Intelligent Control Method and System for Real-time Optimization of High-end Equipment Operation

Abstract

The present invention discloses an intelligent control method and system for real-time optimization of high-end equipment operation. The method is as follows: in the early stage of the working condition cycle, the input and output data of the roasting furnace system are extracted and transmitted to the central control platform; the central control platform performs system identification based on the input and output data of the roasting furnace system, and then calculates the explicit MPC control law offline according to the obtained state space model; the central control platform transmits the state space model and the explicit MPC control law to the edge device bound to the roasting furnace; in the remaining stages of the working condition cycle, the roasting furnace transmits the latest output quantity to the edge device in real time; the edge device calculates the current state quantity of the roasting furnace according to the state space model, substitutes the current state quantity of the roasting furnace into the explicit MPC control law to obtain the optimal control quantity; the edge device transmits the optimal control quantity to the roasting furnace and acts on the control input end of the roasting furnace. The present invention can ensure the real-time performance of the optimization control on the premise of meeting the control accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial intelligent control, and particularly relates to an intelligent control method and system for real-time optimization of high-end equipment operation. Background Art

[0002] At present, the main research directions of industrial intelligent control include model-driven, knowledge-driven, data-driven, etc. Taking the zinc smelting roasting process as an example, the model-driven control method requires sufficient prior knowledge. However, the complexity of the roasting furnace reaction makes it difficult to obtain its mechanism model in reality. At the same time, as an optimization of manual control, scholars such as Rauma et al. naturally proposed a rule-based fuzzy controller, which uses human experience for control and is applied to a 72-square-meter roasting furnace in Kokkola Zinc Plant. With the development of industrial control technology, some complex algorithms have also been proposed by scholars to replace manual control. For example, the trend-based event-triggered fuzzy control algorithm proposed by Feng et al. can be used for temperature stability control of large roasting furnaces. Although the above knowledge-driven control methods have achieved certain application effects, due to the complex process characteristics of the roasting furnace and the large fluctuations in the raw material supply composition, the reaction atmosphere and operating conditions often change frequently, making it difficult to control accurately. On the other hand, due to the complex process and numerous units, the effects of different adjustment variables on on-site control are different. To simplify the treatment, the on-site often only roughly adjusts the air volume and feed rate, while ignoring other key adjustment variables, resulting in limited control effects. Moreover, since the quality of the product zinc calcine cannot be detected in real time, the criterion for judging the control effect of the roasting furnace only considered whether the temperature in the furnace was stable within the set range in the past, and it was impossible to accurately judge the quality of roasting. Therefore, the knowledge-driven control method for roasting furnaces is difficult to ensure control performance.

[0003] With the wide application of information systems such as industrial Internet and distributed control systems, the industrial site already has a good data foundation, enabling the gradual development of data-driven control methods. Model predictive control is a data-driven process control method that achieves control through rolling optimization. With years of development, model predictive control has become a mature control algorithm for controlling controllable variables and output parameters in multivariable systems. In recent years, MPC has developed rapidly in fields such as the automotive and aerospace industries, intelligent energy networks, and financial engineering, especially achieving long-term success in the process industry. Although the MPC method can achieve control, the on-site distributed control system (DCS system) cannot support complex control algorithms, resulting in difficulties in implementing advanced control strategies. Summary of the Invention

[0004] The present invention provides an intelligent control method and system for real-time optimization of high-end equipment operation, which can ensure the real-time nature of optimal control on the premise of meeting control accuracy.

[0005] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0006] An intelligent control method for real-time optimization of high-end equipment operation, comprising:

[0007] In the early stage of the working condition cycle, the input and output data of the roasting furnace system are extracted and transmitted to the central control platform; the central control platform performs system identification based on the input and output data of the roasting furnace system, and then calculates the explicit MPC control law offline according to the obtained state space model; the central control platform transmits the state space model and the explicit MPC control law to the edge device bound to the roasting furnace;

[0008] In the remaining stages of the working condition cycle, the roasting furnace transmits the latest output in real time to the edge device; the edge device calculates the current state quantity of the roasting furnace according to the state space model, substitutes the current state quantity of the roasting furnace into the explicit MPC control law to obtain the optimal control quantity; the edge device transmits the optimal control quantity to the roasting furnace and acts on the control input end of the roasting furnace.

[0009] Further, before performing system identification based on the input and output data, the central control platform first performs standardization and denoising preprocessing on the received data, and then uses the preprocessed data for system identification.

[0010] Further, for the denoising preprocessing, a moving average filtering algorithm is used to weaken the Gaussian noise interference in the input and output data, expressed as:

[0011]

[0012] where k F is the forward sliding window size, and s(t) is a row vector in the input data set matrix or output data set matrix.

