Semi-autogenous mill steel consumption prediction and steel ball supplementing method based on digital twinning

Abstract

The invention discloses a semi-autogenous mill steel consumption prediction and steel ball supplementing method based on digital twinning, and belongs to the technical field of mineral processing equipment. According to the semi-autogenous mill steel consumption prediction and steel ball supplementing method based on digital twinning, the influence of a particle contact model, a particle motion state, an ore crushing probability and a steel ball abrasion mechanism on steel consumption is further considered, and a digital twinning steel consumption prediction model in the ore grinding process is built by fusing machine learning such as an RMSprop genetic neural network; the real-time motion state of the steel ball is simulated, the steel consumption and the real-time steel consumption coefficient under different working conditions are calculated, the steel ball supplementing amount is determined, and the production efficiency is improved. According to the method, the mill and the real-time running state in the mill can be observed, meanwhile, the steel ball supplementing amount can be effectively predicted, and supplementing of the steel balls in actual production is guided.

Description

Technical Field

[0001] The present invention relates to a method for predicting steel consumption of a semi-autogenous grinding mill and adding steel balls based on digital twins, and belongs to the technical field of mineral processing equipment. Background Art

[0002] Semi-autogenous grinding mills (SAG) are essential equipment for pulverizing solid materials in many industries, such as metallurgy, coal, building materials, chemicals, and electric power. Steel balls are the primary grinding medium and are widely used. As the SAG mill operates, the steel balls gradually wear away, and gravity decreases. To compensate for this wear and maintain a consistent ratio of balls of varying sizes within the mill, regular steel ball replenishment is essential.

[0003] Currently, the amount of steel balls added to SAG mills is primarily calculated through manual experience and numerical simulation. Manual experience primarily uses a precise, modified ball-adding method based on the semi-theoretical formula for ball diameter to calculate the amount of steel balls added each time, while numerical simulation primarily relies on the discrete element method, calculating the collision energy between particles and determining the steel ball gradation and addition amount based on energy utilization. However, manual experience can introduce certain errors, and existing numerical simulation methods suffer from high hysteresis and long timescales, making them unable to provide timely guidance on field issues. Therefore, industrial sites urgently need an efficient and accurate method for adding steel balls to SAG mills to guide actual production. Summary of the Invention

[0004] In order to overcome the problems in the background technology, the purpose of the present invention is to provide a method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twin.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0006] A method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twins includes the following steps:

[0007] (1) The actual material parameters of the SAG mill cylinder and the actual parameters of the ore and steel balls inside the SAG mill are obtained to establish the SAG mill cylinder model, ore model and steel ball model. Then, the Hertz-Medellin contact model, particle motion trajectory model, ore crushing probability model and steel ball wear model are integrated, and the RMSprop genetic neural network prediction method is combined to construct a digital twin steel consumption prediction model for the SAG mill.

[0008] (2) Based on the actual working conditions of the SAG mill physical equipment, the boundary conditions required for the simulation process are set, the rotation rate, filling rate and material-ball ratio of the SAG mill digital twin steel consumption prediction model are set, and the numerical simulation calculation of the steel ball wear is started;

[0009] (3) Monitor the real-time operating parameters of the SAG mill, including rotation rate, filling rate, material-to-ball ratio, slurry concentration, slurry viscosity, slurry medium corrosivity, and steel ball quality;

[0010] (4) predicting the steel consumption based on the steel ball wear amount in step (2) and the real-time operating parameters of the physical equipment in step (3), combining the SAG mill digital twin steel consumption prediction model, and obtaining the real-time steel consumption coefficient based on the ore processing capacity of the SAG mill physical equipment;

[0011] (5) Real-time steel consumption coefficient compared with the historical steel consumption coefficient k during the historical maintenance of the SAG mill equipment c If the error is less than or equal to 2%, execute step (6); if it is greater than 2%, modify the weight V in the current SAG mill digital twin steel consumption prediction model. ki 、W jk , re-forecast steel consumption and calculate real-time steel consumption coefficient;

[0012] (6) Calculate the optimal amount of steel balls to be added the next day based on the real-time steel consumption coefficient and ore processing volume.

