CN108931621B - Zinc ore grade soft measurement method based on process texture characteristics - Google Patents

Abstract

A soft measurement method for zinc flotation concentrate grade based on process characteristics integrates expert knowledge and a data modeling method, firstly, single-frame texture characteristics based on image statistical characteristics are provided according to observation key points when field workers watch bubbles to represent foam images, a texture sequence is provided to mathematically transform the current production state according to the characteristic that the field workers need to observe the foam state within a period of time to judge the current production state, and a modeling method for the texture sequence is provided, so that the dimension of a feature vector is reduced. An improved decision tree lifting algorithm is adopted in the prediction algorithm, so that the overfitting problem caused by too fast learning is effectively inhibited, and the generalization capability is improved. Experiments prove that the method has simple calculation, higher execution speed and higher prediction accuracy, is convenient for actual operation on site, can guide the operation on site immediately, optimizes the production process and solves the problem of difficult online detection of the grade of the existing zinc ore.

Description

Zinc ore grade soft measurement method based on process texture characteristics

Technical Field

The invention belongs to the technical field of froth flotation, and particularly relates to a prediction method of zinc flotation concentrate grade.

Background

The froth flotation is one of the most important mineral separation methods in the zinc smelting at present, the flotation is a method for separating the crushed useful minerals and gangue symbiotic with the useful minerals according to the difference of the surface physicochemical properties of the mineral particles according to the difference of the floatability of the minerals, the froth flotation is a process for separating the crushed useful minerals and the gangue symbiotic with the useful minerals, a large number of bubbles with different sizes, shapes, textures and the like are formed in ore pulp through continuous stirring and air blowing in the flotation process, and the bubbles carry the mineral particles to rise to the surface of a flotation tank to form a foam layer, so that the separation of the minerals and the gangue is realized. For a complicated industrial process such as the foam flotation, due to the reasons of long process flow, serious association and coupling of sub-processes, difficult on-line detection of part of key parameters and the like, the working condition state of the flotation process lacks an effective comprehensive perception means, the manual back-and-forth inspection is seriously relied on, and whether the current production is in a normal state or not is roughly judged by virtue of experience so as to further implement a corresponding operation strategy. The single and wild method which seriously depends on artificial experience perception often produces improper production operation and cannot ensure stable and optimal operation of production. Although the plant can determine the production status of the flotation process by obtaining accurate concentrate grade by off-line assay analysis, this often takes several hours, the inspection process is complicated and the cost is high and lags behind the production process. Because the flotation process has long flow and a plurality of influence factors, the on-line detection of the concentrate grade can not be realized, the instant adjustment of the dosing quantity and other parameters is influenced, and the recovery rate of minerals is finally influenced. Therefore, the research on the real-time online detection method of the production indexes in the flotation process has important significance for guiding the production operation and the optimized operation of the process.

With the rapid development of computer technology and digital image processing technology, the application of the machine vision-based soft measurement technology to the flotation process brings a new breakthrough to the real-time monitoring of flotation indexes. Machine vision is an important means for simulating the human visual perception capability to realize the automatic measurement and control of the industrial process, and the online detection of the concentrate grade in the flotation process can be realized due to the advantages of high precision, modularization, intellectualization, nondestructive perception and the like. A large number of foam videos at different grades can be obtained through image acquisition equipment, the videos and the acquired corresponding production data are combined to form an original data set, a data model of the foam images and the concentrate grade is established by a data-driven modeling method, and online detection of the concentrate grade is achieved. The existing concentrate grade prediction methods mainly adopt methods such as B-spline partial least squares regression, support vector machines, neural networks and the like, and all the methods have defects of different degrees, are difficult to process data of large samples, are sensitive to data with noise, and still have certain problems in application.

Disclosure of Invention

Aiming at the defects of difficulty in online detection of concentrate grade, high cost, large delay and prediction of the concentrate grade of zinc flotation in the zinc flotation process, the invention provides a construction method of lead-zinc flotation froth image process characteristics by using experience knowledge of field workers and accumulated production data, and simultaneously constructs a prediction method of the concentrate grade.

