Intelligent reagent adding method for coal flotation with segmental explainability

Abstract

The application discloses a segmented explainable coal flotation intelligent dosing method, which captures a flotation dosing data set based on high-precision sensors and detection equipment, and normalizes input features; then, an XGBoost dosing model of coal flotation is trained by using feature data variables, the model is explained by combining SHAP, a main dependent variable is selected based on SHAP mean, the main dependent variable is directly predicted by using the XGBoost model, a secondary dependent variable is predicted by using a reagent ratio calculated according to dry coal slime content, and a segmented boundary value of added reagent amount is calculated by means of SHAP mean, so that a working condition of insufficient reagent only relying on a machine learning model is avoided when the value is less than a conventional range. The new segmented explainable intelligent dosing method is superior to existing manual operation and machine learning algorithms, the dosing prediction algorithm is more scientific and stable, the combustible recovery rate and clean coal yield of coal slime flotation production can be effectively ensured, and the coal preparation plant is improved in quality and efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent control of coal flotation, and particularly relates to a segmented interpretable coal flotation intelligent dosing method. BACKGROUND

[0002] The flotation process is the most commonly used and best effective technology in fine coal separation, and undertakes about 30% of the separation task in a coal preparation plant. Based on the difference in surface wettability of coal and gangue, an appropriate amount of flotation reagent needs to be added to increase the floatability of coal in the coal slime flotation process, and the dispersion effect and dosing amount of the reagent will directly affect the flotation effect. Studies have shown that when the dosage of the collector is high, the bubbles are unstable, affecting the adsorption of the frother on the mineral surface, and the foam layer is easy to accumulate. When the dosage is low, the hydrophobicity of the coal slime particle surface cannot be fully improved, and the clean coal particles are entrained by the tailings. When the corresponding frother is excessive, large bubbles are easily formed, causing gangue particles to be entrained and float, and the quality of the clean coal is poor. When the frother is less, the bubbles are few and weak, and the clean coal is entrained into the tailings. At present, the coal preparation plant mainly relies on the experience of the flotation post driver to manually adjust the dosing amount, and this dosing method usually maintains a fixed flow rate for several hours or more, which causes the coal preparation plant to face a large loss of materials and benefits in the flotation process every year.

[0003] At present, there are many research and development projects in the field of automatic control dosing technology in flotation, and some methods have been applied to industrial production. They are mainly divided into modeling methods based on the visual features of flotation foam and flotation process variable prediction methods. The consistent point is that these modeling methods are based on machine learning technology, taking image features or flotation process variables as model inputs, and optimizing hyperparameters through training to improve the accuracy of predicted dosing amount. For example, patent CN 113128126 A discloses a modeling method for the flotation dosing process based on a generative adversarial network, which establishes the relationship between the flotation reagent dosage and the foam image features. However, these methods have shortcomings in the application of the flotation site, such as the fact that local single surface foam images cannot accurately identify the over-dosing conditions of the reagent in the ore slurry, and most of them are only for single-tank conditions. The contribution of input variables to the model is not studied in the machine learning modeling process. In addition, the existing continuous flotation dosing model cannot effectively match boundary conditions such as high flow rate and low concentration.

[0004] Obviously, coal slime flotation is a complex and continuous multi-process flow, and the shortcomings in the above methods and technologies have always forced the intelligent dosing technology to be continuously updated. SUMMARY

[0005] In view of the above technical deficiencies, the purpose of the present application is to provide a segmented interpretable coal flotation intelligent dosing method, which takes the tree-based model XGBoost as the basic network structure, determines the optimal model variable feedback and prediction sequence by SHAP to explain the contribution of flotation features to the dosing model, and calculates the segmented boundary value of the added dosage by means of SHAP mean value. This method solves the problems of existing technology, such as difficult dosing, inaccurate drug content identification, and inability to effectively match high flow and low concentration boundary conditions, realizes intelligent dosing, and has wide application value for improving the quality and efficiency of coal preparation plants.