[0013] The historical steel consumption coefficient = historical steel ball consumption / historical ore processing volume.

[0014] More preferably, the actual material parameters of the SAG mill cylinder include the cylinder geometry and material properties; the actual parameters of the ore and steel balls inside the SAG mill include the ore geometry and material properties.

[0015] More preferably, the fused Hertz-Medellin contact model includes: the normal contact force F between particles n and the tangential contact force F between the particle and the cylinder t :

[0016] F n =-K n δ n +C n v n ;

[0017] Among them, K n is the normal spring stiffness, C n is the normal damping coefficient, δ n is the normal overlap between ore particles, v n is the relative normal velocity;

[0018] F t =-min{μF n ,K t δ t +C t v t}

[0019] Among them, F t Indicates the tangential contact force, min{*} indicates taking a smaller set of internal data, K t is the tangential spring stiffness, C t is the tangential damping coefficient, δ t is the tangential overlap between ore particles, v t is the relative tangential velocity, and μ is the static friction coefficient.

[0020] More preferably, the particle motion trajectory model is: the particle motion trajectory is divided into two parts, one part is a circular motion along the mill cylinder, and the expression is as follows:

[0021] (x-Rsinα) 2 +(y+cosα) 2 =R 2 ;

[0022] The other part is the falling motion, which is expressed as follows:

[0023]

[0024] Among them, R is the radius of the cylinder, α is the particle ejection angle, x is the transverse distance of the cylinder section, and y is the longitudinal distance of the cylinder section.

[0025] More preferably, the ore crushing probability model is:

[0026] S=1-exp{-f Mat x0k(W m,kin -W m,min )};

[0027] Among them, S is the probability of ore crushing, exp{*} represents the x-power function of e, and f Mat represents the crushing characteristics of the ore material, x0 represents the initial particle size of the ore, k represents the number of impact crushing, W m，kin Indicates the mass-to-impact kinetic energy, W m，min Indicates the threshold energy at which no impact fragmentation occurs.

[0028] More preferably, the steel ball wear model is

[0029]

[0030] in, Indicates the wear rate of the steel ball, L1 indicates the impact wear amount,

[0031]

[0032] Where K represents the ball filling rate, ψ represents the filling rate, k1 represents the proportional coefficient, which is related to the slurry concentration, slurry viscosity and the impact strength of the steel ball, R represents the cylinder radius, g represents the acceleration of gravity, and m represents the mass of the steel ball.

[0033] L2 represents the amount of abrasive wear,

[0034] L2=k2S a v t ;

[0035] Among them, k2 represents the proportional coefficient, which is related to the shear strength of the ball filling rate, S a Indicates the surface area of the steel ball being worn away, v t represents the tangential motion speed;

[0036] L3 represents the amount of corrosion wear,

[0037] L3=k3S c v r ;

[0038] Among them, k3 represents the proportional coefficient, which is related to the corrosiveness of the slurry medium, S c Indicates the corroded surface area of the steel ball, v r Indicates relative motion speed.

[0039] More preferably, the RMSprop genetic neural network prediction model includes RMSprop algorithm, GA algorithm and BP neural network;

[0040] The RMSprop genetic neural network prediction training process includes: first initializing the grid to generate the initial weights and thresholds of the BP neural network; then optimizing the weights and thresholds through selection, crossover, and mutation of the GA algorithm; secondly finding the optimal weights and thresholds through BP neural network calculation; and finally calculating the state variables and weights of the gradient through the RMSprop algorithm.

