Data stream active sampling and predicting method for product ash content index in coal dense medium separation process

Abstract

The invention provides a data stream active sampling and predicting method for a product ash content index in a coal dense medium separation process. The method comprises the following steps: S1, preprocessing coal dense medium separation data; s2, establishing a dynamic cache window in the sorted data stream to perform synchronous clustering; s3, establishing a feature memory dictionary, and constructing a centralized kernel space by using data in the feature memory dictionary to obtain a dictionary data feature micro-mode direction and a feature main mode direction; s4, calculating the projection of the real-time data point in the feature micro-mode direction, and judging whether active sampling is carried out or not; s5, obtaining the projection of the dictionary data in the main mode direction by using the kernel principal component information in the S4, constructing an evaluation function, and judging whether to sort active sampling or not; s6, taking the maximum sample set of the value function as a new feature memory dictionary; and S7, strengthening the weight of the actively sampled sample, recursively constructing a random weight neural network through a Woodbury formula, and predicting the clean coal ash content in the label data stream until the data stream is finished.

Description

Technical Field

[0001] This invention belongs to the field of soft measurement technology, specifically relating to a data stream active sampling and prediction method for ash content indicators of products in the coal heavy media separation process. Background Technology

[0002] Heavy media coal preparation is a typical chemical production process and one of the most direct and effective coal preparation methods currently available. Its core equipment mainly includes a mixing tank, heavy media hydrocyclone, desliming screen, magnetic separator, magnetite bin, qualified media bin, and related actuators and instruments. The process flow is as follows: First, the deslimed and dewatered raw coal is conveyed by belt to the mixing tank, where it is thoroughly mixed with heavy media to form a slurry, which is then pumped into the heavy media hydrocyclone. Inside the hydrocyclone, the slurry is separated by gravity and centrifugal force. Components with a density lower than the medium are suspended at the top, forming an overflow, while components with a density higher than the medium settle to the bottom, forming an underflow. The overflow and underflow slurries then enter the desliming screen, producing clean coal and tailings waste, respectively. The medium discharged from the desliming screen enters the magnetic separator for purification and recovery. The recovered qualified medium flows into the qualified media bin, where it is mixed with fresh magnetite powder from the magnetite bin and an appropriate amount of make-up water. The density is maintained within a set range by a density adjustment system, ensuring that the recovered medium can be directly recycled.

[0003] Heavy media separation of coal aims to reduce the ash content of refined coal products. The separation process is adjusted in real-time based on the ash content index to improve product quality. However, due to limitations in equipment and operational factors, data collection for refined coal ash content is time-consuming and costly in some scenarios. Furthermore, this process involves complex physicochemical changes, making it difficult to directly establish a predictive model for the ash content mechanism. Therefore, only partial data on refined coal ash content and related parameters can be used to construct a soft sensor model for real-time prediction of refined coal ash content. During measurement, the computer cannot provide real-time results due to on-site computing power constraints; long-period sampling cannot handle sudden concept drift; and uniform sampling strategies struggle to achieve differentiated sampling in regions with varying data flow dynamics. This prevents the soft sensor model from being trained with optimal data, increasing prediction errors. Summary of the Invention

[0004] The purpose of this invention is to overcome the aforementioned problems by providing a method for active sampling and prediction of ash content indicators in coal heavy media separation processes. This method verifies the representativeness and diversity of real-time data points in the separation data stream by synchronizing clustering results within a dynamic cache window. A feature memory dictionary is designed, and a parallel strategy is used to verify the informativeness of real-time data points in the separation data stream in both the micro-pattern and main pattern directions to determine whether active sampling is necessary. A strategy for updating and maintaining the feature memory dictionary is also developed. Furthermore, the weights of sudden concept drift regions are strengthened, and a random weighted neural network is recursively and efficiently modeled using the Woodbury formula to complete the active modeling and prediction of the data stream.

[0005] Technical Solution: To solve the above-mentioned technical problems and achieve the above-mentioned objectives, this invention proposes a data stream active sampling and prediction method for ash content indicators of products in coal heavy media separation processes. This method includes the following steps:

[0006] S1, preprocessing coal heavy media separation data;

[0007] S2, establish a dynamic cache window in the sorted data stream for synchronous clustering, and perform representativeness and diversity tests on the real-time data points of the sorted data stream based on the clustering results to determine whether to proceed to step S3.

