Robust optimization iterative learning control method for unequal-length limited batch process
Through the corrected system error update model and linear matrix inequality technology, iterative learning control is optimized, which solves the problems of slow convergence speed and insufficient robustness in the process of unequal batches, and achieves faster system convergence and better constraint processing ability.
Patent Information
- Application Number
- CN202411200913.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-12-06
AI Technical Summary
The traditional iterative learning control method converges slowly during the process of unequal batches, and lacks robust performance and constraint processing capabilities, so it cannot effectively deal with system model uncertainty and physical constraints.
The corrected system error update model is used to define secondary performance indicators, and the performance indicators are converted into convex optimization problems through linear matrix inequality technology, and the iterative input signals are optimized to improve the system convergence speed and robustness.
The optimized iterative input correction signal is obtained at the end of each batch, which significantly improves the system's convergence speed and constraint processing capabilities, ensuring robust performance and sufficient robustness.
Smart Images

Figure CN119087933A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a robust optimization iterative learning control method for unequal-length restricted batch processes, belonging to the technical field of iterative learning control. Background Art
[0002] With the development of economy and the rapid change of market demand, small batch, multi-variety flexible industrial production mode and high value-added, refined products are becoming more and more popular. Therefore, the position of batch (intermittent) process in industrial production is becoming more and more important. This kind of process follows certain operation procedures within a limited time and obtains more products of the same kind through repeated operations. Batch process has been widely used in food (fermentation), chemical, pharmaceutical, plastic processing and other industries.
[0003] Iterative learning control is a control algorithm based on the repetitive characteristics of the process. It is widely used in batch processes and can improve system performance and product quality. In order to obtain perfect tracking performance, traditional iterative learning control requires that the running time of each batch of the system is the same. This restriction ensures that the ILC algorithm can obtain complete historical information to improve / improve the control performance of the next batch. However, in actual applications, due to safety, physical limitations and environmental changes, the system batch duration will change. For example, the injection molding process continuously injects, holds pressure and cools three operating procedures to manufacture hot plastics into plastic products. The switch from the injection period to the holding period is determined by the nozzle pressure. When the pressure is greater than a certain value, the process automatically switches from the injection period to the holding period, resulting in unequal durations for each batch. Therefore, it is of great significance to study iterative learning control of unequal batch processes.
[0004] For the iterative learning control problem of unequal-length batch processes, most of them adopt the traditional P-type and average operator iterative learning control methods. This fixed-structure control algorithm limits the convergence speed. In addition, in actual industrial processes, model uncertainty and physical constraints are bound to exist. Therefore, the system's robust performance and constraint processing capabilities need to be improved. Summary of the invention
[0005] The purpose of the present invention is to disclose a method for defining a quadratic performance index using a modified system error update model. The present application proposes a method for defining a quadratic performance index using a modified system error update model, and at the same time using linear matrix inequality technology to transform the performance index minimization problem into a convex optimization method containing multiple linear matrix inequality constraints, and obtaining an optimized iterative input correction signal at the end of each batch, thereby improving the system convergence speed. Robust optimization iterative learning control method for unequal length restricted batch processes
[0006] To achieve the above object, a robust optimization iterative learning control method for unequal length restricted batch processes is provided, comprising the following steps:
[0007] Step 1: Establish a state space model in the unequal length batch process control system.
