Data-driven batch production system set point optimization PI learning control method

Abstract

The invention discloses a data-driven batch production system set point optimization PI learning control method, and belongs to the technical field of industrial process control. The invention provides a PI controller parameter setting strategy only based on process input and output data by dynamically linearizing a controlled system model along a time direction. According to an established batch production system model with a closed-loop PI control loop, a self-adaptive set point learning law based on historical batch tracking errors is provided by dynamically linearizing the batch production system model along the batch direction. For unknown gradient vectors and learning gains, an iterative updating algorithm in the batch direction is deduced by optimizing a designed performance index function. Unknown total disturbance formed by non-repetitive load disturbance and initial drift is obtained through online estimation of a data-driven iterative expansion state observer. The method has good theoretical innovation and engineering application values.

Description

Technical Field

[0001] The present invention relates to a control system for a chemical batch production process. Aiming at a batch production system in a complex chemical production process with unknown dynamic response characteristics, non-repetitive load interference and inconsistent initial reset conditions, a data-driven batch production system set point optimization PI learning control method is proposed based on data-driven control and parameter estimation theory. The method is a novel set point optimization PI learning control method based on batch production process data, and belongs to the technical field of industrial process control. Background Art

[0002] With the increasing scale and complexity of modern industrial batch production processes, it is often time-consuming and laborious to establish mathematical models of complex processes, and sometimes even difficult to achieve. This makes it difficult to apply existing control and optimization methods based on process dynamics models. In recent years, data-driven control and batch operation optimization methods that directly use real-time and historical batch data of the controlled process to design controllers have received widespread attention. At present, researchers have been committed to exploring data-driven open-loop iterative learning control methods. However, in practical applications, many batch production processes are often pre-configured with closed-loop control loops, such as P-type and PI-type closed-loop feedback control, or require additional closed-loop feedback control loops to ensure the control performance of the initial batch. For such batch optimization methods for nonlinear production systems with closed-loop feedback control loops, only a few results have been publicly reported or published. For example, the recent paper Enhanced P-type control: indirect adaptive learning from set-point updates (IEEE Transactions on Automatic Control, 2023, 68(3): 1600-1613.) proposes an indirect data-driven set-point learning control method based on a closed-loop P-type feedback loop.

[0003] In addition, there are a wide variety of non-repetitive load disturbances and prominent problems such as inconsistent initial reset conditions during the operation of batch production systems. In recent years, some batch optimization methods have been developed for batch production systems with the above-mentioned non-repetitive uncertainties. For example, the reference Robust optimization-based iterative learning control for nonlinear systems with nonrepetitive uncertainties (IEEE / CAA Journal of Automatica Sinica, 2021, 8(5): 1001-1014.) proposes a data-driven robust iterative learning control method to achieve batch operation optimization by directly adjusting the controlled process control input.

[0004] However, at present, for batch production systems with non-repetitive load disturbances and inconsistent initial reset conditions, how to establish a robust data-driven set point optimization method based on a closed-loop PI feedback control loop needs further research. The relevant results have good theoretical innovation and engineering application value. Summary of the invention

[0005] The technical problem to be solved by the present invention is the data-driven set point optimization PI learning control problem of a batch production system with non-repetitive load disturbances and inconsistent initial reset conditions. To solve the above problems, a data-driven adaptive set point learning control structure based on a closed-loop PI control loop is designed, and a data-driven PI controller parameter tuning strategy and an adaptive set point learning control design method are given.

[0006] The present invention is based on the real-time and historical batch data of the controlled process that can be detected in the actual production engineering, and proposes a data-driven set point learning control method based on a closed-loop PI control loop for a discrete time domain batch production system model with non-repetitive load interference and inconsistent initial reset conditions. By dynamically linearizing the system model along the time direction, a PI controller parameter tuning strategy based only on process input and output data is proposed. According to the established batch production system model with a closed-loop PI control loop, an adaptive set point learning law based on historical batch tracking errors is proposed by dynamically linearizing it along the batch direction. For unknown gradient vectors and learning gains, an update algorithm along the batch direction is derived by optimizing the designed performance index function. For the unknown total disturbance composed of non-repetitive load interference and initial drift, it is estimated online using a data-driven iterative extended state observer. The outstanding advantage of the present invention is that the designed PI controller parameters, gradient vectors, adaptive learning gains, and total disturbances can all be estimated online through real-time and historical batch data of the production process, without relying on the dynamic model of the controlled production system, which is conducive to practical application. Moreover, based on the compression mapping principle and mathematical induction, sufficient conditions are given to ensure the bounded convergence of the closed-loop learning system.

