A method for optimizing and controlling grinding particle size under varying operating conditions

Abstract

This invention provides a method for optimizing grinding particle size control under varying operating conditions, relating to the fields of mineral processing and grinding technology. The method first collects the input and output parameters of the hematite grinding process under changing operating conditions and calculates the performance indicators of the grinding process to determine the optimization control objective for the grinding process under varying operating conditions. Then, based on the optimization control objective, a cascaded neural network model is established to generate preset values ​​for the process variables of the control loop and predict the performance indicators generated by these preset values. Based on the cascaded neural network control method, the output weights of the two networks in the cascaded neural network model are corrected to obtain a cascaded neural network controller. Finally, based on the cascaded neural network controller, the optimal grinding particle size output result for the grinding process under abnormal operating conditions is obtained. This method is helpful for optimizing the grinding particle size control due to changes in ore properties under varying operating conditions.

Description

Technical Field

[0001] This invention relates to the field of mineral processing and grinding technology, and in particular to a method for optimizing and controlling grinding particle size under varying working conditions. Background Technology

[0002] Grinding, as one of the most important procedures in mineral processing, separates different useful minerals. Advanced concentrators commonly employ a grinding process consisting of a single-stage open-circuit rod mill and a two-stage closed-circuit ball mill-cyclone system. This process produces high-grade iron ore with uniform particle size distribution and stable composition and properties, allowing for the establishment of approximate process models. During grinding, the grinding particle size is closely related to the mill feed rate, mill inlet water flow rate, and the concentration of the classifier overflow slurry. Furthermore, the grinding particle size directly affects the concentrate quality in the beneficiation industry, thus influencing the economic indicators of the entire processing.

[0003] Under safe production conditions, optimizing the control of grinding particle size is the goal of grinding production process optimization. Currently, model-based methods such as real-time optimization, model predictive control, and multivariable decoupling control can be used to provide loop setpoints and achieve optimized control of grinding particle size. However, due to the low ore grade and frequent changes in ore particle size distribution and hardness, the system operating conditions are unstable and dynamically changing, making it difficult to establish a process model. Furthermore, the hematite grinding production process mainly relies on operators' experience to adjust the control loop setpoints. However, due to human subjectivity and arbitrariness, operators often cannot timely and comprehensively understand and judge the current grinding process, thus failing to provide correct loop setpoint adjustments. This often leads to grinding particle size exceeding the target range, and sometimes the given loop setpoints deviate significantly from the optimal operating point, resulting in mill underload or overload abnormal conditions, causing equipment damage or even shutdown of the entire grinding process, making it difficult to achieve optimized control of the grinding process. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method for optimizing and controlling grinding particle size under varying working conditions, thereby addressing the shortcomings of the prior art and achieving optimized control of grinding particle size.

[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for optimizing and controlling grinding particle size under varying working conditions, comprising the following steps: Step 1: Collect production information during the hematite grinding process when operating conditions change; the production information includes the input and output parameters of the grinding process; Step 2: Based on the input and output parameters of the grinding process, calculate the performance indicators of the grinding process and determine the operation optimization control target of the grinding process under varying operating conditions; Step 3: Based on the operation optimization control objective of the grinding process under varying operating conditions, establish a series neural network model; the series neural network includes a loop preset value adjustment network and a performance index prediction network; the loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop; the performance index prediction network is used to predict the performance index generated by the preset values ​​of the process variables of the control loop. Step 4: Based on the cascaded neural network control method, the output weights of the two networks in the cascaded neural network model are corrected using the improved LM algorithm; Step 5: Using the corrected output weights, obtain the cascaded neural network controller; based on the cascaded neural network controller, obtain the optimal grinding particle size output result for the grinding process under abnormal operating conditions.

[0006] The input parameters of the grinding process in step 1 are control loop process variables, including mill feed rate, mill inlet water flow rate, and classifier overflow concentration; the output parameters of the grinding process include performance indicators and grinding particle size.