The technical scheme adopted by the invention comprises the following steps:

s1: collecting foam videos and production data of zinc flotation at different grades, and performing data preprocessing on the collected zinc flotation data and production data, wherein the process comprises the following steps:

1) eliminating error data of which the measured data value exceeds the variation range;

2) rejecting unmatched data and data with a vacancy value;

s2: reading RGB foam images by using foam video obtained by a flotation field image acquisition system, converting the foam images from RGB color space to HSI color space, extracting brightness components as source images, and obtaining an image sequence I ═ I₁,I₂,...,I_q]Q is the frame number of the video;

s3: extracting foam image I of ith frame in image sequence I_iTexture feature of (1), noted as T_iAnd performing the same processing on each frame of image in the image sequence I to obtain a texture sequence T ═ T₁,T₂,...,T_q]Wherein T is_i＝[β_ix,μ_ix,γ_ix,β_iy,μ_iy,γ_iy]；

S4: a gaussian-markov autoregressive moving average model is established for the size distribution sequence T obtained in S3, and the function expression thereof is as follows:

wherein: x (k), x (k +1) is an n-dimensional state vector;

y (k) is onem-dimensional output vector, y (k) T_kRepresenting a texture feature;

v (k) is a random variable obeying a gaussian distribution, and the covariance matrix is V;

w (k) is a random variable obeying a gaussian distribution, and the covariance matrix is W;

k＝1,2,3…,q；

estimating the values of the parameters A, C and V, and arranging the values into a column vector F, wherein the column vector F is called as the process texture characteristic of the flotation process in the period of time;

s5: f obtained in S4 and the concentrate grade G corresponding thereto are combined as a sample point D_i＝{F_i,G_i) }; and (F) solving the process texture characteristics of all collected videos, combining the process texture characteristics with the concentrate grade, and solving all sample point sets D { (F)₁,G₁),(F₂,G₂),...,(F_N,G_N) With F }_i ^(j)Is represented by F_iThe jth component of (a);

s6: and (5) training a decision tree model by using the sample set obtained in the S5 through a CART algorithm, and marking as f₀；

S7: establishing a prediction model, designing a loss function Lf to quantitatively calculate the deviation between the generated model output value and the actual measured value, and gradually and incrementally generating a variable f according to the loss function Lf₀The basic lifting tree model comprises the following steps:

1) will f is_iIs updated to f_i'＝f_i+φ_i+1(F) Wherein: i is 0,1,2, …, #_i+1A newly added weak learner;

2) design the loss function Lf, order

Wherein η is a constant less than 1;

3) solving for

Obtaining phi which minimizes the loss function Lf_i+1A numerical solution of (c);

4) is measured by phi_i+1For the target value, a CART algorithm is utilized to establish a decision tree model to obtain f_i+1；

5) And repeatedly generating the decision tree to generate L trees to obtain a promoted decision tree:

f_boost(F)＝f₀(F)+ηf₁(F)+...+ηf_L(F) η is determined by the validation set;

s8: acquiring zinc flotation froth sample data to be detected, inputting the zinc flotation froth sample data into an industrial computer, calculating process texture characteristics by the computer according to the steps S3 and S4, and inputting the obtained process texture characteristics into the model obtained in the S7 to obtain a predicted zinc fine product position;

the Weber distribution function stated in S3 is

C is a normalized constant, β, mu and gamma are parameters;

