Ball mill discharge particle size control method and system based on digital twin technology

Abstract

The present invention belongs to the field of intelligent control of grinding process, specifically a method and system for controlling the discharge particle size of a ball mill based on digital twin technology. It includes: S1: collecting industrial field data, S2: establishing a mechanism model of discharge particle size; S3: establishing a state and output deviation compensation model based on historical data and the mechanism model, and the data-driven deviation compensation model and the mechanism model together constitute a digital twin model of the grinding system; S4: using the digital twin model of the grinding system to train reinforcement learning controller parameters; S5: transmitting the trained controller parameters to the field controller, and the field controller gives control instructions based on the predicted particle size distribution; S6: when the control system performance deteriorates or the process indicators change, the controller parameters are retrained and the data-driven model is updated online. The present invention introduces the discharge particle size distribution as feedback information into the control system by outputting the deviation compensation model.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent control of grinding processes, and specifically provides a ball mill discharge particle size control method and system based on digital twin technology. Background Art

[0002] Ore grinding is a crucial step in the mineral processing process, where the product particle size and output directly impact various indicators of subsequent mineral processing. Currently, most grinding particle size control systems are based on the traditional indicator of cumulative yield at a specific sieve size. This percentage content only represents the approximate distribution range of particle sizes, while the specific distribution of particle sizes is a key factor affecting subsequent processes. Too fine a particle size leads to excessive energy consumption, reduced mineral selectivity, and loss of useful minerals. Too coarse a particle size prevents the full exposure of useful minerals, hindering separation. Therefore, research is needed to develop grinding circuit control systems that use discharge particle size distribution (PSD) as feedback information and control indicators.

[0003] Ore grinding involves a variety of equipment, characterized by nonlinearity, strong multivariable coupling, and large delays, and is easily affected by changes in the properties of the raw ore and the particle size distribution of the feed. Operations based on manual experience often lead to under-grinding and over-grinding, and the operating levels and handover processes between different work teams have a significant impact on the system, leading to deterioration of flotation indicators and increased costs. Currently, most mineral processing companies still use fixed-parameter proportional-integral-differential (PID) controllers for their basic circuits, which have poor control effects. Good performance of model predictive control (MPC) requires high-quality process models. The ore grinding process has strong nonlinear characteristics, and conventional predictive control based on linear models cannot achieve the purpose of optimal control. Nonlinear model predictive control (NMPC) is needed for more accurate prediction and optimization. However, it places high demands on the analytical nature of the optimization target and is prone to problems such as premature convergence, slow convergence, weak local search capabilities, and dependence on the selection of initial parameters.

[0004] Reinforcement learning (RL) solves the problem of sequential decision-making in uncertain environments by training intelligent agents to make optimal decisions through their interactions with uncertain environments. This is highly consistent with the goal of granular control of grinding loops. Reinforcement learning requires a large amount of data for training. If a trained controller is to be successfully transferred to reality, more generalized scenarios must be developed during the training process, which is relatively difficult. Digital twins involve first establishing a digital twin model corresponding to a physical entity, then using the physical entity's action data to change the state of the digital twin model. By completing the mapping in the digital twin model, the entire life cycle of the corresponding physical equipment can be reflected. Using digital twin models during reinforcement learning training eliminates the need to develop more generalized scenarios and effectively transfers the controller to reality. Summary of the Invention

[0005] In order to solve the problems of single information and poor control effect of existing control systems, the present invention provides a ball mill discharge particle size control method and system based on digital twin technology.

[0006] The present invention adopts the following technical solution: a ball mill discharge particle size control method based on digital twin technology, comprising:

[0007] S1: Collect industrial field data,

[0008] S2: Establish a mechanism model for discharge particle size;

[0009] S3: Based on historical data and the mechanism model, a state and output deviation compensation model is established. The data-driven deviation compensation model and the mechanism model together constitute the digital twin model of the grinding system;

[0010] S4: Training reinforcement learning controller parameters using the grinding system digital twin model;

[0011] S5: The trained controller parameters are transmitted to the field controller, and the field controller gives control instructions according to the predicted particle size distribution;

[0012] S6: When the control system performance deteriorates or the process indicators change, retrain the controller parameters and update the data-driven model online.

