operation, the weighted sum of the state error is calculated according to the principle of least squares method, and then the feedback gain matrix of the observer is adjusted. After the optimization of the feedback gain matrix is ​​completed, the observer state equation is iteratively solved and the parameters are corrected. Through iterative solution, it is ensured that the observer state estimation can converge to the true state quickly, and the key parameters in the equation are corrected in multiple iterations to make it more suitable for the operating characteristics of the dynamic system. The result of the iterative process will be output in the form of a system state observation vector, which is a high-precision estimate of the current state of the system. At the same time, the deviation between the output of the system and the output of the system dynamic characteristics prediction model is analyzed for interference characteristics. By identifying the dynamic change law of the deviation, the characteristic parameters of the interference, such as amplitude, frequency and change trend, are extracted. Based on these characteristic parameters, a dynamic model of interference is established, which is used to describe the propagation characteristics of interference in the system and its impact on the system state. The interference dynamic model is combined with the system state observation vector for calculation, and the impact of interference is analyzed in the frequency domain. The interference is converted from the time domain to the frequency domain through Fourier transform and other technologies, so as to analyze the main frequency band distribution of the interference and its response characteristics to the system. Based on the frequency domain analysis, an interference compensation function is constructed, which describes in mathematical form how to eliminate the impact of interference on the system through control input. In order to ensure the real-time and accuracy of the interference compensation function, numerical calculation technology is used to solve the compensation function in real time, and it is dynamically adjusted according to the changes of the current system state and the interference dynamic model. The interference compensation function calculated in real time is applied to the feedback control link of the system to generate an interference compensation vector. The disturbance compensation vector works synergistically with the system control input to significantly reduce the negative impact of disturbance on system performance and improve the stability and dynamic response performance of the system.

[0015] S3, inputting the system state observation vector and the disturbance compensation vector into the fixed time convergence controller to calculate the control parameters and obtain the system control input matrix; It should be noted that by analyzing the deviation between the system state and the target state and designing the feedback gain matrix, the state feedback gain of the system state observation vector is calculated. The design of the feedback gain matrix needs to be combined with the dynamic characteristics of the system and the control target to ensure that the feedback can quickly and accurately guide the system state to converge toward the expected value. At the same time, in order to improve the stability and adaptability of the calculation, the state observation vector is parameter normalized to eliminate the influence between different dimensions, and the state feedback control vector is obtained. The interference compensation vector is input into the compensation control unit to complete the compensation gain optimization and noise suppression. In the compensation control unit, the interference compensation vector is feature analyzed to identify its mode of action and influence intensity in the system dynamics. Based on the analysis results, the compensation gain is optimized so that it can offset the interference to the greatest extent with the minimum control cost. At the same time, for the high-frequency noise in the interference signal, the compensation unit introduces a noise suppression mechanism, and effectively reduces the influence of noise on the control output through filtering and frequency domain adjustment. After the compensation gain optimization and noise suppression are completed, the interference compensation control vector is obtained. The fixed time convergence term design and coefficient optimization calculation are performed on the state observation vector. The core of fixed-time convergence control is to ensure that the system state can reach the expected target state within a fixed time regardless of the initial conditions. This control characteristic can significantly improve the dynamic response speed and control accuracy of the system. By combining the dynamic characteristics in the state observation vector, the fixed-time convergence term is designed, and the coefficient optimization calculation is performed for different working conditions to ensure the balance between convergence speed and control accuracy. The state feedback control vector and the system state observation vector are matrix multiplied to generate the feedback control component. The feedback control component reflects the control instruction part of the current system state deviation after the feedback gain correction, which is directly used to adjust the dynamic behavior of the system. Similarly, the disturbance compensation control vector and the disturbance compensation vector are feedback compensated to generate the compensation control component. The design of the compensation control component aims to compensate the deviation of the system caused by external interference in real time, thereby improving the robustness and anti-interference ability of the system. The feedback control component, the compensation control component and the optimized convergence control parameters are combined by control law to generate the control input matrix of the system. These three control components are combined by weight adjustment and logic to form a comprehensive control strategy, so that the control input matrix can balance the effects of feedback control, disturbance compensation and fixed-time convergence.

[0016] S4, performs adaptive gradient descent operation and momentum term optimization on the system control input matrix, and outputs the optimal control parameter matrix in combination with the system dynamic characteristics prediction model.

