Intelligent ore grinding optimization treatment system based on large model

Abstract

The invention discloses an intelligent ore grinding optimization processing system based on a large model, and the system comprises a large model construction module which is used for constructing an ore grinding process model through taking c improved Transformer blocks as a model basis, inputting fusion association data into the ore grinding process model, and outputting index state data; the intelligent decision optimization module is used for constructing an initial individual population through a differential evolution algorithm on the basis of the index state data, substituting each individual in the initial individual population into an objective function to calculate an adaptive value, selecting operation data corresponding to the individual with the highest adaptive value as optimal ore grinding operation data, and outputting the optimal ore grinding operation data; therefore, the system can screen out optimal ore grinding operation data suitable for different conditions when facing ore characteristics and working condition changes, ore grinding effect fluctuation caused by unbalance of a single index is avoided, and stable optimization treatment under all working conditions is achieved.

Description

Technical Field

[0001] The invention relates to the technical field of grinding processing, and in particular to an intelligent optimization processing system for grinding based on a large model. Background Art

[0002] Grinding is an important link in the mineral processing process. It directly affects the efficiency of subsequent mineral processing operations and the quality of concentrates. Indicators such as the particle size distribution of grinding products and the load status of the mill have a crucial impact on the mineral processing recovery rate and concentrate grade. However, the grinding process is a complex physical and chemical process involving the interaction of multiple factors, such as ore properties (hardness, particle size, mineral composition, etc.), mill equipment parameters (speed, power, volume, etc.), and operating parameters (feed rate, water supply, etc.).

[0003] At present, the existing intelligent optimization processing systems for grinding are usually designed based on specific types of ore and relatively stable working conditions. For example, grinding optimization is usually performed on iron ore with a hardness between 4 and 6 on the Mohs hardness scale and a relatively uniform particle size distribution (such as the particle size is mainly concentrated in the range of 5-20 mm). When the hardness of the ore exceeds this range or the particle size distribution changes significantly (such as the appearance of a large number of fine-grained or coarse-grained ores), the system will be unable to adjust the grinding parameters in a timely and effective manner for execution control, resulting in a sharp decline in the grinding effect. Therefore, a large model-based intelligent optimization processing system for grinding is proposed herein. Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and to achieve the above-mentioned purpose, the present invention proposes the following technical solutions:

[0005] A large-scale model-based intelligent optimization processing system for grinding includes:

[0006] Fusion acquisition module: collects ore property data and equipment operation data, and assigns a rank sequence to the collected data in turn, obtains the correlation factor through the Spearman rank correlation coefficient, takes the correlation factor and the ore property data and equipment operation data as input data, and fuses the data through a data fusion model based on a neural network to obtain fused correlation data;

[0007] Large model building module: The grinding process model is built by using c improved Transformer blocks as the model basis, and the fused correlation data is input into the grinding process model to output the indicator status data;

[0008] Intelligent decision-making optimization module: Based on the indicator status data, the initial individual population is constructed through the differential evolution algorithm, each individual in the initial individual population is brought into an objective function to calculate the fitness value, and the operation data corresponding to the individual with the highest fitness value is selected as the optimal grinding operation data;

[0009] Optimization processing module: sends the optimal operating parameters generated by the intelligent decision-making optimization module as adjustment instructions to the actuator of the grinding equipment. The actuator optimizes and adjusts the equipment operation in real time according to the received optimal grinding operation data;

[0010] The acquisition process of the improved Transformer block is:

[0011] Assume the sequence length of the input data is A, the dimension is d, and create a learnable position encoding matrix P∈R A×d , each element in the position encoding matrix P is initialized to a random value, and the position encoding based on the fixed sine-cosine function in the original Transformer block is replaced by a learnable position encoding matrix P∈R A×d , and obtain the improved Transformer block.