[0008] S3, establish a feature memory dictionary and use the data in the feature memory dictionary to construct a centralized kernel space, perform kernel principal component analysis, and obtain the feature micro-pattern direction and feature principal pattern direction of the dictionary data;

[0009] S4, calculate the projection of real-time data points of the sorted data stream onto the direction of the feature micro-pattern, and then calculate the informational quantitative index to determine whether to actively sample.

[0010] S5. If the real-time data points of the sorting data stream are not actively sampled in S4, the kernel principal component information of S3 is used to obtain the projection of the dictionary data in the main mode direction and construct the evaluation function. The information is calculated by mapping the real-time data points of the sorting data stream onto the evaluation function, and then it is determined whether to actively sample.

[0011] S6. If the real-time data points of the sorting data stream are actively sampled in S4 or S5, the dictionary in S3 is dynamically maintained by using the sample set that maximizes the value function as a new feature memory dictionary.

[0012] S7 strengthens the sample weights of the data actively sampled in S4 or S5, and recursively constructs a random weighted neural network using the Woodbury formula to predict the ash content of clean coal in the labeled data stream until the data stream ends, thus completing the active sampling and prediction of the coal heavy media process data stream.

[0013] Furthermore, step S1 specifically includes the following steps:

[0014] S11, using eight data points—clean coal ash content, raw coal ash content, medium density, medium magnetic content, diversion box opening, medium tank level, feed pressure, and clean coal desliming screen water supply valve opening—to obtain the initial dataset L. old ;

[0015] S12, used For the initial dataset L old Normalization is performed, where μ and σ represent the mean and standard deviation of the data, respectively, and L new This represents the data after normalization.

[0016] S13: Extract the first 10% of data from the normalized dataset in chronological order as the starting data, including the simulated label data stream of clean coal ash content and the simulated sorting data stream of the other seven types of data.

[0017] Furthermore, step S2 specifically includes the following steps:

[0018] S21, establish a dynamic buffer window pw with a capacity of L in the sorting data stream, and store the data of the current time in the sorting data stream. Add the data to the dynamic cache window pw and remove the data that has been in pw for the longest time.

[0019] S22, Perform synchronous clustering in the dynamic cache window, treat the data points in the window as coupled oscillators, and use the phase attraction in its coupled synchronization dynamics to make similar data points spontaneously synchronize to the same phase to achieve clustering;

[0020] S23, n is obtained according to step S22. c Clustering results Sorting data stream data at the current moment 1≤k≤n c ,test Representativeness: C k If the number of data points is greater than m, where m is a set threshold, if the representativeness test passes, proceed to S24; otherwise, proceed to S21 when the next data arrives in the sorted data stream.

[0021] S24, Test the diversity of the data, based on the class distribution within pw, to ensure If the data points are of a different category than those already sampled in pw, and the diversity test passes, proceed to S3; otherwise, proceed to S21 when the next data arrives in the sorted data stream.

[0022] Furthermore, step S3 specifically includes:

[0023] S31, the feature memory dictionary D contains n sampled data points, which do not include the clean coal ash content label in S13. Kernel processing is performed on all data points in D to obtain the kernel matrix K, as shown in the following formula:

[0024] k(d i ,d j )=exp(-γ||d i -d j || 2 )

[0025] Where, k(d) i ,d j ) is the element in the i-th row and j-th column of the kernel matrix K, γ is the parameter controlling the bandwidth, and d i and d j These are the i-th and j-th data points in D;

[0026] The kernel matrix K is centered to obtain the centered kernel matrix K. c The formula is as follows:

[0027]

[0028] Among them, 1 n*n It is a matrix with all elements being 1;

[0029] S32, using kernel principal component analysis to perform eigenvalue decomposition on the centered kernel matrix, obtains n eigenvalues ​​{λ1,λ2,...,λ...} arranged in ascending order. n} and its corresponding unit eigenvector U = [u1, u2, ..., u n The eigenvalues ​​and unit eigenvectors are divided into micro-mode directions {λ1,λ2,...,λ...}. k}with U k =[u1,u2,...,u k ] and main mode direction {λ k+1 ,λ k+2 ,...,λ n}with U n-k =[u k+1 ,u k+2 ,...,u n Two categories, the formula is as follows:

[0030]

[0031] Where η is the set boundary threshold parameter.

