A learning control method and system for complex nonlinear intermittent industrial process systems

Abstract

The present invention discloses a learning control method and system for a complex nonlinear intermittent industrial process system. The method comprises: constructing a strongly nonlinear intermittent process system dynamic model and an offline data set; using the offline data set to train an existing deep Koopman network model to obtain a coefficient matrix of a discrete state equation and an observable function; combining the discrete state equation and the observable function to obtain a linear predictive control model for the strongly nonlinear intermittent process system, calculating the optimal solution and using it as the control variable of the predictive control model to construct an offline predictive control data set; training a deep neural network to replace the predictive controller of the constructed model; and using the deep neural network to output predictive control inputs to achieve real-time state predictive control of the intermittent process system. The present invention realizes intelligent system control while improving the adaptability of the control method; and achieves real-time control of the intermittent industrial process system while retaining control accuracy and performance.

Description

Technical Field

[0001] The present invention relates to the technical field of intermittent industrial process system control, and more particularly to a learning control method and system for a complex nonlinear intermittent industrial process system. Background Art

[0002] In the field of discrete manufacturing industry, intermittent industrial process is an important type of industrial production process. Because of its good flexibility and production efficiency, it has been widely used in discrete manufacturing industry fields such as machinery manufacturing, plastic processing, equipment production, aerospace, electronics and electrical.

[0003] In actual industrial applications, achieving precise, fast, efficient, and real-time operational control of intermittent processes remains a significant challenge due to the complex dynamic characteristics of most intermittent production processes, such as strong nonlinearity, large hysteresis, and multiple time-varying events. This, coupled with the high-performance requirements of modern industrial products for control accuracy and real-time performance, creates a formidable challenge. While intermittent process systems in current industrial environments possess a certain level of basic automation, the optimization and control of intermittent manufacturing processes still largely relies on the experience of skilled on-site operators through trial and error, configuration, and optimization. This process is time-consuming, labor-intensive, and inefficient.

[0004] Currently, several control algorithms have been developed for batch process systems, such as PID control, iterative learning control, model predictive control (MPC), and neural network control. However, these methods also have shortcomings. For example, PID control methods often suffer from difficulties in PID parameter tuning and non-optimal control processes, making them difficult to meet the increasingly demanding requirements for control accuracy and response speed. Batch process algorithms based on iterative learning control often use a linear time-varying model to approximate the nonlinear system. This approach achieves good control results only when the linear model approximates the original nonlinear system model well; if the model approximation is inaccurate, the control effect is mediocre. Common nonlinear model predictive control methods, such as linearizing nonlinear models and using various specialized models, often suffer from certain accuracy losses and limitations. These methods require specialized design for different systems, making them difficult to widely promote and apply in industrial systems. Furthermore, most of these methods rely on offline parameter tuning for typical linear systems, making them difficult to achieve control requirements for batch processes with strong nonlinearities. Because they involve repeated iterations of the offline optimization process, the optimization process is relatively time-consuming and cannot meet the real-time optimal control requirements of practical industrial control systems.

[0005] The prior art discloses a method for using a predictor in a process control system with wireless or intermittent process measurements. The method allows the use of slowly or intermittently received process variable values ​​in a predictor-based control scheme without changing the control algorithm. The control technique includes a controller (e.g., a PID controller) and a predictor (e.g., a model-based predictor), the predictor being coupled to receive intermittent feedback, for example, in the form of a process variable measurement signal from the process. The predictor can be an observer, such as a Kalman filter, or a Smith predictor. The predictor is configured to generate an estimate of the process variable value based on the intermittent or slowly received process feedback signal during each controller execution cycle, and to provide the new process variable estimate to the controller so that the controller can generate a control signal for controlling the process. This scheme does not address the issues of intelligent and adaptive control. Summary of the Invention

[0006] In order to overcome the defects of the above-mentioned prior art in the control method of complex intermittent process system with strong nonlinearity, such as insufficient intelligence and adaptability, and long offline optimization control time, the present invention provides a learning control method and system for complex nonlinear intermittent industrial process system.