[0012] The ore property data is collected by the ore sampling device at a fixed time interval Δt based on the sensor data interface, including ore hardness a and particle size distribution b, which are expressed as

[0013] The equipment operation data is formed by recording the mill speed p, feed amount q, water supply v, and motor current I at a fixed time interval Δt, and is expressed as

[0014] The process of assigning a level sequence to the collected data is as follows:

[0015] Assume that the hierarchical sequence of ore property data is The level sequence of equipment operation data is:

[0016] Hierarchical sequence of ore property data The acquisition process is:

[0017] The hardness values ​​of the ore hardness data and the number of particle size distribution in the ore property data are uniformly quantified and sorted in ascending order. For each data point, a level 1 is assigned according to its position after sorting, the second smallest level is 2, and the largest ore hardness value level is h. The same data is assigned the average level of its position in the sorting to obtain the ore hardness level sequence

[0018] The level sequence of equipment operation data is: The acquisition process is:

[0019] The mill speed, feed rate, water supply and motor current in the equipment operation data are uniformly quantified and sorted from small to large. For each data point, a level 1 is assigned according to its position after sorting. The second smallest level is 2 and the largest data point level is g. The same data is assigned the average level of the position in the sorting to obtain the level sequence of ore hardness.

[0020] The process of obtaining the correlation factor through the Spearman rank correlation coefficient is:

[0021] Grade sequence based on ore property data and level sequence of equipment operating data By Spearman's rank correlation coefficient formula:

[0022]

[0023] Where d is the rank sequence difference between the ore property data and the equipment operation data, n is the number of samples, and r is the correlation factor;

[0024] The process of obtaining fused associated data is as follows:

[0025] Build a neural network-based data fusion model:

[0026] Construct the model structure of the neural network-based data fusion model, including:

[0027] Construct the input layer, and set the number of input layer nodes to i+6;

[0028] Among them, the number of nodes i+6 is the sum of the vector dimension of the ore property data and the vector dimension of the equipment operation data plus the number of associated factors. The vector dimension of the ore property data is 2, the vector dimension of the equipment operation data is 4, and the sum of the number of associated factors is

[0029] Construct a hidden layer, and set the number of hidden layers to 1;

[0030] Construct the output layer, and set the number of output layer nodes to 1, that is, fuse the associated data;

[0031] The activation function uses a linear activation function;

[0032] Collect historical ore property data and equipment operation data, as well as correlation factors and preset fusion correlation data to form a training set. These data are divided into training set and test set according to 7:3. Suppose the number of training samples is N, the actual fusion correlation data of the jth sample is y, and the fusion correlation data predicted by the model is The loss function formula is:

[0033]

[0034] Input the training set data into the neural network, calculate the output of the model through forward propagation, then calculate the loss value according to the loss function, calculate the gradient through the back propagation algorithm, use the optimization algorithm to update the model parameters, repeat this process until the model converges, and obtain a data fusion model based on the neural network;

[0035] Among them, the convergence condition is to reach 50 training rounds;

[0036] The output layer output of the data fusion model of the neural network is the fused associated data e is the number of data points.

[0037] The process of constructing the grinding process model by using c improved Transformer blocks as the model basis is:

[0038] Fusion of linked data As the input layer data of the grinding process model, based on the improved Transformer block, the input data is combined with the position encoding matrix P∈R A×d Perform an addition operation to obtain input data that incorporates position information;

[0039] Stack multiple improved Transformer blocks to obtain the encoding layer of the grinding process model;

[0040] c improved Transformer blocks are stacked in sequence, the output of the previous improved Transformer block is used as the input of the next improved Transformer block, and the output of the cth Transformer block is used as the encoding layer output;

[0041] The output layer of the grinding process model is composed of a fully connected layer. The fully connected layer further maps and integrates the abstract features obtained by the encoding layer and converts them into an output form related to the fused associated data.

[0042] Finally, a grinding process model consisting of an input layer, a coding layer and an output layer is obtained;

[0043] The fused correlation data is input into the grinding process model. The final output of the grinding process model is the particle size distribution index, mill load state index and grinding efficiency index, which constitute the state index data.

[0044] The output form related to the fusion associated data includes a particle size distribution index Mill load status indicators and grinding efficiency index Among them, k1 is the number of particle size distribution index data points, k 2 is the number of mill load status indicator data points, k 3 is the number of data points for the grinding efficiency index.