[0007] Step 2 includes: Step 2.1: Based on the input parameters and output grinding particle size of the grinding process, determine the input-output relationship of the grinding process: (1) In the formula, for k Constant control loop process variables, This refers to the feed rate to the mill. This refers to the feedwater flow rate at the mill inlet. The overflow concentration of the classifier; This is for k Grinding particle size at all times; The properties of raw hematite ore; F ()for k+ Grinding particle size at time 1 and k The nonlinear relationship between grinding particle size, control loop process variables, and raw ore properties at any given time; Step 2.2: Determine the process variable expressions for the control loop: (2) In the formula, Other factors representing the grinding process, ;in, For the electric vibration frequency of the electric vibratory feeder, This refers to the opening degree of the mill inlet feedwater valve. Adjust the opening of the water supply valve for the classifier. for k+ A certain input quantity at time 1 and kThe nonlinear relationship between this input quantity at time t and another factor; Step 2.3: Determine the constraints for the process variables in the control loop: (3) In the formula, and for k The minimum and maximum values ​​of the process variables in the control loop at all times; Step 2.4: Determine the control performance indicators and control objectives of the grinding process under varying operating conditions: (4) (5) (6) In the formula, The performance index (i.e., objective function) for the operation optimization control of the grinding process under varying operating conditions. The target value for grinding particle size. As a weighting factor, The ideal deviation value determined for the grinding process. and These are the minimum and maximum values ​​for grinding particle size, respectively.

[0008] Step 3 includes: Step 3.1: Construct a performance index prediction network and determine the network error to adjust and correct the output weight vector of the performance index prediction network; Step 3.1.1: The performance index prediction network is used to set the preset values ​​of process variables in the control loop. The resulting performance indicators To make predictions, a quadratic performance index model is established as follows: (7) In the formula, For an unknown nonlinear function, Represents the preset value of the process variable in the control loop. For mill current, For the classifier current, This is an estimate of the grinding particle size; Step 3.1.2: Based on the approximation principle of neural networks, a feedforward three-layer neural network is used to fit the quadratic performance index model to obtain the expected value of the operation optimization control performance index of the grinding process under varying working conditions, as shown in the following formula: (8) In the formula, This represents the expected value of the performance index for the operation optimization control of the grinding process under varying operating conditions. The input data vector for the performance metric prediction network; Let be the output weight vector of the performance metric prediction network, where To predict the number of hidden layers in a network for performance metrics; The input weight vector for the performance metric prediction network; The activation function for the performance metric prediction network; Step 3.1.3: Based on the characteristics of the three-layer neural network, and by setting the grinding particle size deviation... The performance index prediction error function can then be obtained: (9) In the formula, This represents the ideal value for grinding particle size deviation; Step 3.1.4: Define the performance index prediction network error according to the performance index prediction error function in formula (9): (10) In the formula, Predict network errors for performance metrics; Step 3.1.5: Predict the network error based on the performance index of formula (10), by minimizing the objective function of the network error shown in formula (11), and using... The output weight vector of the performance metric prediction network is corrected based on the stopping condition. ,in, A threshold for predicting network error as a performance metric; The objective function for the performance metric network error is shown in the following formula: (11) In the formula, The objective function for the performance metric is the network error. Step 3.2: Construct a loop preset value adjustment network and determine the network error to adjust the output weight vector of the correction loop preset value adjustment network; Step 3.2.1: The loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop. Therefore, a loop setpoint optimization model is established, expressed as: (12) In the formula, It is an unknown nonlinear function; Step 3.2.2: Establish the following equation for the loop preset value adjustment network: (13) In the formula, These represent the expected values ​​of the process variables in the control loop, respectively. Adjust the input data vector of the network to the preset values ​​of the loop; Adjust the network output weight vector to the preset values ​​of the loop. Adjust the number of hidden layers in the network to a preset value for the loop; Adjust the input weight vector of the network to the preset values ​​of the loop; Adjust the activation function of the network to the preset values ​​for the loop; Step 3.2.3: Based on the input weights It is fixed, only for the output weights. Adopt Minimization is used for correction, and the error equation of the established loop preset value adjustment network is as follows: (14) If the loop preset value is the optimal value, then the constraint condition for establishing the error equation of the loop preset value adjustment network is: (15) In the formula, Adjust the network's training error using preset values ​​for the loop; This represents the expected value of the performance metric. Step 3.2.4: According to Simply make available The optimized value; The error equation for the minimum established loop preset value adjustment network is: (16) In the formula, Adjust the network error to the preset value of the loop; Step 3.2.5: Adjust the network error equation based on the loop preset value established by formula (16), and minimize the objective function of the loop preset value adjustment network error shown in formula (17), and use... To correct the output weight vector as a stopping condition ; The objective function for adjusting network error using loop preset values ​​is shown in the following formula: (17) In the formula, The objective function is to adjust the network error for the loop preset value.