[0007] The primary purpose of the present invention is to solve the above technical problems, and the technical solutions of the present invention are as follows:

[0008] A first aspect of the present invention provides a learning control method for a complex nonlinear intermittent industrial process system, comprising the following steps:

[0009] S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data;

[0010] S2: Using the offline data set to train the existing deep Koopman network model, obtain the coefficient matrix and observable function of the discrete state equation, and realize the linearization of the strongly nonlinear batch process system;

[0011] S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set.

[0012] S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network;

[0013] S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

[0014] Furthermore, the specific process of step S1 is:

[0015] A strongly nonlinear dynamic model of an intermittent process system is constructed. The mathematical expression of the dynamic model is:

[0016]

[0017] Where u(t) is the control input signal of the batch process system, and x(t) represents all state outputs of the system;

[0018] Inputting different control input signals into the dynamic model to obtain the operation process data of the strongly nonlinear intermittent process system under different input signals, that is, the state vector during the operation;

[0019] Data of the state vector and control input signal are collected respectively according to the preset sampling time, and the collected data are combined into an offline data set.

[0020] Furthermore, the specific process of step S2 is:

[0021] The Koopman operator is used to linearize the strongly nonlinear batch process system and obtain the linearized discrete state equation.

[0022] The Koopman network model is trained using the discrete data set to obtain an observable function and a coefficient matrix of a linearized discrete state equation.

[0023] Furthermore, the linearized expression is:

[0024] z(k+1)=Az(k)+Bu(k) (4a)

[0025] x(k+1)=Cz(k+1) (4b)

[0026] Among them, A, B and C represent coefficient matrices, and z(k), u(k), x(k+1), and z(k+1) represent the system linear state vector at time k, the system control signal, the system nonlinear state vector at time k+1, and the system linear state vector at time k+1, respectively.

[0027] Furthermore, the loss function of the Koopman network model is:

[0028]

[0029] Where X = [x(1), x(2), …, x(q)] is the set of open-loop state vectors of the nonlinear system, Z = z(1), z(2), …, z(q)] = [g(x(1)), g(x(2)), …, g(x(q))] is the set of linear state vectors of the system corresponding to X, q is the size of the collected data set, m is the step size of the forced Koopman operator linear prediction, ||·|| F is the F-norm, ||·||2 is the 2-norm, and k, l, and r are the counting subscripts of the cumulative sum.

[0030] Furthermore, the discrete state equation and the observable function are combined to obtain the linear predictive control model of the strongly nonlinear batch process system, whose mathematical expression is:

[0031] z(k+1)=Az(k)+Bu(k) (6a)

[0032] x(k+1)=Cz(k+1) (6b)

[0033] y(k)=Dx(k) (6c)

[0034] z(k)=g(x(k)) (6d)

[0035] The optimization target is Δu, where the cost function J of the model predictive control is:

[0036]

[0037] Where p and n are the prediction time domain and control time domain respectively, p=n=15, Q and R are the weight matrices of y and Δu;

[0038] Combining formulas (6a), (6b), (6c), (6d) and (7) yields the constrained quadratic programming formula:

[0039]

[0040] st

[0041] z(k+1)=Az(k)+Bu(k),k=1,…,n (14b)

[0042] x(k+1)=Cz(k+1),k=1,…,n (14c)

[0043] y(k)=Dx(k),k=1,…,n (14d)

[0044] z(k)=g(x(k)),k=1,…,n (14e)

[0045] △u(k)=u(k)-u(k-1),k=1,…,n (14f)

[0046] y min ≤y(k)≤y max ,k=1,…,n (14g)

[0047] u min ≤y(k)≤u max ,k=1,…,n (14h)

[0048] Among them, (14b)-(14h) are constraints;

[0049] The optimal solution of the increment △u is obtained by solving the quadratic programming analytical expression, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear intermittent process system.

[0050] Furthermore, the offline predictive control data set is constructed by using process data and control signals as follows:

[0051] Different tracking targets y are set for the linear predictive control model of the strongly nonlinear batch process system. ref , collect the target output y of the controlled object ref , the controlled object output y, the process state x and the control variable u are combined into an offline predictive control data set.

[0052] Furthermore, the input of the deep neural network in step S5 includes: the model state variable x(k) at the current time step, the model output y(k), the model control variable u(k) and the model target output y ref (k), the output of the deep learning network is the optimal control quantity u(k+1) for the next time step.

