CN111709942B - Zinc flotation dosing amount prediction control method based on texture degree optimization - Google Patents

Abstract

The invention discloses a method for predicting and controlling the dosing amount of zinc flotation based on texture degree optimization. The color feature and the texture feature of the selected foam image define the texture complexity according to the texture parameters of the selected foam image and the relationship of concentrate grade; secondly, using the extracted feature parameters, a neural network is used to construct a prediction model of the selected foam texture complexity , and finally, according to the square of the difference between the predicted value of the selected foam texture complexity and the expected optimal value as the objective function, the optimization method is used to calculate the adjustment value of the dosage of medicine to complete the dosing control. The invention combines the characteristics of the selected and roughed foam states, so that the control results are more excellent, the recovery efficiency of minerals is improved, and the consumption of chemicals is reduced.

Description

A predictive control method for zinc flotation dosage based on texture optimization

technical field

The invention belongs to the technical field of froth flotation, in particular to a method for predicting and controlling the dosing amount of zinc flotation process.

Background technique

Froth flotation is a process in which minerals are separated from useless gangue by using the difference in the hydrophilicity and hydrophobicity of minerals. The state of texture, etc. is judged and adjusted according to the production state. However, manual observation has strong subjectivity and randomness of operation, which makes the dosing operation error large and low in efficiency during the flotation process, which may lead to serious waste of chemicals and mineral resources due to inaccurate discrimination. The texture characteristics are closely related to the concentrate grade of the foam, so by controlling the texture complexity in the concentrating tank within a certain range, the final concentrate grade can be controlled to reach the standard.

At present, the existing control methods of dosing amount in the flotation process are mainly based on the research on the size characteristics of the flotation froth. By studying the size distribution probability function (PDF) of the froth size, a Hammerstein-Wiener model is established to add dosing to the roughing tank for 6 minutes. The size distribution curve of the foam is predicted, and the quality of flotation is judged by analyzing the size of the foam. The difference value of the size distribution curve) establishes a calculation model for the adjustment of the dosing amount, and realizes the reasonable prediction and control of the dosing amount. However, the flotation process is a complex multi-technical process, which is completed through roughing, beneficiation and sweeping. This method only analyzes and judges the froth state in the roughing cell, and does not combine the froth state of multiple flotation cells. , in actual industrial applications, there is a problem that the production status cannot be fully monitored and adjusted in a timely manner.

The invention mainly combines the two technological processes of roughing and selection to carry out the research on the prediction control of the dosing amount of the zinc flotation process. Feature information, build a neural network prediction model to predict the foam texture complexity of the selected slot. Taking the square of the difference between the predicted selected foam texture degree and the optimal texture complexity as the control target, an expert experience fuzzy rule base is established, and the optimization method is adopted to solve the problem, and the value of the modified drug amount required to achieve the control target is calculated. By predicting the characteristics of the selection tank, the future working conditions can be judged, the adjustment of the dosage of the medicine can be made in advance, and the production process can be adjusted in time.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for optimizing the prediction control method of dosing amount in zinc flotation process based on texture degree, firstly extracting the features such as size, color, entropy value, energy, inverse moment of foam in roughing tank, and then building a neural network The prediction model predicts the texture degree of the selected foam, and calculates the adjustment value of the drug dose according to the difference between the predicted value and the expected optimal value, so as to realize the control of the drug dose.

The concrete steps of the technical solution adopted in the present invention are as follows:

S1. Extract the froth image features in the zinc flotation roughing cell, including the entropy, energy, inverse moment, froth size and froth color of the flotation froth image: By extracting the features of n rough froth images, the spatial gray The degree co-occurrence matrix method, that is, the SGLCM method, extracts the entropy, energy, and inverse moment of the foam image to obtain the specific data sets as E _r = [E1, E2, E3, E4...En], A _r = [A1, A2 ,A3,A4,...An], C _r =[C1,C2,C3,C4,...Cn], the classical watershed algorithm is used to segment and extract the characteristic data of foam size S=[S1,S2,S3, S4...Sn], the foam color feature is calculated through the HSV color channel to obtain data Co=[Co1, Co2, Co3, Co4...Con], and the five feature data constitute a rough selection feature data set U;