[0009] Step 4 includes: Step 4.1: Minimize the objective function of network error and the loop preset value of the performance indicators of formulas (11) and (17) to correct the output weight vector. and According to the LM algorithm, the weight correction amounts for establishing the performance index network and the loop preset value adjustment network are: (18) (19) In the formula, These are LM parameters and all are greater than 0. ; The learning rate of the network is always greater than 0. for right Jacobian matrix; and These are the weight correction values ​​for the performance index network and the loop preset value adjustment network, respectively. LM parameters The solution expression is: (20) In the formula, It is an adjustable parameter; It is the objective function; Step 4.2: To accelerate the convergence speed, the identity matrix equation is established to replace the identity matrix in the weight correction formulas of the performance index network and loop preset value adjustment network shown in formulas (18) and (19). : (twenty one) In the formula, A diagonal matrix composed of the main diagonal elements; Step 4.3: Substitute formulas (20) and (21) into formulas (18) and (19) to obtain the weight correction amounts of the performance index network and the loop preset value adjustment network. and .

[0010] Step 5 includes: Step 5.1: Adjust the weights of the two neural networks and Substituting these values ​​into the expected value calculation formula (8) and the loop preset value adjustment network equation (13) of the grinding process under variable working conditions, a series neural network controller is obtained. Step 5.2: Substitute the serial neural network controller into the calculation formula (4) for the control performance index of the grinding process under varying working conditions to obtain the optimal grinding particle size output result of the grinding process under abnormal working conditions.

[0011] The beneficial effects of adopting the above technical solution are as follows: This invention provides a method for optimizing grinding particle size control under varying operating conditions. First, the cascaded neural network consists of a performance index prediction network and a loop preset value adjustment network. The Levenberg-Marquardt (LM) algorithm is used to correct the output weights in the neural network. Second, this invention updates the parameters and identity matrix of the traditional LM algorithm to improve the algorithm's stability and accelerate convergence speed and efficiency, obtaining the optimal output weights. Finally, the optimal performance index is obtained using the solution of the output weights, and the optimal output data is obtained based on the optimal performance index. This invention has been verified using specific implementation examples and can provide a new method for the dynamic measurement of grinding particle size during the grinding process in mineral processing plants, contributing to the optimized control of grinding particle size under varying operating conditions due to changes in ore properties. Attached Figure Description

[0012] Figure 1 A flowchart of a grinding particle size optimization control method under varying working conditions provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the grinding process operation optimization control strategy provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the cascaded neural network model structure provided in an embodiment of the present invention; Figure 4 The following is a comparison chart of the changes in the three input loops under over-working conditions within 90 minutes, provided for an embodiment of the present invention. (a) shows the change in mill feed rate, (b) shows the change in mill inlet water flow rate, and (c) shows the change in classifier overflow concentration. Figure 5 A comparison diagram of the grinding particle size change process under different working conditions provided in the embodiments of the present invention. Detailed Implementation

[0013] The specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and are not intended to limit the scope of the invention.

[0014] In this embodiment, a method for optimizing and controlling grinding particle size under varying operating conditions is described, such as... Figure 1 As shown, it includes the following steps: Step 1: Collect production information during the hematite grinding process when operating conditions change; the production information includes the input and output parameters of the grinding process; The input parameters of the grinding process are control loop process variables, including mill feed rate, mill inlet water flow rate, and classifier overflow concentration; the output parameters of the grinding process include performance indicators and grinding particle size.

[0015] In this embodiment, the input and output parameters of the hematite grinding process are as follows: Figure 1 As shown Table 1. Parameters of the grinding process