[0041] More preferably, the RMSprop genetic neural network prediction model training process specifically includes:

[0042] S1: The BP neural network uses the rotation rate, filling rate, ball loading rate, slurry concentration, slurry viscosity, slurry medium corrosiveness, steel ball mass, and steel ball wear as the input layer X, the hidden layer is Z, and the number of nodes is q. The output expressions of the hidden layer Z and the output layer Y are as follows:

[0043]

[0044] Among them, X i The i-th node of the input layer, Z k Indicates the output corresponding to the kth node in the hidden layer, Y jIndicates the jth value output by the kth node in the corresponding hidden layer, V ki Represents the weight of the corresponding node between the input layer X and the hidden layer Z, W jk represents the weight of the corresponding node between the hidden layer Z and the output layer Y, f1 and f2 represent the hidden layer neuron activation function and the output layer neuron activation function respectively;

[0045] S2: After determining the initial weight and threshold, calculate the individual fitness value. The calculation formula is as follows:

[0046]

[0047] Among them, F I is the individual fitness function, z I Predict data for the network; y I is the experimental data, s is the number of individuals;

[0048] S3: The optimal weight and threshold are found through selection, crossover, and mutation of the GA algorithm. The individual selection probability is selected based on the roulette wheel selection method. The specific expression is as follows:

[0049]

[0050] Among them, f I represents the fitness value of the individual, N is the number of populations, P I is the individual selection probability function, F I is the individual fitness function;

[0051] Crossover uses the real crossover method, which is expressed as follows:

[0052]

[0053] where c s represents the crossover of individual s, c r represents the crossover of individual r, b represents a random number in [0,1];

[0054] The mutation operation randomly selects individuals from the population and makes slight changes to their sequence information. Its expression is as follows:

[0055]

[0056] Among them, v s is the sequence gene of individual s, c max and c min c s The upper and lower bounds of is the number of iterations, G max is the maximum number of evolutions, h is a random number in [0,1];

[0057] S4: After the GA algorithm is optimized, the BP neural network is used to calculate the optimal weight and threshold. The error function E of the BP neural network is p The expression is as follows:

[0058]

[0059] Among them, p represents the number of input samples, y jp Represents the output value, y' jp represents the expected value, j is the sequence gene;

[0060] The weight W is calculated by chain rule jk and V ki Make adjustments to obtain the adjusted corresponding weight ΔW jk and ΔV ki , the specific expression is as follows:

[0061]

[0062] Where η represents the learning rate, f'1 and f'2 represent the derivative functions of the hidden layer neuron activation function and the output layer neuron activation function respectively;

[0063] S5: The RMSprop algorithm uses differential square weighted average to process weights. The expression of the RMSprop algorithm is:

[0064]

[0065] Among them, S t represents the t-round state variable, α represents the hyperparameter, g t represents the t-round stochastic gradient, W t represents the weight of round t, S t-1 represents the state variable of round t-1, W t-1 represents the weight of round t-1.

[0066] More preferably, the calculation expression of the real-time steel consumption coefficient k is:

[0067]

[0068] Among them, M represents the consumption of steel balls and Q represents the ore processing volume.

[0069] More preferably, the optimal steel ball supplement amount B the next day is

[0070] B = k × Q;

[0071] Among them, B represents the amount of steel balls added the next day.

[0072] Beneficial effects of the present invention: The digital twin-based semi-autogenous grinding mill steel consumption prediction and steel ball addition method described in the present invention further considers the impact of particle contact model, particle motion state, ore crushing probability, and steel ball wear mechanism on steel consumption. It integrates machine learning such as RMSprop genetic neural network to build a digital twin steel consumption prediction model for the grinding process, simulates the real-time motion state of steel balls, calculates steel consumption and real-time steel consumption coefficient under different working conditions, and determines the amount of steel balls to be added, thereby improving production efficiency. The present invention can observe the real-time operating status of the mill and its internal components, and can effectively predict the amount of steel balls to be added, guiding the addition of steel balls in actual production. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 This is a flow chart of the semi-autogenous grinding mill steel consumption prediction and steel ball replenishment method based on digital twin of the present invention.

[0074] Figure 2 This is a schematic diagram of the operation flow of the digital twin steel consumption prediction model for the semi-autogenous grinding mill of the present invention.

[0075] Figure 3 Schematic diagram of particle motion trajectory of the present invention.

[0076] Figure 4 This is the flow chart of the RMSprop genetic neural network of the present invention.