[0007] The technical solution of the present invention:

[0008] Data-driven PI learning control method for set point optimization of batch production systems, including:

[0009] (1) Data-driven PI controller tuning

[0010] The PI controller tuning strategy provided by the present invention is based only on the input and output measurement data of the controlled process and does not rely on any model information of the controlled process. The controlled production system is dynamically linearized along the time direction, and the original system can be equivalently described as an incremental linearized system with an unknown time-varying parameter vector. By optimizing the performance index function related to the control input and tracking error, an incremental PI controller related to the unknown time-varying parameters is established. The time-varying gradient vector is estimated using a parameter estimation algorithm to determine the feasible tuning interval of the PI controller parameters. Then, the time-invariant PI controller parameters are determined within the feasible interval using an online optimization algorithm.

[0011] (2) Adaptive data-driven set-point learning controller based on closed-loop PI control loop

[0012] The adaptive data-driven set point learning controller provided by the present invention can be used for a batch production system with a closed-loop PI control loop and completely unknown dynamic response characteristics, which is different from the existing data-driven set point learning control method based on a closed-loop P-type control loop and a direct data-driven iterative learning control method designed based on an open-loop batch production system. The set point learning controller designed by the present invention mainly includes a gradient vector estimation part, a total disturbance estimation part, a learning gain estimation part, and a set point instruction update part.

[0013] The gradient vector estimation algorithm designed by the present invention estimates the gradient vector along the batch direction using the input and output data of the controlled process. The convergence speed of the parameter estimation can be adjusted by adjusting the weight parameter and the step size parameter in the algorithm.

[0014] The total disturbance estimation algorithm designed by the present invention is based on the linearized data model of the controlled process. Compared with the existing extended state observer based on fixed input gain, the iterative extended state observer designed by the present invention has a time-varying input gain and is better suitable for batch production systems with non-repetitive process uncertainties. In addition, the gain of the observer can be analytically solved by a simple pole configuration method. By configuring different pole positions of the observer, the estimation accuracy and speed of the unknown time-varying total disturbance can be adjusted.

[0015] The learning gain estimation algorithm designed by the present invention establishes a parameter estimation algorithm along the batch direction by optimizing the performance indicator function related to the learning gain and the tracking error. Its outstanding advantage is that the estimated value of the unknown total disturbance is included in the parameter estimation algorithm, realizing active compensation for the disturbance, thereby improving the estimation accuracy of the adaptive learning gain. This is different from the method given in the existing literature, that is, the disturbance term is directly omitted for simplicity.

[0016] The set point command update designed by the present invention is executed by a proportional learning controller with saturation constraints. Compared with the set point learning control method with fixed learning gain, the learning gain of the present invention is adaptive and can be better applied to batch production systems with unknown dynamic response characteristics, especially for the situation with non-repeatable uncertainty. In addition, the set point learning law also includes an adjustable tuning parameter, which can be used to adjust the convergence speed of the tracking error.

[0017] Beneficial effects of the present invention: Based on the above design, the present invention has better set point tracking performance for batch production systems with unknown dynamic response characteristics, non-repetitive load interference and inconsistent initial reset conditions, and can improve the control effect of existing batch operation optimization methods on batch production systems with unknown dynamic response characteristics, especially improve the tracking performance of the controlled process in the initial batch operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is the principle diagram of the control system of the present invention.