[0013] In step S1, the industrial field data includes:

[0014] Process information x: ball mill power, speed, current, filling rate, slurry pool level, slurry pool concentration, overflow fineness and return sand;

[0015] Process input u: slurry pool water supply, ball mill ore supply, ball supply, water supply and cyclone feed;

[0016] Process output y: the cumulative yield of historical discharge particle size and classifier overflow obtained by manual inspection.

[0017] In step S2, the mechanism model is:

[0018]

[0019] Where, is the total water supply volume, is the solid volume in the feed ore, is the volume of fine powder in the ore, is the volume of coarse ore in the feed, is the volume of the grinding medium steel ball, is the total feedwater volume input to the model, is the belt feeding weight, Give the steel ball weight to the belt, is the ore density, is the density of the steel ball, is the proportion of coarse particles in the ore, is the proportion of fine powder in the ore;

[0020]

[0021] Where, is the volume of water in the ball mill, is the water inlet volume of the ball mill, is the water outlet volume of the ball mill, is the solid volume in the ball mill, is the volume of solid entering the ball mill, is the volume of solid discharged from the ball mill, is the volume of fine powder in the ball mill, is the volume of fine powder entering the ball mill, is the volume of fine powder discharged from the ball mill, is the volume of coarse ore in the ball mill, is the volume of coarse ore entering the ball mill, is the volume of the steel balls in the ball mill, is the volume of steel balls entering the ball mill, To consume the volume of coarse ore for crushing, To crush and produce fine powder volume, The volume of steel balls is consumed for crushing;

[0022]

[0023] Where, is the volume of water in the slurry pool, is the volume of water discharged from the ball mill to the slurry pool, is the effluent volume of the slurry pool, is the water volume of the slurry pool, is the solid volume in the slurry pool, is the volume of solids discharged from the ball mill to the slurry pool, is the volume of solids discharged from the slurry pool, is the volume of fine powder in the slurry pool, is the volume of fine powder discharged from the ball mill to the slurry pool, The volume of fine powder discharged from the slurry pool;

[0024]

[0025] Where, is the volume of water in the returned sand, is the volume of fine powder in the returned sand, is the water inlet volume of the classifier, is the volume of fine powder entering the classifier, is the volume of solids in the returned sand excluding fine powder, is the volume content of solids in the returned sand, is the overflow coarse particle volume, is the overflow fine powder volume, is the cumulative yield;

[0026]

[0027] Where, is the solid volume of a certain particle size in the returned sand, is the volume of solids in the returned sand.

[0028] Step S3 includes:

[0029] A deviation compensation model between state and output is established, and the digital twin model of the grinding system is represented as follows:

[0030]

[0031] in, f Mechanism models representing the state, namely the ball mill model and the slurry pool model; g Represents the output mechanism model, that is, the above and the particle size distribution of the discharge ore, 、 Represent the deviation compensation models of the two respectively.

[0032] The construction process of the bias compensation model is as follows:

[0033] According to process information x, process input u and state mechanism model f , we can get the predicted value of process information, and subtract it from the actual measured process information to get the state error ;

[0034] According to process information x, process input u and output mechanism model g , we can get the predicted value of the process output, and make a difference between it and the actual measured process output to get the output error ;

[0035] State deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the state error The predicted value of The training objective function is: ,in:

[0036]

[0037]

[0038] The least squares method is used to solve it and we get ;

[0039] Output deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the output error The predicted value of The training objective function is: ,in:

[0040]

[0041] The least squares method is used to solve it and we get .

[0042] In step S5, the field controller is a neural network controller, which has the same activation function and structure as the neural network in step S4. Its parameters are the weights and biases of the neurons. After the process of step S4, a trained neural network is obtained. The field controller only accepts the weights and biases of the actor network and does not perform training updates. Its input data is the predicted particle size distribution, and its output is the control instruction.