[0016] Step 2: Based on the input and output parameters of the grinding process, calculate the performance indicators of the grinding process and determine the operation optimization control target of the grinding process under varying operating conditions; Step 2.1: Based on the input parameters and output grinding particle size of the grinding process, determine the input-output relationship of the grinding process: (1) In the formula, for k Constant control loop process variables, This refers to the feed rate to the mill. This refers to the feedwater flow rate at the mill inlet. The overflow concentration of the classifier; This is for k Grinding particle size at all times; The properties of raw hematite ore; F ()for k+ Grinding particle size at time 1 and k The nonlinear relationship between grinding particle size, control loop process variables, and raw ore properties at any given time; Step 2.2: Determine the process variable expressions for the control loop: (2) In the formula, Other factors representing the grinding process, ;in, For the electric vibration frequency of the electric vibratory feeder, This refers to the opening degree of the mill inlet feedwater valve. Adjust the opening of the water supply valve for the classifier. for k+ A certain input quantity at time 1 and k The nonlinear relationship between this input quantity at time t and another factor; Step 2.3: Determine the constraints for the process variables in the control loop: (3) In the formula, and for k The minimum and maximum values ​​of the process variables in the control loop at all times; Step 2.4: To achieve optimized control of grinding particle size, the following steps are introduced: The secondary performance index aims to minimize the sum of squares of grinding particle size deviations at the current and future times, thereby determining the control performance index and control objective of the grinding process under variable operating conditions. (4) (5) (6) In the formula, The performance index (i.e., objective function) for the operation optimization control of the grinding process under varying operating conditions. The target value for grinding particle size. As a weighting factor, The ideal deviation value determined for the grinding process. and These are the minimum and maximum values ​​for grinding particle size, respectively.

[0017] Step 3: According to... Figure 2 To optimize the control objectives of the grinding process under varying operating conditions, a cascaded neural network model is established; the cascaded neural network is as follows: Figure 3 As shown, it includes a loop preset value adjustment network and a performance index prediction network; the loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop; the performance index prediction network is used to predict the performance index generated by the preset values ​​of the process variables of the control loop. Step 3.1: Construct a performance index prediction network and determine the network error to adjust and correct the output weight vector of the performance index prediction network; Step 3.1.1: Since grinding particle size is a function of mill feed rate, mill inlet feedwater rate, classifier overflow concentration, mill current, and classifier current, and because the underlying basic loop control system can control the mill feed rate, mill inlet feedwater rate, and classifier overflow concentration near the setpoint within one sampling period of the operation control layer, the performance index prediction network is used to preset the process variables of the control loop. The resulting performance indicators To make predictions, a quadratic performance index model is established as follows: (7) In the formula, For an unknown nonlinear function, Represents the preset value of the process variable in the control loop. For mill current, For the classifier current, This is an estimate of the grinding particle size; Step 3.1.2: Since both input and output are bounded closed sets, based on the approximation principle of neural networks, a feedforward three-layer neural network is used to fit the quadratic performance index model to obtain the expected value of the operation optimization control performance index of the grinding process under varying working conditions, as shown in the following formula: (8) In the formula, This represents the expected value of the performance index for the operation optimization control of the grinding process under varying operating conditions. The input data vector for the performance metric prediction network; Let be the output weight vector of the performance metric prediction network, where To predict the number of hidden layers in a network for performance metrics; The input weight vector for the performance metric prediction network; The activation function for the performance metric prediction network; Step 3.1.3: Based on the characteristics of a three-layer neural network, the weights from the input layer to the hidden layer and the threshold of the hidden layer can be randomly selected. Only the output weights need to be adjusted to satisfy the approximation requirement. Therefore, this invention sets fixed input weights, corrects the output weights by minimizing the prediction error, and sets a grinding particle size deviation. The performance index prediction error function can then be obtained: (9) In the formula, This represents the ideal value for grinding particle size deviation; Step 3.1.4: According to formula (9), the performance index prediction error depends on the grinding particle size deviation at each moment. Therefore, Minimalness is what makes every moment... The error is extremely small, therefore, according to the performance index prediction error function defined by formula (9), the performance index prediction network error is defined as: (10) In the formula, Predict network errors for performance metrics; Step 3.1.5: Predict the network error based on the performance index of formula (10), by minimizing the objective function of the network error shown in formula (16), and using... The output weight vector of the performance metric prediction network is corrected based on the stopping condition. ,in, A threshold for predicting network error as a performance metric; The objective function for the performance metric network error is shown in the following formula: (11) In the formula, The objective function for the performance metric is the network error. Step 3.2: Construct a loop preset value adjustment network and determine the network error to adjust the output weight vector of the correction loop preset value adjustment network; Step 3.2.1: The loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop. The optimized value based on the setpoint is the optimal performance index value that the generated performance index can approximate. Therefore, a loop setpoint optimization model is established, expressed as: (12) In the formula, It is an unknown nonlinear function; Step 3.2.2: Establish the following equation for the loop preset value adjustment network: (13) In the formula, These represent the expected values ​​of the process variables in the control loop, respectively. Adjust the input data vector of the network to the preset values ​​of the loop; Adjust the network output weight vector to the preset values ​​of the loop. Adjust the number of hidden layers in the network to a preset value for the loop; Adjust the input weight vector of the network to the preset values ​​of the loop; Adjust the activation function of the network to the preset values ​​for the loop; Step 3.2.3: Based on the input weights It is fixed, only for the output weights. Adopt Minimization is used for correction, and the error equation of the established loop preset value adjustment network is as follows: (14) If the loop preset value is the optimal value, then the constraint condition for establishing the error equation of the loop preset value adjustment network is: (15) In the formula, Adjust the network's training error using preset values ​​for the loop; This represents the expected value of the performance metric. Step 3.2.4: According to , ( (As a parameter) simply make available The optimized value; The error equation for the minimum established loop preset value adjustment network is: (16) In the formula, Adjust the network error to the preset value of the loop; Step 3.2.5: Adjust the network error equation based on the loop preset value established by formula (16), and minimize the objective function of the loop preset value adjustment network error shown in formula (17), and use... To correct the output weight vector as a stopping condition ; The objective function for adjusting network error using loop preset values ​​is shown in the following formula: (17) In the formula, The objective function is to adjust the network error for the loop preset value.

