Steel ball mill powder preparation self-adaptive control system and control method
By combining the adaptive control system decoupling compensator and the incremental PID controller, the multivariable coupling problem of the steel ball mill pulverizing system is solved, and the stable automatic operation and efficient energy consumption management of the pulverizing system are realized.
Patent Information
- Application Number
- CN202410814427.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-06-24
- Publication Date
- 2025-12-26
AI Technical Summary
The ball mill grinding system suffers from severe coupling of multiple controlled variables, making it difficult for automatic control to operate stably for a long period of time and to maintain the maximum output operating point.
An adaptive control system is adopted, which utilizes a neural network model estimator and a self-tuning controller to achieve decoupled control of the coal mill outlet temperature, inlet negative pressure, and inlet-outlet differential pressure through a decoupling compensator, and combines an incremental PID controller for adaptive adjustment.
This system enables long-term, stable, and automated operation of the ball mill powder-making system, reducing the amount of manual operation, improving powder-making efficiency, and reducing energy consumption.
Smart Images
Figure CN121198441A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application belongs to the technical field of ball mill pulverizing control of boiler, in particular to a self-adaptive control system and method for ball mill pulverizing. BACKGROUND
[0002] In thermal power plants, the main function of the pulverizing system of the boiler is to grind the coal into coal powder which matches the combustion of the boiler, so as to ensure the stable and safe operation of the boiler and the economy thereof. The ball mill is a key equipment in the pulverizing system, which functions to dry and crush the coal into coal powder. The overall control requirement of the ball mill pulverizing system is to control the outlet temperature of the coal mill, the inlet negative pressure of the coal mill and the pressure difference between the inlet and outlet of the coal mill within the specified range, while maintaining a high pulverizing efficiency. The power consumption of the ball mill is high, and the power consumption of the ball mill accounts for about 20% of the plant power consumption. By means of effective automatic control method, the coal mill is kept running at the maximum output, so as to reduce the power consumption, which is a difficult problem to be solved in thermal power plants.
[0003] The ball mill pulverizing system is a nonlinear, strongly coupled, time-varying and large-inertia multivariable complex control object, and the actual operation process of the coal mill is difficult to be described by an accurate mathematical model. The ball mill pulverizing system has many uncertain interference factors, and any amount of the operation variables such as the opening degree of the hot air door, the opening degree of the recirculation air door and the coal supply amount will affect the outlet temperature, the inlet negative pressure, the coal storage amount and the pressure difference between the inlet and outlet. In the operation process of the ball mill, the opening degree of the hot air door, the opening degree of the recirculation air door and the speed of the coal feeder are adjusted to ensure that the outlet temperature of the coal mill, the inlet negative pressure of the coal mill and the coal storage amount (the load of the coal mill) are within the reasonable range, so that the coal mill is kept running at the maximum output.
[0004] At present, the automatic control of the pulverizing system of most thermal power units in China usually adopts three independent PID loops, i.e. the hot air door controls the outlet temperature of the coal mill, the recirculation air door controls the inlet negative pressure of the coal mill, and the speed of the coal feeder controls the load of the coal mill. This design logic is simple and easy to implement, but there is a serious coupling in the actual control loop, and the system without decoupling cannot obtain good control quality. For such a complex multivariable system, the automatic control cannot be stably put into operation for a long time; due to the time-varying nonlinearity of the controlled object, the system is stable under a certain working condition, but unstable under other working conditions, and cannot obtain satisfactory control effect. SUMMARY
[0005] The application aims to provide a steel ball grinding and powdering adaptive control system and a control method, solve the problem of serious coupling of multiple controlled variables of the steel ball grinding and powdering system, solve the problem of long-term stable automatic operation of the steel ball grinding and powdering system, and solve the problem of keeping the maximum output working point of the steel ball grinding and powdering system.
[0006] The application aims to achieve the above-mentioned purpose through the following technical solutions.
[0007] The steel ball grinding and powdering adaptive control system comprises a self-correcting controller, a steel ball grinding and powdering system, and a neural network model estimator.
[0008] The neural network model estimator comprises an input layer, a hidden layer, and an output layer.
[0009] The self-correcting controller comprises a PID controller, a parameter setting, and a decoupling compensator.
[0010] The decoupling compensator is composed of a characteristic vector space matrix transpose, the decoupling compensator and the control object form a generalized control object, the variables of the generalized control object are decoupled, and the decoupled control variables are compensated.