[0017] Specifically, the system control input matrix is ​​combined with the output value of the system dynamic characteristic prediction model, and the control effect of the system is quantified through the performance evaluation function to obtain the performance index function value. The performance index function comprehensively considers multiple control performance indicators such as steady-state error, dynamic overshoot, and adjustment time, thereby reflecting the influence of control parameters on the overall performance of the system. The derivative of the performance index function value is mapped to the parameter space to generate a gradient vector. The gradient vector describes the change trend of the performance index function in the parameter space, and its direction points to the fastest path for performance improvement. By calculating the components of the gradient vector, a performance parameter gradient matrix is ​​generated to characterize the degree of influence of each control parameter on the performance index. On this basis, the performance parameter gradient matrix and the momentum term of the previous moment are weighted and accumulated to obtain the momentum update vector. The introduction of the momentum term can alleviate the oscillation phenomenon in the gradient descent, smooth the optimization path, and accelerate the convergence speed. The generation of the momentum update vector realizes the comprehensive utilization of historical gradient information by adjusting the momentum coefficient, making the optimization process more stable. The momentum update vector and the performance index function value are subjected to exponential function operation to dynamically generate the adaptive learning rate coefficient. The introduction of the adaptive learning rate coefficient is to adjust the learning rate in real time according to the performance changes during the optimization process. When the gradient is large and the parameter update direction is clear, the learning rate is increased to speed up convergence; when the gradient is small or the optimization path is unstable, the learning rate is reduced to enhance accuracy. Through exponential function operation, the learning rate can be flexibly adjusted throughout the optimization process, thereby improving the efficiency and robustness of gradient descent. The adaptive learning rate coefficient is combined with the momentum update vector to optimize the control parameters. The adaptive learning rate coefficient adjusts the amplitude of the momentum update vector, and applies the adjusted update vector to the current parameter matrix to generate a new iterative parameter matrix. The iterative parameter matrix is ​​converged and corrected according to the set threshold. If the change in the current iterative parameter is lower than the preset convergence threshold, it is considered that the parameter matrix has reached the optimal solution and further optimization is stopped. If the threshold is not reached, the optimization iteration is continued until the convergence condition is met. The purpose of the correction process is to prevent the parameter matrix from falling into the local optimum or causing instability due to numerical errors. By fine-tuning the parameter matrix, it is ensured that the optimal control parameter matrix outputted in the end can meet the performance requirements and have good robustness and generalization capabilities.

[0018] In one example, the input-output characteristic curve of the actuator in the industrial control system is piecewise linearized, and an upper-layer proxy model and a lower-layer proxy model are constructed to obtain a system dynamic characteristic prediction model, including: The slope mutation point detection and interval division are performed on the actuator input-output characteristic curve in the industrial control system to obtain a linearized sub-interval sequence, and each segment of the linearized sub-interval sequence is subjected to linear regression analysis and transfer function fitting to obtain a piecewise linearization model parameter set; The system dynamic equation is established according to the piecewise linearization model parameter set, and the system noise and measurement noise of the system dynamic equation are statistically analyzed to obtain the state space basic equation group; Based on the state space basic equations, the system state matrix, control input matrix and output observation matrix are constructed to obtain the state space model. The steady-state deviation, dynamic overshoot and adjustment time of the system operation data in the industrial control system are input into the multi-objective optimizer for weighted processing to obtain the performance index function. According to the state space model and the performance index function, an upper-layer proxy model of a deep neural network structure and a lower-layer proxy model composed of a local linear model are constructed to obtain a system dynamic characteristics prediction model.