[0008]
[0009] Among them, k is a batch, representing the kth run of the model, t∈[0,N k ] is the sampling time, N k is the actual length of the kth batch, x k (t)∈R n is the state of the model, u k (t)∈R is the input of the model, y k (t)∈R is the output of the model, A, B, C are model matrices with corresponding dimensions, and CB≠0, x0 is the initial state of the model;
[0010] Step 2: Convert the state space model in step 1 into a batch representation lifting model, specifically:
[0011] Based on the state space model in step 1, if each batch length can reach the expected batch length N, then the output expected trajectory of the state space model at this time is set to y d (t), convert it into a batch representation boost model:
[0012] y k =Gu k +d k (2)
[0013] in,
[0014] y k =[y k (1)y k (2)…y k (N)] T , is the output of the model;
[0015] u k =[u k (0)u k (1)…u k (N-1)] T , is the input of the model;
[0016] d k =[CA CA 2 … CA N-1 ] T x0, is the perturbation vector of the model;
[0017] To improve the process matrix;
[0018] Define the error improvement vector as:
[0019]
[0020] Actual batch length N k ∈[N m ,N],N m is the minimum batch length, then there exists τ=NN m +1 batch length, each batch length has a probability of occurrence of p1, p2, ..., p τ If the actual batch length is less than N, the actual batch length needs to be padded to N, and the missing output of each batch is replaced by the expected trajectory y d (t) is filled, and the modified error improvement vector is expressed as:
[0021]
[0022] in, is a random matrix;
[0023] Next, we introduce the random variable γ k (t), indicates whether the model has output at time t in batch k, γ k (t) = 1 means that the state space model has output data at time t of the kth batch, and the probability of this event is assumed to be p(t); γ k (t) = 0 means that the state space model has no output data at time t of the kth batch, and the probability of this event is 1-p(t); p(t) is expressed by the probability of the batch length as:
[0024]
[0025] Further, calculate the random matrix M k The mathematical expectation of is:
[0026]
[0027] The error propagation improvement model of the state space model is obtained:
[0028]
[0029] Among them, Δu k+1 =u k+1 -u k , represents the input increment in the batch direction;
[0030] Step 3: According to the lifting model, the saturation constraint, the incremental constraint in the batch direction, and the incremental constraint in the time direction are input into the lifting model;
[0031] Step 4, determine and obtain the uncertain model existing in the error propagation improvement model;
[0032] Step 5, by defining a performance indicator with the constraints in step 3 and substituting the error transfer improvement model into the performance indicator, the optimization problem to be solved for optimizing the iterative input update signal is obtained;
[0033] Step 6, by substituting the solution of the uncertain model in step 4 into the optimization problem in step 5, a minimization problem containing linear matrix inequality constraints is obtained, and the robust optimization iterative learning control input update signal is obtained by solving the problem.
[0034] As a further improvement of the present invention, in step 3:
[0035] According to the lifting model, the corresponding input constraints include:
[0036] Input Saturation Constraint
[0037] u min ≤u k+1 ≤u max (8)
[0038] Among them, u max and u min u k+1 The upper and lower bounds of
[0039] Batch direction input increment constraint:
[0040] Δu min ≤Δu k+1 ≤Δu max (9)
[0041] Among them, Δu max and Δu min They are Δu k+1 The upper and lower bounds of
[0042] Input increment constraint in time direction:
[0043] δu min ≤δu k+1 ≤δu max (10)
[0044] Among them, δu max and δu min They are δu k+1 The upper and lower bounds of δu k+1 Indicates the input increment in the time direction;
[0045] The input saturation constraint, batch-wise input increment constraint, and time-wise input increment constraint are combined into a unified matrix constraint form:
[0046]
[0047] in,
[0048] As a further improvement of the present invention, in step 4,
[0049] The lifting process matrix in the lifting model in Formula 2 changes according to the uncertain parameter vector θ, specifically:
[0050] G(θ)=G0+θ1G1+θ2G2+…+θ p G p (12)
[0051] Where θ=[θ1θ2θ p ] T is the uncertain parameter vector, G0,G1,…,G p represents the p known lifting process matrices in the control system;
[0052] The uncertain models in the lifting process matrix G(θ) include: a polyhedron uncertain model and an ellipsoid uncertain model;
[0053] Among them, when the uncertain parameter vector is:
[0054]
[0055] At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be a polyhedral uncertainty model;
[0056] When the uncertain parameter vector is:
[0057]
[0058] At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be an ellipsoid uncertainty model;
[0059] in, is the center of the ellipsoid, and the upper and lower bounds of the matrix W represent the size and direction of the ellipsoid.