[0011] The control method of the steel ball grinding and powdering adaptive control system comprises the following steps of historical reference trajectory offline training, neural network parameter online optimization, powdering system control characteristic estimation, controller parameter updating, characteristic vector decoupling compensation, and control calculation output.
[0012] The multivariable adaptive control method is adopted to decouple control the outlet temperature of the mill, the inlet negative pressure of the mill, and the outlet-inlet differential pressure of the mill by taking the mill inlet hot air door, the powder machine recirculation air door, and the mill coal feeder speed as the adjusting variables.
[0013] The neural network parameter online optimization uses the neural network to estimate the model of the steel ball grinding and powdering system, and obtains the sensitivity matrix of the controlled variables of the powdering system to the input variables.
[0014] The radial basis neural network is a three-layer feedforward network with a single hidden layer, simulates the neural network structure of local adjustment and mutual coverage receiving domain in the human brain, is a local approximation neural network, can approximate any continuous function with any precision, and comprises an input layer, a hidden layer, and an output layer to form the neural network model estimator.
[0015] The neural network model estimator, m = 6 input layer neurons, v = 7 hidden layer neurons, n = 3 output layer neurons, network optimization parameters include: radial basis function base width B, center C, weight W; the input layer input vector is denoted as:
[0016]
[0017] The output layer output vector is:
[0018]
[0019] The hidden layer activation vector is:
[0020] H = [h1, h2, h3, h4, h5, h6, h7] T
[0021] Where the neuron activation function is a Gaussian basis function:
[0022]
[0023] Where, j = 1, 2, … 7, b j is the radial basis function base width of the jth neuron in the hidden layer;
[0024] The center point of the radial basis function of the jth neuron in the hidden layer is:
[0025] C j = [c j1 ,c j2 ,c j3 ,c j4 ,c j5 ,c j6 ] T
[0026] The radial basis function base width vector is:
[0027] B = [b1, b2, b3, b4, b5, b6, b7] T
[0028] The weight matrix of the hidden layer and the output layer is:
[0029]
[0030] The output layer output vector is:
[0031]
[0032] The estimation model error is:
[0033]
[0034] The neural network model estimation target is:
[0035]
[0036] i.e. minimizing the error between the estimated output and the true output; neural network parameter iterative updating algorithm:
[0037] Optimization target matrix form:
[0038]
[0039] Weight W updating algorithm:
[0040]
[0041] wherein, model estimated parameter learning rate:
[0042] η∈[0,1] model estimated parameter inertia damping coefficient:
[0043] μ∈[0,1] radial basis width B updating algorithm:
[0044]
[0045] wherein,
[0046]
[0047] Center position C updating algorithm:
[0048]
[0049] Neural network parameter updating formula:
[0050]
[0051] According to the neural network parameters, the prediction output is calculated, the prediction error is calculated, and the neural network parameters are updated and adjusted; the sensitivity matrix of the controlled variable of the controlled object to the input quantity is calculated, and since the model is unknown, an estimated model is used for approximation:
[0052]
[0053] wherein, k=1, 2, 3, i=1, 2, 3
[0054] Sensitivity matrix G of the controlled variable of the controlled object to the input quantity:
[0055]
[0056] The input sensitivity matrix G obtained by the neural network is subjected to eigenvector space decomposition, and control decoupling is realized, and the eigenvectors of the sensitivity matrix constitute a characteristic space matrix V, which satisfies:
[0057] G·V = Λ·V (13)
[0058] where,
[0059]
[0060] The characteristic vector space decomposition and control decoupling are performed, the PID parameters are adjusted according to the model estimated sensitivity characteristics of the controlled variable of the control object to the input variable, the decoupling compensator is composed of a characteristic vector space matrix transpose, the decoupling compensator and the control object form a generalized control object, the variables of the generalized control object are decoupled, and compensation is performed on the control variable after decoupling.
