Process Optimization Method Based on the Self-Organizing Characteristics of the Nonferrous Metallurgy System

Abstract

The present invention relates to the technical field of non-ferrous metal metallurgy, and specifically discloses a process optimization method based on the self-organization characteristics of a non-ferrous metallurgy system. It discovers self-organization phenomena from industrial data, mines self-organization laws, reasonably describes the system operation state and the coupling relationship between associated units by analyzing the self-organization characteristics in the material conversion process within the metallurgy system, accurately estimates the key indicators of the process flow based on different operation states, and finally forms a set of operation optimization schemes to reduce the consumption of industrial impurity removal and replacement agents, providing method support for the dynamic-ordered and collaborative-continuous operation of the process, and solving the problem that the traditional non-ferrous metallurgy reaction process adopting the self-organization cascade collaboration optimization method highly depends on historical production data, and the historical production samples may have problems such as noise, incompleteness or low quality, thus affecting the accuracy and reliability of the model.

Description

Technical Field

[0001] This application relates to the technical field of non-ferrous metal metallurgy, and specifically discloses a process optimization method based on the self-organization characteristics of a non-ferrous metallurgy system. Background Art

[0002] Non-ferrous metals usually exist in the form of alloys and need to be refined through specific process technologies. When refining non-ferrous metals, setting up a purification process can effectively improve the quality and purity of non-ferrous metal products and protect the safe and stable operation of equipment. In the purification process of non-ferrous metals, a non-ferrous metallurgy system that is complex, open, irreversible in process, and far from equilibrium, such as the purification process in a hydrometallurgical zinc smelting system, has inherent self-organization.

[0003] For the process optimization of a non-ferrous metallurgy system with self-organization, in the prior art, a new method for optimizing associated unit processes applied in manufacturing is introduced: the self-organization cascade collaboration optimization method. Based on the theory of self-organized criticality, combining the mechanism knowledge of the system with production data, a coupling relationship model of reactors is constructed. Through an online reaction efficiency estimation model, the reaction participation degree of each unit process is analyzed to obtain its actual operating efficiency. Finally, based on the coupling relationship between associated unit processes and the cascade characteristics within the cascade system and unit processes, a global optimization method for the system is constructed. The results show that this method can achieve the global optimization of the cascade system, improve economic benefits compared with existing production conditions, and ensure the stability of the system.

[0004] However, in the process of optimizing the associated unit process using the self-organization cascade collaboration optimization method, its optimization model and optimization strategy highly rely on historical production data in a highly complex or uncertain system. After the state evolves, the category to which the current operating state belongs is obtained by evaluating the similarity between process variables and historical samples. These historical samples may have problems such as noise, incompleteness, or low quality, and when there are problems with low data quality, it may affect the accuracy and reliability of the model.

[0005] The present invention provides a process optimization method based on the self-organization characteristics of a non-ferrous metallurgy system to solve the above problems. Summary of the Invention

[0006] The object of the present invention is to solve the problem that in the traditional non-ferrous metallurgy reaction process using the self-organization cascade collaboration optimization method, it highly relies on historical production data, and the historical production samples may have problems such as noise, incompleteness, or low quality, which may affect the accuracy and reliability of the model.

[0007] To achieve the above object, the basic solution of the present invention provides a process optimization method based on the self-organization characteristics of a non-ferrous metallurgy system, including the following steps:

[0008] Step A1: Determine the associated reactors in the cascade system;

[0009] Step A2: Collect the process variables of all reactors in Step A1, process them to obtain the graph structure data between historical data and process variables, and calculate the reaction efficiency from the historical data;

[0010] Step A3: Based on the power-law characteristics of the associated reactors, calculate and fit the power-law distribution curve, determine the critical value of the reaction efficiency based on the fitted curve, pre-divide the operation status labels of corresponding quantity types according to the critical value, and label the calculated reaction efficiency corresponding to the process variable acquisition time;

[0011] Step A4: Construct an operation status classification model, and use the process variables and the corresponding operation status labels as inputs to train the model;

[0012] Step A5: Establish a reaction efficiency estimation model. Combining the graph structure data, use the process variables divided according to the operation status labels as inputs to train the model, and obtain the reaction efficiency estimation sub-model one adapted to each operation status label;

[0013] Step A6: Real-time collect the process variables of the associated reactors and input them into the operation status classification model to obtain the real-time operation status labels. Input the process variables and the corresponding real-time operation status labels into the reaction efficiency estimation sub-model one to calculate the estimated reaction efficiency. Taking the compliance of the impurity ion concentration at the outlet of the final reactor and the lowest total amount of impurity removal displacement agent addition as the goal, compare the estimated reaction efficiency through the particle optimization algorithm to obtain the optimal impurity removal rate of the main impurity ions at the outlet of the reactor, and determine the optimal amount of impurity removal displacement agent addition to complete the operation optimization.

[0014] Further, during the process of performing Step A1, it also includes the determination of the remaining reactors in the cascade system;

[0015] During the process of performing Step A5, it will also use the process variables of the remaining reactors as inputs to train the reaction efficiency estimation model, and obtain the reaction efficiency estimation sub-model two adapted to each reactor;

[0016] During the process of performing Step A6, it also includes the real-time collection of the process variables of the remaining reactors. Input the process variables of the remaining reactors into each reaction efficiency estimation sub-model two to calculate the estimated reaction efficiency of each reactor. Taking the compliance of the impurity ion concentration at the outlet of the final reactor and the lowest total amount of impurity removal displacement agent addition as the goal, compare the estimated reaction efficiency through the particle optimization algorithm to obtain the optimal impurity removal rate of the main impurity ions at the outlet of each reactor, and determine the optimal amount of impurity removal displacement agent addition to complete the operation optimization.

