Flotation concentrate grade prediction method based on ISSA-HKLSSVM

Abstract

The invention discloses a flotation concentrate grade prediction method based on ISSA-HKLSSVM, which comprises the following steps: S1, establishing a model for describing a nonlinear relationship between ore element energy intensity and concentrate grade based on data acquired by an X fluorescence analyzer, and establishing a sample database by using channel data measured by the analyzer; s2, introducing a Min-Max data normalization method to pre-process the data, and constructing a data set; s3, optimizing a weight coefficient of a kernel function by adopting the mixed kernel function, and building a concentrate grade prediction model based on an HKLSSVM (Harmonic Kernel Least Square Support Vector Machine); s4, a non-linear exponential logarithm mixed learning factor is introduced to improve a traditional SSA algorithm, kernel function parameters and penalty factors of an HKLSSVM model are optimized through an ISSA algorithm, and a flotation concentrate grade prediction model of ISSA-HKLSSVM is established; and S5, training the grade prediction model based on the ISSA-HKLSSVM by adopting the data of the data set to realize the concentrate grade prediction in the flotation process. According to the method, accurate prediction of the concentrate grade can be effectively achieved, and effective and reliable criteria are provided for optimal control of the flotation process and improvement of flotation efficiency.

Description

Technical Field

[0001] The present invention relates to the field of flotation concentrate grade detection and analysis in the process of mine dressing and smelting, and in particular to a flotation concentrate grade prediction method based on ISSA (Improved Sparrow Search Algorithm)-HKLSSVM (Optimized Mixed Kernel Least Squares Support Vector Machine). Background Art

[0002] my country is a country with abundant mineral resources, but its ores are poor, fine, and impure, and the grade is generally low. In order to meet the demand for high-quality furnace materials, my country has vigorously developed mineral processing technologies represented by flotation. The grade of concentrate is an important economic indicator for measuring flotation production. Stabilizing and improving the grade of concentrate has great economic benefits for enterprises. At present, the measurement of flotation concentrate grade is mainly obtained through manual laboratory testing. The test cycle is long, which makes it difficult for mines to grasp the level of flotation technology in time, resulting in delayed adjustment of flotation operations, unstable products and waste of metals. Therefore, the modeling and prediction of concentrate grade in the flotation process has received more and more attention.

[0003] In actual production, most of the research on the prediction of concentrate grade in flotation process focuses on establishing prediction models using key process parameters of flotation process. This modeling method cannot fully consider the hysteresis and coupling characteristics of the actual flotation process, which will introduce significant errors. In addition, the interpretation and extrapolation capabilities of the neural network modeling method are not strong.

[0004] At present, some mines have placed BOXA grade analyzers on site to achieve real-time monitoring of mineral grades. The BOXA grade analyzer establishes a strength-grade model based on flow channel data, performs regression analysis on variables, and establishes an ore grade regression model based on the least squares method. However, when the training samples of the model are significantly different from the test samples used to predict or verify the model, this limitation may make it difficult for the model to achieve accurate predictions.

[0005] Therefore, it is necessary to design a new method for predicting flotation concentrate grade. Summary of the invention

[0006] The technical problem to be solved by the present invention is to provide a flotation concentrate grade prediction method based on ISSA-HKLSSVM. The flotation concentrate grade prediction method based on ISSA-HKLSSVM constructs a high-precision and high-accuracy flotation concentrate grade prediction model by utilizing the intensity-grade model to obtain channel data of different elements, thereby realizing optimized control of the flotation process and improving flotation efficiency.

[0007] The technical solution of the invention is as follows:

[0008] A flotation concentrate grade prediction method based on ISSA-HKLSSVM comprises the following steps:

[0009] Step 1, using the ore element energy intensity data and concentrate grade data of each energy channel in the X-ray fluorescence analyzer to build a model sample database;

[0010] The model is a model used to describe the nonlinear relationship between energy intensity and grade;

[0011] Step 2, introduce the Min-Max data normalization method to preprocess the data and construct a data set;

[0012] Step 3, using a hybrid kernel function and optimizing the weight coefficient of the kernel function to build a concentrate grade prediction model based on HKLSSVM;

[0013] Step 4, introducing a nonlinear exponential logarithmic hybrid learning factor to improve the SSA algorithm, using the ISSA algorithm to optimize the kernel function parameters and penalty factors of the HKLSSVM model, and establishing an ISSA-HKLSSVM flotation concentrate grade prediction model;

[0014] Step 5: Use the data in the data set to train the ISSA-HKLSSVM flotation concentrate grade prediction model to achieve the flotation process concentrate grade prediction.