[0013] Further, after the input and output data, the central control platform first uses the classical canonical correlation analysis method to screen the input variables, and then uses the data of the input variables as the input data to identify the state space model.

[0014] Further, the specific method for screening the control quantity by using the classical canonical correlation analysis method is as follows:

[0015] (1) In the input data set U and the output data set Y, find the linear combinations of the variables respectively, and denote the correlation coefficients of the linear combinations as: use a pair of correlation coefficients (α (i) , β (i) ) to represent the relationship between the canonical variables (μ i , ν i ) and the input data set U and the output data set Y as:

[0016]

[0017] Then, according to the following calculation formula, based on μ i and ν i the correlation coefficient between them, that is, the canonical correlation coefficient Corr(Uα, Yβ), is solved to obtain the correlation coefficients α (i) and β (i) :

[0018]

[0019] where, Σ UU and Σ YY are the covariance matrices of U and Y respectively, and Σ UY is the covariance matrix of U and Y;

[0020] (2) Select the next pair of canonical variables and repeat step (1) until all the canonical correlations between U and Y are extracted;

[0021] (3) Conduct a significance test on each pair of selected canonical variables, and select several pairs of canonical variables with higher canonical correlation coefficients. The input variables among them are the controlled variables obtained by screening.

[0022] Furthermore, the explicit MPC control law is calculated offline according to the identified state - space model, specifically as follows:

[0023] (1) Introduce the weight matrix R, the weight matrix P, and the weight matrix Q to weight the controlled variable, the output variable, and the terminal output variable y(N p ) in the model predictive control system respectively, and add the roasting index operator μ to set the importance of each index in the output variable, and construct the following optimization problem:

[0024]

[0025] In the formula, N p is the prediction horizon, N c is the control horizon. Let N p = N c = N, and P≥0, Q≥0 are positive semi - definite, and R>0 is positive definite, is the set control target;

[0026] is the controlled variable of the system, which is used as the input data of the system to control the system, is the output variable of the system, is the process state variable of the system, n u , n y and n x are the dimensions of the controlled variable, the output variable, and the state variable respectively, is the system matrix of the state - space model, obtained through system identification;

[0027] h(x t ,u t ,y t ) represents the constraints on the state variables, control variables, and output variables; Denote the current time as time zero, i.e., x(0) is the current state variable;

[0028] (2) Use intermediate variables respectively represent the matrices constructed by the output variables, control variables, and coefficients in the optimization problem (4):

[0029]

[0030] Among them,

[0031] For the constraint h(x t ,u t ,y t ) ≤ 0, since the conversion relationship between the known input - output variables and state variables is known, transform it into an expression only about the control variable; Then, combined with the above - simplified intermediate variables and delete the constants in the objective function, the optimization problem (4) is equivalent to the following optimization problem (5):

[0032]

[0033] (3) Considering the relationship between the state variable and the control variable, combine x(0) and u(0) in the optimization problem (5) into x(1), and denote x(1) as x(0), i.e., x(0) is the known quantity at the current time, and further describe the optimization problem (5) as the following optimization problem (6):

[0034]

[0035] Among them, while the dimensions of the coefficient matrices G, W, and S are determined by the number of constraints;

[0036] (4) Based on the fact that the optimization problem (6) is a differentiable convex optimization problem with constraints, obtain the analytical solution through the KKT conditions, and then the explicit MPC control law can be obtained.

[0037] Furthermore, the method to obtain the explicit MPC control law through the KKT conditions is as follows:

[0038] First, determine the KKT conditions of the optimization problem (6) as follows:

[0039]

[0040] Among them, the control variables with superscript * represent the optimal control sequence, the variables with subscript a represent the effectively activated constraints, and the variables with subscript i represent the ineffective and unactivated constraints. λ is the Lagrange multiplier used to associate the constraints and the objective function;

[0041] Then, the equations in the KKT conditions (7) of the multi-parameter quadratic programming problem are combined to obtain the following equation:

[0042]

[0043] Then, the least squares method is used to solve and in Equation (8), which is expressed as the following optimization problem:

[0044]

[0045] Finally, through the optimization problem (9), and explicit expressions f that are linearly related to x(0) u and f λ are obtained. These two explicit expressions are substituted into the inequality conditions in the KKT conditions to obtain the linear programming region expression for x(0), that is, the feasible region of f u , which is the key region CR of the explicit control law. Then, f u and the key region are correspondingly represented to obtain the explicit MPC control law as:

[0046]

[0047] Among them, for is the coefficient matrix of the control law, k i is the dimension of the i-th key region, and n CR is the number of key regions.

[0048] Furthermore, the output variables of the state space model include the temperature of the roasting furnace and the output sulfur dioxide concentration.

[0049] Furthermore, the edge device is composed of microchips.