[0017] Further, in step A2, the calculated historical data includes impurity ion consumption, the input amount of the impurity removal and replacement agent, and the corresponding change in impurity ion concentration. The reaction efficiency is the ratio of the theoretical amount to the actual amount of the impurity removal and replacement agent consumed by the main impurity ions.

[0018] Further, in step A2, the reaction efficiency is the ratio of the theoretical amount to the actual amount of the impurity removal and replacement agent consumed by the main impurity ions. The cascade system is the cobalt removal and purification subsystem of the hydrometallurgical zinc system, and cobalt ions are the main impurity ions at the reactor outlet. The reaction efficiency γ 1 of the associated reactor is calculated as follows:

[0019] ;

[0020] In the formula, is the theoretical amount of the impurity removal and replacement agent consumed by the associated reactor to remove the remaining copper ions, is the theoretical amount of the impurity removal and replacement agent consumed by the associated reactor to remove cobalt ions, is the actual added amount of the impurity removal and replacement agent of the associated reactor, is the added amount of the impurity removal and replacement agent consumed by the associated reactor to remove hydrogen ions, is the added amount of the impurity removal and replacement agent consumed by the associated reactor to remove nickel ions;

[0021] The reaction efficiency γ i of the i-th reactor in the remaining reactors is calculated as follows:

[0022] ;

[0023] In the formula, is the theoretical amount of the impurity removal and replacement agent consumed by the i-th reactor to remove the remaining cobalt ions, is the actual added amount of the impurity removal and replacement agent of the i-th reactor, is the added amount of the impurity removal and replacement agent consumed by the i-th reactor to remove hydrogen ions, is the added amount of the impurity removal and replacement agent consumed by the i-th reactor to remove nickel ions, i = 2, 3, ….

[0024] Further, in step A3, the calculation formula for fitting the power-law distribution curve is as follows:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, d represents the distance from a point to a line, and D represents the sum of the squares of the distances from a point to a line;

[0030] In the formula, X = lnc I , Y = lnf(c I ), lnf(c I ) = a - b lnc I , c I represents the average concentration of the main impurity ions in the previous process within each cycle k, f(c I ) represents the number of times of operation state switching corresponding to the average concentration of the main impurity ions in each previous process, and a and b are linear fitting coefficients;

[0031] In the formula, m 1 , m 2 and m 3 respectively represent the critical values of the reaction efficiency under different operation states.

[0032] Furthermore, in step A4, the training process of the operation state classification model is as follows:

[0033] Take the process variables associated with the reactor and the corresponding operation state labels as the training set input, with the process variables as samples;

[0034] Autonomously sample the training set, randomly select samples and features;

[0035] Construct a decision tree in a recursive manner, divide the data according to the selected features until the number of decision trees reaches a preset value, thereby forming a random forest;

[0036] Establish a regression sub - decision tree and train it with the goal of minimizing the loss function;

[0037] When the loss function reaches the minimum, the training ends. Integrate the voting results of all decision trees and select the category with the most votes as the final prediction result.

[0038] Furthermore, in step A4, the loss function of the established regression sub - decision tree is as follows:

[0039] ;

[0040] In the formula, x i represents the feature vector of the i - th sample, y i is the label value of the i - th sample, j represents the feature dimension, select the j - th feature among all features as the splitting criterion, s represents the threshold used for splitting on the j - th feature, I 1 and I 2 represent the two subsets divided according to the feature j and the splitting point s, and when x i,jThe set of sample indices ≤ s is I 1 (j, s), and vice versa for I 2 (j, s), x i,j represents the j-th eigenvalue of the i-th sample, d 1 and d 2 represent the optimal predicted values respectively selected from two subsets, and L represents the loss function obtained by selecting the optimal (j, s) and the optimal (d 1 , d 2 ).

[0041] Further, in step A5, the reaction efficiency estimation model includes a number of graph convolutional modules and Transformer modules stacked in sequence, and a fully connected layer is established.

[0042] Further, in step A5, the processing steps of the reaction efficiency estimation model for the input data are as follows:

[0043] The graph convolutional module receives the graph structure data and encodes the graph structure data to obtain node features;

[0044] The extracted node features are input into the Transformer for position encoding to introduce time information, and the correlation between different time steps is calculated through the self-attention mechanism to obtain the spatio-temporal features of the nodes, and finally sent to the fully connected layer for integration to obtain the reaction efficiency prediction result.

[0045] The principle and effect of this solution are as follows:

[0046] 1. Compared with the prior art, the present invention discovers self-organization phenomena and mines self-organization laws from industrial data. By analyzing the self-organization characteristics in the process of material transformation in the metallurgical system, it reasonably describes the system operation state and the coupling relationship between associated units, accurately estimates the key indicators of the process flow based on different operation states, and finally forms a set of operation optimization solutions to reduce the consumption of industrial impurity removal and replacement agents, providing method support for the dynamic-orderly and collaborative-continuous operation of the process.