[0032] Furthermore, step S4 specifically includes the following steps:

[0033] S41, for the data in the sorting data stream at the current moment Calculate its projection k in the high-dimensional kernel space x =[k x1 ,k x2 ,…,k xi The formula is as follows:

[0034]

[0035] Where, k xi It is k x The i-th element, d i It is the i-th data in the feature memory dictionary D;

[0036] k x Substituting this into the centralized kernel space yields the projection into the centralized kernel space, as shown in the following formula:

[0037]

[0038] in, It is k x The centered vector, 1 n*n 1 n*1 1 1*n It is a matrix consisting entirely of 1s;

[0039] S42, Calculation The projection in the micro-mode direction is given by the following formula:

[0040]

[0041] Where, k x(i≤k) yes Projection in the micro-mode direction, U k The S32 micro-mode directional eigenvector matrix is ​​used to calculate the micro-mode directional information quantification index, as follows:

[0042]

[0043] Where α1 is a quantitative indicator measuring the directional informationality of micro-modes, and k i It is k x(i≤k) The value in the i-th dimension;

[0044] S43, Data sorting stream at the current moment The informational quantitative index α1 satisfies:

[0045] α1≥δ

[0046] Here, δ is a set constant. If the condition α1≥δ is not met, proceed to S5; if the condition is met, actively sample the current data from the tag data stream and sort the data stream. tags Switch to S6.

[0047] Furthermore, step S5 specifically includes the following steps:

[0048] S51, based on the eigenvector matrix U obtained in S32 n-k The projection coordinates of the data in the feature memory dictionary D onto the main pattern direction are calculated using the following formula:

[0049] x′ i =K ci U n-k i = 1, 2, 3, ..., n

[0050] Among them, K ci It is the centered kernel matrix K c The i-th row, x′ i These are the projected coordinates of the data in dictionary D in S51 in the main mode direction;

[0051] S52, Establish the evaluation function along the main feature pattern direction, as follows:

[0052]

[0053] Where M is the bandwidth matrix and G is the kernel function;

[0054] S53, according to calculation The projected coordinates in the main pattern direction are used to quantify their similarity to the feature memory dictionary data, as shown in the following formula:

[0055]

[0056] α2=f(k x(k＜i≤n) )

[0057] Where, k x(k＜i≤n) yes Projection in the main mode direction, U n-k α1 is the feature vector matrix in the main mode direction, and α2 is a quantitative indicator that measures the information content in the main mode direction.

[0058] S54, sorting data stream at the current moment. The informational quantitative indicators in S53 satisfy the following:

[0059] α2≤υf(x′ i ) min

[0060] Where υ is a set constant, f(x′) i ) min This represents the minimum value when i = 1, 2, ..., n; if α2 ≤ υf(x′) is not satisfied. i )min If the sorting fails, proceed to step S21 when the next data arrives in the sorting data stream; if the conditions are met, actively sample the data from the label data stream at the current moment in the sorting data stream. tags Switch to S6.

[0061] Furthermore, step S6 specifically includes:

[0062] S61 will be the data points actively sampled in S4 or S5. Add the feature memory dictionary D, and remove data points in D whose retention time exceeds a threshold to obtain D. new The candidate dictionary that maximizes the value function is used as the final new feature memory dictionary. The evaluation function formula is as follows:

[0063]

[0064] in, and T old It is the retention time of the old data being replaced in the dictionary.

[0065] Furthermore, step S7 specifically includes:

[0066] S71, Actively Sampled Data Points Increase the weight of actively sampled sample points and establish a memory enhancement coefficient:

[0067]

[0068] Where κ is the reinforcement value and ξ is the relaxation factor;

[0069] S72, given N previous actively sampled data... The neural network has L hidden layer nodes. The weights and biases of the output layer are randomly generated within a certain interval. The matrix form of its hidden layer output is as follows:

[0070] Y = Hβ

[0071] Where H is the hidden layer output and β is the output layer weight;

[0072] S73, for actively sampled data points The model parameters are updated by using a recursive method and the Woodbury formula to find the inverse, as detailed below:

[0073] Set W N For the memory reinforcement matrix, W0 = I is the identity matrix, H N These are the outputs of the first N sampling points, H0 is the output of the hidden layer of the initial dataset, Y0 is the clean coal ash content of the initial dataset, and w N+1 yes The corresponding memory enhancement coefficient, H N+1 ∈R 1*N yes In the output of the hidden layer, λ is the label of the actively sampled data, and λ is the regularization coefficient.