[0018] Step 4: Based on the cascaded neural network control method, the output weights of the two networks in the cascaded neural network model are corrected using the improved LM algorithm; Step 4.1: Based on the fact that the LM algorithm has advantages over the gradient descent method in terms of learning accuracy and convergence speed, the LM algorithm is used to minimize the objective function of network error and the loop preset value in the performance indicators of formulas (11) and (17) to correct the output weight vector. and According to the LM algorithm, the weight correction amounts for establishing the performance index network and the loop preset value adjustment network are: (18) (19) In the formula, These are LM parameters and all are greater than 0. ; The learning rate of the network is always greater than 0. for right Jacobian matrix; and These are the weight correction values ​​for the performance index network and the loop preset value adjustment network, respectively. The traditional LM algorithm can, to some extent, avoid the ill-conditioned problem of the coefficient matrix in linear equation systems, providing more stable and accurate results. However, the original algorithm may suffer from non-convergence due to the singularity of the Jacobian matrix, leading to larger errors. Therefore, to improve the convergence of the algorithm, the LM parameters are modified. The solution expression is: (20) In the formula, It is an adjustable parameter; It is the objective function; Step 4.2: To accelerate the convergence speed, the identity matrix equation is established to replace the identity matrix in the weight correction formulas of the performance index network and loop preset value adjustment network shown in formulas (18) and (19). : (twenty one) In the formula, A diagonal matrix composed of the main diagonal elements; At this point, when the radius of the convergence region is too large, a larger step size will be generated in the direction with smaller gradient, which speeds up the convergence speed.

[0019] Step 4.3: Substitute formulas (20) and (21) into formulas (18) and (19) to obtain the weight correction amounts of the performance index network and the loop preset value adjustment network. and .

[0020] Step 5: Using the corrected output weights, obtain the cascaded neural network controller; based on the cascaded neural network controller, obtain the optimal grinding particle size output result for the grinding process under abnormal operating conditions.

[0021] Step 5.1: Adjust the weights of the two neural networks and Substituting these values ​​into the expected value calculation formula (8) and the loop preset value adjustment network equation (13) of the grinding process under variable working conditions, a series neural network controller is obtained. Step 5.2: Substitute the serial neural network controller into the calculation formula (4) for the control performance index of the grinding process under varying working conditions to obtain the optimal grinding particle size output result of the grinding process under abnormal working conditions.

[0022] This embodiment also verifies the effectiveness of the method of the present invention through experiments. The main technical parameters and value ranges are as follows: Figure 4 , 5 As shown. Among them, improved LM1: the improved LM method using only formula (20), and improved LM2: the improved LM method using only formula (21); in, Figure 4 This describes the changes in the three input loops—mill feed rate, mill inlet water flow rate, and classifier overflow concentration—over a 90-minute period under operating conditions; and... Figure 5 It describes the process of grinding particle size change under different operating conditions. From Figure 4It can be seen that the stabilization time of the method proposed in this invention is 34 minutes, which is relatively short; the stabilization times of the other three methods are 47 minutes, 45 minutes, and 55 minutes, respectively, which are longer, and the proposed algorithm's change process is more stable. Meanwhile, Table 2 shows the results analysis of the changes in mill feed rate, mill inlet water rate, and classifier overflow concentration at the 25th sampling point. According to Table 2, for each variable, the steady-state error of the proposed algorithm is the smallest compared to the other three algorithms. Meanwhile, Table 3 shows the results analysis at the 90th minute under the over-operational condition. From... Figure 5 As shown in Table 3, the grinding particle size corrected by the loop setpoint optimization algorithms corresponding to the four algorithms can all be reduced to near the actual value. However, the error between the grinding particle size corrected by the algorithm proposed in this invention and the true value is 0.109%, and the performance index error is 0.104%, both of which are less than the experimentally specified 0.2% and less than the other three algorithms. Therefore, the method of this invention can optimize the setpoints of the control loops for mill feed rate, mill inlet water flow rate, and classifier overflow concentration online by utilizing real-time operating data and the interaction between neural networks, enabling the control loops to better track the setpoints.