[0050] An intelligent control system for real-time optimization of high-end equipment operation includes a central control platform and an edge device;

[0051] The central control platform: In the early stage of the working condition cycle, it obtains the input and output data from the roasting furnace system, performs system identification based on the input and output data of the roasting furnace system, then calculates the explicit MPC control law offline according to the identified state space model, and then transmits the state space model and the explicit MPC control law to the edge device bound to the roasting furnace;

[0052] The edge device: In the remaining stages of the operating cycle, obtain the latest output from the roasting furnace in real time, then calculate the current state quantity of the roasting furnace according to the state space model, substitute the current state quantity of the roasting furnace into the explicit MPC control law to obtain the optimal control quantity, and then transmit the optimal control quantity to the roasting furnace and act on the control input end of the roasting furnace.

[0053] Beneficial effects

[0054] The present invention proposes an intelligent control method and system for real-time optimization of high-end equipment operation. The model predictive control method is adopted and the control method is made explicit, and finally deployed on an edge device composed of microchips. In view of the complex situation of the control quantity at the input end of the roasting furnace, a method for extracting key controllable variables is proposed. Compared with the traditional method that only roughly considers the feed quantity, more and more reliable key control variables are selected. And based on the key control variables, the system model is identified, which can better fit the operating conditions of the roasting furnace. Furthermore, an explicit MPC scheme based on data driving is proposed, which simplifies the online calculation process and speeds up the solution of the optimal control quantity while solving the multi-input multi-output model stable control problem. Finally, the solution is deployed to the microchip to achieve real-time control under the condition of small computing power, and the control effect is excellent.

[0055] The present invention does not require an accurate mechanism model of the equipment. Only based on the complex and coupled equipment operation process data, the modeling and control of the equipment can be completed; it can be deployed on a low-cost and weak-computing-power edge device hardware platform, and realize intelligent control for real-time optimization of high-end equipment operation. Description of the drawings

[0056] Figure 1 is the overall architecture diagram of the control method described in the embodiment of the present application;

[0057] Figure 2 is the physical framework diagram of the chip-side implementation solution architecture of the control method described in the embodiment of the present application;

[0058] Figure 3 is the network framework diagram of the chip-side implementation solution architecture of the control method described in the embodiment of the present application;

[0059] Figure 4 is the flowchart based on the chip solution in the experiment described in the embodiment of the present application;

[0060] Figure 5 is the comparison diagram of the output quantity temperature and sulfur dioxide concentration with the corresponding set values in the experiment described in the embodiment of the present application. Detailed implementation manners

[0061] The following is a detailed description of the embodiments of the present invention. Based on the technical solutions of the present invention, detailed implementation manners and specific operation processes are given to further explain and illustrate the technical solutions of the present invention.

[0062] Embodiment 1

[0063] I. Overall framework

[0064] This embodiment provides an intelligent control method for real-time optimization of high-end equipment operation. Referring to Figure 1 as shown, the model predictive control method is adopted and the control method is made explicit, and finally deployed on the edge device composed of microchips to achieve the goal of simultaneously meeting the accuracy and real-time performance of the control of the key quality parameters of the large-scale roasting furnace. Specifically, different from the general roasting furnace control strategy, the controlled object is not only the temperature in the furnace, but also includes the sulfur dioxide concentration at the output end, and the control quantity at the input end not only considers the feeding rate, but the key controllable quantity is selected as the control quantity through canonical correlation analysis (CCA), and then the subspace identification (N4SID) method is used to identify the state space model of the roasting furnace online. In view of the good performance of the model predictive control algorithm (MPC algorithm) in the roasting furnace control, the present invention introduces the MPC algorithm into the operation optimization control of the roasting furnace, and further proposes to use its optimized form - explicit MPC to achieve the purpose of improving the online calculation performance. Moreover, the explicit MPC control law and the control scheme for solving the state quantity are deployed on the microchip hardware platform to simulate the working conditions of the industrial site.

[0065] II. Process data analysis and processing

[0066] There are numerous process parameters in the roasting furnace system, and there are as many as dozens of overall measurable parameters. If no screening is carried out and all parameters are incorporated into the input-output model, it will inevitably affect the accuracy of the model and further lead to the failure of the controller. At the same time, too many parameters and a bloated model will greatly increase the computational complexity and affect the timeliness of control. However, there are mostly coupling relationships among the input parameters of the roasting furnace, which not only increases the difficulty of screening the input parameters for modeling but also is not conducive to the design of the controller. Therefore, in this embodiment, for the following 12 main parameters of the roasting furnace: the set value of the feeding rate of conveyor belt No. 1 (t / h), the set value of the feeding rate of conveyor belt No. 2 (t / h), the measured value of the feeding rate of conveyor belt No. 1 (t / h), the measured value of the feeding rate of conveyor belt No. 2 (t / h), the operating frequency feedback of the first throwing machine (Hz), the operating current of the first throwing machine (A), the operating frequency feedback of the second throwing machine (Hz), the operating current of the second throwing machine (A), the current of the blower (A), the air blowing rate (m3 / h), the air box pressure (Pa), the temperature of the intake air pipe of the roasting furnace (°C), the typical correlation analysis method is used to screen several key controllable variables with the greatest correlation with the roasting quality parameters (standard temperature and sulfur dioxide concentration): the set value of the feeding rate of conveyor belt No. 1, the set value of the feeding rate of conveyor belt No. 2, the operating current of the first throwing machine, the operating current of the second throwing machine, the air blowing rate, and the air box pressure.