[0019] Figure 2 The results of the method of the present invention and the other two control algorithms in the embodiment are compared. Among them, (a) represents the system output response curves of different batches without external interference and initial drift, (b) represents the corresponding average tracking error (ATE) index and maximum set point signal curve, and (c) represents the ATE index and maximum set point signal curve under non-repetitive load interference and initial reset inconsistency. DETAILED DESCRIPTION

[0020] In order to better understand the technical solution of the present invention, the embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0021] The data-driven PI learning control method for batch production system set point optimization has the following steps:

[0022] Step 1: Data-driven PI controller tuning

[0023] The system model is described as follows

[0024]

[0025] in, and are time and batch indices, respectively, N is the total time steps of each batch, f(·) is an unknown linear or nonlinear function, and n u and n y represent the unknown process input and output orders, y k (t) and u k (t) is the process output and input of the kth batch at time t, ω k (t) is a non-repetitive load disturbance, y 0 is the repeatable initial reset condition, δ k is the non-repetitive initial drift. Since the time domain PI controller tuning is independent of the batch index k, it is omitted here for simplicity of notation.

[0026] The controlled process shown in formula (1) is dynamically linearized along the time direction into the following form:

[0027] △ t y(t+1)=Γ(t)△ t H(t) (2)where △ t H(t)=[△ t y(t)△ t u(t)] T ,△ t y(t)=y(t)-y(t-1),△t u(t)=u(t)-u(t-1), y(t) and u(t) are the process output and input in the time domain, Γ(t)=[γ 1 (t)γ 2 (t)] is the unknown time-varying pseudo gradient vector, γ 1 (t) and γ 2 (t) are the pseudo partial derivatives of the unknown dynamic function of the process with respect to y(t) and u(t), respectively.

[0028] In order to establish the specific description form of the PI controller, the following performance index function J is designed: 1 (u(t)):

[0029] J 1 (u(t))=|y d (t+1)-y(t+1)| 2 +λ t |u(t)-u(t-1)| 2 (3)

[0030] Among them, λ t >0 is the setting parameter, which is the weight parameter specified by the user to adjust the change of the time domain control input; d (t) is the desired reference trajectory. Taking the first-order derivative of the performance index function shown in formula (3) and setting the derivative to zero, we can obtain the control input update law △ t u(t) is:

[0031]

[0032] Among them, ρ t ∈(0,2) is the tuning parameter. When the set point command y d When (t) is a constant or a slowly changing signal, the control input update law shown in formula (4) can be approximately described as a PI controller in the following differential form:

[0033]

[0034] in,

[0035] e(t)=y d (t)-y(t),△ t e(t)=e(t)-e(t-1),

[0036]

[0037] In order to tune the PI controller parameter τ P (t) and τ I (t), design the following parameter estimation algorithm to estimate the unknown parameters γ online 1 (t) and γ2 (t):

[0038]

[0039] Among them, η t ∈(0,2) and μ t >0 are two tuning parameters, represents the estimated value of Γ(t), and Respectively represent γ 1 (t) and γ 2 The estimated value of (t), ε t is a sufficiently small threshold for algorithm initialization, and sign(·) is the sign function.

[0040] According to the estimated and The feasible tuning range of the PI controller parameters can be determined as follows:

[0041]

[0042] in

[0043]

[0044] In order to facilitate the subsequent adaptive set point learning control design and analysis, the following iterative algorithm is designed to determine the desired time-invariant PI controller parameters within the feasible tuning interval of the PI controller parameters.

[0045] Step 1: Select the initial PI controller parameters as the minimum value of the feasible tuning interval, that is, and and the appropriate tuning parameter λ t ,ρ t ,η t ,μ t , the search step length △τ of the PI controller parameters P and △τ I , and the initial reset threshold ε in the parameter estimation algorithm shown in formula (6) t and algorithm termination threshold

[0046] Step 2: When and When , the control input update law shown in formula (4) and the parameter estimation algorithm shown in formula (6) are applied to the controlled system shown in formula (1) to calculate the average tracking error

[0047] Step 3: If Then the expected PI controller parameter τ is obtained P and τI ; Otherwise, update the parameter τ I =τ I +△τ I , return to step 2 and execute until

[0048] Step 4: If in step 3 If it still does not hold, update the parameter τ P =τ P +△τ P , and reset Return to step 2 and execute until