[0053] Furthermore, the number of network layers, the number of neurons in a single layer, the learning rate, and the activation function of the deep neural network can be adjusted according to actual needs. After training, the deep neural network can achieve the same control effect as the predictive controller and can predict the optimal control input of the system based on the real-time state of the intermittent process system.

[0054] A second aspect of the present invention provides a complex nonlinear intermittent industrial process system learning control system, the system comprising: a memory and a processor, wherein the memory includes a complex nonlinear intermittent industrial process system learning control method program, and when the complex nonlinear intermittent industrial process system learning control method program is executed by the processor, the following steps are implemented:

[0055] S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data;

[0056] S2: Using the offline data set to train the existing deep Koopman network model, obtain the coefficient matrix and observable function of the discrete state equation, and realize the linearization of the strongly nonlinear batch process system;

[0057] S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set.

[0058] S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network;

[0059] S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

[0060] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0061] The present invention utilizes the Koopman operator to linearize the nonlinear system, thereby reducing the complexity of nonlinear system control and shortening the calculation time. It can retain the model information of the intermittent process system to the greatest extent and is independent of the mathematical model of the intermittent process system. It obtains the system object from the system input and output data through the big data method, realizes the intelligent control of the system, and improves the adaptive ability of the control. The predictive controller that replaces the linear predictive control model of the strongly nonlinear intermittent process system with a deep neural network can compress the calculation time required for controller control and realize real-time control of the intermittent process system while retaining the control accuracy and performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a learning control method for a complex nonlinear intermittent industrial process system according to an embodiment of the present invention.

[0063] Figure 2 Schematic diagram of the deep Koopman network model structure in an embodiment of the present invention.

[0064] Figure 3 Schematic diagram of the deep neural network structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that, in the absence of conflict, the embodiments of the present application and the features therein can be combined with each other.

[0066] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the scope of protection of the present invention is not limited to the specific embodiments disclosed below.

[0067] Example 1

[0068] like Figure 1 As shown, a learning control method for a complex nonlinear intermittent industrial process system includes the following steps:

[0069] S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data;

[0070] It should be noted that the specific process of step S1 is:

[0071] A dynamic model of a strongly nonlinear batch process system dynamic model is constructed, and the mathematical expression of the dynamic model is:

[0072]

[0073] Among them, u(t) is the control input signal of the batch process system, and x(t) represents all state outputs of the system. It should be noted that the internal parameters and constraints of the batch process system can be selected differently according to the actual application scenario and actual performance.

[0074] Inputting different control input signals into the dynamic model to obtain the operation process data of the strongly nonlinear intermittent process system under different input signals, that is, the state vector during the operation;

[0075] Data of the state vector and control input signal are collected respectively according to the preset sampling time, and the collected data are combined into an offline data set.

[0076] like Figure 2 Shown is a schematic diagram of the deep Koopman network model structure.

[0077] S2: Using the offline data set to train the existing deep Koopman network model, obtain the coefficient matrix and observable function of the discrete state equation, and realize the linearization of the strongly nonlinear batch process system;

[0078] It should be noted that the specific process of step S2 is:

[0079] The Koopman operator is used to linearize the strongly nonlinear batch process system and obtain the linearized expression;

[0080] It should be noted that the Koopman operator actually realizes the global linearization of the nonlinear system by performing coordinate transformation on the nonlinear system. The general form of the discrete-time nonlinear system is expressed as follows:

[0081] x(k+1)=f(x(k),u(k))

[0082] Where t is time and f is state space The mapping from state to state is defined as the Koopman operator with infinite dimensions. Acting on observable functions Make

[0083]

[0084] The more commonly used methods such as EDMD adopt the artificial design of observable function g and then obtain the Koopman operator through calculation. In the method of the present invention, a deep learning method based on the Koopman operator is adopted to obtain a finite-dimensional approximation of the observable function g and the Koopman operator by designing a suitable deep neural network training. The deep Koopman network structure is as follows Figure 2 shown.