[0047] 2. Compared with the prior art, the present invention constructs an operating state classification model based on the random forest algorithm and constructs a reaction efficiency estimation model operating based on the operating state classification model. By online estimating the reaction efficiency of each reactor, and then setting the optimization goal as the lowest addition amount of reaction substances and the outlet impurity concentration meeting the standard based on the particle optimization algorithm, combined with the reaction efficiency estimation results of each reactor, redistributing the impurity removal tasks of each reactor to obtain the optimized addition amount of reaction substances, which solves the problem that the traditional non-ferrous metallurgy reaction process using the self-organizing cascade cooperation optimization method highly depends on historical production data, and the historical production samples may have problems such as noise, incompleteness or low quality, thereby affecting the accuracy and reliability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0049] Figure 1 FIG. shows a schematic diagram of the purification and impurity removal process proposed in the embodiment of the present application;

[0050] Figure 2 FIG. shows a schematic diagram of the mutual relationship between process variables proposed in the embodiment of the present application;

[0051] Figure 3 FIG. shows a flowchart of the process optimization method based on the self-organizing characteristics of the non-ferrous metallurgy system proposed in the embodiment of the present application;

[0052] Figure 4 FIG. shows the fitting effect diagram of the power-law distribution curve proposed in the embodiment of the present application;

[0053] Figure 5 FIG. shows a schematic diagram of the impurity removal rate range of five reactors in the second-stage process proposed in the embodiment of the present application;

[0054] Figure 6 FIG. shows the best setting curve diagram of the impurity removal rate of each reactor proposed in the embodiment of the present application;

[0055] Figure 7 FIG. shows the comparison diagram of the addition amount of impurity removal displacement agent before and after optimization proposed in the embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] To further elaborate on the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following, in combination with the drawings and preferred embodiments, details the specific implementation manners, structures, features and their effects according to the present invention as follows.

[0057] The following is an example of applying the process optimization method based on the self-organization characteristics of the non-ferrous metallurgy system provided by the present invention to the purification and impurity removal process of a zinc smelter, specifically applied to the cascade system of the hydrometallurgical zinc system: the cobalt removal and purification subsystem.

[0058] As Figure 1 shown, this cascade system realizes the purification of the main impurity ions at the reactor outlet through the purification and impurity removal process, which is divided into a stage I impurity removal process and a stage II impurity removal process in sequence. The equipment includes 7 reactors connected in sequence and 2 thickeners, and they are arranged in a stepped manner. The inlet of each consecutive reactor is located at the upper part. Specifically, in the stage I impurity removal process, it includes a stage I No. 1 reactor, a stage I No. 2 reactor, and a stage I thickener connected in series in sequence. In the stage II impurity removal process, it includes a stage II No. 1 reactor, a stage II No. 2 reactor... a stage II No. 5 reactor, and a stage II thickener connected in series in sequence. The stage II No. 1 reactor is connected in series with the stage I thickener.

[0059] This cascade system is open to the outside world, and its internal operating state is affected by temperature, air pressure, reactant feeding amount, etc. Since the impurity ions overflowing from the stage I impurity removal process to the stage II impurity removal process mainly affect the internal reaction of the stage II No. 1 reactor, and the stage II No. 1 reactor is related to two consecutive process steps of the stage I impurity removal process and the stage II impurity removal process, therefore, in this embodiment, the stage II No. 1 reactor is used as the associated reactor, where cobalt ions are the main impurity ions at the outlet of each reactor.

[0060] Since the purification and impurity removal process is a process in which the concentration of impurity ions gradually decreases, according to the method provided by the present invention, taking the addition amount of the impurity removal and replacement agent of each reactor applied to the entire process as the final optimization goal, by reasonably setting the impurity removal percentage of each reactor, the input of the impurity removal and replacement agent is reduced, thereby saving process raw materials and improving process efficiency. As Figure 3 shown, the provided process optimization method based on the self-organization characteristics of the non-ferrous metallurgy system specifically includes the following steps:

[0061] Step A1: Determine the associated reactor in the cascade system. In this embodiment, it is also necessary to determine the remaining reactors within the same process.

[0062] Specifically, the stage II No. 1 reactor is used as the associated reactor, and it is determined to select the remaining stage II No. 2 reactor to stage II No. 5 reactor in the stage II impurity removal process.

[0063] Step A2: Real-time collect the process variables of all reactors in Step A1, and obtain the graph structure data between the historical data and the process variables after processing, and then calculate the reaction efficiency of each reactor based on the historical data.

[0064] The acquisition of process variables is carried out at equal time intervals in minutes. The acquired process variables include impurity ion concentration, temperature, flow rate, liquid level, potential, and the material quantity of the feeder. The acquired process variables are used as historical process variables and are historical data. In this embodiment, all the process variables acquired within each 120-minute cycle are sorted as a set of data to be preprocessed. After preprocessing operations such as missing value screening, feature scaling, and calculating mean and variance for data standardization, historical data is obtained. The obtained historical data includes impurity ion consumption, the input amount of impurity removal and replacement agent, and the corresponding change in impurity ion concentration.

[0065] Among them, the process of screening for missing values in process variables includes: deleting the missing values of cobalt-containing copper overflow, and obtaining the indexes of the rows without missing values;

[0066] The process of feature scaling for process variables includes: scaling the units of the feature variables concentration and potential for subsequent calculations;

[0067] The process of calculating the mean and variance of process variables and standardizing includes:

[0068] The mean represents the average value of a set of data and is defined as: , where μ is the sample mean, N is the total number of samples, and x i is the i-th data point of a certain process variable.

[0069] The variance measures the degree of dispersion of the data and is defined as: , where σ 2 represents the population variance.

[0070] Data standardization uses Z-score standardization. After standardization, the mean of the data is 0 and the variance is 1. The calculation formula is:

[0071] ;

[0072] Through these calculations, the central tendency and distribution of the data can be obtained, providing high-quality data input for data analysis and machine learning.