[0043] In step S6, when the difference between the particle size distribution obtained by manual detection and the predicted particle size distribution of the digital twin model is greater than the threshold, the state and output data-driven deviation compensation model is updated online, and then the controller is retrained based on the digital twin model.

[0044] A ball mill discharge particle size control system based on digital twin technology, including:

[0045] Process monitoring and communication modules for collecting industrial field data;

[0046] Digital twin model module, the digital twin model module includes data processing module, parameter identification module and data driven model module,

[0047] The data processing module smoothes the missing values ​​and outliers in the data and then standardizes them;

[0048] The parameter identification module identifies the breakage rate and the breakage distribution constant;

[0049] The data-driven model module establishes a digital twin model of the grinding system based on the crushing rate and crushing distribution constant identified by the parameter identification module;

[0050] Controller learning module, which trains control targets and control rates based on factory indicators and the grinding system digital twin model;

[0051] The process monitoring and communication module transmits the trained controller parameters to the field controller.

[0052] The data processing module smoothes the missing values ​​and outliers in the data: for data collected at industrial sites, the missing values ​​are first filled with the average of the two values ​​above and below the adjacent positions, and then a box plot of the data is drawn. For outliers, the mode of the data is used to replace them.

[0053] The parameter identification module identifies the crushing rate and crushing distribution constant:

[0054] The fragmentation equation is established based on the population balance model:

[0055]

[0056] In the formula is the mass fraction of the material in the i-th particle size, is the crushing distribution function, which represents the mass fraction from the jth to the ith particle size during the crushing process. is the crushing rate of the corresponding particle size, The calculation formula is:

[0057]

[0058] and are the slurry volume and slurry concentration inside the ball mill, J Represents the hardness of the ore. 、 、 、 、 It is the crushing rate constant related to the ore properties and ball mill operating conditions. represents the particle size of the i-th particle size;

[0059] The breakup distribution function is calculated as follows:

[0060]

[0061] , .

[0062] Where i represents the i-th particle size, j represents the j-th particle size, and N represents the total number of ore particle size divisions. represents the mass fraction of the jth particle size broken to the ith particle size, It indicates the percentage of j-th size ore that is smaller than i-th size after crushing. 、 、 is the fragmentation distribution constant, dimensionless, and the final fragmentation rate constant is 、 、 、 、 and the breaking distribution constant 、 、 .

[0063] Compared with the prior art, the present invention has the following beneficial effects:

[0064] The ball mill discharge particle size control system based on digital twin technology proposed in the present invention introduces the discharge particle size distribution as feedback information into the control system through an output deviation compensation model; using reinforcement learning as a controller, the solution problem of the Hamilton-Jacobi-Bellman (HJB) equation is converted into a neural network training problem, which greatly reduces the computational complexity and effectively improves the control accuracy; through digital twin technology, the instability and risk level of migrating and deploying the reinforcement learning controller to the actual industrial site are reduced; the online update of the data-driven model enables the model to have the ability to re-learn. When production conditions change, the particle size of the ball mill discharge product can be quickly stabilized while ensuring production safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is the process flow chart of the grinding circuit;

[0066] Figure 2 This is a schematic diagram of the reinforcement learning controller training proposed by the present invention;

[0067] Figure 3 This is the structural diagram of the ball mill discharge particle size control system based on digital twin technology proposed in this invention;

[0068] Figure 4 This is the flow chart of the ball mill discharge particle size control system based on digital twin technology proposed in this invention. DETAILED DESCRIPTION

[0069] To facilitate understanding of the present invention, the present invention will be described in more comprehensive and detailed form below in conjunction with the accompanying drawings and preferred embodiments. However, the protection scope of the present invention is not limited to the following specific embodiments.

[0070] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to limit the exemplary embodiments according to the present invention.

[0071] Referring to Figures 1-4, the present invention provides a ball mill discharge particle size control system based on digital twin technology, which uses the discharge particle size distribution as feedback information and control indicator, and establishes a digital twin model of the grinding system that combines the mechanism model and the data-driven model, and adopts deep reinforcement learning as the control method.