[0085] During the training process, the state vector x(k) = [x1(k), x2(k), x3(k), x4(k), x5(k)] (X∈R n ) is fed into a multi-layer neural network, and the observable function g in the Koopman theory is simulated by the neural network to output the state vector z(z∈R N ). According to the definition of Koopman operator, the state vector z satisfies the requirements of linear system, A and B are the weight matrices of single-layer unbiased neural network, namely Koopman operator, where C=[I n×n ,O n×(N-n) ], satisfying the following equation,

[0086] z(k+1)=Az(k)+Bu(k) (4a)

[0087] x(k+1)=Cz(k+1) (4b)

[0088] The above (4a) and (4b) are the discrete state equations of the system linearization, where A, B and C represent coefficient matrices, and z(k), u(k), x(k+1), z(k+1) represent the system linear state vector at time k, the system control signal, the system nonlinear state vector at time k+1 and the system linear state vector at time k+1, respectively.

[0089] Next, the Koopman network model is trained using the offline data set obtained in step S1, and the data is normalized by Max-min. The Koopman network model is trained through the back propagation of the neural network. The loss function of the Koopman network model is:

[0090]

[0091] Where X = [x(x1), x(2), …, x(q)] is the set of open-loop state vectors of the nonlinear system, Z = z(1), z(2), …, z(q)] = [g(x1)), g(x(2)), …, g(x(q))] is the set of linear state vectors of the system corresponding to X, q is the size of the collected data set, m is the step size of the forced Koopman operator linear prediction, ||·|| F is the F-norm, ||·||2 is the 2-norm, and k, l, and r are the counting subscripts of the cumulative sum.

[0092] Through the above training, the observable function g and the coefficient matrices A, B and C of the linearized discrete state equation are obtained.

[0093] S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set.

[0094] It should be noted that the linear predictive control model of the strongly nonlinear batch process system is obtained by combining the discrete state equation and the observable function. Its mathematical expression is:

[0095] z(k+1)=Az(k)+Bu(k) (6a)

[0096] x(k+1)=Cz(k+1) (6b)

[0097] y(k)=Dx(k) (6c)

[0098] z(k)=g(x(k))(6d)

[0099] In the present invention, the increment Δu is selected as the optimization target, wherein the cost function J of the model predictive control is:

[0100]

[0101] Where p and n are the prediction time domain and control time domain respectively, p=n=15, Q and R are the weight matrices of y and Δu;

[0102] Combining formulas (6a), (6b), (6c), (6d), and (7) yields the following constrained quadratic programming analytical formula (i.e., a constrained quadratic programming problem):

[0103]

[0104] st

[0105] z(k+1)=Az(k)+Bu(k),k=1,…,n (14b)

[0106] x(k+1)=Cz(k+1),k=1,…,n (14c)

[0107] y(k)=Dx(k),k=1,…,n (14d)

[0108] z(k)=g(x(k)),k=1,…,n (14e)

[0109] △u(k)=u(k)-u(k-1),k=1,…,n (14f)

[0110] y min ≤y(k)≤y max ,k=1,…,n (14g)

[0111] u min ≤y(k)≤u max ,k=1,…,n (14h)

[0112] Among them, (14b)-(14h) are constraints; among them, y min 、y max 、u min and u max are the boundary constraints of y and u respectively. The optimal solution of the increment △u of the above quadratic programming problem can be calculated by using the quadratic programming algorithm, and the optimal solution is used as the control quantity of the predictive control model of the mathematical model of the strongly nonlinear intermittent process system.

[0113] Example 2

[0114] Based on the above embodiment, this embodiment describes the specific process of steps S4 and S5.

[0115] like Figure 3 Shown is a schematic diagram of the deep neural network structure.

[0116] S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network;

[0117] It should be noted that, in the present invention, different tracking targets y are set for the linear predictive control model of the strongly nonlinear intermittent process system. ref , collect the target output y of the controlled object ref , the output of the controlled object y, the process state quantity x and the control quantity u are combined into an offline predictive control data set, and the obtained offline predictive control data set is used to train a deep neural network. The present invention uses a deep neural network to learn the control process of the predictive controller implemented in step S3, and then replaces the original predictive controller to avoid the complex calculations brought about by the optimization process.

[0118] S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

[0119] It should be noted that the input of the deep neural network includes: the model state variable x(k) at the current time step, the model output y(k), the model control variable u(k) and the model target output y ref (k), the output of the deep learning network is the optimal control quantity u(k+1) for the next time step. The network structure is as follows Figure 3 As shown. The number of network layers, number of neurons per layer, learning rate, and activation function of the deep neural network can be adjusted according to actual needs. After training, the deep neural network can achieve the same control effect as a predictive controller and can predict the optimal control input of the system based on the real-time state of the intermittent process system. The present invention achieves the same control effect while consuming less system resources.