[0067] 1) Data preprocessing: There are significant numerical differences among the main parameters of the roasting furnace. When directly using the original data for canonical correlation analysis, the correlation coefficient is greatly affected by the data size. Therefore, it is necessary to normalize the data. Let U = (u 1 , u 2 ,..., u p ) be the observable parameter data at the input end, Y = (y 1 , y2,..., y q ) be the given modeling output parameter data, and S = (s 1 , s 2 ,..., s p+q ) be the overall data set of modeling input and output. Standardize the data set S:

[0068]

[0069] where u i , y i , and s i are column vectors, the vector length is the number of data set samples N, and after calculating the mean of the column vectors and expanding it to the size of the original column vector, it is denoted as The function sum() sums the elements of the column vector, and σ is the standard deviation.

[0070] Moreover, the roasting furnace is a high-noise system, and the proportion of dirty data in the collected on-site data is relatively high, which will affect the accuracy of the identification model. Therefore, the moving average filtering algorithm is used to weaken the Gaussian noise interference in the training data. Denote the size of the forward sliding window as k F , where s(t) is a row vector in the S dataset matrix:

[0071]

[0072] 2) Canonical correlation analysis: CCA respectively finds the linear combinations of variables in the input data U and the output data Y, and denote the correlation coefficients of the linear combinations as α (i) and β (i) :

[0073]

[0074] In the formula, the pair of μ i and ν i is called the canonical variable, and the correlation coefficient between them is called the canonical correlation coefficient.

[0075]

[0076] Among them, Σ UU and Σ YY are the covariance matrices of U and Y respectively, and Σ UY is the covariance matrix of U and Y. Then, similarly, select the next pair of canonical variables (μ i , ν i ), and repeat the above solution process until all the canonical correlations between U and Y are completely extracted. Then, conduct a significance test on each pair of selected canonical variables, and discard the variables with lower canonical correlation coefficients. Since the variables with larger absolute values of the correlation coefficients in the linear combination have a greater impact on their canonical variables. For example, the variables with larger absolute values of the correlation coefficients in μ i have a decisive impact on μ i , and since μ i and ν i are highly correlated, the variables in μ i also have a great impact on ν i . Furthermore, these key variables in μ i are also highly related to ν iThe correlation of the key variables in it is significantly high. Therefore, several input variables with relatively large absolute values of the linear correlation coefficient among the canonical variables with relatively large canonical correlation coefficients should be selected as candidate key variables. Since the variables at the input end of the roasting furnace are coupled, there is a certain correlation among the currently selected candidate key variables, and two variables with a high degree of linear correlation should not be used as input variables for modeling at the same time, and reducing the input variables helps to reduce the complexity of modeling and control. Therefore, the candidate key variables are reduced according to the results of the correlation analysis, and at the same time, uncontrollable variables are removed based on the physical characteristics of the variables. To sum up, in this embodiment, appropriate key controllable quantities are screened out among the complex and coupled input end variables, which can reduce the model complexity while meeting the modeling accuracy requirements.

[0077] III. Explicit Model Predictive Control Implementation

[0078] For the complex reaction process in the roasting furnace, a simple linear model cannot provide a relatively accurate description, while complex intelligent algorithms or nonlinear models will greatly increase the complexity and computational amount of the control algorithm. To balance the model accuracy and the computational complexity of the control algorithm, this embodiment selects the state space model (5) as the system input-output model.

[0079]

[0080] Where is the input control quantity of the system, is the output measurement value of the system, is the process state quantity of the system, n u 、n y and n x are the dimensions of the control quantity, output quantity and state quantity respectively, is the system matrix of the model.

[0081] The mechanism of model predictive control is that at each sampling moment, according to the currently detectable system state value, a finite-time open-loop optimization problem is solved online, and the first value of the obtained control sequence is applied to the controlled system. At the next sampling moment, the above operation is repeated, and the optimization problem is updated with the new state value and solved again to achieve rolling optimization.