[0015] In step 1, the ore element energy intensity data and concentrate grade data of each energy channel are: iron element energy channel, zinc element energy channel, lead element energy channel, scandium element energy channel data and zinc concentrate grade data.

[0016] Some measured data are shown in Table 1 below.

[0017] Table 1 Partial measured data of flotation concentrate grade prediction model

[0018]

[0019] The pretreatment in step 2 specifically includes: for the five variables of flotation concentrate grade, outlier processing is performed based on the 2σ principle. The basic process is as follows:

[0020]

[0021] The 2σ rule is used to screen the data and eliminate outliers whose difference from the mean value is greater than two times the standard deviation; data is the actual value, is the average value, N is the number of values ​​contained in a channel; σ is the standard deviation;

[0022] The data set is normalized and the model is:

[0023]

[0024] Where: z * is the result after data normalization; z is the data to be normalized; z max is the maximum value of the required normalized data; z min is the minimum value of the required normalized data.

[0025] In step 3, the least squares support vector machine (LSSVM) model is:

[0026]

[0027] Among them, f(x i ) is the optimal decision function of the LSSVM model; x i =x i,1 ,x i,2 ,...,x i,D represents sample input data, D is the dimension of input data, N is the number of samples; K(x,x i ) kernel function, where x is the center of the kernel function; α i represents the Lagrange multiplier, which introduces constraints and transforms the original optimization problem into the problem of finding the extreme value of the Lagrangian function; b is the regression deviation, which is used to characterize the gap between the predicted result and the actual result. Its specific value is determined by the training optimization process of the model.

[0028] On the basis of the LSSVM model, the ordinary kernel function is replaced by a hybrid kernel function as a constraint to obtain a hybrid kernel least squares support vector machine model, namely, a HKLSSVM model; (Hybrid Kernel Least Squares Support Vector Machine, HKLSSVM), the HKLSSVM model is as follows:

[0029] Kernel functions are divided into global kernel functions and local kernel functions, which mainly realize the mapping of samples from low-dimensional to high-dimensional feature space. The global kernel function has good generalization performance for long-distance samples, but poor learning performance for close-distance samples; the local kernel function has strong learning performance but weak generalization performance. By constructing a hybrid kernel function, the advantages of both can be effectively combined.

[0030] The Gaussian radial basis function is a kernel function with strong locality, which can map a sample into a higher-dimensional space. Based on this advantage, it is selected as the local kernel function of the concentrate grade prediction model.

[0031] The radial basis function is selected as the local kernel function of the concentrate grade prediction model:

[0032]

[0033] Among them, Σ is the kernel width parameter in the Gaussian kernel function, and its exact value will be determined as a model parameter by the subsequent optimization algorithm.

[0034] The kernel function represented by the polynomial kernel function has good global characteristics, combined with the characteristics of the polynomial kernel function. The polynomial kernel function is selected as the global kernel function of the prediction model:

[0035] K2(x i ,x j )=(x i T ·x j +1) d ,d∈N *

[0036] (Formula 6)

[0037] Where: d is the polynomial kernel function parameter, and its exact value will also be determined by the subsequent optimization algorithm, d∈N * It means that d belongs to the set of natural numbers;

[0038] Using the above classical kernel functions to construct a hybrid kernel function through linear combination, the admissible kernel condition is still satisfied:

[0039] K(x i ,x j )=aK1(x i ,x j )+(1-a)K2(x i ,x j )

[0040] (Formula 7)

[0041] By adjusting the weight coefficient a, the local and global characteristics of the hybrid kernel function can be balanced. Selecting the weight value a as 0.5 can take into account both learning ability and generalization ability.