[0060] As a further improvement of the present invention, in step 5, the following secondary performance index is defined according to the constraint conditions in step 3:
[0061]
[0062] Then, define the following constrained maximum-minimum optimization problem, where Q is the error weighting matrix and R is the input increment weighting matrix.
[0063]
[0064] Next, the error propagation model is substituted into the above performance indicators to obtain the optimization problem formula:
[0065]
[0066] By solving the optimization problem of this formula, the optimized iterative input update signal Δu is obtained k+1 .
[0067] As a further improvement of the present invention, in step 6, the uncertainty set of step 4 is substituted into the optimization problem to be solved in step 5, specifically:
[0068] For the polyhedral uncertainty model Θ a The optimization problem is to use the triangle inequality, which is
[0069]
[0070] Therefore, the optimization problem is rewritten as:
[0071]
[0072] Using Schur's complement lemma, we get the following minimization problem with linear matrix inequality constraints:
[0073]
[0074] By solving the minimization problem of this formula, the robust optimization iterative learning control input update signal Δu under polyhedron uncertainty constraints is obtained: k +1 ;
[0075] For the ellipsoid uncertainty model Θ b The optimization problem is as follows:
[0076]
[0077] in, g(Δu k+1 )=-Q 1 / 2 [G1G2G p ]Δu k+1 ,h(Δu k+1 )=R 1 / 2 Δu k+1 ,
[0078] At this time, the optimization problem that needs to be solved is:
[0079]
[0080] Further using Schur's complement lemma and S-Procedure lemma, we get the following minimization problem with linear matrix inequality constraints:
[0081]
[0082] Finally, the minimization problem of this equation is solved to obtain the robust optimization iterative learning control input update signal Δu under the ellipsoid uncertainty constraint: k+1 .
[0083] Compared with the prior art, the present invention has the following beneficial effects:
[0084] (1) As a control algorithm without a fixed structure, the optimized iterative learning control method combines the information of previous batches and the knowledge of the system uncertainty model to obtain the optimized correction information of the iterative control input after each batch, which can effectively improve the system convergence speed and constraint processing capability.
[0085] (2) By combining the description of input constraints and uncertain models, this method can ensure that the method has sufficient robust performance and constraint processing capabilities. BRIEF DESCRIPTION OF THE DRAWINGS
[0086] Figure 1 Robust optimization of iterative learning control flow graphs for unequal-length batch processes.
[0087] Figure 2 Output the variation curves of the 1st, 3rd and 15th batches for the unequal length batch process.
[0088] Figure 3 For the unequal length batch process, enter the change curves at batches 1, 3, and 15.
[0089] Figure 4 This is the curve of the norm error of the model being improved with batch size changes in the process of unequal batch lengths. DETAILED DESCRIPTION
[0090] The present invention is described in detail below in conjunction with the various embodiments shown in the accompanying drawings, but it should be noted that these embodiments are not limitations of the present invention, and any equivalent transformations or substitutions in functions, methods, or structures made by ordinary technicians in the field based on these embodiments are all within the scope of protection of the present invention.
[0091] according to Figure 1The specific implementation mode of the present invention shown in the figure improves the system convergence speed and constraint processing capability by using the optimized iterative learning control method in the control system of the production process, and ensures its sufficient robust performance for subsequent learning operations; the present application proposes a method for defining a quadratic performance index by updating a modified system error model, and at the same time adopts a linear matrix inequality technology to convert the performance index minimization problem into a convex optimization method containing multiple linear matrix inequality constraints, and obtains an optimized iterative input correction signal at the end of each batch, thereby improving the system convergence speed. The robust optimization iterative learning control method for unequal length restricted batch processes specifically includes the following steps:
[0092] Step 1: Establish a state space model for batch processes of unequal lengths.