[0045] The process of constructing the initial individual population through the differential evolution algorithm based on the indicator state data is as follows:

[0046] Based on indicator status data Assume that the initial individual population size is in, That is, the number of all data points, the dimension of each individual is 3;

[0047] Generate an initial value γ for each individual’s mth dimension (m∈1.2.3), each initial value represents an individual. The initial value of each individual is expressed as Based on the differential evolution algorithm, for each target individual in the population, three different individuals are randomly selected By formula Generate mutant individuals, where α is the mutation factor;

[0048] After the mutation operation, for each individual γ and the corresponding mutant individual v, a random number from 0 to m is generated, where m is the dimension of the individual, m=3. For each dimension, when the random number is less than the preset crossover probability or equal to a dimension index randomly selected from 0 to m, the value of the mutant individual v in that dimension is taken, that is, u=v. Otherwise, the value of the test individual u in that dimension is taken from the value of the initial value individual γ in that dimension, that is, u=γ. These initial value individuals γ, mutant individuals v and test individuals u constitute the initial individual population.

[0049] The process of obtaining the optimal grinding operation data is as follows:

[0050] The particle size distribution index corresponding to each initial value individual γ, variant individual and test individual u in the initial individual population is Mill load status indicators and grinding efficiency index Substituting into the objective function, the objective function is expressed as:

[0051]

[0052] in, They are the particle size distribution index, mill load status index and grinding efficiency index data, is the expected value of the mill load status index, is the expected value of the particle size distribution index, w 1 , w 2 and w3 are weight coefficients, is the fitness value of the individual, Refers to any individual in the initial individual population;

[0053] Compare the fitness values ​​of all individuals to obtain the individual with the maximum fitness value, and combine the operation data that produce these optimal index values, set as Where μ is the number of operating parameters and E represents the operating parameter.

[0054] Based on the optimal operating parameters, the optimal operating parameters are sent to the actuator of the grinding equipment. The actuator of the grinding equipment is equipped with a corresponding communication interface to receive data. The microprocessor or controller at the receiving end decodes and analyzes the optimal operating parameters, extracts each operating parameter and its corresponding value, and the actuator accurately adjusts and optimizes the equipment operation according to these parameters.

[0055] The present invention has the following beneficial effects:

[0056] In the present invention, firstly, the ore properties (covering hardness, particle size distribution) and equipment operation data (mill speed, feed rate, water supply, motor current) are processed through the fusion acquisition module, firstly the rank sequence is assigned and the Spearman rank correlation coefficient is used to obtain the correlation factor, and then the fusion correlation data is generated through the neural network data fusion model. By assigning the rank sequence and the Spearman rank correlation coefficient to obtain the correlation factor, the deep correlation relationship between the ore properties (such as hardness, particle size distribution) and the equipment operation data (such as mill speed, feed rate, water supply, motor current) can be excavated. This correlation information is integrated by the neural network data fusion model to generate the fusion correlation data, and the subsequent large model construction module provides effective high-quality input data;

[0057] Secondly, the large model building module uses c improved Transformer blocks to build a grinding process model. Through position coding and multi-layer stacking structure, it deeply mines and integrates the complex features in the associated data, and can keenly perceive the impact of ore properties and working conditions on the grinding process, ensuring that the particle size distribution, mill load status and grinding efficiency indicators output by the model accurately reflect the actual production status, providing a reliable basis for subsequent optimization decisions;

[0058] Finally, the intelligent decision-making optimization module constructs an initial individual population based on the differential evolution algorithm. The initial population is generated based on the indicator state data and the differential evolution algorithm, covering a variety of different grinding operation parameter combinations, providing a starting point for the optimization process. Different individuals represent possible operation modes under different conditions, avoiding the optimization process from falling into the local optimal solution. The particle size distribution index, mill load state index and grinding efficiency index corresponding to each individual in the population are substituted into the objective function to calculate the fitness value. By combining with the comprehensive objective function, the objective function calculates the individual fitness value based on the key indicators of the grinding process and their deviations from the expected values. This enables each individual to be accurately evaluated on the basis of comprehensive consideration of these important factors, and can clearly distinguish which operating parameter combinations are more conducive to achieving the optimization goal of the grinding process, avoiding the problem that traditional optimization methods may only focus on a single indicator and ignore other important factors. This enables the system to screen out the optimal grinding operation data that adapts to different situations when facing changes in ore characteristics and operating conditions, avoiding fluctuations in grinding effects caused by imbalance of a single indicator, and achieving stable optimization processing under all working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a system block diagram of a large-model-based intelligent optimization processing system for grinding proposed by the present invention;

[0060] Figure 2 This is a code operation diagram of a large-model-based intelligent optimization processing system for grinding proposed in the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] Embodiment 1