[0073] For the calculation of the graph structure data between process variables, it is necessary to first determine the mutual relationship between process variables. As Figure 2 shown, Figure 2 The specific matters of the process variables collected in

[0074] Table 1

[0075]

[0076] Combined with Table 1 and Figure 2, the obtained mutual relationships among process variables include: the impurity ion concentration and the feeder material quantity will affect the reaction efficiency, and at the same time, the change in reaction efficiency will also affect the working state of subsequent reactors. Operating conditions (such as temperature and pH) have a direct impact on both the reaction rate and the removal effect of impurity ions. The material flow rate determines the amount of reactants in each reactor and will affect the balance state of the entire process.

[0077] After determining the mutual relationships among process variables, the construction of the graph structure data among process variables is as follows:

[0078] After preprocessing the data such as screening, cleaning, and standardization, the relationships (edges) among variables are established by combining empirical knowledge and data mining techniques, and they are converted into an adjacency matrix or an adjacency list.

[0079] In a graph structure, the relationships among process variables are usually represented by nodes and edges, and a graph can be represented as:

[0080] (G=(V,E) )

[0081] In the formula, V is the set of nodes, and E is the set of directed edges connecting the nodes.

[0082] In this embodiment, the obtained graph structure data is the adjacency matrix of the graph, and the expression of the adjacency matrix of the graph is as follows:

[0083] ;

[0084] ;

[0085] In the formula, W represents the adjacency matrix, which is a square matrix of size n×n, where n is the number of variable nodes, and w ij is an element of the adjacency matrix, representing the connection relationship between node i and node j, and v ij is the directed edge connecting node i and node j, and v ij ∈E means that there is a directed edge from node i to node j. At this time, w ij =1, otherwise w ij =0.

[0086] The adjacency matrix W characterizes the information transfer direction among node variables and can be used to guide information aggregation in the calculation process of the graph convolutional network, thereby improving the model's learning ability of variable relationships.

[0087] In this embodiment, the ratio of the theoretical amount to the actual amount of the impurity removal and replacement agent consumed by the main impurity ions is used as the reaction efficiency, and the reaction efficiency obtained by calculation describes the operating state of the reactor. Specifically, the calculation formula for the reaction efficiency related to the reactor is as follows:

[0088] ;

[0089] In the formula, γ 1 is the reaction efficiency of the associated reactor, is the theoretical amount of the impurity removal and replacement agent consumed by the associated reactor for removing residual copper ions, is the theoretical amount of the impurity removal and replacement agent consumed by the associated reactor for removing cobalt ions, is the actual addition amount of the impurity removal and replacement agent of the associated reactor, is the addition amount of the impurity removal and replacement agent consumed by the associated reactor for removing hydrogen ions, is the addition amount of the impurity removal and replacement agent consumed by the associated reactor for removing nickel ions.

[0090] For the calculation formula of the reaction efficiency of the second reactor to the fifth reactor in the second stage of impurity removal:

[0091] ;

[0092] In the formula, γ i is the current reaction process, that is, the reaction efficiency of the i-th reactor in the second stage of impurity removal, is the theoretical amount of the impurity removal and replacement agent consumed by the i-th reactor for removing residual cobalt ions, is the actual addition amount of the impurity removal and replacement agent of the i-th reactor, is the addition amount of the impurity removal and replacement agent consumed by the i-th reactor for removing hydrogen ions, is the addition amount of the impurity removal and replacement agent consumed by the i-th reactor for removing nickel ions. In this embodiment, i = 2, 3, 4, 5.

[0093] Step A3: Based on the power-law characteristics of the associated reactor, calculate and fit the power-law distribution curve, determine the critical value of the reaction efficiency based on the obtained curve, pre-divide the corresponding number of types of operation status labels by the critical value, and label the calculated reaction efficiency corresponding to the process variable acquisition time.

[0094] Specifically, the definition of the critical value is judged based on the fitting degree of the fitted power-law distribution curve to the original data.

[0095] For the calculation of the power-law distribution curve, with the fitting effect of the power-law distribution curve as the ultimate goal, the fitting effect is based on the minimum sum of the squares of the distances from points to the line. Traverse all possible value situations to assign values to the operation status, and obtain the final fitted curve. The calculation formula for fitting the power-law distribution curve is as follows:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] In the formula, d represents the distance from a point to a line, and D represents the sum of the squares of the distances from the point to the line;

[0101] In the formula, X = lnc I , Y = lnf(c I ), lnf(c I ) = a - b lnc I , c I represents the median of the main impurity ion concentration in the previous process of each cycle, that is, the median of the main impurity ion concentration in section Ⅰ. The median of the main impurity ion concentration in section Ⅰ within each cycle is regarded as the average concentration of the main impurity ions in section Ⅰ of that cycle. f(c I ) represents the number of operation state switches corresponding to the average concentration of each main impurity ion in section Ⅰ. a and b are linear fitting coefficients. Among them, for the main impurity ion concentration at the inlet of section Ⅰ, it is collected once every 2 hours. Taking k unit times as a cycle, calculate the number of operation state switches corresponding to the main impurity ion concentration in section Ⅰ within each cycle;

[0102] In the formula, m 1 , m 2 and m 3 respectively represent the critical values of the reaction efficiency under different operation states.

[0103] Finally, based on the calculation of the power-law distribution curve, it is determined that m 1 = 0.24, m 2 = 0.35, m 3 = 0.65. And the specific categories of the operation states divided according to the obtained fitted power-law distribution curve are shown in Table 2, and the fitting effect of the power-law distribution curve is as Figure 4 shown.