[0093]
[0094] Among them, k is a batch, representing the kth run of the model, t∈[0,N k ] is the sampling time, N k is the actual length of the kth batch, x k (t)∈R n is the state of the model, u k (t)∈R is the input of the model, y k (t)∈R is the output of the model, A, B, C are model matrices with corresponding dimensions, and CB≠0, x0 is the initial state of the model;
[0095] Step 2: Convert the state space model in step 1 into a batch representation lifting model, specifically:
[0096] Based on the state space model in step 1, if each batch length can reach the expected batch length N, then the output expected trajectory of the state space model at this time is set to y d (t), convert it into a batch representation boost model:
[0097] y k =[y k (1)y k (2)y k (N)] T , is the output of the model;
[0098] u k =[u k (0)u k (1)u k (N-1)] T , is the input of the model;
[0099] dk =[CA CA 2 CA N-1 ] T x0, is the perturbation vector of the model;
[0100] To improve the process matrix;
[0101] Define the error improvement vector as:
[0102]
[0103] Actual batch length N k ∈[N m ,N],N m is the minimum batch length, then there exists τ=NN m +1 batch length, each batch length has a probability of occurrence of p1, p2, ..., p τ If the actual batch length is less than N, the actual batch length needs to be padded to N, and the missing output of each batch is replaced by the expected trajectory y d (t) is filled, and the modified error improvement vector is expressed as:
[0104]
[0105] in, is a random matrix;
[0106] Next, we introduce the random variable γ k (t), indicates whether the model has output at time t in batch k, γ k (t) = 1 means that the state space model has output data at time t of the kth batch, and the probability of this event is assumed to be p(t); γ k (t) = 0 means that the state space model has no output data at time t of the kth batch, and the probability of this event is 1-p(t); p(t) is expressed by the probability of the batch length as:
[0107]
[0108] Further, calculate the random matrix M k The mathematical expectation of is:
[0109]
[0110] The error propagation improvement model of the state space model is obtained:
[0111]
[0112] Among them, Δu k+1 =u k+1 -uk , represents the input increment in the batch direction;
[0113] y k =Gu k +d k (2)
[0114] in,
[0115] y k =[y k (1)y k (2)y k (N)] T , is the output of the model;
[0116] u k =[u k (0)u k (1)u k (N-1)] T , is the input of the model;
[0117] d k =[CA CA 2 …CA N-1 ] T x0, is the perturbation vector of the model;
[0118] To improve the process matrix;
[0119] Define the error improvement vector as:
[0120]
[0121] Actual batch length N k ∈[N m ,N],N m is the minimum batch length, then there exists τ=NN m +1 batch length, each batch length has a probability of occurrence of p1, p2, ..., p τ If the actual batch length is less than N, the actual batch length needs to be padded to N, and the missing output of each batch is replaced by the expected trajectory y d (t) is filled, and the modified error improvement vector is expressed as:
[0122]
[0123] in, is a random matrix;
[0124] Next, we introduce the random variable γ k (t), indicates whether the model has output at time t in batch k, γ k(t) = 1 means that the state space model has output data at time t of the kth batch, and the probability of this event is assumed to be p(t); γ k (t) = 0 means that the state space model has no output data at time t of the kth batch, and the probability of this event is 1-p(t); p(t) is expressed by the probability of the batch length as:
[0125]
[0126] Further, calculate the random matrix M k The mathematical expectation of is:
[0127]
[0128] The error propagation improvement model of the state space model is obtained:
[0129]
[0130] Among them, Δu k+1 =u k+1 -u k , represents the input increment in the batch direction;
[0131] Step 3: Input the saturation constraint, the incremental constraint in the batch direction, and the incremental constraint in the time direction into the system according to the lifting model;
[0132] Considering factors such as safety, actuator limitations and smooth operation, the corresponding input constraints according to the lifting model include:
[0133] Input Saturation Constraint