[0082] For the control problems of the temperature and the output sulfur dioxide concentration of the roasting furnace, it can be described as an optimization problem of minimizing the performance index. To distinguish the importance of the control quantity and the output quantities at different stages, the control weight matrix R, the terminal weight matrix P, and the output weight matrix Q are introduced for weighting. In addition, considering the actual situation of the roasting furnace control, the standard temperature in the output variables directly affects the reaction process in the furnace, while SO 2The concentration directly evaluates the quality of the calcined product. Therefore, a calcination index operator μ is added to set the importance of each index (standard temperature and SO 2 concentration) in the output. Thus, the optimization problem is as follows

[0083]

[0084] where N p is the prediction horizon, N c is the control horizon. Generally, N p ≥N c . To balance accuracy and complexity, N p is selected to be equal to N c , i.e., N and P≥0, Q≥0 are positive semi-definite, and R>0 is positive definite. is the set control target. y - y(set) can be regarded as a whole for solution. h(x t , u t , y t ) represents the constraints on the input control quantity, output quantity, and state quantity. And the current time is denoted as time zero, i.e., x(0) is the current state quantity.

[0085] To simplify the expression, let's assume

[0086]

[0087] where For the constraint h(x t , u t , y t ) ≤ 0, since the transformation relationship between the input, output, and state quantities is known, it can finally be transformed into an expression only about the control quantity. According to the above simplified symbols and deleting the negligible constants in the objective function, the optimization problem is equivalent to

[0088]

[0089] After obtaining this quadratic programming optimization problem, each step of control requires online solving an optimization problem, which requires a relatively large computing power. When the MPC algorithm is deployed on a platform with weak computing power, it is difficult to control the system in a timely manner. To meet the real-time requirements of industrial control, in this embodiment, the MPC algorithm is explicitly processed to reduce the online computing complexity. Simply put, explicit model predictive control is based on traditional model predictive control, transforming the process of online optimizing and solving the control quantity into an offline mode. In other words, the optimization problem is transformed into a function about the state quantity, and this function is a piece-wise affine (PWA) function.

[0090] For simplicity of description and considering the relationship between the state variables and the control variables, this embodiment describes the optimization problem in the following form

[0091]

[0092] Among them, x(0) and u(0) in the optimization problem (8) can be combined into x(1). However, for the convenience of understanding, in the optimization problem (9), x(1) is denoted as x(0), that is, x(0) is a known quantity at the current moment. At the same time The dimensions of G, W, and S are determined by the number of constraints, and the above variables are all known parameter matrices. The roasting furnace control problem, as a differentiable convex optimization problem with constraints, can obtain its analytical solution through the KKT (Karush-Kuhn-Tucker) conditions. The KKT conditions of the above optimization problem (9) are as follows

[0093]

[0094]

[0095] Among them, the control variable with the superscript * represents the optimal control sequence, the variable with the subscript a represents the effectively activated constraint, and the variable with the subscript i represents the ineffective and unactivated constraint. λ is the Lagrange multiplier used to associate the constraint and the objective function. This patent introduces an explicit MPC method to explicitly process the optimization problem, and the goal is to obtain the relational expression of the control sequence with respect to the current state variable. According to the KKT conditions of this multi-parameter quadratic programming problem, the following equations can be obtained

[0096]

[0097] This equation is the combined form of the equality conditions in the KKT conditions and is used to obtain and The expressions of. If the matrix G a is row full rank, the analytical solution of equation (11) can be obtained. And this equation can be solved by the least squares method

[0098]

[0099] Through the optimization problem (12), we can obtain and The explicit expressions f that are linearly related to x(0) u and f λ . Substituting these two explicit expressions into the inequality conditions in the KKT conditions, we can obtain the linear programming region expression with respect to x(0), that is, the feasible region of f u . These feasible regions are called the critical regions (CR) of the explicit control law. Substitute fu The explicit control law of the system can be obtained by correspondingly representing it with CR.

[0100]

[0101] Among them, for is the coefficient matrix of the control law, and k i is the dimension of the i-th key area, and n CR is the number of key areas. During the process of rolling control, at each step, the current state variables are substituted into the explicit control law, and the optimal control sequence is obtained through look-up tables. Then, the first set of control quantities in the optimal sequence is applied to the controlled object to obtain new state variables, and this process is repeated for rolling optimization.

[0102] IV. Chip-side Deployment

[0103] The chip-side implementation scheme architecture of the optimized control of the industrial roasting furnace is as shown in the physical framework diagram and the network framework diagram respectively in Figure 2 and 3 During a working condition cycle, first, the input and output data of the roasting furnace system in the early stage of the current working condition cycle are extracted and transmitted to the central control platform of the factory for system identification. According to the obtained state-space model, its explicit MPC control law is calculated offline. Then, the state-space model and the explicit control law are transmitted to the edge device bound to the roasting furnace for storage. After that, during this working condition cycle, the roasting furnace transmits the latest output quantity to the hardware platform, calculates the current state variables according to the state-space model, substitutes them into the explicit MPC control law to obtain the optimal control quantity, and then applies the control quantity to the control input end of the roasting furnace, and so on in a cycle. The above control scheme can deploy relatively complex control algorithms on chips with limited computing power. This algorithm has a small amount of calculation and can also achieve the goal of real-time control on low-cost microchips.