S2. Select the selected foam images corresponding to the n rough selected foam images, and use the SGLCM method to extract the texture feature parameters of the n selected foam images:

entropy value

energy

Inverse moment

By analyzing the relationship between these three characteristic parameters and the texture of the foam image and the actual concentrate grade, the energy of the selected foam image corresponding to the high concentrate grade is low, and the entropy value and the inverse moment are high, so the texture complexity of the selection tank is defined:

Calculate the texture complexity TC for n selected foam images respectively, and form the selected groove texture complexity data set T=[TC ₁ , TC ₂ , TC ₃ , TC ₄ ... TC _n ], where i, j are foam The pixel value of the image, L is the quantization level of the image, and p _d (i, j) is the element of the ith row and the jth column in the spatial grayscale co-occurrence matrix.

S3. Build a BP neural network prediction model to predict the texture complexity of selected foam images:

a. Entropy value _Er = [E1, E2, E3, E4...En], energy _Ar = [A1, A2, A3, A4,... An] of the extracted rough foam image, inverse difference moment C _r =[C1,C2,C3,C4,...Cn], foam size S=[S1,S2,S3,S4...Sn] and foam color Co=[Co1,Co2,Co3,Co4.. .Con] as the 5 input variables of the input layer; the selected slot texture complexity TC is used as the output variable of the neural network.

b. Assign the trust degree to the five input variables according to the feature importance, and obtain the weights of entropy, energy, inverse moment, foam size and foam color according to the trust degree as w1, w2, w3, w4, w5;

其中m为隐含层节点数，g为输入层节点数，o为输出层节点数，u为常数；由于输入变量为5个，输出变量为1个，故隐含层节点数m确定在2～13之间。c. The hidden layer nodes of the neural network are determined according to the empirical formula:

where m is the number of hidden layer nodes, g is the number of input layer nodes, o is the number of output layer nodes, and u is a constant; since there are 5 input variables and 1 output variable, the number m of hidden layer nodes is determined to be 2 ~13.

d. Input the input variables and output variables of the determined weights into the neural network to train the texture complexity prediction model of the selection slot: select from the feature data set U obtained in step S1 and the texture data set T obtained in step S2 The corresponding M pieces of data constitute the training set of the neural network, which is input into the constructed BP neural network, and the prediction error <Δ is selected as the training end condition to obtain the prediction model.

e. Use the test set to test and revise the prediction model

S4. Use statistical methods to calculate the optimal texture degree T _b corresponding to the concentrate grade obtained by the flotation process at 52% to 56%, and compare the selected foam texture degree predicted value T _p calculated by the BP neural network prediction model with the The square of the difference between the optimal foam texture degrees T _b is selected as the objective function, the expert experience fuzzy rule base is established, and the optimization method is adopted to solve the problem, and the dosing amount x is calculated:

f(x)=min{(T _p -T _b ) ² }

Fuzzy rules are defined as:

if E=e ₀ , A=a ₀ , C=c ₀ , S=s ₀ , Co=Co ₀ , T _p =y ₀ ; thenx=x ₀ ;

if E=e ₁ , A=a ₁ , C=c ₁ , S=s ₁ , Co=Co ₁ , T _p =y ₁ ; thenx=x ₁ ;

...

if E= _en , A=an , C=cn , S= _sn , Co= _Con , _Tp = _yn ; _thenx = _{xn ;}

According to the actual calculation process, among them e ₀ , e ₁ , ... e _n , a ₀ , a ₁ , ... a _n , c ₀ , c ₁ , ... c _n , s ₀ , s ₁ , ... s _n , Co ₀ , Co ₁ ,... _Con are the entropy of the extracted rough selection foam image respectively, the value range is [2.0, 3.0], the energy value interval is [0.2, 0.4], the inverse moment interval is [0.6, 0.8], and the size value interval is [0.5, 0.8], the value range of the color parameter sequence falls within [0.4, 0.7], y ₀ , y ₁ , ... y _n is the predicted value of the selected foam texture, and the value falls within [0.8, 1.0], x ₀ , x ₁ ,...x _n is the calculated dose adjustment amount [2000, 4500], T _b =0.85.