[0023] Table 2. Steady-state error at 90 minutes (steady state) under over-condition conditions

[0024] Table 3. Analysis of grinding particle size at 90 minutes (steady state) under over-working conditions.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for optimizing and controlling grinding particle size under varying operating conditions, characterized in that: Includes the following steps: Step 1: Collect production information during the hematite grinding process when operating conditions change; the production information includes the input parameters and output parameters of the grinding process; the input parameters of the grinding process are process variables of the control loop, including mill feed rate, mill inlet water flow rate, and classifier overflow concentration; the output parameters of the grinding process include grinding particle size; Step 2: Based on the input and output parameters of the grinding process, calculate the performance indicators of the grinding process and determine the operation optimization control target of the grinding process under varying operating conditions; Step 3: Based on the operation optimization control objective of the grinding process under varying operating conditions, establish a cascaded neural network model; the cascaded neural network model includes a loop preset value adjustment network and a performance index prediction network; the loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop; the performance index prediction network is used to predict the performance index generated by the preset values ​​of the process variables of the control loop. Step 4: Based on the cascaded neural network control method, the output weights of the two networks in the cascaded neural network model are corrected using the improved LM algorithm; Step 5: Using the corrected output weights, obtain the cascaded neural network controller; based on the cascaded neural network controller, obtain the optimal grinding particle size output result for the grinding process under abnormal operating conditions; Step 2 includes: Step 2.1: Based on the input parameters and output grinding particle size of the grinding process, determine the input-output relationship of the grinding process: (1) In the formula, for k Constant control loop process variables, This refers to the feed rate to the mill. This refers to the feedwater flow rate at the mill inlet. The overflow concentration of the classifier; This is for k Grinding particle size at all times; It has the properties of a raw hematite ore; F ()for k+ Grinding particle size at time 1 and k The nonlinear relationship between grinding particle size, control loop process variables, and raw ore properties at any given time; Step 2.2: Determine the process variable expressions for the control loop: (2) In the formula, Other factors representing the grinding process, ;in, For the electric vibration frequency of the electric vibratory feeder, This refers to the opening degree of the mill inlet feedwater valve. Adjust the opening of the water supply valve for the classifier. for k+ A certain input quantity at time 1 and k The nonlinear relationship between this input quantity at time t and another factor; Step 2.3: Determine the constraints for the process variables in the control loop: (3) In the formula, and for k The minimum and maximum values ​​of the process variables in the control loop at all times; Step 2.4: Determine the performance indicators and operational optimization control objectives for the grinding process under varying operating conditions: (4) (5) (6) In the formula, Performance indicators for optimizing the control of the grinding process under varying operating conditions. The target value for grinding particle size. As a weighting factor, The ideal deviation value determined for the grinding process. and These are the minimum and maximum values ​​for grinding particle size, respectively. Step 3 includes: Step 3.1: Construct a performance index prediction network and determine the network error to adjust and correct the output weight vector of the performance index prediction network; Step 3.1.1: The performance index prediction network is used to set the preset values ​​of process variables in the control loop. The resulting performance indicators To make predictions, a quadratic performance index model is established as follows: (7) In the formula, For an unknown nonlinear function, Represents the preset value of the process variable in the control loop. For mill current, For the classifier current, This is an estimate of the grinding particle size; Step 3.1.2: Based on the approximation principle of neural networks, a feedforward three-layer neural network is used to fit the quadratic performance index model to obtain the expected value of the performance index of the grinding process under variable operating conditions, as shown in the following formula: (8) In the formula, This represents the expected value of the performance index for the operation optimization control of the grinding process under varying operating conditions. The input data vector for the performance metric prediction network; Let be the output weight vector of the performance metric prediction network, where To predict the number of hidden layers in a network for performance metrics; The input weight vector for the performance metric prediction network; The activation function for the performance metric prediction network; Step 3.1.3: Based on the characteristics of the three-layer neural network, and by setting the grinding