[0019] In this example, slope mutation point detection and interval division are performed on the input-output characteristic curve of the actuator in the industrial control system. Assume that the input-output relationship of the actuator is represented by a set of data points ,in is the input signal, is the output response, which detects the change in the slope of the characteristic curve and identifies the curve's mutation point. Calculate the slope between every two adjacent points If the slope change of a point exceeds the set threshold (Right now ), then the point is considered to be a mutation point. Through this step, the entire curve is divided into several linear intervals. After completing the slope mutation point detection, the data of each linearized sub-interval is analyzed. Assume that the input-output relationship in a sub-interval is expressed by a linear equation Indicates that is the slope, is the intercept. Through linear regression analysis, the following sum of squared errors is minimized to determine and : ; Take the derivative and set the partial derivative to zero, that is Solving for the slope and the intercept The optimal value of is obtained to establish a linear regression model. Based on transfer function fitting, the dynamic characteristics of each sub-interval are represented by the transfer function of the first-order or second-order system, for example: ,in is the gain, is the time constant, is the Laplace transform variable. By minimizing the fitting error of the input and output response data, we can determine and The dynamic equations of the system are established based on the piecewise linearized model parameter set. The dynamic equations are expressed in state space form, where the state variables represents the internal state of the system, and the control input is , the output is The state equation and output equation are written as: ; ; in, is the derivative of the state vector, is the state matrix, is the control input matrix, is the output observation matrix, are direct transfer matrices. The specific form of these matrices is calculated by piecewise linearizing the model parameter set. For example, if the dynamic behavior in a subinterval is a first-order system, its state matrix is , the specific value is determined by the fitting parameters. The statistical characteristics of the system noise and measurement noise of the system dynamic equation are analyzed. Assume that the system noise is Gaussian white noise , the measurement noise is , and their covariance matrices are and Through statistical analysis of historical operating data, it is estimated and Based on the above state space basic equations, a complete state space model is constructed, including the state matrix , control input matrix , output observation matrix and direct transfer matrix . The steady-state deviation of the industrial control system operation data (such as ), dynamic overshoot and adjustment time as the input of the multi-objective optimizer, and define the performance index function: ; in is a weighting coefficient, and the relative importance of each objective is determined by the optimizer. After constructing the performance indicator function, the upper agent model of the deep neural network structure and the lower agent model of the local linear model combination are designed in combination with the state space model. The upper agent model uses a deep neural network to capture the global dynamic characteristics, and its input is the state and control input , the output is the predicted system response. The lower-level proxy model describes the local dynamic behavior through a piecewise linear model and performs collaborative optimization with the upper-level proxy model. By combining the upper-level and lower-level models, a prediction model for the system dynamic characteristics is obtained.

[0020] In one example, an upper-layer proxy model of a deep neural network structure and a lower-layer proxy model of a combination of a local linear model are constructed according to a state space model and a performance indicator function to obtain a system dynamic characteristic prediction model, including: The system state vector, control parameter vector and external disturbance vector in the state space model are input into the deep neural network for feature extraction to obtain the network structure of the upper agent model; The working range of the industrial control system is divided into working sub-intervals, and a local linear model is established for each working sub-interval to obtain a basic model group of the lower-level proxy model; The output results of the basic model group are weighted and fused to obtain the comprehensive output results of the lower-level proxy model; Perform back propagation training on the network structure of the upper-layer proxy model, calculate the loss value based on the performance indicator function and update the network parameters to obtain the trained upper-layer proxy model; The output result of the trained upper-layer proxy model is combined with the comprehensive output result of the lower-layer proxy model to obtain the system state prediction value; According to the deviation between the predicted value of the system state and the actual system state, the parameters of the upper agent model and the lower agent model are adjusted online to obtain the system dynamic characteristics prediction model.

[0021] In this example, the system state vector, control parameter vector, and external disturbance vector in the state space model are input into the deep neural network for feature extraction. Assume that the state space model of the system is represented by the following equation: ; ; in, is the state vector of the system, describing the internal dynamic state of the system; is the control parameter vector representing the adjustment of the input signal; is the external interference vector, reflecting the impact of the environment on the system; They are the state matrix, control input matrix, output observation matrix, direct transfer matrix and interference input matrix. , , As the input of the deep neural network, the deep neural network extracts the dynamic features of these variables through multi-layer nonlinear mapping to obtain the network structure of the upper agent model. Each layer of the deep neural network is calculated by the following formula: ; in, It is The activation value of the layer; and are the weight matrix and bias vector of the layer respectively; is an activation function (such as ReLU or Sigmoid). Through multiple layers of iteration, the nonlinear relationship in the input signal is captured, and the system state prediction value is generated through the output layer. At the same time, in order to enhance the local description ability of the model, the working range of the industrial control system is divided into sub-intervals. Assume that the input signal range of the system is , which is divided into several working sub-intervals ,in In each working subinterval, it is assumed that the dynamic behavior of the system is represented by a local linear model: ; ; in, It is The dynamic matrix parameters of the sub-intervals are obtained by least squares fitting or transfer function approximation. The local linear models are combined to form the basic model group of the lower proxy model. In order to generate the comprehensive output results of the lower proxy model, the outputs of all basic model groups are weighted fused. Assume that The output of the base model is , the corresponding weight is , then the comprehensive output result It is expressed as: ; Among them, the weight The distribution is performed based on the distance between the current working point and the subinterval, for example, using a Gaussian kernel function: ; in is the width parameter of the kernel function. Back propagation training is performed on the network structure of the upper proxy model, and the loss value is calculated and the network parameters are updated through the performance indicator function. Performance indicator function Defined as the error between the system output and the target value, for example: ; in, is the desired system output, is the time window length for performance evaluation, and ||·|| represents the Euclidean norm. By calculating the gradient of the loss value And use gradient descent to update the parameters: ; in are network parameters, is the learning rate. After multiple iterations, the trained upper-layer proxy model is obtained. The output of the trained upper-layer proxy model is The combined output of the underlying proxy model Perform combined operations to generate the final system state prediction value: ; in is the fusion weight of the upper and lower models. The actual system status The online adjustment mechanism combines global and local dynamic characteristics to form a prediction model for system dynamic characteristics.