[0077] Figure 5 This is a curve diagram of the simulation results of the wear of the steel balls of the semi-autogenous grinding mill of the present invention.

[0078] Figure 6 This is a simulated particle velocity cloud diagram of the semi-autogenous grinding mill of the present invention.

[0079] Figure 7 This is a schematic diagram for predicting the consumption of steel balls in the semi-autogenous grinding mill of the present invention. DETAILED DESCRIPTION

[0080] The present invention will be further described in detail below with reference to specific embodiments, but the protection scope of the present invention is not limited thereto.

[0081] Example 1

[0082] like Figure 1 A method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twins includes the following steps:

[0083] S1, such as Figure 2 As shown in the figure, a three-dimensional model of the SAG mill cylinder is established, integrating the particle contact model, particle motion state, ore crushing probability model and steel ball wear mechanism model. The RMSprop genetic neural network prediction method is used to construct a digital twin steel consumption prediction model for the SAG mill, which specifically includes:

[0084] S1.1. Set the actual material parameters of the SAG mill cylinder, including the cylinder geometry and material properties; set the actual parameters of the ore and steel balls inside the SAG mill, including the geometry and material properties of the ore and steel balls.

[0085] The material properties of the cylinder include material, shear modulus, density, and Poisson's ratio; the actual parameters of the ore include diameter, shear modulus, density, and Poisson's ratio; and the actual parameters of the steel ball include diameter, mass, shear modulus, density, and Poisson's ratio.

[0086] Furthermore, the cylinder material is set to structural steel, the length is 3.2m, the diameter is 8.0m, and its shear modulus is set to 7×10 10 Pa, density is set to 7800 kg / m 3 , Poisson's ratio is set to 0.3; the ore includes five particle size grades, with diameters of 260 mm, 150 mm, 100 mm, 75 mm, and 40 mm, and its shear modulus is set to 6.87×10 9 Pa, density is set to 3980kg / m 3 , Poisson's ratio is set to 0.25; the diameter of the steel ball is 150 mm, the mass is 15 kg, and the shear modulus is set to 7×10 10 Pa, density is set to 7800 kg / m 3 , the Poisson’s ratio is set to 0.3.

[0087] Use 3D modeling software to build the semi-autogenous grinding mill cylinder model, ore model and steel ball model and import them into the digital twin model.

[0088] S1.2. By integrating the Hertz-Medellin contact model, the particle motion trajectory model, the ore crushing probability model and the steel ball wear model, a semi-autogenous grinding mill steel wear mechanism model is constructed. The integration of the above models can be regarded as building a physical field, which describes the collision between particles, the particle motion trajectory, the ore crushing in the collision and the steel ball wear in the collision. Determining the boundary conditions is equivalent to the initialization of the physical field, which can simulate the particle motion and calculate the steel ball wear. The digital twin steel consumption prediction model is constructed in combination with the RMSprop genetic neural network prediction method. The steel consumption wear mechanism model is used for numerical simulation to calculate the steel ball wear, and the digital twin steel consumption prediction model is used to predict the steel ball consumption. The steps are as follows:

[0089] S1.2.1. Construct the Hertz-Medellin contact model to calculate the normal contact force F between ore particles and between ore particles and the cylinder. n and tangential contact force F t :

[0090] F n =-K n δ n +Cn v n ;

[0091] Among them, K n is the normal spring stiffness, C n is the normal damping coefficient, δ n is the normal overlap between ore particles, v n is the relative normal velocity.

[0092] F t =-min{μF n ,K t δ t +C t v t};

[0093] Among them, F t Indicates the tangential contact force, min{*} indicates taking a smaller set of internal data, K t is the tangential spring stiffness, C t is the tangential damping coefficient, δ t is the tangential overlap between ore particles, v t is the relative tangential velocity, and μ is the static friction coefficient.

[0094] S1.2.2, Construct the particle motion trajectory equation to describe the motion trajectory, such as Figure 3 , the particles include ore and steel balls; the particle motion trajectory is divided into two parts, one part is the circular motion along the mill cylinder, and its expression is as follows:

[0095] (x-Rsinα) 2 +(y+cosα) 2 =R 2 .