[0074] remember:

[0075]

[0076] Using Woodbury's formula, we get:

[0077]

[0078] The output layer weights are as follows:

[0079] β N+1 =(IK N+1 H N+1 )β N +K N+1 y N+1

[0080] in,

[0081] Let H DS It is the output of the sorted data stream in the hidden layer of the random weighted neural network. It is The predicted clean coal ash content is updated after the output layer weights are updated; it is then transferred back to S2 for continuous processing until the data stream ends, thus completing the active sampling and prediction of the data stream.

[0082] Beneficial effects: Compared with the prior art, the technical solution of the present invention has the following beneficial technical effects:

[0083] This invention considers the characteristics of real-time data streams in industrial processes, the kernel space features of the data, and computing power. It proposes an active sampling and prediction method for ash content indicators in coal heavy media separation processes. A dynamic buffer window performs synchronous clustering, utilizing the relative density of the real-time data stream to assess its representativeness and diversity. A feature memory dictionary is designed using kernel space, and features are decomposed into micro-patterns and principal patterns through kernel principal component analysis. The feature differences between real-time data points and sampled data points are analyzed using two parallel methods: difference accumulation and evaluation functions, enabling informative evaluation of real-time data. A dynamic maintenance method for the feature memory dictionary is designed to prevent infinite expansion and increased computational complexity. A recursive modeling of a stochastic weighted neural network improves model learning efficiency and the accuracy of real-time prediction of clean coal ash content. Attached Figure Description

[0084] To more clearly illustrate the technical solutions of the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0085] Figure 1 This is a flowchart of the method of the present invention.

[0086] Figure 2 A diagram illustrating the MAE process for constructing a predictive model for the ash content of refined coal.

[0087] Figure 3 A diagram illustrating the RMSE (Real-Time Sequence) process in constructing a model for predicting the ash content of refined coal.

[0088] Figure 4 R for the construction of a coal ash content prediction model 2 Display image. Detailed Implementation

[0089] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0090] like Figure 1 As shown, this invention proposes a data stream active sampling and prediction method for ash content indicators of products in coal heavy media separation processes. This method includes the following steps:

[0091] S1, preprocessing coal heavy media separation data;

[0092] Furthermore, step S1 specifically includes the following steps:

[0093] S11, using eight data points—clean coal ash content, raw coal ash content, medium density, medium magnetic content, diversion box opening, medium tank level, feed pressure, and clean coal desliming screen water supply valve opening—to obtain the initial dataset L. old ;

[0094] S12, used For the initial dataset L old Normalization is performed, where μ and σ represent the mean and standard deviation of the data, respectively, and L new This represents the data after normalization.

[0095] S13: Extract the first 10% of data from the normalized dataset in chronological order as the starting data, including the simulated label data stream of clean coal ash content and the simulated sorting data stream of the other seven types of data.

[0096] S2, establish a dynamic cache window in the sorted data stream for synchronous clustering, and perform representativeness and diversity tests on the real-time data points of the sorted data stream based on the clustering results to determine whether to proceed to step S3.

[0097] Furthermore, step S2 specifically includes the following steps:

[0098] S21, establish a dynamic buffer window pw with a capacity of L in the sorting data stream, and store the data of the current time in the sorting data stream. Add the data to the dynamic cache window pw and remove the data that has been in pw for the longest time.

[0099] S22, Perform synchronous clustering in the dynamic cache window, treat the data points in the window as coupled oscillators, and use the phase attraction in its coupled synchronization dynamics to make similar data points spontaneously synchronize to the same phase to achieve clustering;

[0100] S23, n is obtained according to step S22. c Clustering results Sorting data stream data at the current moment 1≤k≤n c ,test Representativeness: C k If the number of data points is greater than m, where m is a set threshold, if the representativeness test passes, proceed to S24; otherwise, proceed to S21 when the next data arrives in the sorted data stream.

[0101] S24, Test the diversity of the data, based on the class distribution within pw, to ensure If the data points are of a different category than those already sampled in pw, and the diversity test passes, proceed to S3; otherwise, proceed to S21 when the next data arrives in the sorted data stream.