[0022] In one example, the working range of the industrial control system is divided into working sub-intervals, and a local linear model is established for each working sub-interval to obtain a basic model group of the lower-level proxy model, including: Cluster analysis is performed on the operating data of the industrial control system according to the distribution of working points to obtain the target working point set within the working range, and the target working point set is input into the K-means clustering algorithm for automatic partition calculation to obtain m working sub-intervals; Establish state equations and output equations for m working sub-intervals respectively, obtain the dynamic model matrix of each working sub-interval, and perform continuity constraint calculation on the dynamic model matrix according to the sub-interval boundary conditions to obtain the state transfer matrix of the local linear model; The state transfer matrix is ​​subjected to zero-pole analysis and stability judgment to obtain the effective working range of each working sub-interval. Based on the effective working range, the local linear model is subjected to parameter optimization and model fusion to obtain the basic model group of the lower-level proxy model.

[0023] In this example, the system operation data is collected, including the input signals of the system (such as the control parameter vector and the output response (such as the system state vector and the output vector ). Assume that the input signal range is , preprocess it to remove noise and outliers, and then perform cluster analysis based on the distribution of working points. By analyzing the distribution characteristics of the input signal and the corresponding output state, the different operating modes of the system are preliminarily identified. The K-means clustering algorithm is used for automatic partition calculation. Assume that the sampling point set of the input signal is , each input signal Indicates a dimensional vector. The K-means algorithm achieves clustering by minimizing the following objective function: ; in, is the number of clusters (i.e., the number of working subintervals), It is A collection of clusters, It is The center of the cluster, Represents the Euclidean norm. By iteratively updating the cluster center and assigning sample points, K-means divides the input signal into After clustering, the state equation and output equation are established for each subinterval. The dynamic relationship between the input and output of each subinterval is described by the state space model, which is in the form of: ; ; in, is the state vector, is the input vector, is the output vector, They are The state matrix, control input matrix, and output matrix of each subinterval are used. The parameters of these matrices are estimated by the least squares method using the operating data in the subinterval. For example, Elements , solved by the following optimization problem: ; in, is the time derivative of the state vector, is the number of samples in the subinterval. In order to ensure the continuity between subintervals, continuity constraints are imposed at the subinterval boundaries. and The boundaries of the subintervals are , the following conditions must be met: ; ; in, is the state vector of the boundary point. The matrix parameters of the subinterval are adjusted through the optimization algorithm to make the state transfer matrix continuous at the boundary and improve the consistency of the model. Zero-pole analysis and stability judgment are performed on the state transfer matrix of each subinterval. State matrix The eigenvalues ​​(i.e., zeros and poles) of determine the dynamic behavior of the system in the subinterval. Assume that the eigenvalues ​​are ( ), stability requires that the real parts of all eigenvalues ​​satisfy Re . If the state matrix of a subinterval does not meet the stability conditions, its parameters need to be readjusted or rebuilt. After completing the zero-pole analysis and stability judgment, the local linear model is optimized and the model is fused according to the effective working range of the subinterval. The determination of the effective working range is based on the eigenvalue distribution and the dynamic performance within the subinterval. For example, if a subinterval has poor stability within a specific input range, the range of the subinterval is reduced, or the weights of the adjacent subintervals are adjusted. By weighted fusion of the outputs of all subintervals, a basic model group of the lower-level proxy model is formed. Assume that The output of each subinterval is , whose weight is , then the fused output is: ; Among them, the weight Dynamic adjustment is made through the dynamic performance of the sub-interval and the position of the current working point.