[0120] Example 3

[0121] A second aspect of the present invention provides a complex nonlinear intermittent industrial process system learning control system, the system comprising: a memory and a processor, wherein the memory includes a complex nonlinear intermittent industrial process system learning control method program, and when the complex nonlinear intermittent industrial process system learning control method program is executed by the processor, the following steps are implemented:

[0122] S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data;

[0123] It should be noted that the specific process of step S1 is:

[0124] A strongly nonlinear dynamic model of an intermittent process system is constructed. The mathematical expression of the dynamic model is:

[0125]

[0126] Where u(t) is the control input signal of the batch process system, and x(t) represents all state outputs of the system;

[0127] Among them, the internal parameters and constraints of the batch process system can be selected differently according to the actual application scenario and actual performance.

[0128] Inputting different control input signals into the dynamic model to obtain the operation process data of the strongly nonlinear intermittent process system under different input signals, that is, the state vector during the operation;

[0129] Data of the state vector and control input signal are collected respectively according to the preset sampling time, and the collected data are combined into an offline data set.

[0130] S2: Using the offline data set to train the existing deep Koopman network model, obtain the coefficient matrix and observable function of the discrete state equation, and realize the linearization of the strongly nonlinear batch process system;

[0131] It should be noted that the specific process of step S2 is:

[0132] The Koopman operator is used to linearize the strongly nonlinear batch process system and obtain the linearized discrete state equation.

[0133] It should be noted that the Koopman operator actually realizes the global linearization of the nonlinear system by performing coordinate transformation on the nonlinear system. The general form of the discrete-time nonlinear system is expressed as follows:

[0134] x(k+1)=f(x(k),u(k))

[0135] Where t is time and f is state space The mapping from state to state is defined as the Koopman operator with infinite dimensions. Acting on observable functions Make

[0136]

[0137] The more commonly used methods such as EDMD adopt the artificial design of observable function g and then obtain the Koopman operator through calculation. In the present invention, a deep learning method based on the Koopman operator is adopted to obtain the finite dimensional approximation of the observable function g and the Koopman operator by designing a suitable deep neural network training. The deep Koopman network structure is as follows Figure 2 shown.

[0138] During the training process, the state vector x(k) = [x1(k), x2(k), x3(k), x4(k), x5(k)] (X∈R n ) is fed into a multi-layer neural network, and the observable function g in the Koopman theory is simulated by the neural network to output the state vector z(z∈R N ). According to the definition of Koopman operator, the state vector z satisfies the requirements of linear system, A and B are the weight matrices of single-layer unbiased neural network, namely Koopman operator, where C=[I n×n ,O n×(N-n) ], satisfying the following equation,

[0139] z(k+1)=Az(k)+Bu(k) (4a)

[0140] x(k+1)=Cz(k+1) (4b)

[0141] The above formulas (4a) and (4b) are the discrete state equations after the system is linearized, where z(k), u(k), x(k+1), and z(k+1) represent the system linear state vector at time k, the system control signal, the system nonlinear state vector at time k+1, and the system linear state vector at time k+1, respectively.

[0142] Next, the Koopman network model is trained using the offline data set obtained in step S1, and the data is normalized by Max-min. The Koopman network is trained by back propagation of the neural network. The loss function of the Koopman network model is:

[0143]

[0144] Where X = [x(1), x(2), …, x(q)] is the set of open-loop state vectors of the nonlinear system, Z = [z(1), z(2), …, z(q)] = [g(x(1)), g(x(2)), …, g(x(q))] is the set of linear state vectors of the system corresponding to X, q is the size of the collected data set, m is the step size of the forced Koopman operator linear prediction, ‖·‖ F is the F-norm, ‖·‖2 is the 2-norm, and k, l, and r are the counting subscripts of the cumulative sum.

[0145] Through the above training, the observable function g and the coefficient matrices A, B and C of the linearized discrete state equation are obtained.

[0146] S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set.