[0102] S3, establish a feature memory dictionary and use the data in the feature memory dictionary to construct a centralized kernel space, perform kernel principal component analysis, and obtain the feature micro-pattern direction and feature principal pattern direction of the dictionary data;

[0103] Furthermore, step S3 specifically includes:

[0104] S31, the feature memory dictionary D contains n sampled data points, which do not include the clean coal ash content label in S13. Kernel processing is performed on all data points in D to obtain the kernel matrix K, as shown in the following formula:

[0105] k(di ,d j )=exp(-γ||d i -d j || 2 )

[0106] Where, k(d) i ,d j ) is the element in the i-th row and j-th column of the kernel matrix K, γ is the parameter controlling the bandwidth, and d i and d j These are the i-th and j-th data points in D;

[0107] The kernel matrix K is centered to obtain the centered kernel matrix K. c The formula is as follows:

[0108]

[0109] Among them, 1 n*n It is a matrix with all elements being 1;

[0110] S32, using kernel principal component analysis to perform eigenvalue decomposition on the centered kernel matrix, obtains n eigenvalues ​​{λ1,λ2,...,λ...} arranged in ascending order. n} and its corresponding unit eigenvector U = [u1, u2, ..., u n The eigenvalues ​​and unit eigenvectors are divided into micro-mode directions {λ1,λ2,...,λ...}. k}with U k =[u1,u2,...,u k ] and main mode direction {λ k+1 ,λ k+2 ,...,λ n}with U n-k =[u k+1 ,u k+2 ,...,u n Two categories, the formula is as follows:

[0111]

[0112] Where η is the set boundary threshold parameter.

[0113] S4, calculate the projection of real-time data points of the sorted data stream onto the direction of the feature micro-pattern, and then calculate the informational quantitative index to determine whether to actively sample.

[0114] Furthermore, step S4 specifically includes the following steps:

[0115] S41, for the data in the sorting data stream at the current moment Calculate its projection k in the high-dimensional kernel space x =[kx1 ,k x2 ,…,k xi The formula is as follows:

[0116]

[0117] Where, k xi It is k x The i-th element, d i It is the i-th data in the feature memory dictionary D;

[0118] k x Substituting this into the centralized kernel space yields the projection into the centralized kernel space, as shown in the following formula:

[0119]

[0120] in, It is k x The centered vector, 1 n*n 1 n*1 1 1*n It is a matrix consisting entirely of 1s;

[0121] S42, Calculation The projection in the micro-mode direction is given by the following formula:

[0122]

[0123] Where, k x(i≤k) yes Projection in the micro-mode direction, U k The S32 micro-mode directional eigenvector matrix is ​​used to calculate the micro-mode directional information quantification index, as follows:

[0124]

[0125] Where α1 is a quantitative indicator measuring the directional informationality of micro-modes, and k i It is k x(i≤k) The value in the i-th dimension;

[0126] S43, Data sorting stream at the current moment The informational quantitative index α1 satisfies:

[0127] α1≥δ

[0128] Here, δ is a set constant. If the condition α1≥δ is not met, proceed to S5; if the condition is met, actively sample the current data from the tag data stream and sort the data stream. tags Switch to S6.

[0129] S5. If the real-time data points of the sorting data stream are not actively sampled in S4, the kernel principal component information of S3 is used to obtain the projection of the dictionary data in the main mode direction and construct the evaluation function. The information is calculated by mapping the real-time data points of the sorting data stream onto the evaluation function, and then it is determined whether to actively sample.

[0130] Furthermore, step S5 specifically includes the following steps:

[0131] S51, based on the eigenvector matrix U obtained in S32 n-k The projection coordinates of the data in the feature memory dictionary D onto the main pattern direction are calculated using the following formula:

[0132] x′ i =K ci U n-k i = 1, 2, 3, ..., n

[0133] Among them, K ci It is the centered kernel matrix K c The i-th row, x′ i These are the projected coordinates of the data in dictionary D in S51 in the main mode direction;

[0134] S52, Establish the evaluation function along the main feature pattern direction, as follows:

[0135]

[0136] Where M is the bandwidth matrix and G is the kernel function;

[0137] S53, according to calculation The projected coordinates in the main pattern direction are used to quantify their similarity to the feature memory dictionary data, as shown in the following formula:

[0138]

[0139] α2=f(k x(k＜i≤n) )

[0140] Where, k x(k＜i≤n) yes Projection in the main mode direction, U n-k α1 is the feature vector matrix in the main mode direction, and α2 is a quantitative indicator that measures the information content in the main mode direction.