the invention provides a soft measurement method for the grade of zinc flotation concentrate based on process texture characteristics, which solves the problem that the on-site antimony ore grade is difficult to detect on line; aiming at the problem that the traditional method only depends on the image characteristics extracted from one picture to represent the current foam state, the invention provides a time-dependent process size characteristic by combining the experience of field workers. According to observation key points of field workers when watching bubbles, single-frame texture features based on image statistical features are provided to represent foam features, expert knowledge is fused, and the current state of the foam can be represented more accurately; according to the characteristic that field workers need to observe the foam state within a period of time to judge the current production state, a modeling method for the size distribution sequence is provided, the dimension of the characteristic vector is reduced, and the rule of dynamic change of the size distribution sequence can be reflected; aiming at the problems that the traditional grade prediction method is difficult to process data of a large sample and is sensitive to data with noise, a promotion decision tree model for grade prediction is established in a data-driven mode; the tree model is a nonlinear model, so that a strong nonlinear function of grade prediction can be well fitted, and the tree model is of a branch structure, so that the operation speed is high, the concentrate grade corresponding to the current foam state can be quickly solved, and the online detection is convenient to realize; aiming at the problem that the generalization capability of a trained model is poor due to overfitting easily occurring when a decision tree model is applied to zinc concentrate grade prediction, the learning rate eta is introduced, and a constant less than 1 is added in the forward increasing process each time, so that the overfitting problem caused by too fast learning can be effectively inhibited, and the generalization capability is improved; the method has the advantages of simple calculation, high execution speed, high prediction accuracy, convenience for actual field operation, capability of immediately guiding field operation and optimization of production process.

Drawings

Fig. 1 is a flow chart of the zinc flotation concentrate grade prediction method implemented by the invention.

Detailed Description

The technical solutions adopted in the present invention are described and explained in more detail and clearly with reference to the accompanying drawings. The invention provides a time-related process size feature extraction method aiming at the limitation that the traditional method is difficult to accurately reflect the foam state only by depending on a single-frame picture, and realizes the online detection of the zinc concentrate grade by utilizing an improved decision tree promotion model. It should be apparent that the described embodiments are only some of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be obtained by a person skilled in the relevant art without any inventive step based on the embodiments of the present invention, shall be within the scope of the present invention.

As shown in fig. 1, a flow chart of a soft measurement method for lead-zinc ore grade based on process size distribution characteristics in an embodiment of the present invention is shown, and the method includes the following specific steps:

s1: collecting foam video and production data of zinc flotation under different grades, and collecting zinc flotation data and production

Data preprocessing is carried out on the data, and the process is as follows:

1) eliminating error data of which the measured data value exceeds the variation range;

2) rejecting unmatched data and data with a vacancy value;

s2: reading RGB foam images by using foam video obtained by a flotation field image acquisition system, converting the foam images from RGB color space to HSI color space, extracting brightness components as source images, and obtaining an image sequence I ═ I₁,I₂,...,I_q]Q is the frame number of the video;

s3: for foam image I in image sequence_iExtracting the texture features, wherein the flow of extracting the texture features is as follows:

1) using median filter pair I_iPerforming median filtering and removing noise to obtain I_i'; a 3 x 3 template median filter is used in the method,

2) using a Gaussian derivative filter pair I in the x-direction_i' Filtering to obtain an image B containing edges of image details_ix；

3) Statistics B_ixProbability histogram of gray levels H_ixTraversing all the pixel points to count the number of the pixel points belonging to different brightness, dividing by the image resolution, H_ix＝{(-255,H_-255),(-254,H_-254),…,(255,H₂₅₅)}；

4) H is to be_ixIs linearly transformed to the interval [ -1, +1 [)]Obtaining a normalized probability histogram H_ixFitting H by using a Weber distribution function_ixSolving for

Obtain parameters β of the Weber distribution function_ix,μ_ix,γ_ix；

5) Using a pair of Gaussian derivative filters I in the y-direction_i' Filtering to obtain an image B containing edges of image details_iy；

6) Statistics B_iyProbability histogram of gray levels H_iyTraversing all the pixel points to count the pixel points belonging to different brightness, and dividing the pixel points by the image resolution to obtain H_iy；

7) H is to be_iyAsh of (2)Linear transformation of degree to [ -1, +1]Obtaining a normalized probability histogram H_ixFitting H with a Weber distribution function_iySolving for