[0006] To solve the above technical problems, the present application provides a segmented interpretable coal flotation intelligent dosing method, comprising the following steps:

[0007] S1: Capture coal flotation dosing process feature data and flotation indicators, and establish a coal slime flotation dosing prediction data set, the specific process is:

[0008] S11: Use sensors and tailings ash content meters to capture concentration, flow rate, air charge and tailings ash content as characteristic variables, and use combustible recovery rate and clean coal yield as indicators to mark the indicators of feed flow rate, feed concentration, tailings ash content, collector addition amount and frother addition amount when the combustible recovery rate and clean coal yield meet the requirements of the preparation plant;

[0009] S12: Remove abnormal data, and normalize the values of each variable feature in the interval [0, 1] based on the maximum and minimum values of a single variable, the normalization formula is as follows:

[0010] v(x) = [a(x)-mina(x)] / [maxa(x)-mina(x)]

[0011] Wherein, a(x) is the flotation process variable feature or tailings ash content, and v(x) is the model input feature;

[0012] S13: Divide the data into a training set (80%) and a validation set (20%) according to the ratio of 80:20, and establish a coal slime flotation dosing prediction data set;

[0013] S2: Design and build an XGBoost flotation dosing regression prediction model and calculate the training error, the specific process is:

[0014] S21: Read in the flotation dosing prediction data set by using 10-fold cross-validation method;

[0015] S22: The XGBoost model in S2 is a number integration model, which is composed of multiple regression trees. In order to optimize the loss function, the objective function for further training is:

[0016]

[0017] wherein

[0018] g i is the first derivative of the loss function, h i is the second derivative of the loss function, f represents the optimal dosage of the flotation reagent, Ω is a regularization term, wherein γ is a penalty parameter of the leaf node, and T is the number of leaf nodes.

[0019] S23: Grid SearchCV is used to optimize the hyperparameters to be optimized in XGBoost, including the number of weak reagent learning machines n_estimators, the maximum tree depth max_depth, the penalty coefficient gamma, and the learning rate learning_rate;

[0020] S24: A group of flotation characteristic variables are encoded and the flotation reagent dosage is calculated, the root mean square error RMSE of the loss function is optimized, and the reagent addition model is continuously optimized according to the direction of loss function reduction. The calculation formula of RMSE is as follows:

[0021]

[0022] wherein Y obj is the actual flotation reagent dosage, Y model is the predicted flotation reagent dosage of the model, and n is the number of variable samples.

[0023] S3: SHAP algorithm is used to explain the contribution and importance of the flotation process characteristic variables to the flotation reagent dosage, and the most important characteristics for the collector dosage and the frother dosage are calculated, and the specific process is as follows:

[0024] S31: Taking the i-th group of training data as an example, each group corresponds to four characteristic variables of flotation feed concentration, flow rate, aeration amount and tailing ash content. The SHAP value prediction of the i-th training sample can be expressed as:

[0025]

[0026] wherein

[0027] wherein is the k-th characteristic, is the average prediction of the sample, Z represents the sample set, S is the sample set excluding the k-th characteristic, G represents the prediction model, and G S represents a model trained only with the set S.

[0028] S32: Each flotation feature variable in the training data has a corresponding SHAP value, and the SHAP mean value of the sample is selected to represent the importance of the feature to the prediction of the flotation reagent addition amount, and finally the importance ranking of the flotation process features is obtained;

[0029] S33: Select the variable feature with the largest SHAP mean value as the most important feature for predicting the collector or frother;

[0030] S4: The proportion of reagents in the flotation process is very important, and it is very unfavorable for the flotation system to rely on the model to frequently change the reagent addition ratio. This step mainly relies on SHAP mean value to select one of the flotation reagents as the main dependent variable, and adds according to the prediction value of the XGBoost model in step S2, and the other reagent as the secondary dependent variable, and adds according to the calculated reagent ratio, The detailed steps are as follows:

[0031] S41: Train the flotation reagent addition model by the method in step S2, and then calculate the SHAP mean value (collector) S b and SHAP (frother) S q according to the method in step S3;

[0032] S42: If Sb>Sq, select the collector addition amount as the main dependent variable, and take the frother addition amount as the secondary dependent variable, the main dependent variable is directly predicted by the XGBoost model, and the secondary dependent variable = main dependent variable / k, where k is the reagent ratio; otherwise, the frother addition amount is the main dependent variable, and the collector addition amount is the secondary dependent variable.

[0033] S43: Take the real-time dry slime content of the flotation system to determine the optimal reagent ratio (collector addition amount / frother addition amount), and the calculation formula of the dry slime content is as follows:

[0034] m(g) = Q(l)·p(l)

[0035] Where Q is the feed slurry flow rate, and p is the feed slurry concentration; then under the premise that the flotation index meets the production requirements of the coal preparation plant, the relationship between the flotation reagent ratio and the dry slime content is fitted by SPSS using flotation experiments, k = f(m(g)), where k is the reagent ratio;

[0036] S5: From step S1 to step S4, the flotation reagent can be added intelligently according to the coal slime flotation production process, but due to the unstable factors of industrial flotation feed conditions, when the variable feature is less than the conventional range, only relying on the machine learning model will still exist the condition of insufficient reagent, This step mainly relies on SHAP mean value to segment the boundary value of the added reagent amount, and the detailed steps are as follows:

[0037] S51: SHAP dependence_plot and the relationship between yield, combustible recovery and flotation variables are plotted in turn according to the ranking of the flotation variable features in step S2;

[0038] S52: In the SHAP dependence_plot graph, find the predicted dosage of the corresponding variable so that the flotation index does not meet the concentration value p l , flow value Q l , aeration amount A l and tailing ash content Ad l , set the minimum dosage of the segmented point as p l , Q l , A l and Ad l , and the corresponding collector dosage is the minimum dosage Y b ;

[0039] S6: Set the dosing rules by comprehensively considering all the above steps:

[0040] Read the flotation variables p, Q, A and Ad;

[0041] then calculate the collector dosage Y as Y model , and the foaming agent dosage P as Y model / k;

[0042] if the concentration value p< p l or the flow value Q<Q l or the tailing ash content Ad<Ad l or the aeration amount A<A l ;

[0043] then Y=Y b , and P=Y b / k;

[0044] else Y=Y model and P=Y model / k.

[0045] In summary, the present application discloses a segmented interpretable coal flotation intelligent dosing method, which has the following beneficial effects compared with existing methods and technologies:

[0046] The application is based on high-precision sensors and detection equipment to capture the flotation reagent data set, and the input characteristics of feed concentration, flow rate, tailings ash content and aeration amount are normalized; then the XGBoost reagent model of coal flotation is trained by using the characteristic data variables, and the XGBoost reagent prediction model is explained by combining SHAP, including selecting the main dependent variable based on the SHAP mean value, directly predicting the main dependent variable by using the XGBoost model, predicting the reagent ratio by using the dry coal slime content, in order to avoid the condition of insufficient reagent when less than the conventional range, and adding the segmented boundary value of the reagent amount by means of SHAP mean value calculation; the new segmented interpretable intelligent reagent method proposed in the application is superior to the existing manual operation and machine learning algorithm, the reagent prediction algorithm of the method is more scientific and stable, and the combustible recovery rate and clean coal yield of coal slime flotation production can be effectively ensured, so as to improve the quality and efficiency of the coal preparation plant. BRIEF DESCRIPTION OF DRAWINGS

[0047] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or the prior art description will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0048] Figure 1 A specific flow chart of the segmented interpretable coal flotation intelligent reagent method implemented in embodiment 1 of the present application is shown in the figure.