[0141] S54, sorting data stream at the current moment. The informational quantitative indicators in S53 satisfy the following:

[0142] α2≤υf(x′ i ) min

[0143] Where υ is a set constant, f(x′) i ) min This represents the minimum value when i = 1, 2, ..., n; if α2 ≤ υf(x′) is not satisfied. i ) min If the sorting fails, proceed to step S21 when the next data arrives in the sorting data stream; if the conditions are met, actively sample the data from the label data stream at the current moment in the sorting data stream. tags Switch to S6.

[0144] S6. If the real-time data points of the sorting data stream are actively sampled in S4 or S5, the dictionary in S3 is dynamically maintained by using the sample set that maximizes the value function as a new feature memory dictionary.

[0145] Furthermore, step S6 specifically includes:

[0146] S61 will be the data points actively sampled in S4 or S5. Add the feature memory dictionary D, and remove data points in D whose retention time exceeds a threshold to obtain D. new The candidate dictionary that maximizes the value function is used as the final new feature memory dictionary. The evaluation function formula is as follows:

[0147]

[0148] in, and T old It is the retention time of the old data being replaced in the dictionary.

[0149] S7 strengthens the sample weights of the data actively sampled in S4 or S5, and recursively constructs a random weighted neural network using the Woodbury formula to predict the ash content of clean coal in the labeled data stream until the data stream ends, thus completing the active sampling and prediction of the coal heavy media process data stream.

[0150] Furthermore, step S7 specifically includes:

[0151] S71, Actively Sampled Data Points Increase the weight of actively sampled sample points and establish a memory enhancement coefficient:

[0152]

[0153] Where κ is the reinforcement value and ξ is the relaxation factor;

[0154] S72, given N previous actively sampled data... The neural network has L hidden layer nodes. The weights and biases of the output layer are randomly generated within a certain interval. The matrix form of its hidden layer output is as follows:

[0155] Y = Hβ

[0156] Where H is the hidden layer output and β is the output layer weight;

[0157] S73, for actively sampled data points The model parameters are updated by using a recursive method and the Woodbury formula to find the inverse, as detailed below:

[0158] Set W N For the memory reinforcement matrix, W0 = I is the identity matrix, H N These are the outputs of the first N sampling points, H0 is the output of the hidden layer of the initial dataset, Y0 is the clean coal ash content of the initial dataset, and w N+1 yes The corresponding memory enhancement coefficient, H N+1 ∈R 1*N yes In the output of the hidden layer, λ is the label of the actively sampled data, and λ is the regularization coefficient.

[0159] remember:

[0160]

[0161] Using Woodbury's formula, we get:

[0162]

[0163] The output layer weights are as follows:

[0164] β N+1 =(IK N+1 H N+1 )β N +K N+1 y N+1

[0165] in,

[0166] Let H DS It is the output of the sorted data stream in the hidden layer of the random weighted neural network. It is The predicted clean coal ash content is updated after the output layer weights are updated; it is then transferred back to S2 for continuous processing until the data stream ends, thus completing the active sampling and prediction of the data stream.

[0167] like Figure 2 , Figure 3 and Figure 4As shown in the figure, the performance comparison results of the Active Modeling (AM) method and the Long Period Sampling (PS) soft measurement modeling method in the prediction of clean coal ash content are presented. Analysis of the results shows that, compared with the long period sampling method, the Active Modeling method of this invention exhibits a significant advantage in the accuracy of clean coal ash content prediction: its ash content prediction error not only decreases faster but also has a lower final convergence error level; simultaneously, the ash content prediction value obtained by the Active Modeling method has a higher coefficient of determination (R²) compared with the actual value. 2 Specifically, regarding the quantitative evaluation of soft measurement results for clean coal ash content, the active modeling method focuses on key performance indicators—root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²)—for predicting clean coal ash content. 2 In terms of ), it is significantly better than long-period sampling methods. This is mainly because long-period sampling mechanisms are easily affected by aliasing effects, making it impossible for them to effectively capture the rapid dynamic changes of key process variables affecting the ash content of clean coal.