[0042] The Improved Sparrow Search Algorithm (ISSA) is introduced to optimize the parameters to be tested of the HKLSSVM model to solve the problem that the traditional Sparrow Search Algorithm cannot jump out of the local extreme value in the later stage of the search. The improvement of the ISSA algorithm in the finder position update method is specifically manifested as follows:

[0043] In order to improve the problem in the traditional sparrow algorithm that, with the number of iterations, the finder's search space gradually decreases and the probability of falling into the local space increases, a nonlinear exponential logarithmic hybrid learning factor is introduced in the finder position update method. This factor is required to have a larger value in the early stage of the search, which helps to quickly conduct global exploration and improve search efficiency; in the later stage of the search, it has a smaller value, which can further enhance the local exploration ability and improve accuracy.

[0044] The nonlinear exponential logarithmic mixed learning factor is introduced, and its calculation formula is:

[0045]

[0046] In the formula, w min With w max They represent the maximum and minimum values ​​of the hybrid learning factors obtained during the search process; iter max The maximum number of iterations is a constant, which indicates the maximum number of iterations. Generally, 100 iterations are selected as the benchmark.

[0047] The improved method of updating the finder's location is as follows:

[0048]

[0049] Where, X ij represents the position information of the i-th sparrow in the j-th dimension; t represents the current iteration number; w represents the hybrid learning factor calculated by formula 8; r1 is a random number in [0,2π], r2 is a random number in [0,2], R2∈(0,1] represents the warning value, ST∈(0.5,1] represents the safety value, X best Represents the current global optimal position, and its value will be updated according to the optimal solution found by the sparrow algorithm.

[0050] The Improved Sparrow Search Algorithm (ISSA) is introduced to optimize the parameters to be tested of the HKLSSVM model to solve the problem that the traditional Sparrow Search Algorithm cannot jump out of the local extreme value in the later stage of the search. The improvement of the ISSA algorithm in the follower position update method is specifically manifested as follows:

[0051] When the discoverer iterates a certain number of times and the fitness value remains unchanged, the follower becomes the discoverer and the algorithm will fall into a local optimum. To avoid this, the levy flight strategy is introduced into the follower update formula to improve the global search capability. The levy flight strategy is described as follows:

[0052]

[0053] Where: r3 and r4 are random numbers in [0,1], the value of ζ is 1.5, and (i)! refers to the gamma function.

[0054] The follower position update formula after the introduction of the improved Levy flight strategy is as follows:

[0055]

[0056] Where, X P Indicates the optimal position currently occupied by the discoverer; X worst Indicates the current global worst position; is the best position currently occupied by the discoverer; Q is a random number that follows a normal distribution; Represents element-wise multiplication.

[0057] The data set is divided into a training set and a test set. The model performance is tested using the training set and verified using the test set. The fitting coefficient R is used. 2 , mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are used as indicators to evaluate the performance of the model.

[0058] Beneficial effects:

[0059] The present invention relates to a flotation concentrate grade prediction method based on ISSA-HKLSSVM, which comprises: step S1, building a model for describing the nonlinear relationship between ore element energy intensity and concentrate grade based on data collected by an X-ray fluorescence analyzer, and using the channel data measured by the analyzer to establish a sample database; step S2, introducing a Min-Max data normalization method to pre-process the data and construct a data set; step S3, using a mixed kernel function to optimize the weight coefficient of the kernel function and build a concentrate grade prediction model based on HKLSSVM; step S4, introducing a nonlinear exponential logarithmic mixed learning factor to improve the traditional SSA algorithm, using the ISSA algorithm to optimize the kernel function parameters and penalty factors of the HKLSSVM model, and establishing an ISSA-HKLSSVM flotation concentrate grade prediction model; step S5, using the data set data to train the grade prediction model based on ISSA-HKLSSVM to realize the prediction of the concentrate grade in the flotation process. The method can effectively realize the accurate prediction of the concentrate grade, and provide an effective and reliable criterion for optimizing the control of the flotation process and improving the flotation efficiency.