[0096] The other part is the falling motion, which is expressed as follows:

[0097]

[0098] Among them, R is the radius of the cylinder, α is the particle ejection angle, x is the transverse distance of the cylinder section, and y is the longitudinal distance of the cylinder section.

[0099] S1.2.3. Construct an ore crushing probability model, expressed as follows:

[0100] S=1-exp{-f Mat x0k(W m,kin -W m,min )};

[0101] Among them, exp{*} represents the power function of e, f Matrepresents the crushing characteristics of the ore material, x0 represents the initial ore particle size, k represents the number of ore impact crushing, W m，kin Indicates the mass-to-impact kinetic energy, W m，min It represents the threshold energy at which no impact crushing occurs, and S is the probability of ore crushing.

[0102] S1.2.4. Construct a steel ball wear mechanism model to calculate steel ball wear. The expression is as follows:

[0103]

[0104] in, represents the wear rate of the steel ball, L1 represents the impact wear, L2 represents the abrasive wear, and L3 represents the corrosive wear. The calculation formulas for each part are as follows;

[0105]

[0106] Among them, K represents the ball filling rate, which is set to 33.3% in this example; ψ represents the filling rate, which is set to 30% in this example; k1 represents the proportional coefficient, which is related to the slurry concentration, slurry viscosity and steel ball impact strength, and is set to 0.1×10 -4 , R represents the radius of the cylinder, which is set to 1600 mm in this example, g represents the acceleration due to gravity, which is set to 9.81 in this example, and m represents the mass of the steel ball.

[0107] L2=k2S a v t ;

[0108] Where k2 represents the proportional coefficient, which is related to the shear strength of the ball filling rate. In this example, it is set to 33.53×10 -4 , S a Indicates the surface area of the steel ball being worn away, v t Indicates the tangential motion speed.

[0109] L3=k3S c v r ;

[0110] Among them, k3 represents the proportional coefficient, which is related to the corrosiveness of the slurry medium. In this example, it is set to 2.93×10 -4 , S c Indicates the corroded surface area of the steel ball, v r Indicates relative motion speed.

[0111] S1.3, such as Figure 4 As shown, the RMSprop genetic neural network prediction model is constructed;

[0112] The RMSprop genetic neural network prediction model includes the RMSprop algorithm, the GA algorithm, and the BP neural network. The training process includes first initializing the grid to generate the initial weights and thresholds of the BP neural network; then optimizing the weights and thresholds through selection, crossover, and mutation of the GA algorithm; secondly finding the optimal weights and thresholds through neural network calculation; and finally calculating the state variables and weights of the gradient through the RMSprop algorithm to improve the convergence speed of the loss function. Specifically, the training process includes:

[0113] S1.3.1. The BP neural network uses the rotation rate, filling rate, ball loading rate, slurry concentration, slurry viscosity, slurry medium corrosivity, steel ball mass and steel ball wear as the input layer X, the hidden layer is Z, and the number of nodes is q. The output expressions of the hidden layer Z and the output layer Y are as follows:

[0114]

[0115] Among them, X i represents the i-th node of the input layer, Z k represents the output corresponding to the kth node in the hidden layer, Y j Indicates the jth value output by the kth node in the corresponding hidden layer, V ki Represents the weight of the corresponding node between the input layer X and the hidden layer Z, W jk represents the weight of the corresponding node between the hidden layer Z and the output layer Y, f1 and f2 represent the hidden layer neuron activation function and the output layer neuron activation function respectively.

[0116] After determining the initial weight and threshold, the individual fitness value is calculated. The calculation formula is as follows:

[0117]

[0118] Among them, z I Predict data for the network; y I is the experimental data, s is the number of individuals, F I is the individual fitness value function.

[0119] S1.3.2. Next, the optimal weights and thresholds are found through selection, crossover, and mutation of the GA algorithm:

[0120] The individual selection probability is selected based on the roulette wheel selection method, and the specific expression is as follows:

[0121]

[0122] Among them, f I represents the fitness value of the individual, N is the number of populations, P I Choose a probability function for an individual.