[0063] like Figure 1 to Figure 2 As shown, the present invention proposes an intelligent optimization processing system for grinding based on a large model, comprising: a fusion acquisition module: collecting ore property data and equipment operation data, and assigning a rank sequence to the collected data in turn, obtaining a correlation factor through the Spearman rank correlation coefficient, taking the correlation factor and the ore property data and the equipment operation data as input data, and performing data fusion through a data fusion model based on a neural network to obtain fused correlation data;

[0064] Ore property data is collected by ore sampling equipment at fixed time intervals Δt based on the sensor data interface, including ore hardness a and particle size distribution b, which are expressed as

[0065] Equipment operation data, by recording the mill speed p, feed amount q, water supply v, motor current I at a fixed time interval Δt, forms equipment operation data, expressed as

[0066] The process of assigning a level sequence to the collected data is as follows:

[0067] Assume that the hierarchical sequence of ore property data is The level sequence of equipment operation data is:

[0068] Specifically, the hierarchical sequence of ore property data The acquisition process is:

[0069] The hardness values ​​of the ore hardness data in the ore property data and the number values ​​of the particle size distribution (the proportion of particles of different particle sizes) are uniformly quantified and sorted in order from small to large. For each data point, a level 1 is assigned according to its position after sorting, the second smallest level is 2, and so on. The largest ore hardness value level is h. For multiple identical data situations (such as multiple identical ore hardness values ​​or the same number of multiple particle size distributions), the average level of these identical values ​​in the sorting position is assigned, so that the ore hardness level sequence is obtained.

[0070] Specifically, the level sequence of equipment operation data is: The acquisition process is:

[0071] The mill speed, feed rate, water supply and motor current in the equipment operation data are uniformly quantified and sorted from small to large. For each data point, a level 1 is assigned according to its position after sorting, the second smallest level is 2, and so on. The largest data point level is g. For multiple identical data situations (such as multiple identical feed rate values ​​or motor current values), the average level of these identical values ​​in the sorting position is assigned, so that the level sequence of ore hardness is obtained.

[0072] The process of obtaining the correlation factor through the Spearman rank correlation coefficient is:

[0073] Grade sequence based on ore property data and level sequence of equipment operating data By Spearman's rank correlation coefficient formula:

[0074]

[0075] Where d is the rank sequence difference between the ore property data and the equipment operation data, n is the number of samples, and r is the correlation factor;

[0076] The process of obtaining fused associated data is:

[0077] Construct a data fusion model based on neural network. The process is:

[0078] Construct the model structure of the neural network-based data fusion model, including:

[0079] Construct the input layer, and set the number of input layer nodes to i+6;

[0080] Specifically, the number of nodes is the sum of the vector dimension of the ore property data and the vector dimension of the equipment operation data plus the number of associated factors. The vector dimension of the ore property data is 2 (2 types of data), the vector dimension of the equipment operation data is 4 (4 types of data), and the sum of the number of associated factors is

[0081] Construct a hidden layer, and set the number of hidden layers to 1;

[0082] Construct the output layer, and set the number of output layer nodes to 1, that is, fuse the associated data;

[0083] The activation function uses a linear activation function;

[0084] Collect historical ore property data and equipment operation data, as well as the calculated correlation factors and preset fusion correlation data to form a training set. These data are divided into training set and test set according to 7:3. The difference between the fusion correlation data predicted by the model and the actual fusion correlation data is measured by the loss function. Suppose the number of training samples is N, the actual fusion correlation data of the jth sample is y, and the fusion correlation data predicted by the model is The loss function formula is:

[0085]

[0086] Input the training set data into the neural network, calculate the output of the model through forward propagation, then calculate the loss value according to the loss function, calculate the gradient through the back propagation algorithm, use the optimization algorithm to update the model parameters, repeat this process until the model converges, and obtain a data fusion model based on the neural network;

[0087] Specifically, the convergence condition is a preset number of training rounds, which is set to 50 here;

[0088] The output layer output of the data fusion model based on neural network is the fused associated data e is the number of data points.