[0072] A ball mill discharge particle size control method based on digital twin technology, comprising:

[0073] S1: Collect industrial field data;

[0074] Process information x: ball mill power, speed, current, filling rate, slurry pool level, slurry pool concentration, overflow fineness and return sand;

[0075] Process input u: water supply to the slurry pool, ore supply to the ball mill, ball supply, water supply, and material supply to the classifier;

[0076] Process output y: the cumulative yield of historical discharge particle size and classifier overflow obtained by manual inspection.

[0077] S2: Establish a mechanism model for discharge particle size;

[0078] The mechanism model is:

[0079]

[0080] Where, is the total water supply volume, is the solid volume in the feed ore, is the volume of fine powder in the ore, is the volume of coarse ore in the feed, is the volume of the grinding medium steel ball, is the total feedwater volume input to the model, is the belt feeding volume, Give the steel ball volume for the belt, is the ore density, is the density of the steel ball, is the proportion of coarse particles in the ore, The proportion of fine powder in the ore.

[0081]

[0082] Where, is the volume of water in the ball mill, is the water inlet volume of the ball mill, is the water outlet volume of the ball mill, is the solid volume in the ball mill, is the volume of solid entering the ball mill, is the volume of solid discharged from the ball mill, is the volume of fine powder in the ball mill, is the volume of fine powder entering the ball mill, is the volume of fine powder discharged from the ball mill, is the volume of coarse ore in the ball mill, is the volume of coarse ore entering the ball mill, is the volume of the steel balls in the ball mill, is the volume of steel balls entering the ball mill, To consume the volume of coarse ore for crushing, To crush and produce fine powder volume, The volume of the steel balls is consumed for crushing.

[0083]

[0084] Where, is the volume of water in the slurry pool, is the volume of water discharged from the ball mill to the slurry pool, is the effluent volume of the slurry pool, is the water volume of the slurry pool, is the solid volume in the slurry pool, is the volume of solids discharged from the ball mill into the slurry pool, is the volume of solids discharged from the slurry pool, is the volume of fine powder in the slurry pool, is the volume of fine powder discharged from the ball mill to the slurry pool, The volume of fine powder discharged from the slurry pool.

[0085]

[0086] Where, is the volume of water in the returned sand, is the volume of fine powder in the returned sand, is the water inlet volume of the classifier, is the volume of fine powder entering the classifier, is the volume of solids in the returned sand excluding fine powder, is the volume content of solids in the returned sand, is the overflow coarse particle volume, is the overflow fine powder volume, is the cumulative yield.

[0087] S3: Based on historical data and mechanism models, a data-driven deviation compensation model for state and output is established. The data-driven deviation compensation model and the mechanism model together constitute the digital twin model of the grinding system.

[0088] A deviation compensation model between state and output is established, and the digital twin model of the grinding system is represented as follows:

[0089]

[0090] in, f Mechanism models representing the state, namely the ball mill model and the slurry pool model; g Represents the output mechanism model, that is, the above and the particle size distribution of the discharge ore, 、 Represent the deviation compensation models of the two respectively.

[0091] The construction process of the bias compensation model is as follows:

[0092] According to process information x, process input u and state mechanism model f , we can get the predicted value of process information, and subtract it from the actual measured process information to get the state error ;

[0093] According to process information x, process input u and output mechanism model g , we can get the predicted value of the process output, and make a difference between it and the actual measured process output to get the output error ;

[0094] State deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the state error The predicted value of The training objective function is: ,in:

[0095]

[0096]

[0097] The least squares method is used to solve it and we get ;

[0098] Output deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the output error The predicted value of The training objective function is: ,in:

[0099]

[0100] The least squares method is used to solve it and we get .

[0101] S4: Use the grinding system digital twin model to train reinforcement learning controller parameters.