[0060] The present invention proposes a flotation concentrate grade prediction method which is easy to use and has high accuracy. The flotation concentrate grade prediction method based on ISSA-HKLSSVM solves the problems that the existing method is difficult to achieve fast and accurate measurement of small samples of concentrate grade, the optimization process is prone to fall into local optimal solutions, the training time is too long, and the grade prediction accuracy is low. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is the basic structure diagram of the present invention;

[0062] Figure 2 The principle flow chart of the ISSA-HKLSSVM algorithm constructed by the present invention is as follows;

[0063] Figure 3 It is an intensity-grade model diagram established according to the working principle of X-ray fluorescence grade analyzer;

[0064] Figure 4 A comparison chart of the absolute error results of the test set prediction obtained by training the data with different algorithms. DETAILED DESCRIPTION

[0065] The following will be combined Figure 1-Figure 4 The technical solution of the present invention is described in further detail.

[0066] Embodiment 1

[0067] The present invention provides a method for predicting the grade of flotation concentrate based on ISSA-HKLSSVM, which is aimed at the problems that the ore is poor, fine, impure, and generally low in grade, the parameters in the flotation process are coupled and lagging, and the small sample of concentrate grade is difficult to measure quickly and accurately, and the existing optimization method is prone to fall into the local optimal solution, the training time is too long, and the grade prediction accuracy is low.

[0068] A flotation concentrate grade prediction method based on ISSA-HKLSSVM is as follows: Figure 1 As shown. First, an intensity-grade model is built based on the data measured by the X-ray fluorescence analyzer, and a sample database is established using the channel data measured by the analyzer to construct a concentrate grade prediction data set. The 2σ principle is used to process the data for outliers, and outliers that differ from the average value by more than two standard deviations are eliminated. Then the Z-Score method is used for normalization to obtain a data set between 0 and 1. The HKLSSVM model is further built, in which the local kernel function uses the Gaussian radial basis function, which can map a sample to a higher-dimensional space; the kernel function represented by the polynomial kernel has good global characteristics. Combined with the characteristics of the polynomial kernel function, it is selected as the global kernel function of the prediction model. The advantages of the local kernel function and the global kernel function are combined by constructing a hybrid kernel function. The kernel function parameters and penalty factors of the HKLSSVM model are optimized using the ISSA algorithm, and a grade prediction method based on ISSA-HKLSSVM is constructed. Online prediction of the concentrate grade in the flotation process is achieved.

[0069] The specific implementation steps of the concentrate grade prediction method based on ISSA-HKLSSVM are as follows:

[0070] Step S1, build an intensity-grade model based on X-ray fluorescence analyzer Figure 3 As shown, the content of elements in the pulp will affect the fluorescence intensity of the elements. Therefore, this method predicts the grade of the pulp by measuring the fluorescence intensity of the elements. The sample database is established using the channel data measured by the analyzer;

[0071] Step S2, introduce the Min-Max data normalization method to pre-process the data and construct a data set. The specific details of step S2 are as follows:

[0072] Obtain usage data from the database and divide the data into a test set and a training set, where the ratio of the test set to the training set is 2:8;

[0073] In order to eliminate the dimensional influence between indicators, the data is normalized, and the normalization function model is:

[0074]

[0075] Where: z * is the result after data normalization; z is the data to be normalized; z max is the maximum value of the required normalized data; z min is the minimum value of the required normalized data.

[0076] Step S3, using a hybrid kernel function, optimizing the weight coefficient of the kernel function, and building a concentrate grade prediction model based on HKLSSVM, the specific steps of step S3 are as follows:

[0077] LSSVM is an improvement based on support vector machine. It converts complex quadratic programming problems into linear equations for solution and replaces inequality constraints in SVM optimization problems with equality constraints, which improves the algorithm's running speed and reduces the complexity of calculation. The LSSVM model is:

[0078]

[0079] Among them, K(x i ,x j ) is a kernel function, which is used to replace the inner product operation in high-dimensional space, x i is the center of the kernel function, x is the input value of the training sample, α i represents the Lagrange multiplier and b is the deviation.

[0080] Regarding the selection of local kernel functions and global kernel functions, the HKLSSVM model is as follows:

[0081] The Gaussian radial basis function is a kernel function with strong locality, which can map a sample into a higher-dimensional space. Based on this advantage, it is selected as the local kernel function of the concentrate grade prediction model.