[0089] Large model building module: The grinding process model is built by using c improved Transformer blocks as the model basis, and the fused correlation data is input into the grinding process model to output the indicator status data;

[0090] The acquisition process of the improved Transformer block is:

[0091] Assume input data (fused with related data ) has a sequence length of A and a dimension of d, creating a learnable position encoding matrix P∈R A×d , each element in the position encoding matrix P is initialized to a random value, and the position encoding based on the fixed sine-cosine function in the original Transformer block is replaced by a learnable position encoding matrix P∈R A×d , get the improved Transformer block;

[0092] The process of constructing the grinding process model using c improved Transformer blocks as the model basis is:

[0093] Fusion of linked data As the input layer data of the grinding process model, based on the improved Transformer block, the input data (fused with associated data ) and the position encoding matrix P∈R A×d Perform an addition operation to obtain input data that incorporates position information;

[0094] Specifically, the associated fusion data received by the input layer contains rich information, including ore property data (ore hardness, particle size distribution vector) and equipment operation data (such as mill speed, feed rate, water feed rate, motor current), as well as associated factors. These data are combined to form an input vector and input into the model.

[0095] The encoding layer of the grinding process model is composed of multiple improved Transformer blocks stacked together. Each improved Transformer block contains two parts: a multi-head attention mechanism and a feedforward neural network.

[0096] Specifically, the multi-head attention mechanism is one of the core components of the Transformer block, which allows the model to focus on different parts of the input sequence from different representation subspaces, and after the multi-head attention mechanism, a feed-forward neural network is connected, which contains two linear transformations and an activation function, which can be directly used as components of the grinding process model;

[0097] c improved Transformer blocks are stacked in sequence, and the output of the previous improved Transformer block is used as the input of the next improved Transformer block. Through the multi-layer structure, the model can gradually extract more complex and abstract feature representations in the input data (fused with associated data), and use the output of the cth Transformer block as the encoding layer output;

[0098] Specifically, in this grinding process model, the output of the previous improved Transformer block will be used as the input of the next improved Transformer block. Through the stacking of this multi-layer structure, the model can gradually and deeply mine the input data (i.e., fused associated data, which includes various parameters in the grinding process, such as mill speed, feed rate, water feed rate, particle size distribution, mill load, etc.) More complex and abstract feature representations, for example, the first layer of improved Transformer blocks may extract some simple features, such as the preliminary association between different parameters; as the number of layers increases, subsequent improved Transformer blocks can further extract more advanced features on this basis, such as the deep-level pattern of the impact of certain parameter combinations on grinding efficiency. Finally, the output of the th improved Transformer block is used as the output of the encoding layer, and the output of this encoding layer contains rich abstract feature information obtained after being processed by multiple layers of improved Transformer blocks;

[0099] The output layer of the grinding process model consists of a fully connected layer, which maps and integrates the abstract features extracted by the encoding layer and converts them into an output form related to the fused associated data;

[0100] Specifically, after receiving the abstract features, the fully connected layer will use its own weight matrix and activation function to map these abstract features from difficult-to-understand high-dimensional vectors to meaningful low-dimensional vectors, and integrate them with relevant features such as mill speed, feed rate, and particle size distribution;

[0101] The output forms related to the fusion of linked data include particle size distribution indicators Mill load status indicators and grinding efficiency index Among them, k 1 is the number of particle size distribution index data points, k 2 is the number of mill load status indicator data points, k 3 is the number of data points for the grinding efficiency index;

[0102] Finally, a grinding process model consisting of an input layer, a coding layer and an output layer is obtained;

[0103] Specifically, the encoding layer is stacked in sequence through c improved Transformer blocks. Based on the feature representation extracted from the input fusion associated data (ore property data, equipment operation data, etc.), the fully connected layer is responsible for mapping these high-dimensional abstract features to specific output dimensions to meet the prediction needs of different indicator status data;

[0104] The fused correlation data is input into the grinding process model. The final output of the grinding process model is the particle size distribution index, mill load state index and grinding efficiency index, which constitute the state index data.

[0105] Intelligent decision-making optimization module: Based on the indicator status data, the initial individual population is constructed through the differential evolution algorithm, each individual in the initial individual population is brought into an objective function to calculate the fitness value, and the operation data corresponding to the individual with the highest fitness value is selected as the optimal grinding operation data;

[0106] Based on the indicator status data, the process of constructing the initial individual population through the differential evolution algorithm is:

[0107] Based on indicator status data Assume that the initial individual population size is in, That is, the number of all data points, the dimension of each individual is 3, that is, the individual type, which is divided into 3 types of indicator data (particle size distribution index, mill load status index and grinding efficiency index);