The weights corresponding to each feature of the roughly selected foam image in step a in S3 are w1=0.1, w2=0.3, w3=0.2, w4=0.3, and w5=0.1.

The prediction model error in step d in S3 is Δ=0.01.

In step c of S3, the size of u is 0<u<10, and the value of the number of hidden layer nodes m is 7.

Compared with the prior art, the beneficial effect of the present invention is that compared with the traditional flotation process dosage control method, the production state between multiple production processes can be considered, and the foam state between the concentrating tank and the roughing tank can be considered. Carry out analysis and research, and predict the actual production process, and make timely and effective calculations for the dosing amount control of the flotation process. The flotation process has a long process flow, and the coupling of multiple processes is serious. The traditional drug dosage control method only involves the study of the influence and effect of a single process or a single foam feature on the dosage of the flotation process, which is easy to increase the misjudgment of the production state and the dosage of the drug. Insufficient or excessive problems cause waste of ore resources and reduce the economic benefits of enterprises. Therefore, the present invention considers the importance of multiple features, establishes a neural network prediction model and utilizes a drug dose calculation method, and effectively combines the double processes of selection and sweeping to realize reasonable and efficient drug dose prediction control.

Description of drawings

Fig. 1 is the schematic flow sheet of the method for the predictive control of dosing amount in zinc flotation process of the present invention;

FIG. 2 is a structural block diagram of the present invention.

Detailed ways

In order to describe the technical solutions and advantages of the present invention in more detail and concretely, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

S1. Extract the froth image features in the zinc flotation roughing cell, including the entropy, energy, inverse moment, froth size and froth color of the flotation froth image: By extracting the features of n rough froth images, the spatial gray The degree co-occurrence matrix method, that is, the SGLCM method, extracts the entropy, energy, and inverse moment of the foam image to obtain the specific data sets as E _r = [E1, E2, E3, E4...En], A _r = [A1, A2 ,A3,A4,...An], C _r =[C1,C2,C3,C4,...Cn], the classical watershed algorithm is used to segment and extract the characteristic data of foam size S=[S1,S2,S3, S4...Sn], the foam color feature is calculated through the HSV color channel to obtain data Co=[Co1, Co2, Co3, Co4...Con], and the five feature data constitute a rough selection feature data set U;

S2. Select the selected foam images corresponding to the n rough selected foam images, and use the SGLCM method to extract the texture feature parameters of the n selected foam images:

entropy value

energy

Inverse moment

By analyzing the relationship between these three characteristic parameters and the texture of the foam image and the actual concentrate grade, the energy of the selected foam image corresponding to the high concentrate grade is low, and the entropy value and the inverse moment are high, so the texture complexity of the selection tank is defined:

Calculate the texture complexity TC for n selected foam images respectively, and form the selected groove texture complexity data set T=[TC ₁ , TT ₂ , TC ₃ , TC ₄ ... TC _n ], where i, j are foam The pixel value of the image, L is the quantization level of the image, and p _d (i, j) is the element of the ith row and the jth column in the spatial grayscale co-occurrence matrix;

S3: Build a BP neural network prediction model to predict the texture complexity of selected foam images. The specific steps are as follows:

Step 1. Extract the entropy value _Er = [E1, E2, E3, E4...En], energy _Ar = [A1, A2, A3, A4,... An], Moments _Cr =[C1,C2,C3,C4,...Cn], foam size S=[S1,S2,S3,S4...Sn] and foam color Co=[Co1,Co2,Co3,Co4. ..Con] as the 5 input variables of the input layer; select the slot texture complexity TC as the output variable of the neural network;

Step 2. According to the feature importance, the trust degree is assigned to the five input variables, and the weight values of the entropy value, energy, inverse difference moment, foam size, and foam color are obtained according to the trust degree. w1, w2, w3, w4, w5;

其中m为隐含层节点数，g为输入层节点数，o为输出层节点数；由于输入变量为5个，输出变量为1个，故隐含层节点数m确定在2～13之间，为了加快计算速度和降低计算误差，m选为7。Step 3. The hidden layer nodes of the neural network are determined according to the empirical formula:

where m is the number of hidden layer nodes, g is the number of input layer nodes, and o is the number of output layer nodes; since there are 5 input variables and 1 output variable, the number m of hidden layer nodes is determined to be between 2 and 13 , in order to speed up the calculation speed and reduce the calculation error, m is selected as 7.