[0049] Step 5: If Then we get the expected τ P and τ I ; Otherwise, select the PI controller parameter corresponding to the minimum ATE as its expected value;

[0050] Considering the actual implementation, the initial value of the adjustment parameter can be selected as: t ,μ t ∈[0.1,5],ρ t ,η t ∈[0.01,1.99],ε t ∈[10 -6 ,10 -8 ],

[0051] Step 2: Adaptive data-driven set-point learning control design based on closed-loop PI control loop

[0052] The controlled process shown in formula (1) is dynamically linearized along the batch direction into the following form:

[0053]

[0054] in, and They are the dynamic response function of the unknown process to u k (t),ω k (t-1) and y k The partial derivative of (0), is the unknown total disturbance consisting of the non-repetitive load disturbance and the initial drift

[0055]

[0056] △u k (t)=[△u k (0),△u k (1),...,△u k(t)] T ,△u k (t) = u k (t)-u k-1 (t)

[0057] △ω k (t)=[△ω k (0),△ω k (1),...,△ω k (t)] T ,△ω k (t) = ω k (t)-ω k-1 (t)

[0058] △δ k =δ k -δ k-1

[0059] According to the PI controller parameters determined in step 1, the control input of the controlled process is designed to be

[0060]

[0061] Where τ = τ P +τ I , is the set point tracking error, is the set point signal.

[0062] In order to achieve batch performance optimization, the following adaptive set point learning update law is designed

[0063]

[0064] Among them, k (t) = αθ k (t)e k-1 (t+1), θ k (t) is the adaptive set point learning gain to be designed, α is the user-specified tuning parameter used to adjust the convergence speed of the learning system, and δ sp >0 is the user-specified upper bound on the rate of change of the setpoint signal in the batch direction.

[0065] To estimate the learning gain θ k (t), design the following performance indicator function

[0066] J 2 (θ k (t))=|e k (t+1)| 2 +λ|θ k (t)-θ k-1 (t)|2 (11)

[0067] Where λ>0 is a tuning parameter used to adjust the learning gain θ k (t) The rate of change along the batch direction. Taking the first-order derivative of the performance index function shown in formula (11) and setting the derivative to zero, we can get

[0068]

[0069] in, ρ∈(0,2) is a tuning parameter.

[0070] For the unknown gradient vector φ in formula (12) k,k-1 (t), can be obtained by the following parameter estimation algorithm

[0071]

[0072] in, Yes k,k-1 (t), μ>0 and η∈(0,2) are two tuning parameters, ε 1 is a sufficiently small threshold for algorithm initialization.

[0073] In addition, in order to estimate the unknown total disturbance χ in formula (12): k (t), the dynamic linearized data model shown in formula (8) is described as follows:

[0074]

[0075] Where, C = [1 0], Y k (t+1)=[y k (t+1)χ k (t)] T ,

[0076]

[0077] For the generalized system shown in formula (14), the following iterative extended state observer is designed:

[0078]

[0079] Where L = [l 1 l 2 ] T is the observer gain vector, l 1 and l 1 are two adjustable observer gains, which can be analytically obtained by configuring the characteristic roots of A-LC at the desired positions. Yes k-1The estimated value of (t+1).

[0080] Based on estimates and The implementable form of the parameter estimation algorithm shown in formula (12) is described as:

[0081]

[0082] in, yes The estimated value of 2 is a sufficiently small threshold for algorithm initialization.

[0083] The specific implementation process of adaptive data-driven set point learning control based on closed-loop PI control loop is given below:

[0084] First, select the initial learning gain θ 0 (t), the gradient vector Control input u 0 (t), and the process output y 0 (t), iterative extended state observer state Expected output trajectory y d , PI controller parameter τ P and τ I , observer gain l 1 and l 2 , tuning parameters η, μ, ρ and α, weight parameter λ, user-specified threshold ε 1 ,ε 2 and δ sp , batch length N and maximum batch size k max ;

[0085] Then, the gradient vector is estimated by formula (13), formula (15) and formula (16) in turn. Total disturbance And the adaptive learning gain θ k (t), and then update the set point command signal through formula (10) Simultaneously calculates the set point tracking error

[0086] Finally, the PI controller shown in formula (9) is applied to the batch process shown in formula (1) to calculate the output tracking error e k (t) and ATE performance indicators.