[0123] Crossover uses the real crossover method, which is expressed as follows:

[0124]

[0125] where c s represents the crossover of individual s, c r represents the crossover of individual r, and b represents a random number in [0,1].

[0126] The mutation operation randomly selects individuals from the population and makes slight changes to their sequence information. Its expression is as follows:

[0127]

[0128] Among them, c s is the sequence gene of individual s, c max and c min c s The upper and lower bounds of is the number of iterations, G max is the maximum number of evolutions, and h is a random number in [0,1].

[0129] S1.3.3. After the GA algorithm is finished, the BP neural network calculation is started, and its error function E p The expression is as follows:

[0130]

[0131] Among them, p represents the number of input samples, y jp Represents the output value, y' jp Denotes the expected value. By using the chain rule, the weight W jk and V ki Make adjustments to obtain the adjusted corresponding weight ΔW jk and ΔV ki , the specific expression is as follows:

[0132]

[0133] Where η represents the learning rate, f'1 and f'2 represent the derivative functions of the hidden layer neuron activation function and the output layer neuron activation function, respectively.

[0134] S1.3.3, RMSprop algorithm uses the weighted average of differential squares to process weights, thereby solving the problem of excessive swing in the loss function during iteration and further improving the convergence speed of the loss function. Taking the tth iteration as an example, the expression of the RMSprop algorithm is:

[0135]

[0136] Among them, S t represents the t-round state variable, α represents the hyperparameter, g t represents the t-round stochastic gradient, W t represents the weight of round t, and η represents the learning rate.

[0137] S2. Based on the actual working conditions, the boundary conditions required for the grinding process are set. The filling rate of the digital twin model is set to 30%, the rotation rate is set to 75%, and the material-ball ratio is set to 2:1. Determining the boundary conditions is equivalent to initializing the physical field. The SAG mill steel wear mechanism model in the digital twin model simulates particle movement and performs numerical simulation calculations on the steel ball wear. The simulation results of the steel ball wear are shown as follows: Figure 5 As shown, the particle movement trajectory in the cylinder is as follows Figure 6 shown.

[0138] S3. Monitor the real-time operating parameters of the SAG mill, including rotation rate, filling rate, material-to-ball ratio, slurry concentration, slurry viscosity, slurry medium corrosivity, and steel ball quality.

[0139] S4. Combining the steel ball wear amount in step S2 and the real-time operating parameters in step S3, the digital twin model predicts the steel consumption. The prediction results are as follows: Figure 7 As shown, combined with the ore processing capacity of the physical equipment, the real-time steel consumption coefficient k is obtained, and the calculation expression is as follows:

[0140]

[0141] Among them, M represents the consumption of steel balls and Q represents the ore processing volume.

[0142] S5. Real-time steel consumption coefficient k compared with historical steel consumption coefficient k during historical maintenance of physical equipment c If the error is less than or equal to 2%, execute S6. If it is greater than 2%, modify the weight V in the current neural network prediction model. ki 、W jk , re-forecast steel consumption and calculate real-time steel consumption coefficient.

[0143] S6. Calculate the optimal steel ball addition amount B for the next day based on the real-time steel consumption coefficient and ore processing volume. The calculation expression is as follows:

[0144] B=k×Q

[0145] Among them, B represents the amount of steel balls added the next day.