[0108] Generate an initial value γ for each individual’s mth dimension (m∈1.2.3), each initial value represents an individual. The initial value of each individual is expressed as Based on the differential evolution algorithm, for each target individual in the population, three different individuals are randomly selected By formula Generate mutant individuals, where α is the mutation factor (a constant between 0 and 2);

[0109] Specifically, it is mentioned here that each individual contains a set of indicator status data, which actually reflects a specific combination of grinding operation parameters. For example, an individual contains specific operating parameter values ​​such as mill speed of 250 rpm, feed rate of 20 tons / hour, and water supply of 12 cubic meters / hour. These operating parameters are set in actual production to achieve a certain specific grinding effect (i.e., the corresponding indicators);

[0110] For example, for a grinding efficiency indicator The grinding efficiency index is 1.0 ton / kWh. When the speed is increased to 300 rpm, the grinding media can crush the material more effectively, and the grinding efficiency is 1.0 ton / kWh.

[0111] After the mutation operation, for each individual γ and the corresponding mutant individual v (a set of new grinding operation parameters obtained through the mutation operation), a random number (0 to m, m is the dimension of the individual, m = 3) is generated. For each dimension, if the random number is less than the preset crossover probability (the probability of obtaining dimension information from the mutant individual), or is equal to a dimension index randomly selected from 0 to m (to ensure that at least one dimension of the test individual comes from the mutant individual), then the value of the test individual u that meets the rule on this dimension is the value of the mutant individual v on this dimension, that is, u = v. Otherwise, the value of the test individual u on this dimension is the value of the initial value individual γ on this dimension, that is, u = γ. These initial value individuals γ, mutant individuals v and test individuals u form an initial individual population.

[0112] The process of obtaining the optimal grinding operation data is as follows:

[0113] The particle size distribution index corresponding to each initial value individual γ, variant individual and test individual u in the initial individual population is Mill load status indicators and grinding efficiency index Substituting into the objective function, the objective function is expressed as:

[0114]

[0115] in, They are the particle size distribution index, mill load status index and grinding efficiency index data, is the expected value of the mill load status index, is the expected value of the particle size distribution index, w 1 , w 2 and w 3 are weight coefficients, is the fitness value of the individual, Refers to any individual in the initial individual population;

[0116] Compare the fitness values ​​of all individuals to obtain the individual with the maximum fitness value, and combine the operation data that produce these optimal index values, set as Among them, μ is the number of operating parameters, and E represents the operating parameters, such as mill speed, feed rate, water supply, etc.;

[0117] Specifically, for example, suppose that in a grinding optimization scenario, there are 100 different operating parameter combinations (i.e., 100 individuals), each of which contains three operating parameters: mill speed, feed rate, and water supply. By calculating the fitness value of each individual (calculated based on indicators such as particle size distribution, mill load status, and grinding efficiency), it is found that one of the individuals has the largest fitness value. The mill speed corresponding to this individual is 300 rpm, the feed rate is 20 tons / hour, and the water supply is 15 cubic meters / hour. Then the optimal operating data combination can be expressed as In actual production, the operating parameters of the grinding equipment can be set according to this optimal operating data combination in order to achieve the best grinding effect and production efficiency.

[0118] Optimization processing module: sends the optimal operating parameters generated by the intelligent decision-making optimization module as adjustment instructions to the actuator of the grinding equipment. The actuator optimizes and adjusts the equipment operation in real time according to the received optimal grinding operation data;

[0119] Based on the optimal operating parameters, these parameters include specific values ​​such as mill speed, feed rate, water supply, etc. For example, the optimal operating parameters are mill speed of 350 rpm, feed rate of 25 tons / hour, and water supply of 18 cubic meters / hour. These data are sent to the actuator of the grinding equipment. The actuator of the grinding equipment is equipped with a corresponding communication interface (such as Ethernet interface card, field bus interface module), which receives data from the optimization module. The microprocessor or controller at the receiving end decodes and analyzes the optimal operating parameters, extracts each operating parameter and its corresponding value, such as converting the speed parameter into the frequency setting value of the inverter, converting the feed rate parameter into the speed or opening control signal of the feeder, and converting the water supply parameter into the flow control signal of the water pump, etc.;

[0120] Specifically, by analyzing the optimal operating parameters, they can be accurately transmitted to the grinding equipment actuators, and the actuators can accurately adjust and optimize the equipment operation based on these parameters, ensuring that the grinding process is always in an efficient and stable operating state, reducing energy consumption and production costs.