[0104] Table 2

[0105]

[0106] As Figure 4 shown, the fitting line can be approximated as a monotonically decreasing straight line, which proves the power-law characteristics of the associated reactor. And based on Figure 4 it is concluded that the higher the average concentration of the main impurity ions in section Ⅰ, the fewer the number of operation state switches. When the concentration of the main impurity ions in section Ⅰ is high, the reaction competes for a large amount of impurity removal and displacement agents, so that the operation state continues in the third and second operation states, and the number of switches is less than that of other concentrations.

[0107] According to the calculated fitted power-law distribution curve for division, the historical operating states of the associated reactor are divided into the following four types based on the division boundary, corresponding to four different ranges of the impurity removal and replacement agent utilization rate:

[0108] The first type of operating state: The reaction efficiency is in the range of (0, m 1 ), indicating that a very small amount of the impurity removal and replacement agent added to the reactor can effectively participate in the impurity removal reaction;

[0109] The second type of operating state: The reaction efficiency is in the range of (m 1 , m 2 ), indicating that a small amount of the impurity removal and replacement agent added to the reactor can effectively participate in the impurity removal reaction;

[0110] The third type of operating state: The reaction efficiency is in the range of (m 2 , m 3 ), indicating that most of the impurity removal and replacement agent added to the reactor can effectively participate in the impurity removal reaction;

[0111] The fourth type of operating state: The reaction efficiency is in the range of (m 3 , 1), indicating that almost all of the impurity removal and replacement agent added to the reactor can effectively participate in the impurity removal reaction.

[0112] Step A4: Construct an operating state classification model, and use the process variables of the associated reactor and the corresponding operating state labels as inputs to train the model.

[0113] Specifically, the operating state classification model is constructed based on the random forest algorithm, and the operating steps of the constructed operating state classification model are as follows:

[0114] Data preparation: Prepare the data sets for training and testing, and classify them into training sets and testing sets. The data sets include process variables and corresponding operating state labels;

[0115] Randomly select samples and features: Perform self-sampling on the training set, randomly select a sample subset from the training set with replacement sampling, repeat the extraction multiple times to create a new data set with the same size as the original data set, and randomly extract the features of the original data set to form a feature subset;

[0116] Construct decision trees: Use a recursive method to construct decision trees, divide the data according to the selected features until the number of decision trees reaches the preset value, thereby forming a random forest. Then use the sample subset and the feature subset as the training set of the decision tree, and train with the goal of minimizing the loss function. By randomly extracting features and the segmentation optimal cut vector x j and the cut point S, establish the loss function of the regression sub-decision tree as follows:

[0117] ;

[0118] In the formula, x i represents the feature vector of the i-th sample, i.e., the input data, and y i is the label value of the i-th sample. j represents the feature dimension (index), that is, the j-th feature is selected as the splitting criterion among all features, and s represents the threshold used for splitting on the j-th feature. I 1 and I 2 represent two subsets divided according to the feature j and the splitting point s. Specifically: when the sample index set that satisfies x i,j ≤ s is I 1 (j, s), and vice versa is I 2 (j, s). x i,j represents the j-th feature value of the i-th sample, and d 1 and d 2 represent the optimal prediction values respectively selected from the two subsets. L represents the loss function obtained by selecting the optimal (j, s) and the optimal (d 1 , d 2 ).

[0119] Ensemble decision tree: When the loss function reaches the minimum, the training ends. Integrate the voting results of all decision trees and select the category with the most votes as the final prediction result.

[0120] Evaluate the model: Use the test set to evaluate the performance of the trained random forest model and calculate the accuracy rate.

[0121] In this embodiment, the offline training of the operating state classification model is only performed on the associated reactor: In this training, after inputting the process variables and operating state labels of each period of the associated reactor into the operating state classification model, the random forest algorithm classifies according to the established critical values, divides its operating state categories according to the reaction efficiency range it falls into, and makes it match the operating state that conforms to the operating state label.

[0122] Step A5: Establish a reaction efficiency estimation model. Combine the graph structure data, use the process variables of the associated reactor divided according to the operating state label as the input to train the model, and obtain a reaction efficiency estimation sub-model one adapted to each reactor.

[0123] In this embodiment, the process variables of the remaining reactors are also used as the input to train the reaction efficiency estimation model, and a reaction efficiency estimation sub-model two adapted to each reactor is obtained.

[0124] Specifically, the reaction efficiency estimation model is constructed based on a graph convolutional network (GCN) and a Transformer model, including several sequentially stacked graph convolutional modules and Transformer modules, and a fully connected layer is established. The reaction efficiency estimation model is also equipped with a Huber loss function and an Adam optimizer.

[0125] Taking the graph-structured data constructed according to the mutual relationship between the process flow and process variables as input, and generating an adjacency matrix to represent the spatial association between process variables.

[0126] Through which the GCN (graph convolutional network) learns the topological structure relationship between variables. In each layer of graph convolution, the new feature of node v i is obtained by weighted averaging the features of its neighbor nodes, expressed as:

[0127] ;

[0128] In the formula, N(i) represents the set of neighbor nodes of node x i , c i is the normalization constant of node x i , W (l) is the weight matrix of the l-th layer, and σ is the activation function.

[0129] Then stack multiple graph convolutional layers so that the network can extract higher-level node features, thereby learning more effective graph structure information.

[0130] The Transformer module is responsible for capturing time series features, introducing time information through position encoding, and calculating the correlation between different time steps using the self-attention mechanism to extract high-quality spatio-temporal features. The calculation formula of the self-attention mechanism (Self-Attention) of the Transformer module is as follows:

[0131] ;

[0132] In the formula, Q, K, V are the projections of the input data, which are the query vector, key vector, and value vector respectively, d k represents the dimension of the key vector, and Self-Attention is used to calculate the importance of different time steps.