[0082]

[0083] Among them, Σ is the kernel width parameter in the Gaussian kernel function.

[0084] The kernel function represented by the polynomial kernel has good global characteristics. Combining the characteristics of the polynomial kernel function, it is selected as the global kernel function of the prediction model:

[0085] K2(x i ,x j )=(x i T ·x j +1) d ,d∈N *

[0086] (Formula 4)

[0087] Where: d is the polynomial kernel function parameter.

[0088] Using the above classical kernel functions to construct a hybrid kernel function through linear combination, the admissible kernel condition is still satisfied:

[0089] K(x i ,x j )=aK1(x i ,x j )+(1-a)K2(x i ,x j )

[0090] (Formula 5)

[0091] By adjusting the weight coefficient a, the local and global characteristics of the hybrid kernel function can be balanced. The weight value a is selected as 0.5 to take into account both learning ability and generalization ability.

[0092] Step S4, introducing a nonlinear exponential logarithmic hybrid learning factor to improve the traditional SSA algorithm, and using the ISSA algorithm to optimize the kernel function parameters and penalty factors of the HKLSSVM model, the specific details of step S4 are as follows;

[0093] In order to improve the problem that the search space of the discoverer gradually decreases with the number of iterations, which increases the probability of falling into the local space, a nonlinear exponential logarithmic hybrid learning factor is introduced in the discoverer position update method. A larger value in the early stage of the search helps to quickly conduct global exploration and improve search efficiency; a smaller value in the later stage of the search can further improve local development capabilities and improve accuracy.

[0094] The calculation formula of the nonlinear exponential logarithmic hybrid learning factor and the improved discoverer position update method are as follows:

[0095]

[0096] Where, X ij represents the position information of the i-th sparrow in the j-th dimension; t represents the current iteration number; w represents the hybrid learning factor calculated by formula 8; r1 is a random number in [0,2π], r2 is a random number in [0,2], R2∈(0,1] represents the warning value, ST∈(0.5,1] represents the safety value, X best Represents the current global optimal position, and its value will be updated according to the optimal solution found by the sparrow algorithm.

[0097] The Levy flight strategy and the improved follower position update method are as follows:

[0098] When the discoverer iterates a certain number of times and the fitness value remains unchanged, the follower becomes the discoverer and the algorithm will fall into a local optimum. To avoid this, the levy flight strategy is introduced into the follower update formula to improve the global search capability. The levy flight strategy is described as follows:

[0099]

[0100] Where, X P Indicates the optimal position currently occupied by the discoverer; X worst Indicates the current global worst position; A represents a matrix, each element is randomly assigned a value of 1 or -1. is the best position currently occupied by the discoverer. Q is a random number that follows a normal distribution.

[0101] Step S5, in order to verify the accuracy and effectiveness of the method proposed in the present invention for predicting the concentrate grade in the flotation process, verification is performed based on the measured data of the X-ray fluorescence analyzer in the workshop of the Guosen Mining Concentrator.

[0102] Different algorithms are used to predict the flow channel data, and the comparison results of the indicators are shown in Table 1. Figure 4 This is a comparison chart of the prediction absolute error results.

[0103] Table 2 Comparison of grade prediction indicators by different methods

[0104]

[0105] Among them, the fitting coefficient R2, mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are used as indicators to evaluate the performance of the model.

[0106] It can be seen from Table 1 that compared with the OLS and PLS methods, the model established by BP neural network, LSSVM and its optimization method has better performance. The reason is that the model of concentrate grade is multivariable and has strong nonlinearity and uncertainty. Compared with BP neural network, LSSVM method is more suitable for small sample prediction of concentrate grade and has stronger generalization ability. Compared with the unimproved LSSVM method, the regression coefficient of LSSVM algorithm optimized by improved sparrow search algorithm and new hybrid kernel function reached 0.9069, and the mean square error, root mean square error and mean absolute error were 1.1946, 1.0929 and 0.8144 respectively, indicating that the previous sparrow search algorithm can find suitable grade prediction model parameters, which is helpful to improve the adaptability of the model. It can be seen from the experimental results of single kernel and mixed kernel that the mixed kernel method is better than the single kernel method. Comprehensive experimental statistical results show that ISSA-HKLSSVM can efficiently predict the concentrate grade and meet the needs of grade prediction in actual flotation site.