[0104] V. Detection and Experimental Verification Data

[0105] In order to verify the feasibility, accuracy, and real-time performance of the proposed solution of the present invention, an industrial experiment of a roasting furnace based on a microchip was designed in this experiment.

[0106] This experiment used the historical operation data of the on-site roasting furnace of Zhuzhou Smelter in Hunan Province, China to simulate the actual operation conditions inside the roasting furnace during a certain period. According to expert knowledge, when the temperature inside the roasting furnace is 929.5 degrees Celsius and the sulfur dioxide concentration at the output end is 6.73%, the operation condition of the roasting furnace is in a good state. Therefore, the control target was set to the above indicators. At the same time, different control frequencies will affect the control performance. Without considering the large time delay problem of the roasting furnace, the present invention can adjust the control frequency to 1 min, which is the same as the sampling frequency, to achieve better control effects. Considering the complexity of solving the optimization problem by model predictive control, the control step size N was set cand the prediction step length N p Both are 5. In the explicit MPC objective function, the output weight Q = 40I, the control weight R = 4I, and the terminal weight P = lqr. The roasting index operator μ is set to diag(1, 2).

[0107] Using the method of canonical correlation analysis to analyze 1440 groups of experimental data, 2 pairs of canonical variables with correlation coefficients of 0.7432 and 0.4906 are obtained. The results of canonical correlation analysis are shown in the following table.

[0108] Table Results of Canonical Correlation Analysis

[0109]

[0110] Select the input variables at the input end of the roasting furnace that have a greater impact on the output quality parameters of the roasting furnace in the table as the modeling candidate input variables, namely the set value of the feeding rate of conveyor belt No. 1, the set value of the feeding rate of conveyor belt No. 2, the operating current of the first throwing machine, the operating current of the second throwing machine, the blowing rate, the air box pressure, and the temperature of the air inlet pipe of the roasting furnace. Then, through correlation analysis, the linear correlation of these modeling candidate input parameters is judged, and the variable pairs with large correlation are removed. However, the correlation coefficient between the blowing rate and the air box pressure, which has the relatively largest correlation coefficient among these variables, is 0.7838, which is not large, and both the blowing rate and the air box pressure can be controlled by adjusting the blower, so neither of them can be discarded. And the temperature of the air inlet pipe of the roasting furnace is an uncontrollable quantity and is discarded. Finally, the adjustable and controllable input variables of the zinc roasting ore quality model for modeling are determined as the set value of the feeding rate of conveyor belt No. 1, the set value of the feeding rate of conveyor belt No. 2, the operating current of the first throwing machine, the operating current of the second throwing machine, the blowing rate, and the air box pressure. Therefore, the input dimension of the model is determined to be 6, the output dimension is 2, and the model form is a state space equation with 6 inputs and 2 outputs.

[0111] The hardware platform selected for this experiment is the VisionFive single-board computer produced by StarFive. It is an economical RISC-V computer designed to run Linux and is completely open-source, with open software, open hardware design, and an open RISC-V architecture.

[0112] This experiment adopts the idea of separating the control algorithm from the controlled object. The control algorithm is deployed on the VisionFive development board, while the controlled object is encapsulated in MATLAB on another computer. Data transmission between the two is carried out through Ethernet based on the UDP communication protocol. This solution develops the explicit model predictive control algorithm code based on C / C++, with a short development cycle and good code reusability. The flow chart of this chip-based solution is as Figure 4 shown, where the roasting furnace body is replaced by the MATLAB simulation system.

[0113] With the operation of the hardware-in-the-loop simulation platform, the state variables, control variables, output variables of each iteration of the system, as well as the running time, can be obtained, as shown in Table 1 below.

[0114] Table 1 Online calculation time of the control algorithm

[0115]

[0116] Although the running time of each iteration of the controller on the hardware-in-the-loop simulation platform is relatively long, objectively speaking, this running time is still much less than the sampling period and the control period. Therefore, this controller meets the real-time requirements of the system. In terms of accuracy, as Figure 5 shown, there is a slight deviation between the control stable value and the set value, but the deviation is small. The furnace internal standard temperature error is within ±0.88°C, and the output SO 2 concentration error is within ±0.01%. Therefore, it is still within an acceptable range. In summary, the experiments on the hardware-in-the-loop simulation platform verify that the solution of the present invention has good control tracking performance and real-time performance.