[0133] ​After the joint feature extraction by GCN and Transformer, the data enters the fully connected layer for prediction. At the same time, the Huber loss function is used for error calculation. This loss function combines the advantages of mean squared error (MSE) and mean absolute error (MAE) to improve the robustness to outliers, thereby avoiding the influence of extreme data on the model. Finally, the target prediction result is obtained. The calculation formula of its Huber loss function is:

[0134] ;

[0135] In the formula, a is the error between the true value and the predicted value, and δ is a hyperparameter that determines the threshold of the error range.

[0136] In addition, the reaction efficiency estimation model also uses the Adam optimizer for parameter update. By adjusting the weights with an adaptive learning rate, the training efficiency and convergence speed are improved. The specific update process is as follows:

[0137] ;

[0138] In the formula, θ t is the parameter after the t-th iteration, is the parameter to be optimized, θ t-1 is the parameter of the previous iteration, α is the learning rate that determines the step size of each parameter update, m t is the exponentially weighted moving average of the current gradient, v t is the exponentially weighted moving average of the square of the current gradient, and ε is a very small number used to prevent division by zero errors.

[0139] The entire training process is continuously iteratively optimized, so that the reaction efficiency estimation model can finally accurately estimate the reaction efficiency prediction results of each reactor: the reaction efficiency γ, providing an optimization basis for industrial production, and then improving production efficiency and resource utilization rate.

[0140] As can be seen from step A4, only the associated reactors are labeled with operation status labels. Therefore, the offline training of the reaction efficiency estimation model includes two parts:

[0141] Offline training based on the historical data of associated reactors. Based on the division of the operation status of associated reactors by the operation status classification model, the historical data of associated reactors in different operation statuses are used as inputs for offline training of associated reactors, and a reaction efficiency estimation sub-model one for different operation status categories of associated reactors is obtained;

[0142] Offline training based on the reactors from the second reactor to the fifth reactor in section II. The historical data of the second reactor to the fifth reactor in section II are used as inputs for offline training of the second reactor to the fifth reactor in section II respectively, and a reaction efficiency estimation sub-model two for different reactors is obtained;

[0143] Specifically, the offline training of the reaction efficiency estimation model includes the following steps:

[0144] After inputting the graph structure data obtained in step 2 into the reaction efficiency estimation model, the original input variables of the associated reactor obtained through the graph convolution module include 7 features. The number of input variable features of the second reactor in section II is the same as the number of process variables of the associated reactor. The original input variables of the third reactor in section II include 6 features, and the original input variables of the fourth reactor and the fifth reactor in section II include 5 features.

[0145] There are 1099 groups of valid data for the associated reactor for training and testing. There are 2325, 2052, 1708, and 1442 groups of valid data for the second to fifth reactors in section II for training and testing respectively. And the reaction efficiency is calculated to label the effective data with the utilization rate of the impurity removal replacement agent. To reduce the influence of outliers on model training, the mean value is taken for every 15 out of 120 data in each group. After preliminary data preprocessing and dataset partitioning, the effective process data is integrated and input into the estimation model.

[0146] And the reaction efficiency estimation model is trained respectively with four operating states of the associated reactor. Among them, the feature dimension of the input vector of the GCN-TS model is 7, and the output vector dimension is 1. Through repeated experiments, the connection situation of the graph network nodes is defined as [[1,2,3,0,1,2,3,4,6],[3,3,4,5,5,5,5,5,5]], the batch size is 32, the number of iterations is selected as 100, the number of nodes in the graph is the feature dimension of the input variable, the feature dimension of each node is set to 8, and the number of attention heads is set to 8, and the number of stacked layers is set to 5.

[0147] The reaction efficiency estimation model is trained respectively for the second to fifth reactors in the impurity removal section II. The feature dimensions of the input vectors of the GCN-TS model are 7, 6, 5, and 5 respectively, and the output vector dimensions are all 1. The connection situations of the graph network nodes are defined as [[1,2,3,0,1,2,3,6],[4,4,4,5,5,5,5,5]], [[1,1,2,0,5,2],[2,4,3,4,4,4]], [[1,1,0,4],[2,3,3,3]], and [[1,1,0,4],[2,3,3,3]] respectively. The settings of the remaining hyperparameters are the same as those of the associated reactor.

[0148] Thus far, a non-linear relationship between the process variables and the reaction efficiency has been constructed based on historical data, and the offline training of the reaction efficiency estimation model is completed.

[0149] Step A6: Collect the process variables of the associated reactor in real time and input them into the operating state classification model to obtain the real-time operating state label. Input the process variables and the corresponding real-time operating state label into the reaction efficiency estimation sub-model 1 to calculate the estimated reaction efficiency. Taking the compliance of the impurity ion concentration at the outlet of the final reactor and the lowest overall addition amount of the impurity removal replacement agent as the goal, compare the estimated reaction efficiency through the particle optimization algorithm to obtain the optimal impurity removal rate of cobalt ions in the reactor, determine the optimal addition amount of the impurity removal replacement agent, and complete the operation optimization of the associated reactor.

[0150] In this embodiment, it also includes the real-time collection of the process variables of the remaining reactors, input the process variables of the remaining reactors into each reaction efficiency estimation sub-model 2 to calculate the estimated reaction efficiency of each reactor. Taking the compliance of the impurity ion concentration at the outlet of the final reactor and the lowest overall addition amount of the impurity removal replacement agent as the goal, compare the estimated reaction efficiency through the particle optimization algorithm to obtain the optimal impurity removal rate of cobalt ions in each reactor, determine the optimal addition amount of the impurity removal replacement agent, and complete the operation optimization of the remaining reactors.