Obtain parameters β of the Weber distribution function_iy,μ_iy,γ_iy；

Will T_i＝[β_ix,μ_ix,γ_ix,β_iy,μ_iy,γ_iy]As an image I_iThe texture feature of (1) is to perform the same processing on each frame of image in the image sequence I to obtain a texture sequence T ═ T₁,T₂,...,T_q]；

S4: a gaussian-markov autoregressive moving average model is established for the texture sequence T obtained in S3, and the function expression thereof is as follows:

wherein: x (k), x (k +1) is an n-dimensional state vector;

y (k) is an m-dimensional output vector, and y (k) is T_kRepresenting a texture feature;

v (k) is a random variable obeying a gaussian distribution, desirably 0, and the covariance matrix is V;

w (k) is a random variable obeying a gaussian distribution, desirably 0, and the covariance matrix is W;

k＝1,2,3…,q；

estimating the values of the parameters A, C and V, arranging the values into a column vector F, and referring to F as the process texture characteristic of the flotation process in the period of time, wherein the steps are as follows:

1) writing Y (k) in the gaussian markov sliding autoregressive model in S4 in the form Y ═ Y (1), Y (2), Y (3),.., Y (q);

2) determining the average y (k) of all the outputs y (k) of the model, k 1,2_m,

Let Y^q＝Y-y_m；

3) And simulating y (k) by using a mapping function from m × q dimensional space to a real number domain as a probability density function, wherein the mathematical expression is as follows:

wherein: mu is one and Y^qThe same shape of the matrix, ∑ is a symmetric matrix with positive elements, z is a normalization constant;

4) solving so that p (Y)^q) Maximum value of the parameter, i.e.

The method comprises the following specific steps:

① the original form is written as Y₁ ^τ＝[y(1),y(2),…,y(k)],

W₁ ^τ＝[w(1),w(2),…,W(k)]Then, then

② pairs of Y₁ ^τSingular value decomposition to obtain Y₁ ^τ＝U∑V^TThen, then

③ solving for

Obtaining an estimated value of A;

④

wherein

5) Arranging A, C and V in rowsArranged as a column vector F_iSetting the dimension as n;

s5: f obtained in S4 and the concentrate grade G corresponding thereto are combined as a sample point D_i＝{F_i,G_i) }; and (F) solving the process texture characteristics of all collected videos, combining the process texture characteristics with the concentrate grade, and solving all sample point sets D { (F)₁,G₁),(F₂,G₂),...,(F_N,G_N)}；

S6: training a decision tree model by using the sample set obtained in the S5 and adopting a CART algorithm, and marking as f₀；

S7: establishing a prediction model, designing a loss function Lf to quantitatively calculate the deviation between the generated model output value and the actual measured value, and gradually and incrementally generating a variable f according to the loss function Lf₀The basic lifting tree model comprises the following steps:

1) will f is_iIs updated to f_i'＝f_i+φ_i+1(F) Wherein: i is 0,1,2, …, #_i+1A newly added weak learner;

2) designing a loss function Lf, taking the loss function as L2 loss function in the method, and enabling

Wherein η is a constant less than 1;

3) solving for

Obtaining phi which minimizes the loss function Lf_i+1The numerical solution of

Find phi_i+1；

4) Is measured by phi_i+1For the target value, a CART algorithm is utilized to establish a decision tree model to obtain f_i+1；

5) And (3) repeatedly generating the decision tree, if the situation that the prediction error is not reduced by continuously adding new trees for 10 times occurs, considering that the stop condition is met, and generating L trees to obtain a promoted decision tree:

f_boost(F)＝f₀(F)+ηf₁(F)+...+ηf_L(F)；

s8: and (4) acquiring the zinc flotation froth sample data to be detected, inputting the zinc flotation froth sample data into a computer, calculating the process texture characteristics by the computer according to the steps S3 and S4, and inputting the obtained process texture characteristics into the model obtained in the step S7 to obtain the predicted zinc fine product position.