particle size deviation... The error function of the performance index prediction network can then be obtained: (9) In the formula, This represents the ideal value for grinding particle size deviation; Step 3.1.4: Based on the performance index prediction network error function defined in formula (9), the performance index prediction network error is defined as follows: (10) In the formula, Predict network errors for performance metrics; Step 3.1.5: Predict network error based on the performance index of formula (10), and minimize the objective function of predicting network error using the performance index shown in formula (11), and use... The output weight vector of the performance metric prediction network is corrected based on the stopping condition. ,in, A threshold for predicting network error as a performance metric; The objective function for predicting network error based on performance metrics is shown in the following formula: (11) In the formula, The objective function for predicting network error for performance metrics; Step 3.2: Construct a loop preset value adjustment network and determine the network error to adjust the output weight vector of the correction loop preset value adjustment network; Step 3.2.1: The loop preset value adjustment network is used to generate preset values ​​for the process variables of the control loop. Therefore, a loop preset value optimization model is established, expressed as: (12) In the formula, It is an unknown nonlinear function; Step 3.2.2: Establish the loop preset value adjustment network equation as follows: (13) In the formula, These represent the expected values ​​of the process variables in the control loop, respectively. Adjust the input data vector of the network to the preset values ​​of the loop; Adjust the network output weight vector to the preset values ​​of the loop. Adjust the number of hidden layers in the network to a preset value for the loop; Adjust the input weight vector of the network to the preset values ​​of the loop; The activation function of the network is adjusted to a preset value for the loop; Step 3.2.3: Based on the input weight vector It is fixed, only for the output weight vector. Adopt Minimization is used for correction, and the error equation of the established loop preset value adjustment network is as follows: (14) If the loop preset value is the optimal value, then the constraint condition for establishing the error equation of the loop preset value adjustment network is: (15) In the formula, Adjust the network's training error by setting preset values ​​for the loop; This represents the expected value of the performance metric. Step 3.2.4: According to Simply make available The optimized value; The minimum established loop preset value adjusts the network error as follows: (16) In the formula, Adjust the network error to the preset value of the loop; Step 3.2.5: Adjust the network error according to the loop preset value established by formula (16), and minimize the objective function of adjusting the network error by making the loop preset value shown in formula (17) the minimum. To correct the output weight vector as a stopping condition ; The objective function for adjusting network error using loop preset values ​​is shown in the following formula: (17) In the formula, The objective function for adjusting network error to a preset value for the loop; Step 4 includes: Step 4.1: Minimize the objective function of network error and the loop preset value to adjust the performance index of formulas (11) and (17) to correct the output weight vector. and According to the LM algorithm, the weight correction amounts for the performance index prediction network and the loop preset value adjustment network are as follows: (18) (19) In the formula, These are LM parameters and all are greater than 0. ; The learning rate of the network is always greater than 0. for right Jacobian matrix; and These are the weight correction values ​​for the performance index prediction network and the loop preset value adjustment network, respectively. LM parameters The solution expression is: (20) In the formula, It is an adjustable parameter; It is the objective function; Step 4.2: To accelerate the convergence speed, the identity matrix equation is established to replace the identity matrix in the weight correction formulas of the performance index prediction network and the loop preset value adjustment network shown in formulas (18) and (19). : (21) In the formula, A diagonal matrix composed of the main diagonal elements; Step 4.3: Substitute formulas (20) and (21) into formulas (18) and (19) to obtain the weight correction amounts of the performance index prediction network and the loop preset value adjustment network. and .

2. The method for optimizing and controlling grinding particle size under varying operating conditions according to claim 1, characterized in that: Step 5 includes: Step 5.1: Adjust the weights of the two neural networks and Substituting these into the expected value calculation formula (8) for the performance index of the grinding process under variable working conditions and the loop preset value adjustment network equation (13), a series neural network controller is obtained. Step 5.2: Substitute the serial neural network controller into the performance index calculation formula (4) for the operation optimization control of the grinding process under varying working conditions to obtain the optimal grinding particle size output result of the grinding process under abnormal working conditions.

Images