[0087] Considering the actual execution, the initial values ​​of the adjustment parameters can be selected as: λ,μ∈[0.1,5],ρ,η∈[0.01,1.99],ε 1 ,ε 2∈[10 -6 ,10 -8 ],δ sp ∈[10,50],α∈[0.01,10].

[0088] The control system corresponding to the above control method is as follows: Figure 1 shown. Figure 1 In, τ P and τ I Respectively represent the proportional and integral gains of the PI controller; ∑ represents the summation operation; u k (t),y k (t), and e k (t) represent the control input signal, process output, set point signal, set point tracking error and output tracking error of the kth batch at time t, respectively; d (t) represents the desired reference trajectory; θ k (t) represents the learning gain of the kth batch at time t; Represents the estimated value of the gradient vector at time t of the kth batch; represents the total disturbance estimate at time t of the k-1th batch; △y k (·) represents the difference between the output of the kth batch process and the output of the k-1th batch process; the memory is used to record the output tracking error (e) of the current batch (k) k (t)), set point command Process input (u k (t)), process output (y k (t)), gradient estimate Also provides historical batch information and information about the current batch The gradient estimation algorithm uses the information stored in memory △u k-1 (t) and △y k-1 (t+1) Estimation along the batch direction The iterative extended state observer estimates the And the information stored in the memory △u k (t) and y k-1 (t+1) Estimate the total disturbance along the batch direction The learning gain estimation algorithm uses the information stored in the memory △y k (t),△u k (t-1),θ k-1 (t) and e k-1 (t+1), and the total disturbance estimated in the iterative extended state observer Update the learning gain θ along the batch directionk-1 (t); The set point learning update law uses the historical batch tracking error information stored in the memory e k-1 (t+1) and learning gain θ k-1 (t) Adjust the set point command of the current batch in the batch direction With the system output y k (t) The difference produces a set point tracking error Then the control signal u is generated by the closed-loop PI controller k (t) Act on the controlled process.

[0089] Step 3: Simulation Verification

[0090] Consider the following nonlinear discrete batch production process with time-varying parameters, time-varying order, non-repetitive load disturbances, and inconsistent initial conditions:

[0091]

[0092] Among them, the time-varying parameter a(t) = round(t / 50), and the load interference is The initial reset condition is y k (0) = -1 + δ k , δ 1,k (t) and δ k represents a random variable in the interval [-0.05, 0.05].

[0093] Given the following desired tracking trajectory

[0094]

[0095] And the tuning parameter λ t =0.15,ρ t =0.75,η t =1,μ t =0.2, and ε t =10 -8 , using the parameter algorithm (6) and the control input update law (4), the feasible tuning interval of the PI controller parameters can be determined as τ P ∈[0.01,0.87] and τ I ∈[0.3,0.97]. Select the step size parameter △τ P =△τ I =0.01, using the iterative algorithm in step 1, we can get the time-invariant PI controller parameter τ P =0.01, τ I =0.46.

[0096] In addition, the parameters are selected as η=0.1,μ=0.6,ρ=1,λ=1,α=0.7,ε1 =ε 2 =10 -7 , the initial values ​​of gradient and learning gain are set to θ 0 (t) = 1. The specific implementation process of the adaptive data-driven set point learning control based on the closed-loop PI control loop in step 2 is applied to the above nonlinear batch production system, and the corresponding control results are as follows: Figure 2 shown.

[0097] For comparison, the control methods IDL-iAILC (Enhanced P-type control: indirect adaptive learning from set-point updates (IEEE Transactions on Automatic Control, 2023, 68(3): 1600-1613.) and OAILC (Design and analysis of data-driven learning control: An optimization-based approach. IEEE Transactions on Neural Networks and Learning Systems, 2022, 33(10): 5527-5541.) in the references are also applied to the above nonlinear batch production system, where the parameters are set as:

[0098] IDL-iAILC: η = 0.1, μ = 0.6, η nl =1,μ nl =1,α=3.6,ε 2 =ε 3 =10 -7 ;

[0099] OAILC:λ=1,μ 1 =1,μ 2 =0.001,γ 1 =0.8,γ 2 =0.16,γ 3 =0.04, ε=0.01 .