[0121] In the application, several formulas involved are calculated by removing dimensions and taking their numerical values. The formulas are established by collecting a large amount of data and performing software simulation to obtain a formula for the most recent actual situation. Some coefficients or weights in the formulas are set by technicians in this field according to actual conditions, so they will not be elaborated here.

[0122] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented using software, the above embodiments may be implemented in whole or in part in the form of a computer program product. A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented in electronic hardware, or in a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution.

[0123] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A grinding intelligent optimization processing system based on a large model, characterized in that: include: Fusion acquisition module: collects ore property data and equipment operation data, and assigns a rank sequence to the collected data in turn, obtains the correlation factor through the Spearman rank correlation coefficient, takes the correlation factor and the ore property data and equipment operation data as input data, and fuses the data through a data fusion model based on a neural network to obtain fused correlation data; Large model building module: The grinding process model is built by using c improved Transformer blocks as the model basis, and the fused correlation data is input into the grinding process model to output the indicator status data; Intelligent decision-making optimization module: Based on the indicator status data, the initial individual population is constructed through the differential evolution algorithm, each individual in the initial individual population is brought into an objective function to calculate the fitness value, and the operation data corresponding to the individual with the highest fitness value is selected as the optimal grinding operation data; Optimization processing module: sends the optimal operating parameters generated by the intelligent decision-making optimization module as adjustment instructions to the actuator of the grinding equipment. The actuator optimizes and adjusts the equipment operation in real time according to the received optimal grinding operation data; The acquisition process of the improved Transformer block is: Assume the sequence length of the input data is A, the dimension is d, and create a learnable position encoding matrix P∈R A×d , each element in the position encoding matrix P is initialized to a random value, and the position encoding based on the fixed sine-cosine function in the original Transformer block is replaced by a learnable position encoding matrix P∈R A×d , and obtain the improved Transformer block.

2. According to the large model-based intelligent optimization processing system for grinding according to claim 1, it is characterized in that: The ore property data is collected by the ore sampling device at a fixed time interval Δt based on the sensor data interface, including ore hardness a and particle size distribution b, which are expressed as The equipment operation data is formed by recording the mill speed p, feed amount q, water supply v, and motor current I at a fixed time interval Δt, and is expressed as The process of assigning a level sequence to the collected data is as follows: Assume that the hierarchical sequence of ore property data is The level sequence of equipment operation data is: Hierarchical sequence of ore property data The acquisition process is: The hardness values ​​of the ore hardness data and the number of particle size distribution in the ore property data are uniformly quantified and sorted in ascending order. For each data point, a level 1 is assigned according to its position after sorting, the second smallest level is 2, and the largest ore hardness value level is h. The same data is assigned the average level of its position in the sorting to obtain the ore hardness level sequence The level sequence of equipment operation data is: The acquisition process is: The mill speed, feed rate, water supply and motor current in the equipment operation data are uniformly quantified and sorted from small to large. For each data point, a grade of 1 is assigned according to its position after sorting. The second smallest grade is 2 and the largest data point grade is g. The same data is assigned the average grade of its position in the sorting, which results in a grade sequence of ore hardness.

3. The large-model-based intelligent optimization processing system for grinding according to claim 2 is characterized in that: The process of obtaining the correlation factor through the Spearman rank correlation coefficient is: Grade sequence based on ore property data and level sequence of equipment operating data By Spearman's rank correlation coefficient formula: Where d is the rank sequence difference between the ore property data and the equipment operation data, n is the number of samples, and r is the correlation factor; The process of obtaining fused associated data is as follows: Build a neural network-based data fusion model: Construct the model structure of the neural network-based data fusion model, including: Construct the input layer, and set the number of input layer nodes to i+6; Among them, the number of nodes i+6 is the sum of the vector dimension of the ore property data and the vector dimension of the equipment operation data plus the number of associated factors. The vector dimension of the ore property data is 2, the vector dimension of the equipment operation data is 4, and the sum of the number of associated factors is i. Construct a hidden layer, and set the number of hidden layers to 1; Construct the output layer, and set the number of output layer nodes to 1, that is, fuse the associated data; The activation function uses a linear activation function; Collect historical ore property data and equipment operation data, as well as correlation factors and preset fusion correlation data to form a training set. These data are divided into training set and test set according to 7:

3. Suppose the number of training samples is N, the actual fusion correlation data of the jth sample is y, and the fusion correlation data predicted by the model is The loss function formula is: Input the training set data into the neural network, calculate the output of the model through forward propagation, then calculate the loss value according to the loss function, calculate the gradient through the back propagation algorithm, use the optimization algorithm to update the model parameters, repeat this process until the model converges, and obtain a data fusion model based on the neural network; Among them, the convergence condition is to reach 50 training rounds; The output layer output of the data fusion model of the neural network is the fused associated data e is the number of data points.