[0024] In one example, a state observer design and gain matrix optimization are performed on an industrial control system based on a system dynamic characteristic prediction model to obtain a system state observation vector and a disturbance compensation vector, including: Inputting the state prediction data generated by the system dynamic characteristic prediction model and the actual system state data into the observer design unit to obtain the observer basic structure model; Perform weighted least squares operation and numerical optimization on the state error in the observer infrastructure model to obtain the observer feedback gain matrix, and iteratively solve the observer state equation and perform parameter correction based on the observer feedback gain matrix to obtain the system state observation vector; Perform interference characteristic analysis on the deviation between the industrial control system output and the system dynamic characteristic prediction model output to obtain an interference dynamic model; The disturbance dynamic model and the system state observation vector are combined for calculation and frequency domain analysis to obtain a disturbance compensation function, and the disturbance compensation function is subjected to real-time numerical calculation and feedback compensation to obtain a disturbance compensation vector.

[0025] In this example, the state prediction data generated by the system dynamic characteristics prediction model and the actual system state data of the industrial control system are input into the observer design unit. Assume that the state prediction data is , the actual system status data is , the difference between the two is defined as the state error The observer infrastructure model is described based on the state equation and output equation in the following form: ; ; in, is the control input, is the system output, is the feedback gain matrix of the observer, are the state matrix, input matrix and output matrix of the system respectively. In order to optimize the feedback gain matrix of the observer , the state errors in the observer infrastructure model are weighted least squares and numerically optimized. The weighted least squares method optimizes the gain matrix by minimizing the following weighted sum of squared errors: ; in, is a weight matrix used to emphasize the relative importance of errors in different states. Derivative and set the gradient to zero to obtain the analytical solution of the optimal feedback gain matrix or iteratively solve it through numerical optimization methods. After the optimization is completed, according to the feedback gain matrix , iteratively solve the observer's state equation and correct the parameters. Using the real-time sampled state data and model prediction values, update the observer's parameters and correct the state estimate to make it closer to the actual system state. The core of the iterative solution is to update the state estimate based on the following recursive formula through technologies such as the Kalman filter: ; The result of this process is the system state observation vector , which can estimate the internal state of the system in real time with high accuracy. And the system dynamic characteristics prediction model output The deviation between Perform interference characteristic analysis. Through statistical analysis of deviation data, extract the main characteristics of interference, such as amplitude , frequency distribution and phase characteristics. These characteristics are used to establish a dynamic model of interference, which has the general form: ; ; in, is the disturbance state vector, is the parameter matrix of the disturbance dynamics model. These parameters are estimated by minimizing the error between the disturbance model prediction and the actual deviation. Perform combined calculation and frequency domain analysis to generate an interference compensation function. Assume that the interference compensation function is in the form of: ; in, is the interference compensation gain matrix. Through frequency domain analysis, optimization The value of can effectively suppress the interference within the main frequency range of the interference. The interference compensation function is numerically calculated in real time and applied to the system control input to achieve feedback compensation. The compensated control input is: ; Real-time calculation of disturbance compensation vector And feeding it back to the system helps to significantly reduce the impact of interference on system output, thereby improving control accuracy and system stability.

[0026] In one example, the system state observation vector and the disturbance compensation vector are input into a fixed time convergence controller to calculate control parameters, and a system control input matrix is ​​obtained, including: The state feedback gain calculation and parameter normalization processing are performed on the system state observation vector to obtain the state feedback control vector, and the interference compensation vector is input into the compensation control unit to perform compensation gain optimization and noise suppression to obtain the interference compensation control vector; Perform fixed time convergence term design and coefficient optimization calculation on the state observation vector to obtain convergence control parameters; Performing matrix multiplication operation on the state feedback control vector and the system state observation vector to obtain the feedback control component, and performing feedback compensation calculation on the disturbance compensation control vector and the disturbance compensation vector to obtain the compensation control component; The feedback control component, compensation control component and convergence control parameter are combined by control law operation to obtain the system control input matrix.