[0134] u min ≤u k+1 ≤u max (8)
[0135] Among them, u max and u min u k+1 The upper and lower bounds of
[0136] Batch direction input increment constraint:
[0137] Δu min ≤Δu k+1 ≤Δu max (9)
[0138] Among them, Δu max and Δu min They are Δu k+1 The upper and lower bounds of
[0139] Input increment constraint in time direction:
[0140] δu min ≤δu k+1 ≤δu max (10)
[0141] Among them, δu max and δu min They are δu k+1 The upper and lower bounds of δu k+1 Indicates the input increment in the time direction;
[0142] The input saturation constraint, batch-wise input increment constraint, and time-wise input increment constraint are combined into a unified matrix constraint form:
[0143]
[0144] in,
[0145] Step 4, solving the uncertainty model in the error transfer lifting model;
[0146] The system's promotion model contains the following models:
[0147] The lifting process matrix G in the lifting model in Formula 2 changes according to the uncertain parameter vector θ, specifically:
[0148] G(θ)=G0+θ1G1+θ2G2+…+θ p G p (12)
[0149] Here G(θ) is the G in Equation 2, which is the lifting process matrix, where θ = [θ1 θ2 … θ p ] T is the uncertain parameter vector, G0,G1,…,G p Represents p known lifting process matrices in the control system; according to the variation range of the uncertain parameter vector θ, the uncertain model form of the lifting process matrix G(θ) is determined;
[0150] The uncertain models in the lifting process matrix G(θ) include: a polyhedron uncertain model and an ellipsoid uncertain model;
[0151] Among them, when the uncertain parameter vector is:
[0152]
[0153] At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be a polyhedral uncertainty model;
[0154] When the uncertain parameter vector is:
[0155]
[0156] At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be an ellipsoid uncertainty model;
[0157] in, is the center of the ellipsoid, and the upper and lower bounds of the matrix W represent the size and direction of the ellipsoid.
[0158] Step 5, by defining the performance index with the constraints in step 3 and substituting the error propagation improvement model into it, the optimization problem to be solved for optimizing the iterative input update signal is obtained;
[0159] According to the constraints in step 3, the following secondary performance indicators are defined:
[0160]
[0161] Then, define the following maximum-minimum optimization problem with constraints:
[0162]
[0163] Next, the error propagation model is substituted into the above performance indicators to obtain the optimization problem formula:
[0164]
[0165] By solving the optimization problem of this formula, the optimized iterative input update signal Δu is obtained k+1 .
[0166] Step 6: Substitute the solution of the uncertain model in step 4 into the optimization problem in step 5 to obtain a minimization problem with linear matrix inequality constraints. Solve the problem to obtain the robust optimization iterative learning control input update signal, which is:
[0167] when θ i ≥0,
[0168] The polyhedron uncertainty model Θ a Substituting into the optimization problem and using the triangle inequality, we get
[0169]
[0170] Therefore, the optimization problem is rewritten as:
[0171]
[0172] Using Schur's complement lemma, we get the following minimization problem with linear matrix inequality constraints:
[0173]
[0174] By solving the minimization problem of this formula, the robust optimization iterative learning control input update signal Δu under polyhedron uncertainty constraints is obtained: k+1 ;
[0175] when hour,
[0176] The ellipsoid uncertainty model Θ b Substituting into the optimization problem, we get:
[0177]
[0178] in, g(Δu k+1 )=-Q 1 / 2 [G1G2…G p ]Δu k+1 ,h(Δu k+1 )=R 1 / 2 Δu k+1 ,
[0179] At this time, the optimization problem that needs to be solved is:
[0180]
[0181] Further using Schur's complement lemma and S-Procedure lemma, we get the following minimization problem with linear matrix inequality constraints:
[0182]
[0183] Finally, the minimization problem of this equation is solved to obtain the robust optimization iterative learning control input update signal Δu under the ellipsoid uncertainty constraint: k+1 .