[0061] Control error:
[0062]
[0063] where, j = 1, 2, 3
[0064] An incremental PID controller is used:
[0065]
[0066] where t is the control sampling time, and is defined
[0067]
[0068] Then
[0069]
[0070] PID controller parameter matrix
[0071]
[0072] Let
[0073]
[0074] Then
[0075]
[0076] Decoupling is performed using the decoupling compensation matrix obtained by using a neural network estimation model:
[0077] U = V T ·U d (19)
[0078] The closed-loop control error optimization goal, that is, minimizing the closed-loop control error mean square value:
[0079]
[0080] PID closed-loop control parameter update algorithm:
[0081]
[0082] Wherein the closed-loop optimization target is the output change amount:
[0083]
[0084] The sensitivity matrix G of the controlled variable of the control object to the input variable is estimated using the neural network model:
[0085]
[0086] The decoupling compensation of the controlled variable is obtained:
[0087]
[0088] The change of the controlled variable to the control parameter is:
[0089]
[0090] The arrangement is obtained:
[0091]
[0092] The controller parameter update algorithm is:
[0093]
[0094] Wherein, the controller parameter learning rate is:
[0095] Alpha is in [0, 1]
[0096] The controller parameter inertia damping coefficient is:
[0097] Beta is in [0, 1]
[0098] The above model parameter update algorithm and the control parameter update algorithm are used to realize adaptive control of the steel ball grinding and powder system.
[0099] The advantages of the present application are:
[0100] The steel ball grinding and powder adaptive control system and control method of the present application uses a neural network model, estimates the sensitivity matrix of the controlled variable of the control object to the input variable using historical data, designs an adaptive PID controller based on the method of decoupling control in the space of the characteristic vector of the control sensitivity matrix, realizes decoupling control of the outlet temperature of the coal mill, the inlet negative pressure of the coal mill and the differential pressure between the inlet and outlet of the coal mill of the steel ball grinding and powder system, and the control method proposed by the present application can effectively overcome the multivariable coupling problem existing in the conventional independent multi-loop control mode, and realize long-term stable operation of the automatic control of the powder system. BRIEF DESCRIPTION OF DRAWINGS
[0101] Figure 1 The adaptive control system structure block diagram of the steel ball mill powder production system.
[0102] Figure 2 The adaptive control flow chart of the steel ball mill powder production system.
[0103] Figure 3 The neural network structure diagram of the application.
[0104] Figure 4 The adaptive control system structure block diagram of the steel ball mill powder production system. DETAILED DESCRIPTION
[0105] The specific embodiments of the application are further described below with reference to the accompanying drawings.
[0106] As shown in Figures 1-4 The adaptive control system of the steel ball mill powder production system, characterized in that it comprises a self-correcting controller, a steel ball mill powder production system, a neural network model estimator,
[0107] The neural network model estimator comprises an input layer, a hidden layer and an output layer,
[0108] The self-correcting controller comprises a PID controller, a parameter setting and a decoupling compensator. The PID parameters are adjusted according to the sensitivity characteristics of the controlled variables of the model-estimated controlled object to the input variables. The decoupling compensator is composed of the transpose of the characteristic vector space matrix. The decoupling compensator and the controlled object form a generalized controlled object, and the variables of the generalized controlled object are decoupled to realize compensation for the decoupled controlled variables.
[0109] A control method using the adaptive control system of the steel ball mill powder production system, characterized in that the adaptive control flow comprises: offline training of historical reference trajectories, online optimization of neural network parameters, estimation of the control characteristics of the powder production system, updating of the controller parameters, decoupling compensation based on the characteristic vectors, and control calculation output.
[0110] A multivariable adaptive control method is adopted to decouple control the mill outlet temperature, the mill inlet negative pressure and the mill outlet-inlet differential pressure by taking the mill inlet hot air door, the powder mill recirculation air door and the mill coal feeder speed as the regulating variables. A radial basis neural network is used to estimate the model parameters of the control characteristics of the steel ball mill powder production system by using historical data, and the decoupling compensation control is performed based on the characteristic vector space of the sensitivity matrix of the controlled variables of the controlled object to the input variables.
[0111] The adaptive control system structure of the steel ball mill powder production system designed by the application is shown in Figure 1
[0112] The given variable vector is:
[0113] S = [s1, s2, s3]T
[0114] s1 is the coal mill outlet temperature given value, s2 is the mill inlet negative pressure given value, and s3 is the mill outlet-inlet differential pressure (mill coal storage amount) given value.
[0115] The controlled quantity of the controlled object of the steel ball mill pulverizing system is:
[0116] Y=[y1, y2, y3] T
[0117] y1 is the coal mill outlet temperature measured value, y2 is the mill inlet negative pressure measured value, and y3 is the mill outlet-inlet differential pressure (mill coal storage amount) measured value.