[0100] Figure 2 The control effect of the method of the present invention on a controlled object case is compared with the other two control algorithms, and the results are obtained by simulation using MATLAB software. Note that the comparison method and the method of the present invention select the same gradient and initial value of the learning gain to make an objective and fair comparison. Figure 2 In the example, the desired reference trajectory is the following piecewise time-varying signal

[0101]

[0102] Non-repetitive load disturbance and initial reset condition is ω k (t) = 0.2sin(t / 70+k / 50)+δ 1,k (t) and y k (0) = -1 + δ k , δ 1,k (t) and δ k Represents the uncertainty of random changes in the interval [-0.05, 0.05]. Among them, (a) represents the system output response curves of different batches without external disturbance and initial drift, (b) represents the corresponding average tracking error (ATE) index and maximum set point signal curve, and (c) represents the ATE index and maximum set point signal curve under non-repetitive load disturbance and initial reset inconsistency.

[0103] from Figure 2 It can be seen from the control results shown that the batch optimization method of the present invention has a faster reference trajectory tracking speed and a minimum steady-state tracking error.

Claims

1. A data-driven PI learning control method for set point optimization of batch production systems, characterized in that: Here are the steps: Step 1: Data-driven PI controller tuning The system model is described as follows in, and are time and batch indices, respectively, N is the total time steps of each batch, f(·) is an unknown linear or nonlinear function, and n u and n y represent the unknown process input and output orders, y k (t) and u k (t) is the process output and input of the kth batch at time t, ω k (t) is the non-repetitive load disturbance, y0 is the repeatable initial reset condition, δ k is the non-repetitive initial drift; the controlled process shown in formula (1) is dynamically linearized along the time direction into the following form: D t y(t+1)=Γ(t)Δ t H(t) (2) Among them, Δ t H(t)=[Δ t y(t)Δ t u(t)] T , Δ t y(t)=y(t)-y(t-1), Δ t u(t) = u(t) - u(t-1), y(t) and u(t) are the process output and input in the time domain, respectively, Γ(t) = [γ1(t)γ2(t)] is the unknown time-varying pseudo gradient vector, γ1(t) and γ2(t) are the pseudo partial derivatives of function f with respect to y(t) and u(t), respectively; In order to establish the specific description form of the PI controller, the following performance index function J1(u(t)) is designed: J1(u(t))=|y d (t+1)-y(t+1)| 2 +λ t |u(t)-u(t-1)| 2 (3) Among them, λ t >0 is the setting parameter, which is the weight parameter specified by the user to adjust the change of the time domain control input; d (t) is the desired reference trajectory; taking the first-order derivative of the performance index function shown in formula (3) and setting the derivative to zero, we get the control input update law Δ t u(t) is: Among them, ρ t ∈(0,2) is the tuning parameter; when the set point command y d When (t) is a constant or a slowly changing signal, the control input update law shown in formula (4) is described as a PI controller in the following differential form: in, e(t)=y d (t)-y(t),Δ t e(t)=e(t)-e(t-1), In order to tune the PI controller parameter τ P (t) and τ I (t), the following parameter estimation algorithm is designed to estimate the unknown parameters γ1(t) and γ2(t) online: Among them, η t ∈(0,2) and μ t >0 are two setting parameters, represents the estimated value of Γ(t), and denote the estimated values ​​of γ1(t) and γ2(t), ε t is the threshold used to initialize the algorithm, sign(·) is the sign function; According to the estimated and Determine the feasible tuning range of the following PI controller parameters in In order to facilitate the subsequent adaptive set point learning control design and analysis, the following iterative algorithm is designed to determine the desired time-invariant PI controller parameters within the feasible tuning interval of the PI controller parameters: Step 1: Select the initial PI controller parameters as the minimum value of the feasible tuning interval, that is, and And the tuning parameter λ t ,ρ t ,η t ,μ t , the search step length Δτ of the PI controller parameters P and Δτ I , and the initial reset threshold ε in the parameter estimation algorithm