4. The intelligent optimization processing system for grinding based on a large model according to claim 1 is characterized in that: The process of constructing the grinding process model by using c improved Transformer blocks as the model basis is: Fusion of linked data As the input layer data of the grinding process model, based on the improved Transformer block, the input data is combined with the position encoding matrix P∈R A×d Perform an addition operation to obtain input data that incorporates position information; Stack multiple improved Transformer blocks to obtain the encoding layer of the grinding process model; c improved Transformer blocks are stacked in sequence, the output of the previous improved Transformer block is used as the input of the next improved Transformer block, and the output of the cth Transformer block is used as the encoding layer output; The output layer of the grinding process model is composed of a fully connected layer. The fully connected layer maps and integrates the abstract features obtained by the encoding layer and converts them into an output form related to the fused associated data. Finally, a grinding process model consisting of an input layer, a coding layer and an output layer is obtained; The fused correlation data is input into the grinding process model. The final output of the grinding process model is the particle size distribution index, mill load state index and grinding efficiency index, which constitute the state index data.

5. The large-model-based intelligent optimization processing system for grinding according to claim 4 is characterized in that: The output form related to the fusion associated data includes a particle size distribution index Mill load status indicators and grinding efficiency index Among them, k1 is the number of data points of particle size distribution index, k2 is the number of data points of mill load status index, and k3 is the number of data points of grinding efficiency index.

6. The large-model-based intelligent optimization processing system for grinding according to claim 1 is characterized in that: The process of constructing the initial individual population through the differential evolution algorithm based on the indicator state data is as follows: Based on indicator status data Assume that the initial individual population size is in, That is, the number of all data points, the dimension of each individual is 3; Generate an initial value γ for each individual’s mth dimension (m∈1.2.3), each initial value represents an individual. The initial value of each individual is expressed as Based on the differential evolution algorithm, for each target individual in the population, three different individuals are randomly selected By formula Generate mutant individuals, where α is the mutation factor; After the mutation operation, for each individual γ and the corresponding mutant individual v, a random number from 0 to m is generated, where m is the dimension of the individual, m=3. For each dimension, when the random number is less than the preset crossover probability or equal to a dimension index randomly selected from 0 to m, the value of the mutant individual v in that dimension is taken, that is, u=v. Otherwise, the value of the test individual u in that dimension is taken from the value of the initial value individual γ in that dimension, that is, u=γ. These initial value individuals γ, mutant individuals v and test individuals u constitute the initial individual population.

7. The large-model-based intelligent optimization processing system for grinding according to claim 6 is characterized in that: The process of obtaining the optimal grinding operation data is as follows: The particle size distribution index corresponding to each initial value individual γ, variant individual and test individual u in the initial individual population is Mill load status indicators and grinding efficiency index Substituting into the objective function, the objective function is expressed as: in, They are the particle size distribution index, mill load status index and grinding efficiency index data, is the expected value of the mill load status index, is the expected value of the particle size distribution index, w1, w2 and w3 are weight coefficients, is the fitness value of the individual, Refers to any individual in the initial individual population; Compare the fitness values ​​of all individuals to obtain the individual with the maximum fitness value, and combine the operation data that produce these optimal index values, set as Where μ is the number of operating parameters and E represents the operating parameter.

8. The large-model-based intelligent optimization processing system for grinding according to claim 7 is characterized in that: The process of real-time optimization and adjustment of equipment operation is as follows: Based on the optimal operating parameters, the optimal operating parameters are sent to the actuator of the grinding equipment. The actuator of the grinding equipment is equipped with a corresponding communication interface to receive data. The microprocessor or controller at the receiving end decodes and analyzes the optimal operating parameters, extracts each operating parameter and its corresponding value, and the actuator accurately adjusts and optimizes the equipment operation according to these parameters.

Images