[0118] The controlled object adjustment vector is:
[0119] U=[u1, u2, u3] T
[0120] u1 is the mill inlet hot air door opening degree control instruction value, u2 is the pulverizer recirculation air door opening degree control instruction value, and u3 is the mill coal feeder speed control instruction value.
[0121] The ball mill outlet temperature affects the coal powder temperature, and the temperature is too high to easily cause the coal powder in the ball mill cylinder to explode, and the temperature is too low to reduce the drying output of the ball mill. In order to prevent the coal powder in the cylinder from leaking, the mill inlet pressure is controlled in a slightly negative pressure state. The ball mill inlet and outlet pressure difference is controlled, and since the coal storage amount in the cylinder cannot be directly obtained, the inlet and outlet pressure difference is used to represent the coal storage amount of the ball mill. In order to make the ball mill work near the maximum output point, the ball mill inlet and outlet differential pressure needs to be controlled in a reasonable range.
[0122] The adaptive control flow of the steel ball mill pulverizing system is as shown in Figure 2 The offline training and online rolling optimization are used to estimate the model of the pulverizing system online, the estimated parameters, the controller parameter update, the control decoupling, and the control calculation output are outputted by the self-correcting controller.
[0123] The neural network is used to estimate the model of the steel ball mill pulverizing system, and the sensitivity matrix (G matrix in the figure) of the controlled quantity of the pulverizing system to the input quantity is obtained. The neural network structure designed by the application is as shown in Figure 3 .
[0124] The radial basis neural network is a three-layer feedforward network with a single hidden layer, simulates the neural network structure of local adjustment and mutual coverage receiving domain in the human brain, is a local approximation neural network, and can approximate any continuous function with arbitrary precision. The neural network model estimator structure is as shown in Figure 3 . The neural network model estimator is composed of an input layer, a hidden layer, and an output layer.
[0125] In the neural network model estimator, m = 6 input layer neurons, v = 7 hidden layer neurons, n = 3 output layer neurons, and the network optimization parameters include: radial basis function base width B, center C, and weight W. The input layer input vector is denoted as:
[0126]
[0127] The output layer output vector is:
[0128]
[0129] The hidden layer activation vector is:
[0130] H = [h1, h2, h3, h4, h5, h6, h7] T
[0131] Where the neuron activation function is a Gaussian basis function:
[0132]
[0133] Where, j = 1, 2, … 7, b j is the radial basis function base width of the jth neuron in the hidden layer.
[0134] The center point of the radial basis function of the jth neuron in the hidden layer is the center vector:
[0135] C j = [c j1 , c j2 , c j3 , c j4 , c j5 , c j6 ] T
[0136] The radial basis function base width vector is:
[0137] B = [b1, b2, b3, b4, b5, b6, b7] T
[0138] The weight matrix of the hidden layer and the output layer is:
[0139]
[0140] The output layer output vector is:
[0141]
[0142] The estimation model error is:
[0143]
[0144] The estimation target of the neural network model is:
[0145]
[0146] i.e. minimize the error between the estimated output and the true output.
[0147] Neural network parameter iterative update algorithm:
[0148] Optimization objective matrix form:
[0149]
[0150] Weight W update algorithm:
[0151]
[0152] wherein the model estimated parameter learning rate:
[0153] η∈[0,1] model estimated parameter inertia damping coefficient:
[0154] μ∈[0,1] radial basis width B update algorithm:
[0155]
[0156] wherein,
[0157]
[0158] Center position C update algorithm:
[0159]
[0160] Neural network parameter update formula:
[0161]
[0162] According to the neural network parameters, the prediction output is calculated, the prediction error is calculated, and the neural network parameters are updated and adjusted. The sensitivity matrix of the controlled variable of the controlled object to the input quantity is calculated, and since the model is unknown, an estimated model is used for approximation:
[0163]
[0164] wherein, k = 1, 2, 3, i = 1, 2, 3
[0165] Sensitivity matrix G of controlled variable of controlled object to input quantity:
[0166]
[0167] The control sensitivity matrix G obtained by the neural network is subjected to eigenvector space decomposition to realize control decoupling. The eigenvectors of the sensitivity matrix constitute a characteristic space matrix V, which satisfies:
[0168] G·V=Λ·V (13)
[0169] in,
[0170]
[0171] The structure of the self-calibrating controller is as follows: Figure 4 As shown, the PID parameters are adjusted based on the sensitivity characteristics of the controlled variable to the input variable estimated by the model. The decoupling compensator is composed of the transpose of the eigenvector space matrix. The decoupling compensator and the controlled object form a generalized controlled object. The variables of the generalized controlled object are decoupled, thereby compensating for the decoupled control variable.