shown in formula (6) t and algorithm termination threshold Step 2: When and When , the control input update law shown in formula (4) and the parameter estimation algorithm shown in formula (6) are applied to the controlled system shown in formula (1) to calculate the average tracking error Step 3: If Then the desired PI controller parameter τ is obtained P and τ I ; Otherwise, update the parameter τ I =τ I +Δτ I , return to step 2 of the algorithm and execute until Step 4: If in step 3 If it still does not hold, update the parameter τ P =τ P +Δτ P , and reset Return to step 2 and execute until Step 5: If Then we get the expected τ P and τ I ; Otherwise, select the PI controller parameter corresponding to the minimum ATE as its expected value; Step 2: Adaptive data-driven set point learning control design based on closed-loop PI control loop The controlled process shown in formula (1) is dynamically linearized along the batch direction as follows: in, and They are the dynamic response function of the unknown process to u k (t),ω k (t-1) and y k The partial derivative of (0), is the unknown total disturbance consisting of the non-repetitive load disturbance and the initial drift Δu k (t)=[Δu k (0),Δu k (1),...,Δu k (t)] T ,Δu k (t)=u k (t)-u k-1 (t) Give k (t)=[See k (0),See k (1),...,Do k (t)] T ,Here k (t)=ω k (t)-ω k-1 (t) According to the PI controller parameters determined in step 1, the control input of the controlled process is designed to be Where τ = τ P +τ I , is the set point tracking error, is the set point signal; in order to achieve batch performance optimization, the following adaptive set point learning update law is designed Among them, k (t) = αθ k (t)e k-1 (t+1), θ k (t) is the adaptive set point learning gain to be designed, α is the user-specified tuning parameter used to adjust the convergence speed of the learning system, and δ sp >0 is the upper limit of the rate of change of the set point signal in the batch direction specified by the user; To estimate the learning gain θ k (t), design the following performance indicator function J2(θ k (t))=|e k (t+1)| 2 +λ|θ k (t)-θ k-1 (t)| 2 (11) Where λ>0 is a tuning parameter used to adjust the learning gain θ k (t) The rate of change along the batch direction; Taking the first-order derivative of the performance index function shown in formula (11) and setting the derivative to zero, we can get in, is the setting parameter; For the unknown gradient vector φ in formula (12) k,k-1 (t), obtained by the following parameter estimation algorithm in, Yes k,k-1 (t), μ>0 and η∈(0,2) are two tuning parameters, and ε1 is a sufficiently small threshold for algorithm initialization; In addition, in order to estimate the unknown total disturbance χ in formula (12): k (t), the dynamic linearized data model shown in formula (8) is described as follows: where C = [1 0], Y k (t + 1) = [y k (t + 1)χ k (t)] T , For the generalized system shown in formula (14), the following iterative extended state observer is designed: Where L = [l1 l2] T is the observer gain vector, l1 and l2 are two adjustable observer gains, which are obtained analytically by configuring the characteristic roots of A-LC at the desired positions; Yes k-1 The estimated value of (t+1); Based on estimates and The implementable form of the parameter estimation algorithm shown in formula (12) is described as: in, yes An estimate of , ε2 is a sufficiently small threshold for algorithm initialization; The specific implementation process of adaptive data-driven set point learning control based on closed-loop PI control loop is given below: First, select the initial learning gain θ0(t), the gradient vector Control input u0(t), and process output y0(t), iterative extended state observer state Expected output trajectory y d , PI controller parameter τ P and τ I , observer gains l1 and l2, tuning parameters η, μ, ρ and α, weight parameter λ, user-specified thresholds ε1, ε2 and δ sp , batch length N and maximum batch size k max ; Then, the gradient vector is estimated by formula (13), formula (15) and formula (16) in turn. Total disturbance And the adaptive learning gain θ k (t), and then update the set point command signal through formula (10) Simultaneously calculates the set point tracking error Finally, the PI controller shown in formula (9) is applied to the batch process shown in formula (1) to calculate the output tracking error e k (t) and ATE performance indicators.

Images