[0102] like Figure 2 As shown, an actor-critic network is constructed as the controller:

[0103] Define an experience buffer with a capacity of 8000, which stores the process data generated by the digital twin model simulation, that is, y t , u t , r t , y t+1 , done ,in y t is the particle size distribution at the current moment y (·) and setting value y *:

[0104] ;

[0105] ;

[0106] ;

[0107] done It is the simulation flag bit, which is 1 when the simulation duration is reached, and 0 otherwise;

[0108] Define the actor network, which consists of a three-layer linear neural network. The weights and biases are initialized in a uniformly distributed manner. The first two layers use ReLU as the activation function, and the last layer does not use an activation function. Adam is used as the parameter optimizer. The target actor network copies the actor network;

[0109] The commentator network is defined as a three-layer linear neural network. The weights and biases are initialized in a uniform distribution. The first two layers use relu as the activation function, and the last layer uses tanh to constrain the output to , and then linearly changes to the range of the control input, using adam as the parameter optimizer, and the target commentator network copies the commentator network;

[0110] Define OU-noise as exploration noise, where the mean , , , the decay constant is set to 10000;

[0111] The constructed digital twin model trains the deep reinforcement learning controller described:

[0112] Initialize the exploration noise and y t , and then obtain the action based on the actor network and noise u t , get rewards based on the system dynamic model and reward function y t+1 and r, and store the data in the experience buffer;

[0113] The critic network is updated based on the minimum mean square error, and the actor network is updated based on the sampled gradient;

[0114] The loss function of the critic network in the actor-critic network structure is:

[0115]

[0116]

[0117] The loss function of the actor network in the actor-critic network structure is:

[0118]

[0119]

[0120] Target network soft update:

[0121]

[0122] in, For the commentator network, For the target reviewer network, is the discount factor, , Indicates the number of samples required for each iteration of the neural network parameters, 、 represents the network parameters of the actor network and the critic network, 、 Represents the parameters of the target network.

[0123] For the convenience of description, the parameters in the deep learning controller are defined as follows:

[0124]

[0125] S5: The trained controller parameters are transmitted to the field controller, and the field controller gives control instructions according to the predicted particle size distribution.

[0126] The scene controller is a neural network controller with the same activation function and structure as the actor network described above. Its parameters are the neuron weights and biases. After the above process, a trained actor network and commentator network are obtained. The commentator network only assists in the training of the actor network. The scene controller only accepts the weights and biases of the actor network and does not perform training updates. Its input data is the predicted particle size distribution, and its output is the control command.

[0127] S6: When the difference between the particle size distribution obtained by manual inspection and the particle size distribution predicted by the digital twin model is greater than a threshold, the data-driven deviation compensation model of the state and output is updated online, and then the controller is retrained based on the digital twin model.

[0128] A ball mill discharge particle size control system based on digital twin technology, including:

[0129] Process monitoring and communication modules for collecting industrial field data;

[0130] Digital twin model module, the digital twin model module includes data processing module, parameter identification module and data driven model module,

[0131] The data processing module smoothes the missing values ​​and outliers in the data and then standardizes them;

[0132] The parameter identification module identifies the breakage rate and the breakage distribution constant;

[0133] The data-driven model module establishes a digital twin model of the grinding system based on the crushing rate and crushing distribution constant identified by the parameter identification module;

[0134] a controller learning module, the controller learning module using the method of S4 in claim 6 to train the control target and control rate according to the plant indicators and the grinding system digital twin model;

[0135] The process monitoring and communication module transmits the trained controller parameters to the field controller.

[0136] The data processing module smoothes the missing values ​​and outliers in the data:

[0137] For the data collected at the industrial site, the missing values ​​are first filled with the average of the two values ​​above and below the adjacent positions, and then a box plot of the data is drawn. For outliers, the mode of the data is used to replace them.

[0138] The parameter identification module identifies the crushing rate and crushing distribution constant:

[0139] The fragmentation equation is established based on the population balance model:

[0140]

[0141] In the formula is the mass fraction of the material in the i-th particle size, is the crushing distribution function, which represents the mass fraction from the jth to the ith particle size during the crushing process. is the crushing rate of the corresponding particle size, The calculation formula is:

[0142]

[0143] and are the slurry volume and slurry concentration inside the ball mill, J Represents the hardness of the ore. 、 、 、 、 It is the crushing rate constant related to the ore properties and ball mill operating conditions. represents the particle size of the i-th particle size;

[0144] The breakup distribution function is calculated as follows:

[0145]

[0146] , .