[0172] Control error:
[0173]
[0174] Where j = 1, 2, 3
[0175] Incremental PID controller is used:
[0176]
[0177] Where t is the control sampling time, defined as...
[0178]
[0179] but
[0180]
[0181] PID controller parameter matrix
[0182]
[0183] remember
[0184]
[0185] but
[0186]
[0187] Decoupling is performed using the decoupling compensation matrix obtained from the neural network estimation model.
[0188] U = V T ·U d (19)
[0189] The objective of closed-loop control error optimization is to minimize the mean square value of the closed-loop control error.
[0190]
[0191] PID closed loop control parameter updating algorithm:
[0192]
[0193] Wherein the closed loop optimization target is the output change:
[0194]
[0195] Using the neural network model to estimate the parameters, the control object controlled variable input quantity sensitivity matrix G can be obtained:
[0196]
[0197] From the control variable decoupling compensation:
[0198]
[0199] The control variable changes with the control parameter:
[0200]
[0201] The arrangement can be obtained:
[0202]
[0203] Controller parameter updating algorithm:
[0204]
[0205] Wherein, the controller parameter learning rate:
[0206] Alpha is in [0, 1]
[0207] Controller parameter inertia damping coefficient:
[0208] Beta is in [0, 1]
[0209] The above model parameter updating algorithm and control parameter updating algorithm are used to realize the adaptive control of the steel ball grinding system.
[0210] The sensitivity matrix characteristics of the controlled variable of the automatic control object to the input quantity are estimated by combining the historical trajectory offline training and online learning method, the controlled variable of the control object to the input quantity sensitivity matrix is adjusted according to the model estimation, the controlled variable of the control object to the input quantity sensitivity matrix is decoupled and compensated based on the characteristic vector space, which can effectively solve the automatic control problem of the steel ball grinding system.
[0211] Embodiment:
[0212] The present application is applied successfully in the coal pulverizing system of No.5 boiler of Daqing Petrochemical Company, the coal pulverizing system realizes full automatic operation, the personnel operation amount is reduced by more than 95%, the coal mill works in the optimization area, and the coal pulverizing efficiency is improved. The automatic coal pulverizing system of the present application has been continuously operated for more than 6 months in No.5 boiler of Daqing Petrochemical Company, meets the production demand of Daqing Petrochemical Company, and realizes the main technical indexes as follows:
[0213] 1) the steady-state deviation of the negative pressure at the inlet of the coal mill is less than or equal to ±30Pa;
[0214] 2) the transient-state deviation of the negative pressure at the inlet of the coal mill is less than or equal to ±50Pa;
[0215] 3) the temperature change range of the outlet of the coal mill is 60℃-70℃;
[0216] 4) the automatic control loop investment rate of the coal pulverizing system is 100%;
[0217] The power saving rate of the coal pulverizing system is greater than or equal to 10%.
Claims
1. An adaptive control system for steel ball mill powder production, characterized in that, This includes a self-calibrating controller, a ball mill powder production system, and a neural network model estimator. The neural network model estimator includes an input layer, a hidden layer, and an output layer. The self-tuning controller includes a PID controller, parameter tuning, and a decoupling compensator.
2. The adaptive control system for steel ball milling according to claim 1, wherein the decoupling compensator is composed of the transpose of the eigenvector space matrix, and the decoupling compensator and the controlled object form a generalized control object, wherein the variables of the generalized control object are decoupled, thereby realizing compensation for the decoupled control quantity.
3. A control method for an adaptive control system for steel ball mill powder production, characterized in that, The adaptive control process includes: offline training of historical reference trajectories, online optimization of neural network parameters, estimation of control characteristics of the pulverizing system, controller parameter updating, decoupling compensation based on feature vectors, and control calculation output. A multivariable adaptive control method is adopted, using the mill inlet hot air damper, the pulverizer recirculation damper, and the mill feeder speed as adjustment variables to decouple the mill outlet temperature, mill inlet negative pressure, and mill inlet-outlet differential pressure. A radial basis function neural network is used to estimate the control characteristic model parameters of the ball mill pulverizing system using historical data, and decoupled compensation control is performed based on the eigenvector space of the sensitivity matrix of the controlled variable to the input variable.