[0027] In this example, the state feedback gain calculation and parameter normalization processing are performed on the system state observation vector to obtain the state feedback control vector. Assume that the system state observation vector is , whose elements represent the current state values ​​of the system. By introducing the feedback gain matrix , calculate the state feedback control vector : ; in, Each column of corresponds to the feedback weight for a specific state. In order to ensure that the impact of different state variables is calculated in a unified dimension, the state observation vector Normalize the parameters. Normalization is done by introducing a diagonal matrix , adjust the state vector to normalized form , thereby avoiding the influence of the scale difference between states on the feedback gain calculation. At the same time, the interference compensation vector A compensation control unit is introduced to calculate the disturbance compensation control vector. The compensation control unit optimizes the compensation gain matrix Generate effective compensation control signal: ; In order to improve the compensation effect and suppress the existing noise, the gain weight is adjusted in combination with the noise characteristics during the compensation gain optimization process. For example, the filter matrix is ​​introduced Perform low-pass filtering on the interfering signal to reduce the impact of high-frequency noise: ; ; After the calculation of the state feedback control vector and the disturbance compensation control vector is completed, the fixed-time convergence term is designed for the state observation vector. The purpose of fixed-time convergence control is to ensure that the system state converges to the target state within a finite time, regardless of the initial conditions. Assume that the target state is , the convergence term is designed in the following nonlinear form: ; in, is the convergence gain matrix, is the nonlinear convergence coefficient, which is used to accelerate the reduction of state error. and , to achieve the fixed convergence time. The state feedback control vector and the system state observation vector Perform matrix multiplication to generate feedback control components : ; At the same time, the disturbance compensation control vector Perform feedback compensation calculations to generate compensation control components Feedback compensation uses an iterative optimization strategy to adjust the compensation signal in real time, thereby enhancing the compensation effect: ; in, is the feedback compensation gain. , compensation control component and convergence control parameters Generate the control input matrix of the system through the combination of control laws : ; This control input matrix includes the comprehensive effects of feedback control of state error, compensation control of external disturbance and fixed time convergence control.

[0028] In one example, an adaptive gradient descent operation and momentum term optimization are performed on the system control input matrix, and an optimal control parameter matrix is ​​output in combination with the system dynamic characteristic prediction model, including: The system control input matrix and the output value of the system dynamic characteristic prediction model are used to calculate the performance index to obtain the performance index function value; The derivative of the performance index function value is mapped into parameter space and a gradient vector is generated to obtain a performance parameter gradient matrix, and the performance parameter gradient matrix is ​​weightedly accumulated with the momentum term at the previous moment to obtain a momentum update vector; Perform exponential function operation on the momentum update vector and the performance indicator function value to obtain the adaptive learning rate coefficient; The adaptive learning rate coefficient and the momentum update vector are used to perform control parameter optimization calculation to obtain an iterative parameter matrix, and the iterative parameter matrix is ​​subjected to convergence judgment and correction processing according to a set threshold to obtain an optimal control parameter matrix.

[0029] In this example, the system control input matrix is ​​combined with the output value of the system dynamic characteristics prediction model, and the control effect of the system is quantitatively evaluated through the performance index function. Assume that the system control input matrix is , whose elements represent the input intensity of each control variable; the output value of the system dynamic characteristics prediction model is , whose elements represent the response values ​​of the system under the prediction model; the actual system output is , then the performance index function The definition of comprehensively considers factors such as steady-state error, dynamic overshoot and adjustment time, and is in the form of: ; in, is a weight factor used to adjust the relative importance of different performance indicators. Represents the Euclidean norm. By calculating the performance index function value, the impact of the current control input on the system performance is quantitatively evaluated. After obtaining the performance index function value, its derivative in the parameter space is calculated to generate a gradient vector. Assume that the control parameter is , performance indicator function about Gradient It is expressed as: ; In actual calculations, the gradient is derived from the expression of the performance indicator function through the chain rule and automatic differentiation techniques. The gradient vector is mapped to the parameter space and the performance parameter gradient matrix is ​​generated by , quantifying the sensitivity of the control parameters in different directions. In order to stabilize the gradient update process, the momentum term optimization method is introduced. Assume that the momentum term at the previous moment is , the momentum update vector Calculated by the following formula: ; in, is the momentum coefficient, which is used to control the influence of historical gradients. By weighted accumulation of current and historical gradients, the momentum term can effectively suppress the oscillation phenomenon in gradient descent and accelerate convergence. After obtaining the momentum update vector, the adaptive learning rate coefficient is calculated by performing an exponential function operation with the performance indicator function value. Assume that the performance indicator function value is , adaptive learning rate The calculation formula is: ; in, is the initial learning rate, is the learning rate attenuation factor. This formula ensures that when the performance indicator function value is large, the learning rate decreases to enhance the stability of the optimization; when the performance indicator function value is small, the learning rate increases to accelerate convergence. The adaptive learning rate coefficient is combined with the momentum update vector to complete the optimization calculation of the control parameters. The update rule of the control parameters is: ; This formula indicates that the control parameters are adjusted through learning rate adjustment and momentum update to generate a new iterative parameter matrix The iterative parameter matrix is ​​judged to converge according to the set threshold value if the following conditions are met: ; in, If is the convergence threshold, the parameters are considered to have converged to the optimal solution and the optimization process is stopped; otherwise, the iteration continues until the convergence condition is met. In order to prevent the parameters from falling into the local optimal solution, random perturbations or other global optimization strategies are combined for correction. The optimal control parameter matrix obtained in the end is It is directly used to generate the system control input matrix, thereby achieving accurate optimization of industrial control systems.