[0147] Where i represents the i-th particle size, j represents the j-th particle size, and N represents the total number of ore particle size divisions. represents the mass fraction of the jth particle size broken to the ith particle size, It indicates the percentage of j-th size ore that is smaller than i-th size after crushing. 、 、 is the fragmentation distribution constant, dimensionless, and the final fragmentation rate constant is 、 、 、 、 and the breaking distribution constant 、 、 .

[0148] The actual grinding process divides the ore into 12 grades according to the particle size. The three coarsest grades are defined as coarse ore, and the three finest grades are defined as fine powder, which is the mechanism model mentioned above. RC , FP It is replaced by a population balance model.

[0149] The optimization objective of particle swarm optimization is defined as the mean square error between the real data and the simulation data of the ball mill, and the parameter to be optimized is the above-mentioned crushing rate constant 、 、 、 、 , the breaking distribution constant 、 、 , ore hardness J, finally get:

[0150] .

Claims

1. A ball mill discharge particle size control method based on digital twin technology, characterized in that: include: S1: Collect industrial field data, S2: Establish a mechanism model for discharge particle size; The mechanism model is: Where, is the total water supply volume, is the solid volume in the feed ore, is the volume of fine powder in the ore, is the volume of coarse ore in the feed, is the volume of the grinding medium steel ball, is the total feedwater volume input to the model, is the belt feeding weight, Give the steel ball weight to the belt, is the ore density, is the density of the steel ball, is the proportion of coarse particles in the ore, is the proportion of fine powder in the ore; Where, is the volume of water in the ball mill, is the water inlet volume of the ball mill, is the water outlet volume of the ball mill, is the solid volume in the ball mill, is the volume of solid entering the ball mill, is the volume of solid discharged from the ball mill, is the volume of fine powder in the ball mill, is the volume of fine powder entering the ball mill, is the volume of fine powder discharged from the ball mill, is the volume of coarse ore in the ball mill, is the volume of coarse ore entering the ball mill, is the volume of the steel balls in the ball mill, is the volume of steel balls entering the ball mill, To consume the volume of coarse ore for crushing, To crush and produce fine powder volume, The volume of steel balls is consumed for crushing; Where, is the volume of water in the slurry pool, is the volume of water discharged from the ball mill to the slurry pool, is the effluent volume of the slurry pool, is the water volume of the slurry pool, is the solid volume in the slurry pool, is the volume of solids discharged from the ball mill into the slurry pool, is the volume of solids discharged from the slurry pool, is the volume of fine powder in the slurry pool, is the volume of fine powder discharged from the ball mill to the slurry pool, The volume of fine powder discharged from the slurry pool; Where, is the volume of water in the returned sand, is the volume of fine powder in the returned sand, is the water inlet volume of the classifier, is the volume of fine powder entering the classifier, is the volume of solids in the returned sand excluding fine powder, is the volume content of solids in the returned sand, is the overflow coarse particle volume, is the overflow fine powder volume, is the cumulative yield; Where, is the solid volume of a certain particle size in the returned sand, is the volume of solids in the returned sand; S3: Based on historical data and the mechanism model, a state and output deviation compensation model is established. The data-driven deviation compensation model and the mechanism model together constitute the digital twin model of the grinding system; Step S3 includes: A deviation compensation model between state and output is established, and the digital twin model of the grinding system is represented as follows: in, f Mechanism models representing the state, namely the ball mill model and the slurry pool model; g Represents the output mechanism model, that is, the above and the particle size distribution of the discharge ore, 、 Represent the deviation compensation models of the two respectively; The construction process of the bias compensation model is as follows: According to process information x, process input u and state mechanism model f , we can get the predicted value of process information, and subtract it from the actual measured process information to get the state error ; According to process information x, process input u and output mechanism model g , we can get the predicted value of the process output, and make a difference between it and the actual measured process output to get the output error ; State deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the state error The predicted value of The training objective function is: ,in: The least squares method is used to solve it and we get ; Output deviation compensation model It is an extreme learning machine neural network, which consists of three layers of linear neural networks. The weights and biases are initialized in a uniform distribution. The input data is process information x and process input u. The hidden layer output is recorded as , the output is the output error The predicted value of The training objective function is: ,in: The least squares method is used to solve it and we get ; S4: Training reinforcement learning controller parameters using the grinding system digital twin model; S5: The trained controller parameters are transmitted to the field controller, and the field controller gives control instructions according to the predicted particle size distribution; S6: When the control system performance deteriorates or the process indicators change, retrain the controller parameters and update the data-driven model online.