[0117] The explicit MPC control solution for optimizing the operation of the roasting furnace based on data-driven proposed by the present invention can ensure the real-time performance of the optimal control on the premise of meeting the control accuracy. This method does not require an accurate mechanism model, and only based on the complex and coupled equipment operation process data, the modeling and control of the equipment can be completed. This solution can be deployed on a low-cost and low-computing-power edge device hardware platform, and realize intelligent control for the real-time optimization of high-end equipment operation.

[0118] The above embodiments are the preferred embodiments of the present application. Those of ordinary skill in the art can also make various transformations or improvements on this basis. Without departing from the general concept of the present application, these transformations or improvements should all fall within the scope of protection required by the present application.

[0119] References:

[0120] [1] Feng Z, Li Y, Sun B, et al. A trend-based event-triggering fuzzy controller for the stabilizing control of a large-scale zinc roaster[J]. Journal of Process Control, 2021, 97(9): 59-71.

[0121] [2]Ping, Zhou, Heda, et al. Data-Driven Nonlinear Subspace Modeling for Prediction and Control of Molten Iron Quality Indices in Blast Furnace Ironmaking[J]. IEEE Transactions on Control Systems Technology, 2017.

[0122] [3]Bemporad A, Morari M, Dua V, et al. The explicit linear quadratic regulator for constrained systems[J]. Automatica, 2002, 38(1): 3-20.

[0123] [4]Karg B, Lucia S. Efficient representation and approximation of model predictive control laws via deep learning[J]. 2018.

[0124] [5]Hardoon D R, Szedmak S, Shawe-Taylor J. Canonical Correlation Analysis: An Overview with Application to Learning Methods[J]. Neural Computation, 2004, 16(12): 2639-2664.

[0125] [6]Feng Z, Liu G, Li Y, et al. A novel mechanism model of the fluidized bed roaster in the zinc roasting process[J]. 2021.

[0126] [7]Chen H,Xu F,Xi Y.Field programmable gate array / system on aprogrammable chip-based implementation of model predictive controller[J].IETControl Theory and Applications,2012,6(8):1055-1063.

[0127] [8]Ferreau H J,Bock H G,Diehl M.An online active set strategy toovercome the limitations of explicit MPC[J].International Journal of Robust&Nonlinear Control,2008,18(8):816-830.

[0128] [9]Pannocchia G,Rawlings J B,Wright S J.Fast,large-scale modelpredictive control by partial enumeration[J].Automatica,2007,43(5):852-860.

Claims

1. An intelligent control method for real-time optimization of high-end equipment operation, characterized in that, it includes: In the early stage of the working condition cycle, extract the input and output data of the roasting furnace system and transmit it to the central control platform; The central control platform performs system identification based on the input and output data of the roasting furnace system, and then calculates the explicit MPC control law offline according to the obtained state space model; The central control platform transmits the state space model and the explicit MPC control law to the edge device bound to the roasting furnace; In the remaining stages of the working condition cycle, the roasting furnace transmits the latest output in real time to the edge device; The edge device calculates the current state quantity of the roasting furnace according to the state space model, substitutes the current state quantity of the roasting furnace into the explicit MPC control law to obtain the optimal control quantity; the edge device transmits the optimal control quantity to the roasting furnace and acts on the control input end of the roasting furnace; Among them, the explicit MPC control law is: Among them, for is the coefficient matrix of the control law, k i is the dimension of the i-th critical area, n CR is the number of critical areas, x(0) is the current state quantity, n x is the dimension of the state quantity.

2. The intelligent control method for real-time optimization of high-end equipment operation according to claim 1, characterized in that, Before performing system identification based on the input and output data, the central control platform first performs standardization and denoising preprocessing on the received data, and then uses the preprocessed data for system identification.

3. The intelligent control method for real-time optimization of high-end equipment operation according to claim 2, characterized in that, For the denoising preprocessing, the moving average filtering algorithm is used to weaken the Gaussian noise interference in the input and output data, expressed as: where k F is the size of the forward sliding window, and s(t) is a row vector in the input data set matrix or the output data set matrix.

4. The intelligent control method for real-time optimization of high-end equipment operation according to claim 1, characterized in that, After the input and output data, the central control platform first uses the classical canonical correlation analysis method to screen the input variables, and then uses the data of the input variables as the input data to identify the state space model.

5. The intelligent control method for real-time optimization of high-end equipment operation according to claim 4, characterized in that, The method of screening control quantities by using the classical canonical correlation analysis method is specifically: (1) Find the linear combinations of variables in the input data set U and the output data set Y respectively, and denote the correlation coefficients of the linear combinations as: Use a pair of correlation coefficients (α (i) , β (i) ) to represent the relationship between the canonical variables (μ i , ν i ) and the input data set U and the output data set Y as: Then, according to the following calculation formula, solve the correlation coefficient α i and ν i between the correlation coefficient, that is, the canonical correlation coefficient Corr(Uα, Yβ), and solve the correlation coefficients α (i) and β (i) : Among them, Σ UU and Σ YY are the covariance matrices of U and Y respectively, and Σ UY is the covariance matrix of U and Y; (2) Select the next pair of canonical variables, and repeat step (1) until all the canonical correlations between U and Y are extracted; (3) Perform a significance test on each pair of selected canonical variables, and select several pairs of canonical variables with higher canonical correlation coefficients, and the input variables among them are the screened control quantities.