[0184] The purpose of this application is to design a robust iterative learning controller under multi-objective requirements. This method uses a modified system error update model to define a quadratic performance index, and uses linear matrix inequality technology to transform the performance index minimization problem into a convex optimization method containing multiple linear matrix inequality constraints, and obtains an optimized iterative input correction signal at the end of each batch, thereby improving the system convergence speed. This method can also combine the description of input constraints and uncertain models, thereby ensuring certain robust performance and constraint processing capabilities. Specific embodiments
[0186] The injection molding system is a typical intermittent process, which generally includes three stages: filling, holding pressure and cooling. In the filling stage, the transfer function of the injection velocity (IV) relative to the valve opening (VO) and the transfer function of the nozzle pressure (NP) relative to the injection velocity (IV) can be obtained by identification method:
[0187]
[0188] Define state x = [IV 0.03191IV-5.617VO NP], the input and output are valve opening (VO) and injection speed (IV), and get the state space model of formula 1
[0189]
[0190] Further consider that the injection molding process system matrix B contains polyhedral uncertainty, that is, The batch length of the injection molding process is randomly selected from the set {45,46,…,50}, and the probability of each batch length event follows a uniform distribution, that is, p i =1 / 6, i = 1, 2, ... 6. The expected trajectory of the batch process is y d (t)=40,t∈{1,2,…,50}, the initial state of the system is x k (0) =
[000] T , the initial input is u0(t)=0,t∈{0,1,,49}, and the input constraint is u max =-u min =6*
[111] T and Δu max =-Δu min =δu max =-δu min =5*
[111] T .
[0191] In order to evaluate the tracking performance, the following error 2 norm index is defined
[0192]
[0193] For the above situation, the controller parameter is selected as Q = 0.1*I 50 , R=10*I 50 , I is the unit matrix, according to Δu k+1 =u k+1 -u k The controller data was adjusted and the system was run iteratively and adjusted continuously. After running 15 batches, the results are shown in the table below. Figure 2-4 .
[0194] Figure 2 Output the change curves of the 1st, 3rd and 15th batches for the batch process, from Figure 2 It can be seen that the system length varies randomly with batches, and the 15th batch output can track the expected trajectory well;
[0195] Figure 3 The change curves of the batch process input in the 1st, 3rd and 15th batches show that the input signals of each batch meet the restricted conditions;
[0196] Figure 4 This is the curve of the change of the model norm error with the batch in the batch process. It can be seen that the system 2 norm error converges quickly to a smaller range, showing the superiority of the designed controller in convergence speed.
[0197] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.
Claims
1. A robust optimization iterative learning control method for unequal length constrained batch processes, characterized in that: The following steps are involved: Step 1: Establish a state space model in the unequal length batch process control system. Among them, k is a batch, representing the kth run of the model, t∈[0,N k ] is the sampling time, N k is the actual length of the kth batch, x k (t)∈R n is the state of the model, u k (t)∈R is the input of the model, y k (t)∈R is the output of the model, A, B, C are model matrices with corresponding dimensions, and CB≠0, x0 is the initial state of the model; Step 2: Convert the state space model in step 1 into a batch representation lifting model, specifically: Based on the state space model in step 1, if each batch length can reach the expected batch length N, then the output expected trajectory of the state space model at this time is set to y d (t), convert it into a batch representation boost model: the k =To k +d k (2) where y k =[y k (1) y k (2) … y k (N)] T , is the output of the model; u k =[u k (0) u k (1) … u k (N-1)] T , is the input of the model; d k =[CA CA 2 … CA N-1 ] T x0, is the perturbation vector of the model; To improve the process matrix; Define the error improvement vector as: Actual batch length N k ∈[N m ,N],N m is the minimum batch length, then there exists τ=NN m +1 batch length, each batch length has a probability of occurrence of p1, p2, ..., p τ If the actual batch length is less than N, the actual batch length needs to be padded to N, and the missing output of each batch is replaced by the expected trajectory y d (t) is filled, and the modified error improvement vector is expressed as: in, is a random matrix; Next, we introduce the random variable γ k (t), indicates whether the model has output at time t in batch k, γ k (t) = 1 means that the state space model has output data at time t of the kth batch, and the probability of this event occurring is assumed to be p(t); γ k (t) = 0 means that the state space model has no output data at time t of the kth batch, and the probability of this event is 1-p(t); p(t) is expressed by the probability of the batch length as: Further, calculate the random matrix M k The mathematical expectation of is: M=diag{E{γ k (1)}E{c k (2)}…E{c k (N m -1)}E{c k (N m )}…E{c k (N)}} =diag{1 1 … 1 p(N m ) … p(Nm)…p(N)} T (6), The error propagation improvement model of the state space model is obtained: Among them, Δu k+1 =u k+1 -u k , represents the input increment in the batch direction; Step 3, according to the lifting model, the saturation constraint, the increment constraint in the batch direction, and the increment constraint in the time direction are input into the lifting model; Step 4, determine and obtain the uncertain model existing in the error propagation improvement model; Step 5, by defining a performance indicator with the constraints in step 3 and substituting the error transfer improvement model into the performance indicator, the optimization problem to be solved for optimizing the iterative input update signal is obtained; Step 6, by substituting the solution of the uncertain model in step 4 into the optimization problem in step 5, a minimization problem containing linear matrix inequality constraints is obtained, and the robust optimization iterative learning control input update signal is obtained by solving the problem.