2. The ball mill discharge particle size control method based on digital twin technology according to claim 1 is characterized in that: In step S1, the industrial field data includes: Process information x: ball mill power, speed, current, filling rate, slurry pool level, slurry pool concentration, overflow fineness and return sand; Process input u: slurry pool water supply, ball mill ore supply, ball supply, water supply and cyclone feed; Process output y: the cumulative yield of historical discharge particle size and classifier overflow obtained by manual inspection.

3. The ball mill discharge particle size control method based on digital twin technology according to claim 1 is characterized in that: In step S5, the scene controller is a neural network controller, which has the same activation function and structure as the neural network in step S4, and its parameters are the weights and biases of the neurons. After the step S4 process, a trained neural network is obtained. The scene controller only accepts the weights and biases of the actor network and does not perform training updates. Its input data is the predicted particle size distribution, and its output is a control instruction.

4. The ball mill discharge particle size control method based on digital twin technology according to claim 1 is characterized in that: In step S6, when the difference between the particle size distribution obtained by manual detection and the predicted particle size distribution of the digital twin model is greater than a threshold, the state and output data-driven deviation compensation model is updated online, and then the controller is retrained based on the digital twin model.

5. A ball mill discharge particle size control system based on digital twin technology, characterized in that: include: Process monitoring and communication modules for collecting industrial field data; Digital twin model module, the digital twin model module includes data processing module, parameter identification module and data driven model module, The data processing module smoothes the missing values ​​and outliers in the data and then standardizes them; The parameter identification module identifies the breakage rate and the breakage distribution constant; The data-driven model module establishes a digital twin model of the grinding system in claim 1 based on the crushing rate and crushing distribution constant identified by the parameter identification module; Controller learning module, which trains control targets and control rates based on factory indicators and the grinding system digital twin model; The process monitoring and communication module transmits the trained controller parameters to the field controller.

6. The ball mill discharge particle size control system based on digital twin technology according to claim 5 is characterized in that: The data processing module smoothes the missing values ​​and outliers in the data: for data collected at industrial sites, the missing values ​​are first filled with the average of the two values ​​above and below the adjacent positions, and then a box plot of the data is drawn. For outliers, the mode of the data is used to replace them.

7. The ball mill discharge particle size control system based on digital twin technology according to claim 5 is characterized in that: The parameter identification module identifies the crushing rate and crushing distribution constant: The fragmentation equation is established based on the population balance model: In the formula is the mass fraction of the material in the i-th particle size, is the crushing distribution function, which represents the mass fraction from the jth to the ith particle size during the crushing process. is the crushing rate of the corresponding particle size, The calculation formula is: and are the slurry volume and slurry concentration inside the ball mill, J Represents the hardness of the ore. 、 、 、 、 It is the crushing rate constant related to the ore properties and ball mill operating conditions. represents the particle size of the i-th particle size; The breakup distribution function is calculated as follows: ， Where i represents the i-th particle size, j represents the j-th particle size, and N represents the total number of ore particle size divisions. represents the mass fraction of the jth particle size broken to the ith particle size, It indicates the percentage of j-th size ore that is smaller than i-th size after crushing. 、 、 is the fragmentation distribution constant, dimensionless, and the final fragmentation rate constant is 、 、 、 、 and the breaking distribution constant 、 、 .

Images