4. The control method for the adaptive control system of steel ball mill powder production according to claim 3, characterized in that, The aforementioned online optimization of neural network parameters involves using a neural network to perform model estimation of the ball mill pulverizing system, resulting in a sensitivity matrix of the controlled variable to the input variable. Radial basis function neural networks (RBNs) are three-layer feedforward networks with a single hidden layer. They simulate the neural network structure in the human brain where the receptive fields are locally adjusted and overlapped. They are a type of locally approximating neural network that can approximate any continuous function with arbitrary precision. The neural network model estimator consists of an input layer, a hidden layer, and an output layer.
5. The control method for an adaptive control system for steel ball mill powder production according to claim 4, characterized in that, In the neural network model estimator, m = 6 input layer neurons, v = 7 hidden layer neurons, and n = 3 output layer neurons. The network optimization parameters include: radial basis function basis width B, center C, and weight W. The input vector of the input layer is denoted as: Output layer output vector: AND p =[and p1 ,and p2 ,and p3 ] T Hidden layer activation vectors: H=[h1,h2,h3,h4,h5,h6,h7] T The neuron activation function used is the Gaussian function. Where j = 1, 2, ..., 7, b j Here, represents the basis width of the radial basis function of the j-th neuron in the hidden layer; and the center vector of the radial basis function center point of the j-th neuron in the hidden layer is: C j =[c j1 ,c j2 ,c j3 ,c j4 ,c j5 ,c j6 ] T Radial basis function basis width vector: B=[b1,b2,b3,b4,b5,b6,b7] T Hidden layer and output layer weight matrices: Output layer output vector: The estimated model error is: The neural network model estimates the target: That is, minimizing the error between the estimated output and the true output; neural network parameter iterative update algorithm: Optimize the form of the objective matrix: Weight W update algorithm: Among them, the learning rate of the model estimation parameters: The estimated parameter for the model η∈[0,1] is the inertial damping coefficient. Algorithm for updating radial base width B in μ∈[0,1]: in, Center position C update algorithm: Neural network parameter update formula: Based on the neural network parameters, predict the output, calculate the prediction error, and update and adjust the neural network parameters. Calculate the sensitivity matrix of the controlled variable to the input variable. Since the model is unknown, an estimated model is used for approximation. Where, k = 1, 2, 3, i = 1, 2, 3 The sensitivity matrix G of the controlled variable to the input variable:
6. The control method for an adaptive control system for steel ball mill powder production according to claim 5, characterized in that, The sensitivity matrix G of the input quantity obtained by the neural network is decomposed into eigenspace to achieve control decoupling. The eigenvectors of the sensitivity matrix form the feature space matrix V, which satisfies the following: G·V=Λ·V (13) in, 7. The control method for an adaptive control system for steel ball mill powder production according to claim 6, characterized in that, The aforementioned eigenvector space decomposition and control decoupling are performed by adjusting PID parameters based on the sensitivity characteristics of the controlled variable to the input variable estimated by the model. The decoupling compensator is composed of the transpose of the eigenvector space matrix. The decoupling compensator and the controlled object form a generalized controlled object. The variables of the generalized controlled object are decoupled, thereby compensating the decoupled control variable. Control error: Where j = 1, 2, 3 Incremental PID controller is used: Where t is the control sampling time, defined as... but PID controller parameter matrix remember but Decoupling is performed using the decoupling compensation matrix obtained from the neural network estimation model. U=V T ·U d (19) The objective of closed-loop control error optimization is to minimize the mean square value of the closed-loop control error. PID closed-loop control parameter update algorithm: The change in output is related to the closed-loop optimization objective: By estimating the parameters using a neural network model, we can obtain the sensitivity matrix G of the controlled variable to the input variable: From the decoupling compensation of the control quantity, we can obtain: The change in control parameter in relation to the control quantity: After sorting, we can obtain: Controller parameter update algorithm: Wherein, the controller parameter learning rate is: α∈[0,1] Controller parameter: inertia damping coefficient β∈[0,1] The above-mentioned model estimation parameter update algorithm and control parameter update algorithm are used to realize the adaptive control of the steel ball mill powder production system.