[0030] Reference Figure 2 , this embodiment provides an intelligent industrial control parameter self-tuning device, including: Building module 1, used for performing piecewise linearization processing on the input-output characteristic curve of the actuator in the industrial control system, and building an upper-layer proxy model and a lower-layer proxy model, and combining them to obtain a system dynamic characteristic prediction model; Optimization module 2, used to design the state observer and optimize the gain matrix of the industrial control system based on the system dynamic characteristic prediction model, and obtain the system state observation vector and the interference compensation vector; A calculation module 3 is used to input the system state observation vector and the disturbance compensation vector into the fixed time convergence controller to calculate the control parameters and obtain the system control input matrix; The output module 4 is used to perform adaptive gradient descent operation and momentum term optimization on the system control input matrix, and output the optimal control parameter matrix in combination with the system dynamic characteristic prediction model.

[0031] In this embodiment, for the specific implementation of each unit in the above device embodiment, please refer to the above method embodiment, which will not be repeated here.

[0032] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0033] The above description is only a preferred embodiment of the present invention, and does not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for self-tuning intelligent industrial control parameters, characterized in that: The following steps are involved: The input-output characteristic curve of the actuator in the industrial control system is processed piecewise linearly, and the upper-layer agent model and the lower-layer agent model are constructed to obtain the system dynamic characteristic prediction model; Based on the system dynamic characteristic prediction model, the state observer design and gain matrix optimization are performed on the industrial control system to obtain a system state observation vector and a disturbance compensation vector; Inputting the system state observation vector and the disturbance compensation vector into a fixed time convergence controller to calculate control parameters, thereby obtaining a system control input matrix; An adaptive gradient descent operation and momentum term optimization are performed on the system control input matrix, and an optimal control parameter matrix is ​​output in combination with the system dynamic characteristic prediction model.

2. The method for self-tuning intelligent industrial control parameters according to claim 1, characterized in that: The step of performing piecewise linearization on the input-output characteristic curve of the actuator in the industrial control system, and constructing an upper-layer proxy model and a lower-layer proxy model to combine and obtain a system dynamic characteristic prediction model includes: Perform slope mutation point detection and interval division on the actuator input-output characteristic curve in the industrial control system to obtain a linearized sub-interval sequence, and perform linear regression analysis and transfer function fitting on each segment of data in the linearized sub-interval sequence to obtain a piecewise linearization model parameter set; Establishing a system dynamic equation according to the piecewise linearization model parameter set, and performing statistical characteristic analysis on system noise and measurement noise of the system dynamic equation to obtain a state space basic equation group; Based on the state space basic equation group, a system state matrix, a control input matrix and an output observation matrix are constructed to obtain a state space model, and the steady-state deviation, dynamic overshoot and adjustment time of the system operation data in the industrial control system are input into a multi-objective optimizer for weighted processing to obtain a performance index function; An upper-layer proxy model of a deep neural network structure and a lower-layer proxy model of a combination of a local linear model are constructed according to the state space model and the performance indicator function to obtain a system dynamic characteristic prediction model.

3. The method for self-tuning intelligent industrial control parameters according to claim 2 is characterized in that: The upper layer proxy model of the deep neural network structure and the lower layer proxy model of the local linear model combination are constructed according to the state space model and the performance index function to obtain the system dynamic characteristics prediction model, including: Inputting the system state vector, control parameter vector and external disturbance vector in the state space model into a deep neural network for feature extraction to obtain a network structure of an upper-layer proxy model; Dividing the working range of the industrial control system into working sub-intervals, and establishing a local linear model for each working sub-interval to obtain a basic model group of the lower-layer proxy model; Performing a weighted fusion operation on the output results of the basic model group to obtain a comprehensive output result of the lower-level proxy model; Performing back-propagation training on the network structure of the upper-layer proxy model, calculating the loss value based on the performance indicator function and updating the network parameters to obtain a trained upper-layer proxy model; Combining the output result of the trained upper-layer proxy model with the comprehensive output result of the lower-layer proxy model to obtain a system state prediction value; The upper agent model and the lower agent model are adjusted online in terms of parameters according to the deviation between the system state prediction value and the actual system state, so as to obtain a system dynamic characteristic prediction model.