[0147] It should be noted that the linear predictive control model of the strongly nonlinear batch process system is obtained by combining the discrete state equation and the observable function. Its mathematical expression is:

[0148] z(k+1)=Az(k)+Bu(k) (6a)

[0149] x(k+1)=Cz(k+1) (6b)

[0150] y(k)=Dx(k) (6c)

[0151] z(k)=g(x(k)) (6d)

[0152] In the present invention, the increment Δu is selected as the optimization target, wherein the cost function J of the model predictive control is:

[0153]

[0154] Where p and n are the prediction time domain and control time domain respectively, p=n=15, Q and R are the weight matrices of y and Δu;

[0155] Combining formulas (6a), (6b), (6c), (6d), and (7) yields the following constrained quadratic programming analytical formula (i.e., a constrained quadratic programming problem):

[0156]

[0157] st

[0158] z(k+1)=Az(k)+Bu(k),k=1,…,n (14b)

[0159] x(k+1)=Cz(k+1),k=1,…,n (14c)

[0160] y(k)=Dx(k),k=1,…,n (14d)

[0161] z(k)=g(x(k)),k=1,…,n(14e)

[0162] △u(k)=u(k)-u(k-1),k=1,…,n (14f)

[0163] y min ≤y(k)≤y max ,k=1,…,n (14g)

[0164] u min ≤y(k)≤u max ,k=1,…,n (14h)

[0165] Among them, (14b)-(14h) are constraints; among them, y min 、y max 、u min and u max are the boundary constraints of y and u respectively. The optimal solution of the increment △u of the above quadratic programming problem can be calculated by using the quadratic programming algorithm, and the optimal solution is used as the control quantity of the predictive control model of the mathematical model of the strongly nonlinear intermittent process system.

[0166] S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network;

[0167] It should be noted that, in the present invention, different tracking targets y are set for the linear predictive control model of the strongly nonlinear intermittent process system. ref , collect the target output y of the controlled object ref , the output of the controlled object y, the process state quantity x and the control quantity u are combined into an offline predictive control data set, and the obtained offline predictive control data set is used to train a deep neural network. The present invention uses a deep neural network to learn the control process of the predictive controller implemented in step S3, and then replaces the original predictive controller to avoid the complex calculations brought about by the optimization process.

[0168] S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

[0169] It should be noted that the input of the deep neural network includes: the model state variable x(k) at the current time step, the model output y(k), the model control variable u(k) and the model target output y ref (k), the output of the deep learning network is the optimal control quantity u(k+1) for the next time step. The network structure is as follows Figure 3 As shown. The number of network layers, number of neurons per layer, learning rate, and activation function of the deep neural network can be adjusted according to actual needs. After training, the deep neural network can achieve the same control effect as a predictive controller and can predict the optimal control input of the system based on the real-time state of the intermittent process system. The present invention achieves the same control effect while consuming less system resources.

[0170] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A learning control method for a complex nonlinear intermittent industrial process system, characterized in that: The following steps are involved: S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data; S2: Using the offline data set to train the existing deep Koopman network model, the coefficient matrix and observable function of the discrete state equation are obtained to achieve linearization of the strongly nonlinear batch process system; S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set. S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network; S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

2. A learning control method for a complex nonlinear intermittent industrial process system according to claim 1, characterized in that: The specific process of step S1 is: A strongly nonlinear dynamic model of the batch process system is constructed. The mathematical expression of the dynamic model is: Where u(t) is the control input signal of the batch process system, and x(t) represents all state outputs of the system; Inputting different control input signals into the dynamic model to obtain the operation process data of the strongly nonlinear intermittent process system under different input signals, that is, the state vector during the operation; The state vector and control input signal are collected respectively according to the preset sampling time, and the collected data are combined into an offline data set.

3. The learning control method for a complex nonlinear intermittent industrial process system according to claim 1, characterized in that: The specific process of step S2 is: The Koopman operator is used to linearize the strongly nonlinear batch process system and obtain the linearized discrete state equation. The Koopman network model is trained using the offline data set to obtain an observable function and a coefficient matrix of a linearized discrete state equation.

4. A learning control method for a complex nonlinear intermittent industrial process system according to claim 3, characterized in that: The linearized expression is: z(k+1)=Az(k)+Bu(k) (4a) x(k+1)=Cz(k+1) (4b) Among them, A, B and C represent coefficient matrices, and z(k), u(k), x(k+1), and z(k+1) represent the system linear state vector at time k, the system control signal, the system nonlinear state vector at time k+1, and the system linear state vector at time k+1, respectively.