[0146] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. A method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twins, characterized by: The steps include: (1) The actual material parameters of the SAG mill cylinder and the actual parameters of the ore and steel balls inside the SAG mill are obtained to establish the SAG mill cylinder model, ore model, and steel ball model respectively. Then, the Hertz-Medellin contact model, particle motion trajectory model, ore crushing probability model, and steel ball wear model are integrated, and the RMSprop genetic neural network prediction method is combined to construct a digital twin steel consumption prediction model for the SAG mill. (2) Based on the actual working conditions of the SAG mill physical equipment, the boundary conditions required for the simulation process are set, including the rotation rate, filling rate, and material-ball ratio. Then, the SAG mill digital twin steel consumption prediction model begins to perform numerical simulation calculations on the steel ball wear. (3) Monitor the real-time operating parameters of the SAG mill; these parameters include rotation rate, filling rate, material-to-ball ratio, slurry concentration, slurry viscosity, slurry medium corrosivity, and steel ball quality; (4) predicting the steel consumption based on the steel ball wear amount in step (2) and the real-time operating parameters of the SAG mill physical equipment in step (3), combining the SAG mill digital twin steel consumption prediction model, and then obtaining the real-time steel consumption coefficient based on the ore processing capacity of the SAG mill physical equipment; (5) Real-time steel consumption coefficient compared with the historical steel consumption coefficient k during the historical maintenance of the SAG mill equipment c If the error is less than or equal to 2%, execute step (6); if it is greater than 2%, modify the weight V in the current SAG mill digital twin steel consumption prediction model. ki 、W jk , re-forecast steel consumption and calculate real-time steel consumption coefficient; (6) Calculate the optimal amount of steel balls to be added the next day based on the real-time steel consumption coefficient and ore processing volume.

2. The method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twinning according to claim 1 is characterized in that: The actual material parameters of the SAG mill cylinder include the cylinder geometry and material properties; the actual parameters of the ore and steel balls inside the SAG mill include the ore geometry and material properties.

3. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1 is characterized in that: The fused Hertz-Medellin contact model includes: the normal contact force F between particles n and the tangential contact force F between the particle and the cylinder t : F n =-K n δ n +C n v n ; Among them, K n is the normal spring stiffness, C n is the normal damping coefficient, δ n is the normal overlap between ore particles, v n is the relative normal velocity; F t =-min{μF n ,K t d t +C t v t }; Among them, F t Indicates the tangential contact force, min{*} indicates taking a smaller set of internal data, K t is the tangential spring stiffness, C t is the tangential damping coefficient, δ t is the tangential overlap between ore particles, v t is the relative tangential velocity, and μ is the static friction coefficient.

4. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1 is characterized in that: The particle motion trajectory model is as follows: the particle motion trajectory is divided into two parts, one part is the circular motion along the mill cylinder, and the expression is as follows: (x-Rsinα) 2 +(y+cosα) 2 =R 2 ; The other part is the falling motion, which is expressed as follows: Among them, R is the radius of the cylinder, α is the particle ejection angle, x is the transverse distance of the cylinder section, and y is the longitudinal distance of the cylinder section.

5. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1 is characterized in that: The ore crushing probability model is: S=1-exp{-f Mat x0k(W m,kin -W m,min )}; Among them, S is the probability of ore crushing, exp{*} represents the x-power function of e, and f Mat represents the crushing characteristics of the ore material, x0 represents the initial particle size of the ore, k represents the number of impact crushing, W m，kin Indicates the mass-to-impact kinetic energy, W m，min Indicates the threshold energy at which no impact fragmentation occurs.

6. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1 is characterized in that: The steel ball wear model is in, Indicates the wear rate of the steel ball, L1 indicates the impact wear amount, Where K represents the ball filling rate, ψ represents the filling rate, k1 represents the proportional coefficient, which is related to the slurry concentration, slurry viscosity and the impact strength of the steel ball, R represents the cylinder radius, g represents the acceleration of gravity, and m represents the mass of the steel ball. L2 represents the amount of abrasive wear, <h2 style=";text-align:left;direction:ltr">L2 = k2S<h2 style=";text-align:left;direction:ltr"> a <h2 style=";text-align:left;direction:ltr"> v<h2 style=";text-align:left;direction:ltr"> t <h2 style=";text-align:left;direction:ltr"> ; Among them, k2 represents the proportional coefficient, which is related to the shear strength of the ball filling rate, S a Indicates the surface area of the steel ball being worn away, v t represents the tangential motion speed; L3 represents the amount of corrosion wear, <h2 style=";text-align:left;direction:ltr">L3=k3S<h2 style=";text-align:left;direction:ltr"> c <h2 style=";text-align:left;direction:ltr"> v<h2 style=";text-align:left;direction:ltr"> r <h2 style=";text-align:left;direction:ltr"> ; Among them, k3 represents the proportional coefficient, which is related to the corrosiveness of the slurry medium, S c represents the corroded surface area of the steel ball, v r Indicates relative motion speed.

7. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1, characterized in that: The RMSprop genetic neural network prediction method includes: first initializing the grid to generate the initial weights and thresholds of the BP neural network; then optimizing the weights and thresholds through selection, crossover, and mutation of the GA algorithm; secondly finding the optimal weights and thresholds through BP neural network calculation; and finally calculating the state variables and weights of the gradient through the RMSprop algorithm.

8. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 7, characterized in that: The RMSprop genetic neural network prediction method specifically includes: S1: The BP neural network uses the rotation rate, filling rate, ball loading rate, slurry concentration, slurry viscosity, slurry medium corrosiveness, steel ball mass and steel ball wear as the input layer X, the hidden layer is Z, and the number of nodes is q. The output expressions of the hidden layer Z and the output layer Y are as follows: Among them, X i The i-th node of the input layer, Z k Indicates the output corresponding to the kth node in the hidden layer, Y j Indicates the jth value output by the kth node in the corresponding hidden layer, V ki Represents the weight of the corresponding node between the input layer X and the hidden layer Z, W jk represents the weight of the corresponding node between the hidden layer Z and the output layer Y, f1 and f2 represent the hidden layer neuron activation function and the output layer neuron activation function respectively; S2: After determining the initial weight and threshold, calculate the individual fitness value. The calculation formula is as follows: Among them, F I is the individual fitness function, z I Predict data for the network; y I is the experimental data, s is the number of individuals; S3: The optimal weight and threshold are found through selection, crossover, and mutation of the GA algorithm. The individual selection probability is selected based on the roulette wheel selection method. The specific expression is as follows: Among them, f I represents the fitness value of the individual, N is the number of populations, P I is the individual selection probability function, F I is the individual fitness function; Crossover uses the real crossover method, which is expressed as follows: where c s represents the crossover of individual s, c r represents the crossover of individual r, b represents a random number in [0,1]; The mutation operation randomly selects individuals from the population and makes slight changes to their sequence information. Its expression is as follows: Among them, v s is the sequence gene of individual s, c max and c min c s The upper and lower bounds of is the number of iterations, G max is the maximum number of evolutions, h is a random number in [0,1]; S4: After the GA algorithm is optimized, the BP neural network is used to calculate the optimal weight and threshold. The error function E of the BP neural network is p The expression is as follows: Among them, p represents the number of input samples, y jp Represents the output value, y' jp represents the expected value, m is the sequence gene; The weight W is calculated by chain rule jk and V ki Make adjustments to obtain the adjusted corresponding weight ΔW jk and ΔV ki , the specific expression is as follows: Among them, η represents the learning rate, f ’ 1 and f ’ 2 represents the derivative function of the hidden layer neuron activation function and the output layer neuron activation function, q is the number of hidden layer nodes, and n is the number of output layer nodes; S5: The RMSprop algorithm uses differential square weighted average to process weights. The expression of the RMSprop algorithm is: Among them, S t represents the t-round state variable, α represents the hyperparameter, g t represents the t-round stochastic gradient, W t represents the weight of round t, S t-1 represents the state variable of round t-1, W t-1 represents the weight of the t-1 round, and η represents the learning rate.

9. The method for predicting steel consumption and adding steel balls for a SAG mill based on digital twinning according to claim 1, characterized in that: The calculation expression of the real-time steel consumption coefficient k is: Among them, M represents the consumption of steel balls and Q represents the ore processing volume.

10. The method for predicting steel consumption and adding steel balls for a semi-autogenous grinding mill based on digital twinning according to claim 1, characterized in that: The optimal steel ball supplement amount B for the next day is B = k × Q; Among them, B represents the amount of steel balls added the next day.

Images