4. The method for self-tuning intelligent industrial control parameters according to claim 3 is characterized in that: The working range of the industrial control system is divided into working sub-intervals, and a local linear model is established for each working sub-interval to obtain a basic model group of the lower-layer proxy model, including: Perform cluster analysis on the operation data of the industrial control system according to the distribution of working points to obtain a target working point set within the working range, and input the target working point set into the K-means clustering algorithm for automatic partition calculation to obtain m working sub-intervals; Establishing state equations and output equations for the m working sub-intervals respectively, obtaining a dynamic model matrix for each working sub-interval, and performing continuity constraint calculation on the dynamic model matrix according to the sub-interval boundary conditions to obtain a state transfer matrix of a local linear model; The state transfer matrix is ​​subjected to zero-pole analysis and stability judgment to obtain the effective working range of each working sub-interval, and the local linear model is subjected to parameter optimization and model fusion based on the effective working range to obtain the basic model group of the lower-level proxy model.

5. The method for self-tuning intelligent industrial control parameters according to claim 1, characterized in that: The state observer design and gain matrix optimization of the industrial control system based on the system dynamic characteristic prediction model to obtain the system state observation vector and the interference compensation vector include: Inputting the state prediction data generated by the system dynamic characteristic prediction model and the actual system state data into the observer design unit to obtain the observer basic structure model; Performing weighted least squares operation and numerical optimization on the state error in the observer infrastructure model to obtain an observer feedback gain matrix, and iteratively solving and parameter correction on the observer state equation according to the observer feedback gain matrix to obtain a system state observation vector; Performing interference characteristic analysis on the deviation between the output of the industrial control system and the output of the system dynamic characteristic prediction model to obtain an interference dynamic model; The disturbance dynamic model and the system state observation vector are combined and calculated and analyzed in the frequency domain to obtain a disturbance compensation function, and the disturbance compensation function is numerically calculated and feedback compensated in real time to obtain a disturbance compensation vector.

6. The method for self-tuning intelligent industrial control parameters according to claim 1, characterized in that: The step of inputting the system state observation vector and the interference compensation vector into a fixed time convergence controller to calculate control parameters to obtain a system control input matrix includes: Performing state feedback gain calculation and parameter normalization processing on the system state observation vector to obtain a state feedback control vector, and inputting the interference compensation vector into a compensation control unit to perform compensation gain optimization and noise suppression to obtain an interference compensation control vector; Perform fixed time convergence term design and coefficient optimization calculation on the state observation vector to obtain convergence control parameters; Performing matrix multiplication operation on the state feedback control vector and the system state observation vector to obtain a feedback control component, and performing feedback compensation calculation on the disturbance compensation control vector and the disturbance compensation vector to obtain a compensation control component; The feedback control component, the compensation control component and the convergence control parameter are subjected to a control law combination operation to obtain a system control input matrix.

7. The method for self-tuning intelligent industrial control parameters according to claim 1, characterized in that: The method of performing adaptive gradient descent operation and momentum term optimization on the system control input matrix and outputting an optimal control parameter matrix in combination with the system dynamic characteristic prediction model includes: Performing performance index calculation on the system control input matrix and the output value of the system dynamic characteristic prediction model to obtain a performance index function value; Perform parameter space mapping and gradient vector generation on the derivative of the performance indicator function value to obtain a performance parameter gradient matrix, and perform weighted accumulation calculation on the performance parameter gradient matrix and the momentum term at the previous moment to obtain a momentum update vector; Performing an exponential function operation on the momentum update vector and the performance indicator function value to obtain an adaptive learning rate coefficient; The adaptive learning rate coefficient and the momentum update vector are used to perform control parameter optimization calculation to obtain an iterative parameter matrix, and the iterative parameter matrix is ​​subjected to convergence judgment and correction processing according to a set threshold to obtain an optimal control parameter matrix.

8. An intelligent industrial control parameter self-tuning device, characterized in that: For implementing the steps of the method according to any one of claims 1 to 7, the device comprises: A construction module is used to perform piecewise linearization processing on the input-output characteristic curve of the actuator in the industrial control system, and to construct an upper-layer proxy model and a lower-layer proxy model, and to combine them to obtain a system dynamic characteristic prediction model; An optimization module, used for designing a state observer and optimizing a gain matrix of the industrial control system based on the system dynamic characteristic prediction model, to obtain a system state observation vector and a disturbance compensation vector; A calculation module, used for inputting the system state observation vector and the interference compensation vector into a fixed time convergence controller to calculate control parameters and obtain a system control input matrix; The output module is used to perform adaptive gradient descent operation and momentum term optimization on the system control input matrix, and output the optimal control parameter matrix in combination with the system dynamic characteristic prediction model.

Images