Step 4: Input the input variables and output variables of the determined weights into the neural network to train the texture complexity prediction model of the selected slot: from the feature data set U obtained in step S1 and the texture data set T obtained in step S2 Select the corresponding M pieces of data to form the training set of the neural network, input it into the constructed BP neural network, select the prediction error <Δ as the training end condition, and obtain the prediction model;

Step 5: According to the actual selected foam image texture complexity and the error of the predicted value, use the test set to test and correct the prediction model.

S4.: Use statistical methods to calculate the optimal texture degree T _b when the concentrate grade obtained by the flotation process is about 54%, and compare the selected foam texture degree predicted value T _p calculated by the BP neural network prediction model with the refined ore texture degree T p . The square of the difference between the optimal foam texture degrees T _b is selected as the objective function, the expert experience fuzzy rule base is established, the optimization method is adopted to solve the problem, and the dosage x is calculated:

f(x)=min{(T _p -T _b ) ² }

Fuzzy rules are defined as:

if E=e ₀ , A=a ₀ , C=c ₀ , S=s ₀ , Co=Co ₀ , T _b =y ₀ ; thenx=x ₀ ;

if E=e ₁ , A=a ₁ , C=c ₁ , S=s ₁ , Co=Co ₁ , T _b =y ₁ ; thenx=x ₁ ;

...

if E= _{en , A=an , C=c n ,} S=s _n , Co=Con , T _b = _yn ; _thenx = _{x n}

According to the actual calculation process, among them e ₀ , e ₁ , ... e _n , a ₀ , a ₁ , ... a _n , c ₀ , c ₁ , ... c _n , s ₀ , s ₁ , ... s _n , Co ₀ , Co ₁ ,... _Con are the entropy of the extracted rough selection foam image respectively, the value range is [2.0, 3.0], the energy value interval is [0.2, 0.4], the inverse moment interval is [0.6, 0.8], and the size value interval is [0.5, 0.8], the value range of the color parameter sequence falls within [0.4, 0.7], y ₀ , y ₁ , ... y _n is the predicted value of the selected foam texture, and the value falls within [0.8, 1.0], x ₀ , x ₁ ,...x _n is the calculated dose adjustment amount [2000, 4500], T _b =0.85.

Claims (6)

1. a zinc flotation dosage prediction control method based on texture degree optimization, is characterized in that comprising the following steps: S1. Extract the froth image features in the zinc flotation roughing cell, including the entropy, energy, inverse moment, froth size and froth color of the flotation froth image: By extracting the features of n rough froth images, the spatial gray The degree co-occurrence matrix method, that is, the SGLCM method, extracts the entropy, energy, and inverse moment of the foam image to obtain the specific data sets as _Er =[E1, E2, E3, E4...En], Ar=[A1, A2, A3, A4, ... An], C _r = [C1, C2, C3, C4, ... Cn], using the classical watershed algorithm to segment and extract the characteristic data of foam size S = [S1, S2, S3, S4 ...Sn], the foam color feature is calculated through the HSV color channel to obtain data Co=[Co1, Co2, Co3, Co4...Con], and the five feature data constitute a rough selection feature data set U; S2. Select the selected foam images corresponding to the n rough selected foam images, and use the SGLCM method to extract the texture feature parameters of the n selected foam images: entropy value

energy

Inverse moment

By analyzing the relationship between these three characteristic parameters and the texture of the foam image and the actual concentrate grade, the energy of the selected foam image corresponding to the high concentrate grade is low, and the entropy value and the inverse moment are high, so the texture complexity of the selection tank is defined:

The texture complexity TC is calculated for n selected foam images respectively, and the selected groove texture complexity data set T=[TC ₁ , TC ₂ , TC ₃ , TC ₄ ... TC _n ] is constructed, where i, j are foam The pixel value of the image, L is the quantization level of the image, p _d (i, j) is the element of the ith row and the jth column in the spatial grayscale co-occurrence matrix; S3. Build a BP neural network prediction model to predict the texture complexity of selected foam images: a. Entropy value Er=[E1, E2, E3, E4...En] of the extracted rough foam image, energy _Ar =[A1, A2, A3, A4,...An], inverse difference moment C _r = [C1, C2, C3, C4,...Cn], foam size S=[S1, S2, S3, S4...Sn] and foam color Co=[Co1, Co2, Co3, Co4... Con] as the 5 input variables of the input layer; the selected slot texture complexity TC is used as the output variable of the neural network; b. Assign the trust degree to the five input variables according to the feature importance, and obtain the weight values of entropy, energy, inverse moment, foam size and foam color according to the trust degree as w1, w2, w3, w4, w5; c. The hidden layer nodes of the neural network are determined according to the empirical formula:

Among them, m is the number of hidden layer nodes, g is the number of input layer nodes, and o is the number of output layer nodes; since there are 5 input variables and 1 output variable, choose the number of hidden layer nodes m between 2 and 13; d. Input the input variables and output variables of the determined weights into the neural network to train the selected slot texture complexity prediction model: select from the feature data set U obtained in step S1 and the texture data set T obtained in step S2 The corresponding M pieces of data constitute the training set of the neural network, which is input into the constructed BP neural network, and the prediction error <Δ is selected as the training end condition to obtain the prediction model; e. Use the test set to test and revise the prediction model S4. Use statistical methods to calculate the optimal texture degree T _b corresponding to the concentrate grade obtained by the flotation process at 52% to 56%, and compare the selected foam texture degree predicted value T _p calculated by the BP neural network prediction model with the The square of the difference between the optimal foam texture degrees T _b is selected as the objective function, the expert experience fuzzy rule base is established, and the optimization method is adopted to solve the problem, and the dosing amount x is calculated: f(x)=min{(T _p -T _b ) ² } Fuzzy rules are defined as: if E=e ₀ , A=a ₀ , C=c ₀ , S=s ₀ , Co=Co ₀ , T _p =y ₀ ; then x=x ₀ ; if E=e ₁ , A=a ₁ , C=c ₁ , S=s ₁ , Co=Co ₁ , T _p =y ₁ ; then x=x ₁ ; ... if E= _en , A=an, C=cn, S= _sn , Co= _Con , _Tp = _yn ; then _{x = xn} ; According to the actual calculation process, where e ₀ , e ₁ ... e _n , a ₀ , a ₁ , ... a _n , c ₀ , c ₁ , ... c _n , s ₀ , s ₁ , ... s _n , Co ₀ , Co ₁ , ... Co _n are the entropy of the extracted rough selected foam image respectively, the value range is [2.0, 3.0], the energy value interval is [0.2, 0.4], and the inverse moment interval is [ 0.6, 0.8], the size range is [0.5, 0.8], the value range of the color parameter sequence is [0.4, 0.7], y ₀ , y ₁ , ... y _n is the predicted value of the selected foam texture , the value falls in [0.8, 1.0], x ₀ , x ₁ , . . . x _n is the calculated dose adjustment amount [2000, 4500], T _b =0.85. 2 . The method for predicting and controlling zinc flotation dosing dosage based on texture degree optimization according to claim 1 , wherein the total number n of the rough selection foam images is 3000; the number M of training set data is 2000. 3 . . 3 . The method for predicting and controlling the dosing amount of zinc flotation based on texture degree optimization according to claim 1 , wherein the weights corresponding to each feature of the rough selected froth image are w1=0.1, w2=0.3, w3=0.2, w4=0.3, w5=0.1. 4 . The method for predicting and controlling the dosing amount of zinc flotation based on texture optimization according to claim 1 , wherein the prediction model error is Δ=0.01. 5 . 5 . The method for predicting and controlling the dosing amount of zinc flotation based on texture degree optimization according to claim 1 , wherein, in step c of S3 , the value of u is 0<u<10. 6 . 6 . The method for predicting and controlling the dosing amount of zinc flotation based on texture degree optimization according to claim 1 , wherein the value of the number m of nodes in the hidden layer is 7. 7 .

Images