Claims (2)

1. A soft measurement method for lead-zinc ore grade based on process texture characteristics is characterized by comprising the following steps:

s1: collecting foam videos and production data of zinc flotation at different grades, and performing data preprocessing on the collected zinc flotation data and production data, wherein the process comprises the following steps:

1) eliminating error data of which the measured data value exceeds the variation range;

2) rejecting unmatched data and data with a vacancy value;

s2: reading RGB foam images by using foam video obtained by a flotation field image acquisition system, converting the foam images from RGB color space to HSI color space, extracting brightness components as source images, and obtaining an image sequence I ═ I₁,I₂,...,I_q]Q is the frame number of the video;

s3: extracting foam image I of ith frame in image sequence I_iTexture feature of (1), noted as T_iObtaining a texture sequence T ═ T₁,T₂,...,T_q]Wherein T is_i＝[β_ix,μ_ix,γ_ix,β_iy,μ_iy,γ_iy]；

S4: a gaussian-markov autoregressive moving average model is established for the size distribution sequence T obtained in S3, and the function expression thereof is as follows:

wherein: x (k), x (k +1) is an n-dimensional state vector;

y (k) is an m-dimensional output vector, y (k) T_kRepresenting a texture feature;

v (k) is a random variable obeying a gaussian distribution, whose covariance matrix is V;

w (k) is a random variable obeying a gaussian distribution, and the covariance matrix is W;

k＝1,2,3...,q；

estimating the values of the parameters A, C and V, and arranging the parameters A, C and V into a column vector F, wherein the column vector F is called as the process texture characteristic of the flotation process in the period of time;

s5: f obtained in S4 and the concentrate grade G corresponding thereto are combined as a sample point D_i＝{F_i,G_i) }; and (F) solving the process texture characteristics of all collected videos, combining the process texture characteristics with the concentrate grade, and solving all sample point sets D { (F)₁,G₁),...,(F_i,G_i),...,(F_N,G_N) With F }_i ^(j)Is represented by F_iThe jth component of (a);

s6: training a decision tree model by using the sample set obtained in the S5 and adopting a CART algorithm, and marking as f₀；

S7: establishing a prediction model, designing a loss function Lf to quantitatively calculate the deviation between the generated model output value and the actual measured value, and gradually and incrementally generating a variable f according to the loss function Lf₀The basic lifting tree model comprises the following steps:

1) will f is_iIs updated to f_i'＝f_i+φ_i+1(F) Wherein: i is 0,1,2_i+1A newly added weak learner;

2) design the loss function Lf, order

Wherein η is a constant less than 1;

3) solving for

Obtaining phi which minimizes the loss function Lf_i+1A numerical solution of (c);

4) is measured by phi_i+1For the target value, f is obtained by establishing a decision tree model according to the step S6_i+1；

5) Repeating 1) to 4), and generating L trees together to obtain a lifting decision tree: f. of_boost(F)＝f₀(F)+ηf₁(F)+...+ηf_L(F)；

S8: and (4) acquiring the zinc flotation froth sample data to be detected, inputting the zinc flotation froth sample data into a computer, calculating the process texture characteristics by the computer according to the steps S3 and S4, and inputting the obtained process texture characteristics into the model obtained in the step S7 to obtain the predicted zinc fine product position.

2. The soft measurement method for lead-zinc ore grade based on process texture features according to claim 1, characterized in that:

the values of the parameters a, C, V are estimated and arranged into a column vector F, as follows:

1) writing Y (k) in the gaussian markov sliding autoregressive model in S4 in the form Y ═ Y (1), Y (2), Y (3),.., Y (q);

2) determining the average y (k) of all the outputs y (k) of the model, k 1,2_m,

Let Y^q＝Y-y_m；

3) And simulating y (k) by using a mapping function from m × q dimensional space to a real number domain as a probability density function, wherein the mathematical expression is as follows:

wherein: mu is one and Y^qThe same shape of the matrix, ∑ is a symmetric matrix with positive elements, z is a normalization constant;

4) solving so that p (Y)^q) Maximum value of the parameter, i.e.

Arranging A, C, V into a column vector F_iLet its dimension be n, F_i ^(j)Is represented by F_iThe jth component of (a).

Images