2. The robust optimization iterative learning control method for unequal length restricted batch processes according to claim 1, characterized in that: In step 3: According to the lifting model, the corresponding input constraints include: Input Saturation Constraint in min in k+1 in max (8), Among them, u max and u min u k+1 The upper and lower bounds of Batch direction input increment constraint: Δu min ≤Δu k+1 ≤Δu max (9), Among them, Δu max and Δu min They are Δu k+1 The upper and lower bounds of Input increment constraint in time direction: δu min ≤δu k+1 ≤δu max (10), Among them, δu max and δu min They are δu k+1 The upper and lower bounds of δu k+1 Indicates the input increment in the time direction; The input saturation constraint, batch-wise input increment constraint, and time-wise input increment constraint are combined into a unified matrix constraint form: in, 3. The robust optimization iterative learning control method for unequal length restricted batch processes according to claim 1, characterized in that: In step 4, The lifting process matrix G in the lifting model in Formula 2 changes according to the uncertain parameter vector θ, specifically: G(θ)=G0+θ1G1+θ2G2+…+…+θ p G p (12), where θ=[θ1 θ2 … θ p ] T is the uncertain parameter vector, G0,G1,…,G p represents the p known lifting process matrices in the control system; The uncertain models in the lifting process matrix G(θ) include: a polyhedron uncertain model and an ellipsoid uncertain model; Among them, when the uncertain parameter vector is: At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be a polyhedral uncertainty model; When the uncertain parameter vector is: At this time, the uncertainty model of the lifting process matrix G(θ) is determined to be an ellipsoid uncertainty model; in, is the center of the ellipsoid, and the upper and lower bounds of the matrix W represent the size and direction of the ellipsoid.
4. The robust optimization iterative learning control method for unequal length restricted batch processes according to claim 3, characterized in that: In step 5, the following secondary performance indicators are defined according to the constraints in step 3: Then, define the following constrained maximum-minimum optimization problem, where Q is the error weighting matrix and R is the input increment weighting matrix. Next, the error propagation model is substituted into the above performance indicators to obtain the optimization problem formula: By solving the optimization problem of this formula, the optimized iterative input update signal Δu is obtained k+1 .
5. The robust optimization iterative learning control method for unequal length restricted batch processes according to claim 4, characterized in that: In step 6, the uncertainty set of step 4 is substituted into the optimization problem to be solved in step 5, which is specifically: For the polyhedral uncertainty model Θ a The optimization problem is to use the triangle inequality, which is Therefore, the optimization problem is rewritten as: Using Schur's complement lemma, we get the following minimization problem with linear matrix inequality constraints: By solving the minimization problem of this formula, the robust optimization iterative learning control input update signal Δu under polyhedron uncertainty constraints is obtained: k+1 ; For the ellipsoid uncertainty model Θ b The optimization problem is as follows: Among them, g(Du k+1 )=-Q 1 / 2 [G1G2…G p ]No k+1 ,h(Δu k+1 )=R 1 / 2 Thu k+1 , At this time, the optimization problem that needs to be solved is: Further using Schur's complement lemma and S-Procedure lemma, we get the following minimization problem with linear matrix inequality constraints: Finally, the minimization problem of this equation is solved to obtain the robust optimization iterative learning control input update signal Δu under the ellipsoid uncertainty constraint: k+1 .