6. The intelligent control method for real-time optimization of high-end equipment operation according to claim 1, characterized in that, The method of calculating the explicit MPC control law offline according to the obtained state space model is specifically: (1) Introduce the weight matrices \(R\), \(P\), and \(Q\) to weight the control quantity, output quantity, and terminal output quantity \(y(N p ) in the model predictive control system respectively, and add the roasting index operator \(\mu\) to set the importance of each index within the output quantity, and construct the following optimization problem: where N p is the prediction horizon, N c is the control horizon, and let N p = N c = N, and P≥0, Q≥0 are positive semi - definite, R>0 is positive definite, is the set control objective; is the control quantity of the system, which is used to control the system as the input data of the system, is the output quantity of the system, is the process state quantity of the system, n u n, y n, x and n are the dimensions of the control quantity, output quantity and state quantity respectively, is the system matrix of the state space model, which is obtained through system identification; h(x t ,u t ,y t ) represents the constraints on the state variables, control variables, and output variables; denote the current moment as the zero moment, that is, x(0) is the current state variable; (2) Use an intermediate variable respectively represent each matrix constructed by the output quantity, control quantity, and coefficient in the optimization problem (4): Among them, For the constraint h(x t , u t , y t ) ≤ 0, since the conversion relationship between the known input and output quantities and the state quantity is known, it is transformed into an expression only about the control quantity; then combined with the above intermediate variables and deleting the constants in the objective function, the optimization problem (4) is equivalent to the following optimization problem (5): (3) Considering the relationship between the state quantity and the control quantity, merge x(0) and u(0) in the optimization problem (5) into x(1), and denote x(1) as x(0), that is, x(0) is the known quantity at the current moment, and further describe the optimization problem (5) as the following optimization problem (6): Among them, while the dimensions of the coefficient matrices G, W, and S are determined by the number of constraints; (4) Based on the fact that the optimization problem (6) is a constrained differentiable convex optimization problem, obtain the analytical solution through the KKT conditions, and the explicit MPC control law can be obtained.

7. The intelligent control method for real-time optimization of high-end equipment operation according to claim 6, characterized in that, The method of obtaining the explicit MPC control law through the KKT conditions is: First, the KKT conditions of the optimization problem (6) are determined as follows: Among them, the control variable with the superscript * represents the optimal control sequence, the variable with the subscript a represents the effectively activated constraint, and the variable with the subscript i represents the ineffectively unactivated constraint. λ is the Lagrange multiplier used to associate the constraint and the objective function; Then, the equations in the KKT conditions (7) of the multi-parameter quadratic programming problem are combined to obtain the following equation: Then use the least squares method to solve for and in Equation (8), which is expressed as the following optimization problem: Finally, by optimizing problem (9), we obtain and explicit expressions f that are linearly related to x(0) u and f λ . Substitute these two explicit expressions into the inequality conditions in the KKT conditions to obtain the linear programming region expression for x(0), that is, the feasible region of f u , which is the key region CR of the explicit control law; then represent f u and the key region correspondingly to obtain the explicit MPC control law.

8. The intelligent control method for real-time optimization of high-end equipment operation according to claim 1, characterized in that The output variables of the state space model include the temperature of the roasting furnace and the output sulfur dioxide concentration.

9. The intelligent control method for real-time optimization of high-end equipment operation according to claim 1, characterized in that The edge device is composed of microchips.

10. An intelligent control system for real-time optimization of high-end equipment operation, characterized in that It includes a central control platform and an edge device; The central control platform: In the early stage of the working condition cycle, it obtains the input and output data from the roasting furnace system, and performs system identification based on the input and output data of the roasting furnace system. Then, it calculates the explicit MPC control law offline according to the identified state space model, and then transmits the state space model and the explicit MPC control law to the edge device bound to the roasting furnace; The edge device: In the remaining stages of the working condition cycle, it obtains the latest output from the roasting furnace in real time, then calculates the current state quantity of the roasting furnace according to the state space model, substitutes the current state quantity of the roasting furnace into the explicit MPC control law to obtain the optimal control quantity, and then transmits the optimal control quantity to the roasting furnace and acts on the control input end of the roasting furnace; Among them, the explicit MPC control law is: Among them, for is the coefficient matrix of the control law, k i is the dimension of the i-th critical area, n CR is the number of critical areas, x(0) is the current state quantity, n x is the dimension of the state quantity.

Images