[0151] Specifically, it includes the following steps:

[0152] First, calculate the difference between the inlet and outlet impurity ion concentrations of each reactor at all times, and calculate the ratio of the obtained difference to the inlet impurity ion concentration. Based on statistical methods, obtain the impurity removal rate range of the five reactors in the second stage of impurity removal, as Figure 5 shown.

[0153] Next, use the trained reaction efficiency estimation sub-model corresponding to each reactor to estimate the reaction efficiency of each reactor inputting the real-time collected process variables.

[0154] Finally, using the estimated reaction efficiency, taking the compliance of the impurity ion concentration at the outlet of the final reactor and the lowest overall addition amount of the impurity removal replacement agent as the goal, combine the particle optimization algorithm to obtain the optimal impurity removal rate of cobalt ions in each reactor, and determine the optimal addition amount of the impurity removal replacement agent.

[0155] Specifically, when the impurity removal rate of each reactor is t i (i = 1, 2, 3, 4, 5), combined with the reaction efficiency, the formula for estimating the addition amount of the impurity removal replacement agent for removing impurities in all reactors is as follows:

[0156] ;

[0157] In the formula, T i is the addition amount of the impurity removal replacement agent for each reactor, α is an empirical coefficient or proportionality factor, M A is the relative atomic mass of the main metal, M B is the relative atomic mass of the impurity metal, V is the reactor flow rate, t iThe impurity removal rate for each reactor, γ i is the estimated reaction efficiency for each reactor.

[0158] The goal of the particle optimization algorithm is to find an optimal impurity removal rate allocation that minimizes the total impurity removal agent consumption. The goal and constraints can be expressed as:

[0159] ;

[0160] ;

[0161] ;

[0162] where c out is the set value of the impurity ion concentration at the final outlet of the impurity removal process, and c outmax represents the upper limit requirement of the impurity ion concentration at the final outlet of the impurity removal process. t imin and t imax respectively represent the upper and lower limit requirements of the impurity removal rate of the i-th reactor during the impurity removal process.

[0163] The particle swarm optimization algorithm is a global optimization algorithm that simulates the foraging behavior of a bird flock. Each particle represents a possible solution, that is, an impurity removal rate allocation scheme for a reactor. The particle swarm optimization algorithm continuously adjusts the position and velocity of the particles through iteration to make it approach the optimal solution. The process includes the following steps:

[0164] Input the particle swarm and initialize the particle swarm parameters. The input data includes the particle swarm size, particle dimension, number of iterations, inertia weight, learning factor, and iteration step range;

[0165] Randomly initialize the position and velocity of each particle;

[0166] Iteratively update the velocity and position of each particle, calculate the fitness value of each particle, compare and update the individual historical optimal fitness value and position of each particle, compare and update the group historical optimal fitness value and position, and correspondingly update other parameters such as inertia weight and number of iterations;

[0167] Iteratively obtain the individual historical optimal position, group historical optimal position, individual historical optimal fitness value, and group historical optimal fitness value;

[0168] When the maximum number of iterations is reached or the minimum difference in fitness values between two iterations is reached, the iteration stops, and the optimal solution obtained by the iteration is output, and the optimization ends;

[0169] Time intervals were randomly selected in the test set, and the best setting curve of the impurity removal rate of each reactor obtained using the operation optimization strategy based on the reactor reaction efficiency estimation with self-organization characteristic analysis is shown as Figure 6as shown

[0170] The comparison results of the addition amounts of the impurity removal replacement agent before and after optimization are as Figure 7 shown, and it can be further obtained from Figure 7 that allocating more impurity removal tasks to the associated reactors with high reaction efficiency can reduce the workload of impurity removal in the subsequent Reactor No. 2 to Reactor No. 5 in Section II, which helps to improve the operation efficiency of the process. By calculating the estimated total addition amount of the impurity removal replacement agent after optimization and comparing it with the addition amount of the impurity removal replacement agent in the actual production process, it can be seen that the proposed strategy can reduce the consumption of the impurity removal replacement agent and reduce the waste of industrial raw materials.

[0171] The above are only the preferred embodiments of the present invention, and do not impose any form of limitation on the present invention. Although the present invention has been disclosed above with the preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes. However, as long as it does not depart from the technical content of the present invention, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A process optimization method based on the self-organizing characteristics of nonferrous metallurgical systems, characterized in that: The steps include: Step A1: determining an associated reactor in a cascade system, wherein the cascade system is a cobalt removal purification subsystem of a hydrometallurgical zinc smelting system, and cobalt ions are main impurity ions at the reactor outlet; Step A2: collecting the process variables of all reactors in step A1, obtaining historical data and graph structure data between the process variables through processing, and calculating the reaction efficiency from the historical data; The collected process variables are historical process variables, which are historical data. In step A2, the historical data obtained include impurity ion consumption, input amount of impurity removal replacement agent, and corresponding impurity ion concentration changes. The reaction efficiency is the ratio of the theoretical amount of impurity removal replacement agent consumed by the main impurity ions to the actual amount. Step A3: Based on the power law characteristics of the associated reactor, a power law distribution curve is calculated and fitted, and a critical value of the reaction efficiency is determined based on the fitted curve. The critical value is pre-divided into operation status labels of corresponding quantity types, and the calculated reaction efficiency is marked corresponding to the process variable collection time; In step A3, the calculation formula for fitting the power law distribution curve is as follows: ； ； ； ； In the formula, d represents the distance from the point to the line, and D represents the sum of the squares of the distances from the point to the line; In the formula, , , , c I represents the average concentration of the main impurity ion concentration of the previous process in each cycle k, f(c I ) represents the number of operation state switching corresponding to the average concentration of main impurity ions in each previous process, and a and b are linear fitting coefficients; In the formula, m1, m2 and m3 represent the critical values ​​of reaction efficiency under different operating conditions; Step A4: construct an operation status classification model, and use the process variables and the corresponding operation status labels as input to train the model; Step A5: Establish a reaction efficiency estimation model, combine the graph structure data, use the process variables divided according to the operation state labels as input training models, and obtain a reaction efficiency estimation sub-model 1 adapted to each operation state label; In step A5, the reaction efficiency estimation model includes a plurality of graph convolution modules and Transformer modules stacked in sequence, and a fully connected layer is established; Step A6: Collect the process variables of the associated reactor in real time and input them into the operation status classification model to obtain the real-time operation status label, input the process variables and the corresponding real-time operation status label into the reaction efficiency estimation sub-model to calculate and obtain the estimated reaction efficiency, with the final impurity ion concentration at the reactor outlet meeting the standard and the lowest overall impurity removal and replacement agent addition amount as the goal, compare the estimated reaction efficiency through the particle optimization algorithm to obtain the optimal impurity removal rate of the main impurity ions at the reactor outlet, determine the optimal impurity removal and replacement agent addition amount, and complete the operation optimization.