5. The learning control method for a complex nonlinear intermittent industrial process system according to claim 4, characterized in that: The loss function of the Koopman network model is: Where X = [x(1), x(2), …, x(q)] is the set of open-loop state vectors of the nonlinear system, Z = [z(1), z(2), …, z(q)] = [g(x(1)), g(x(2)), …, g(x(q))] is the set of linear state vectors of the system corresponding to X, q is the size of the collected data set, m is the step size of the forced Koopman operator linear prediction, ||·|| F is the F-norm, ||·||2 is the 2-norm, and k, l, and r are the counting subscripts of the cumulative sum.

6. A learning control method for a complex nonlinear intermittent industrial process system according to claim 5, characterized in that: The linear predictive control model of the strongly nonlinear batch process system is obtained by combining the discrete state equation and the observable function. Its mathematical expression is: z(k+1)=Az(k)+Bu(k) (6a) x(k+1)=Cz(k+1) (6b) y(k)=Dx(k) (6c) z(k)=g(x(k)) (6d) The optimization target is △u, where The cost function J of the model predictive control is: Where p and n are the prediction time domain and control time domain respectively, p=n=15, Q and R are the weight matrices of y and Δu; Combining formulas (6a), (6b), (6c), (6d) and (7) yields the constrained quadratic programming formula: st z(k+1)=Az(k)+Bu(k),k=1,…,n (14b) x(k+1)=Cz(k+1),k=1,…,n (14c) y(k)=Dx(k),k=1,…,n (14d) z(k)=g(x(k)),k=1,…,n (14e) △u(k)=u(k)-u(k-1),k=1,…,n (14f) y min ≤y(k)≤y max ,k=1,…,n (14g) u min ≤y(k)≤u max ,k=1,…,n (14h) Among them, (14b)-(14h) are constraints, y min 、y max 、u min and u max are the boundary constraints of y and u respectively. The offline predictive control data set is constructed using process data and control signals as follows: Different tracking targets y are set for the linear predictive control model of the strongly nonlinear batch process system. ref , collect the target output y of the controlled object ref , the controlled object output y, the process state x and the control variable u are combined into an offline predictive control data set; The input of the deep neural network includes: the model state variable x(k) at the current time step, the model output y(k), the model control variable u(k), and the model target output y ref (k), the output of the deep learning network is the optimal control quantity u(k+1) for the next time step; The optimal solution of the increment △u is obtained by solving the quadratic programming analytical expression, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear intermittent process system.

7. The learning control method for a complex nonlinear intermittent industrial process system according to claim 1, characterized in that: The number of network layers, the number of neurons in a single layer, the learning rate, and the activation function of the deep neural network are adjusted according to actual needs. After training, the deep neural network can achieve the same control effect as the predictive controller and can predict the optimal control input of the system according to the real-time state of the intermittent process system.

8. A learning control system for a complex nonlinear intermittent industrial process system, characterized in that: The system includes: a memory and a processor. The memory includes a complex nonlinear intermittent industrial process system learning control method program. When the complex nonlinear intermittent industrial process system learning control method program is executed by the processor, the following steps are implemented: S1: Construct a dynamic model of a strongly nonlinear intermittent process system, input a number of input signals into the model, obtain process data of the open-loop operation of the strongly nonlinear intermittent process system under different input signals, and construct an offline data set using the process data and input signal data; S2: Using the offline data set to train the existing deep Koopman network model, the coefficient matrix and observable function of the discrete state equation are obtained to achieve linearization of the strongly nonlinear batch process system; S3: The discrete state equation and the observable function are combined to obtain a linear predictive control model for the strongly nonlinear batch process system. The optimization goal is set, the optimal solution of the optimization goal is calculated, and the optimal solution is used as the control variable of the predictive control model of the mathematical model of the strongly nonlinear batch process system. The process data and control signals are used to construct an offline predictive control data set. S4: Using the offline predictive control data set to train a deep neural network, and replacing the predictive controller of the linear predictive control model of the strongly nonlinear batch process system in step S3 with the trained deep neural network; S5: Use the trained deep neural network output to predict the control input and achieve real-time state predictive control of the batch process system.

Images