Claims

1. A method for active data stream sampling and prediction of ash content indicators in coal heavy media separation processes, characterized in that, The method includes the following steps: S1, preprocessing coal heavy media separation data; S2, establish a dynamic cache window in the sorted data stream for synchronous clustering, and perform representativeness and diversity tests on the real-time data points of the sorted data stream based on the clustering results to determine whether to proceed to step S3. S3, establish a feature memory dictionary and use the data in the feature memory dictionary to construct a centralized kernel space, perform kernel principal component analysis, and obtain the feature micro-pattern direction and feature principal pattern direction of the dictionary data; S4, calculate the projection of real-time data points of the sorted data stream onto the direction of the feature micro-pattern, and then calculate the informational quantitative index to determine whether to actively sample. S5. If the real-time data points of the sorting data stream are not actively sampled in S4, the kernel principal component information of S3 is used to obtain the projection of the dictionary data in the main mode direction and construct the evaluation function. The information is calculated by mapping the real-time data points of the sorting data stream onto the evaluation function, and then it is determined whether to actively sample. S6. If the real-time data points of the sorting data stream are actively sampled in S4 or S5, the dictionary in S3 is dynamically maintained by using the sample set that maximizes the value function as a new feature memory dictionary. S7 strengthens the sample weights of the data actively sampled in S4 or S5, and recursively constructs a random weighted neural network using the Woodbury formula to predict the ash content of clean coal in the labeled data stream until the data stream ends, thus completing the active sampling and prediction of the coal heavy media process data stream.

2. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11, using eight data points—clean coal ash content, raw coal ash content, medium density, medium magnetic content, diversion box opening, medium tank level, feed pressure, and clean coal desliming screen water supply valve opening—to obtain the initial dataset L. old ; S12, used For the initial dataset L old Normalization is performed, where μ and σ represent the mean and standard deviation of the data, respectively, and L new This represents the data after normalization. S13: Extract the first 10% of data from the normalized dataset in chronological order as the starting data, including the simulated label data stream of clean coal ash content and the simulated sorting data stream of the other seven types of data.

3. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 1, characterized in that, Step S2 specifically includes the following steps: S21, establish a dynamic buffer window pw with a capacity of L in the sorting data stream, and store the data of the current time in the sorting data stream. Add the data to the dynamic cache window pw and remove the data that has been in pw for the longest time. S22, Perform synchronous clustering in the dynamic cache window, treat the data points in the window as coupled oscillators, and use the phase attraction in its coupled synchronization dynamics to make similar data points spontaneously synchronize to the same phase to achieve clustering; S23, n is obtained according to step S22. c Clustering results The data stream at the current moment is being sorted. 1≤k≤n c ,test Representativeness: C k If the number of data points is greater than m, where m is a set threshold, if the representativeness test passes, proceed to S24; otherwise, proceed to S21 when the next data arrives in the sorted data stream. S24, Test the diversity of the data, based on the class distribution within pw, to ensure If the data points are of a different category than those already sampled in pw, and the diversity test passes, proceed to S3; otherwise, proceed to S21 when the next data arrives in the sorted data stream.

4. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 3, characterized in that, Step S3 specifically includes: S31, the feature memory dictionary D contains n sampled data points, which do not include the clean coal ash content label in S13. Kernel processing is performed on all data points in D to obtain the kernel matrix K, as shown in the following formula: k(d i ,d j )=exp(-γ||d i -d j || 2 ) Where, k(d) i ,d j ) is the element in the i-th row and j-th column of the kernel matrix K, γ is the parameter controlling the bandwidth, and d i and d j These are the i-th and j-th data points in D; The kernel matrix K is centered to obtain the centered kernel matrix K. c The formula is as follows: Among them, 1 n*n It is a matrix with all elements being 1; S32, using kernel principal component analysis to perform eigenvalue decomposition on the centered kernel matrix, obtains n eigenvalues ​​{λ1,λ2,...,λ...} arranged in ascending order. n } and its corresponding unit eigenvector U = [u1, u2, ..., u n The eigenvalues ​​and unit eigenvectors are divided into micro-mode directions {λ1,λ2,...,λ...}. k }with U k =[u1,u2,...,u k ] and main mode direction {λ k+1 ,λ k+2 ,...,λ n }with U n-k =[u k+1 ,u k+2 ,...,u n Two categories, the formula is as follows: Where η is the set boundary threshold parameter.

5. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 4, characterized in that, Step S4 specifically includes the following steps: S41, for the data in the sorting data stream at the current moment Calculate its projection k in the high-dimensional kernel space x =[k x1 ,k x2 ,…,k xi The formula is as follows: Where, k xi It is k x The i-th element, d i It is the i-th data in the feature memory dictionary D; k x Substituting this into the centralized kernel space yields the projection into the centralized kernel space, as shown in the following formula: in, It is k x The centered vector, 1 n*n 1 n*1 1 1*n It is a matrix consisting entirely of 1s; S42, Calculation The projection in the micro-mode direction is given by the following formula: Where, k x(i≤k) yes Projection in the micro-mode direction, U k The S32 micro-mode directional eigenvector matrix is ​​used to calculate the micro-mode directional information quantification index, as follows: Where α1 is a quantitative indicator measuring the directional informationality of micro-modes, and k i It is k x(i≤k) The value in the i-th dimension; S43, Data sorting stream at the current moment The informational quantitative index α1 satisfies: α1≥δ Here, δ is a set constant. If the condition α1≥δ is not met, proceed to S5; if the condition is met, actively sample the current data from the tag data stream and sort the data stream. tags Switch to S6.

6. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 5, characterized in that, Step S5 specifically includes the following steps: S51, based on the eigenvector matrix U obtained in S32 n-k The projection coordinates of the data in the feature memory dictionary D onto the main pattern direction are calculated using the following formula: x′ i =K ci U n-k ,i=1,2,3,...,n Among them, K ci It is the centered kernel matrix K c The i-th row, x′ i These are the projected coordinates of the data in dictionary D in S51 in the main mode direction; S52, Establish the evaluation function along the main feature pattern direction, as follows: Where M is the bandwidth matrix and G is the kernel function; S53, according to calculation The projected coordinates in the main pattern direction are used to quantify their similarity to the feature memory dictionary data, as shown in the following formula: α2=f(k x(k＜i≤n) ) Where, k x(k＜i≤n) yes Projection in the main mode direction, U n-k α1 is the feature vector matrix in the main mode direction, and α2 is a quantitative indicator that measures the information content in the main mode direction. S54, sorting data stream at the current moment. The informational quantitative indicators in S53 satisfy the following: α2≤υf(x′ i ) min Where υ is a set constant, f(x′) i ) min This represents the minimum value when i = 1, 2, ..., n; if α2 ≤ υf(x′) is not satisfied. i ) min If the condition is met, proceed to step S21 when the next data arrives in the sorting data stream; if the condition is met, actively sample the data from the tag data stream at the current moment in the sorting data stream. tags Switch to S6.

7. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 6, characterized in that, Step S6 specifically includes: S61 will be the data points actively sampled in S4 or S5. Add the feature memory dictionary D, and remove data points in D whose retention time exceeds a threshold to obtain D. new The candidate dictionary that maximizes the value function is used as the final new feature memory dictionary. The evaluation function formula is as follows: in, and T old It is the retention time of the old data being replaced in the dictionary.

8. The method for active data stream sampling and prediction of ash content indicators in coal heavy media separation process according to claim 7, characterized in that, Step S7 specifically includes: S71, Actively Sampled Data Points Increase the weight of actively sampled sample points and establish a memory enhancement coefficient: Where κ is the reinforcement value and ξ is the relaxation factor; S72, given N previous actively sampled data... The neural network has L hidden layer nodes. The weights and biases of the output layer are randomly generated within a certain interval. The matrix form of its hidden layer output is as follows: Y = Hβ Where H is the hidden layer output and β is the output layer weight; S73, for actively sampled data points The model parameters are updated by using a recursive method and the Woodbury formula to find the inverse, as detailed below: Set W N For the memory reinforcement matrix, W0 = I is the identity matrix, H N These are the outputs of the first N sampling points, H0 is the output of the hidden layer of the initial dataset, Y0 is the clean coal ash content of the initial dataset, and w N+1 yes The corresponding memory enhancement coefficient, H N+1 ∈R 1*N yes In the output of the hidden layer, λ is the label of the actively sampled data, and λ is the regularization coefficient. remember: Using Woodbury's formula, we get: The output layer weights are as follows: b N+1 =(IK N+1 H N+1 )b N +K N+1 y N+1 in, Let H DS It is the output of the sorted data stream in the hidden layer of the random weighted neural network. It is The predicted clean coal ash content is updated after the output layer weights are updated; it is then transferred back to S2 for continuous processing until the data stream ends, thus completing the active sampling and prediction of the data stream.

Images