[0107] In summary, the present invention relates to a flotation concentrate grade prediction method based on ISSA-HKLSSVM, which constructs a data set by acquiring data of different element channels using a strength-grade model; the traditional SSA algorithm is improved by introducing a Levy flight strategy and a nonlinear exponential logarithmic hybrid learning factor to obtain an ISSA algorithm; the key parameters in the HKLSSVM model are optimized using the ISSA algorithm, and an ISSA-HKLSSVM flotation concentrate grade prediction model is constructed to further improve the prediction accuracy; the method described in the present invention can effectively realize the online prediction of the concentrate grade in the flotation process, realize the optimization control of the flotation process, and improve the flotation efficiency.

[0108] Although the present invention has been disclosed as above with preferred embodiments, the embodiments do not limit the present invention. Any equivalent changes or modifications made without departing from the spirit and scope of the present invention also belong to the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the content defined by the claims of this application.

Claims

1. A flotation concentrate grade prediction method based on ISSA-HKLSSVM, characterized in that: The following steps are involved: Step 1, using the ore element energy intensity data and concentrate grade data of each energy channel in the X-ray fluorescence analyzer to build a model sample database; The model is a model used to describe the nonlinear relationship between energy intensity and grade; Step 2, introduce the Min-Max data normalization method to preprocess the data and construct a data set; Step 3, using a hybrid kernel function and optimizing the weight coefficient of the kernel function to build a concentrate grade prediction model based on HKLSSVM; Step 4, introducing a nonlinear exponential logarithmic hybrid learning factor to improve the SSA algorithm, using the ISSA algorithm to optimize the kernel function parameters and penalty factors of the HKLSSVM model, and establishing an ISSA-HKLSSVM flotation concentrate grade prediction model; Step 5: Use the data in the data set to train the ISSA-HKLSSVM flotation concentrate grade prediction model to achieve the flotation process concentrate grade prediction.

2. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: In step 1, the ore element energy intensity data and concentrate grade data of each energy channel are: iron element energy channel, zinc element energy channel, lead element energy channel, scandium element energy channel data and zinc concentrate grade data. Some measured data are shown in Table 1 below. Table 1 Partial measured data of flotation concentrate grade prediction model 3. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: The pretreatment in step 2 specifically includes: for the five variables of flotation concentrate grade, outlier processing is performed based on the 2σ principle. The basic process is as follows: The 2σ rule is used to screen the data and eliminate outliers whose difference from the mean value is greater than two times the standard deviation; data is the actual value, is the average value, N is the number of values ​​contained in a channel; σ is the standard deviation; The data set is normalized and the model is: Where: z * is the result after data normalization; z is the data to be normalized; z max is the maximum value of the required normalized data; z min is the minimum value of the required normalized data.

4. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: In step 3, the least squares support vector machine (LSSVM) model is: Among them, f(x i ) is the optimal decision function of the LSSVM model; x i =x i,1 ,x i,2 ,...,x i,D represents sample input data, D is the dimension of input data, N is the number of samples; K(x,x i ) kernel function, where x is the center of the kernel function; α i represents the Lagrange multiplier, which introduces constraints and transforms the original optimization problem into the problem of finding the extreme value of the Lagrangian function; b is the regression deviation, which is used to characterize the gap between the predicted result and the actual result. Its specific value is determined by the training optimization process of the model.

5. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 4 is characterized in that: Based on the LSSVM model, the ordinary kernel function is replaced by a hybrid kernel function as a constraint to obtain a hybrid kernel least squares support vector machine model, namely the HKLSSVM model. Support Vector Machine, HKLSSVM), the HKLSSVM model is as follows: Kernel functions are divided into global kernel functions and local kernel functions, which mainly realize the mapping of samples from low-dimensional to high-dimensional feature space. The global kernel function has good generalization performance for long-distance samples, but poor learning performance for close-distance samples; the local kernel function has strong learning performance but weak generalization performance. By constructing a hybrid kernel function, the advantages of both can be effectively combined. The Gaussian radial basis function is a kernel function with strong locality, which can map a sample into a higher-dimensional space. Based on this advantage, it is selected as the local kernel function of the concentrate grade prediction model. The radial basis function is selected as the local kernel function of the concentrate grade prediction model: Among them, Σ is the kernel width parameter in the Gaussian kernel function, and its exact value will be determined as a model parameter by the subsequent optimization algorithm. The kernel function represented by the polynomial kernel function has good global characteristics, combined with the characteristics of the polynomial kernel function. The polynomial kernel function is selected as the global kernel function of the prediction model: K2(x i ,x j )=(x i T ·x j +1) d ,d∈N * (Formula 6) Where: d is the polynomial kernel function parameter, and its exact value will also be determined by the subsequent optimization algorithm, d∈N * It means that d belongs to the set of natural numbers; Using the above classical kernel functions to construct a hybrid kernel function through linear combination, the admissible kernel condition is still satisfied: K(x i ,x j )=aK1(x i ,x j )+(1-a)K2(x i ,x j ) (Formula 7) By adjusting the weight coefficient a, the local and global characteristics of the hybrid kernel function can be balanced. Selecting the weight value a as 0.5 can take into account both learning ability and generalization ability.

6. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: The Improved Sparrow Search Algorithm (ISSA) is introduced to optimize the parameters to be tested of the HKLSSVM model to solve the problem that the traditional Sparrow Search Algorithm cannot jump out of the local extreme value in the later stage of the search. The improvement of the ISSA algorithm in the finder position update method is specifically manifested as follows: In order to improve the problem in the traditional sparrow algorithm that, with the number of iterations, the finder's search space gradually decreases and the probability of falling into the local space increases, a nonlinear exponential logarithmic hybrid learning factor is introduced in the finder position update method. This factor is required to have a larger value in the early stage of the search, which helps to quickly conduct global exploration and improve search efficiency; in the later stage of the search, it has a smaller value, which can further enhance the local exploration ability and improve accuracy. The nonlinear exponential logarithmic mixed learning factor is introduced, and its calculation formula is: In the formula, w min With w max They represent the maximum and minimum values ​​of the hybrid learning factors obtained during the search process; iter max The maximum number of iterations is a constant, which indicates the maximum number of iterations. Generally, 100 iterations are selected as the benchmark. The improved method of updating the finder's location is as follows: In the formula, X ij Represents the position information of the i-th sparrow in the j-th dimension; t represents the current iteration number; w represents the hybrid learning factor calculated by Formula 8; r1 is a random number in [0,2π], r2 is a random number in [0,2], R2∈(0,1] represents the warning value, ST∈(0.5,1] represents the safety value, X best Represents the current global optimal position, and its value will be updated according to the optimal solution found by the sparrow algorithm.

7. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: The Improved Sparrow Search Algorithm (ISSA) is introduced to optimize the parameters to be tested of the HKLSSVM model to solve the problem that the traditional Sparrow Search Algorithm cannot jump out of the local extreme value in the later stage of the search. The improvement of the ISSA algorithm in the follower position update method is specifically manifested as follows: When the discoverer iterates a certain number of times and the fitness value remains unchanged, the follower becomes the discoverer and the algorithm will fall into a local optimum. To avoid this, the levy flight strategy is introduced into the follower update formula to improve the global search capability. The levy flight strategy is described as follows: Where: r3 and r4 are random numbers in [0,1], the value of ζ is 1.5, and (i)! refers to the gamma function. The follower position update formula after the introduction of the improved Levy flight strategy is as follows: In the formula, X P Indicates the optimal position currently occupied by the discoverer; X worst Indicates the current global worst position; is the best position currently occupied by the discoverer; Q is a random number that follows a normal distribution; Represents element-wise multiplication.

8. The flotation concentrate grade prediction method based on ISSA-HKLSSVM according to claim 1 is characterized in that: The data set is divided into a training set and a test set. The model performance is tested using the training set and verified using the test set. The fitting coefficient R is used. 2 , mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE) are used as indicators to evaluate the performance of the model.

Images