2. The process optimization method based on the self-organizing characteristics of nonferrous metallurgical system according to claim 1 is characterized in that: The process of performing step A1 also includes determining the remaining reactors in the cascade system; During step A5, the reaction efficiency estimation model is trained based on the process variables of the remaining reactors as input to obtain a second reaction efficiency estimation sub-model adapted to each reactor; The process of performing step A6 also includes real-time collection of process variables of the remaining reactors, inputting the process variables of the remaining reactors into each reaction efficiency estimation sub-model 2 to calculate the estimated reaction efficiency of each reactor, with the goal of achieving the impurity ion concentration at the final reactor outlet and the lowest overall impurity removal and replacement agent addition amount, and comparing the estimated reaction efficiencies through the particle optimization algorithm to obtain the optimal impurity removal rate of the main impurity ions at the outlet of each reactor, determine the optimal impurity removal and replacement agent addition amount, and complete the operation optimization.

3. The process optimization method based on the self-organizing characteristics of nonferrous metallurgical system according to claim 1 is characterized in that: In step A2, the reaction efficiency is the ratio of the theoretical amount of impurity removal agent consumed by the main impurity ions to the actual amount. The cascade system is a cobalt removal purification subsystem of the hydrometallurgical zinc smelting system. Cobalt ions are the main impurity ions at the reactor outlet. The reaction efficiency γ1 of the associated reactor is calculated as follows: ； In the formula, is the theoretical amount of impurity removal agent consumed by the associated reactor to remove the remaining copper ions. is the theoretical amount of impurity removal agent consumed by the associated reactor to remove cobalt ions, is the actual amount of impurity removal agent added to the associated reactor, is the amount of impurity removal agent added to the associated reactor to remove hydrogen ions. is the amount of impurity removal agent consumed by the associated reactor to remove nickel ions; The reaction efficiency of the i-th reactor among the remaining reactors is γ i The formula is as follows: ； In the formula, is the theoretical amount of impurity removal agent consumed by the ith reactor to remove the remaining cobalt ions, is the actual amount of impurity removal agent added to the ith reactor, is the amount of impurity removal agent added to the ith reactor to remove hydrogen ions, is the amount of impurity removal agent added to the i-th reactor for removing nickel ions, i=2,3,….

4. The process optimization method based on the self-organizing characteristics of nonferrous metallurgical system according to claim 1 is characterized in that: In step A4, the training process of the running status classification model is as follows: The process variables of the associated reactor and the corresponding operating status labels are input as the training set, and the process variables are used as samples; Autonomous sampling of the training set, randomly selecting samples and features; Decision trees are constructed recursively to partition the data according to the selected features until the number of decision trees reaches a preset value, thus forming a random forest. Establish a regression sub-decision tree and train it with the goal of minimizing the loss function; The training ends when the loss function reaches the minimum, the voting results of all decision trees are integrated, and the category with the most votes is selected as the final prediction result.

5. The process optimization method based on the self-organizing characteristics of nonferrous metallurgical system according to claim 4 is characterized in that: In step A4, the loss function of the established regression sub-decision tree is as follows: ； In the formula, x i represents the feature vector of the i-th sample, y i is the label value of the i-th sample, j represents the feature dimension, the j-th feature is selected as the segmentation criterion among all features, s represents the threshold used for segmentation on the j-th feature, I1 and I2 represent the two subsets divided according to feature j and segmentation point s, and when x is satisfied i,j The sample index set with ≤s is I1 (j,s), otherwise it is I2 (j,s), x i,j represents the jth eigenvalue of the i-th sample, d1 and d2 represent the optimal prediction values ​​selected from the two subsets respectively, and L represents the loss function obtained by selecting the optimal (j, s) and the optimal (d1, d2).

6. The process optimization method based on the self-organizing characteristics of nonferrous metallurgical system according to claim 1 is characterized in that: In step A5, the reaction efficiency estimation model processes the input data as follows: The graph convolution module receives the graph structure data and encodes the graph structure data to obtain node features; The extracted node features are input into the Transformer for position encoding to introduce time information. The correlation between different time steps is calculated through the self-attention mechanism to obtain the spatiotemporal features of the nodes, and finally sent to the fully connected layer for integration to obtain the reaction efficiency prediction results.

Images