[0049] Figure 2 A feature variable importance diagram for collector prediction in embodiment 1 of the present application is shown in the figure.

[0050] Figure 3 A feature variable importance diagram for frother prediction in embodiment 1 of the present application is shown in the figure.

[0051] Figure 4 A dependence_plot diagram of pulp concentration contribution to the XGBoost model of collector in embodiment 1 of the present application is shown in the figure. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be described clearly and completely in combination with the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0053] Embodiment 1

[0054] A segmented explainable coal flotation intelligent dosing method, Figure 1 A specific flowchart for implementing the embodiment 1 of the present application is shown, and the specific implementation flow of the present application is as follows:

[0055] S11: The concentration of flotation feed ρ, the feed flow rate Q and the aeration amount A in the industrial flotation process are collected by using the control module and the sensor, the ash content Ad of the flotation tailings is monitored by using the tailings visual ash content instrument, and the sample data are collected while the clean coal ash content, the collector addition amount and the frother addition amount are collected, wherein the clean coal ash content is tested in the laboratory, the collector and frother addition amounts are calibrated by using a measuring cylinder and a stopwatch, the clean coal ash content and the tailings ash content are used to calculate the clean coal yield and the combustible material recovery rate, and the calculation formula is as follows:

[0056]

[0057]

[0058] wherein γ i is the clean coal yield, η is the combustible material recovery rate, A w is the tailings ash content, A y is the raw coal ash content, and A j is the clean coal ash content.

[0059] S12: The data of abnormal conditions such as incomplete indicators or no production are removed, and the variable characteristics are normalized by using the python programming language, that is, each characteristic value is placed between [0, 1];

[0060] S13: The flotation data is divided into a training set (80%) and a validation set (20%), and the data distribution is shown in the following table:

[0061] Table 1: Coal flotation characteristic variable data statistics

[0062]

[0063] S21: The flotation dosing training set is divided into 10 parts, and 9 parts are used as training data and 1 part is used as test data in turn for training;

[0064] S22: Many hyperparameters are included in the XGBoost training, and during the training, the flotation reagent addition amount is randomly initialized, a random initial value is calculated according to y i = random(1b i ,nb i ), wherein 1b i and nb i are the upper and lower boundaries of the i-th optimization hyperparameter search variable, and one of them is returned after the training in the interval [1b i ,nb iThe hyperparameters in the XGBoost model are set, the fitness function value of the reagent addition is calculated according to the set fitness function, including setting an initial value, inputting a training set to train the model, and calculating the loss value and training error of the loss function of the model;

[0065] S23: Specifically, the optimal hyperparameters of the flotation reagent addition model obtained by the grid search method are n_estimators = 80, maximum tree depth max_depth = 7, penalty coefficient gamma = 0.21, and learning rate learning_rate = 0.01;

[0066] S24: The XGBoost reagent addition model optimized by inputting the training set is evaluated by RMSE, and the experimental results are shown in the following table:

[0067] Model RMSE Collector XGBoost 1.16 Frother XGBoost 0.94

[0068] S3: The SHAP mean value of the optimized XGBoost model is calculated, Figure 2 Feature variable importance diagram for collector prediction in the embodiment, Figure 3 Feature variable importance diagram for frother prediction in the embodiment;

[0069] As can be seen from the figure, in the embodiment, the four coal flotation characteristic process variables have the largest contribution to the collector, and the importance in the XGBoost reagent addition model is in the order of p > Q > A > Ad, and the corresponding SHAP values are SHAP_p = 4.82, SHAP_Q = 4.08, SHAP_Ad = 3.94, and SHAP_A = 0.89. From the SHAP values, it can be seen that the feed concentration, feed flow rate and feed ash content have a greater contribution to the XGBoost reagent addition model in the embodiment;

[0070] S4: In order to ensure that the reagent addition ratio in the coal flotation process is more stable, the reagent ratio parameter k is introduced in the embodiment;

[0071] From the results of S3, the SHAP mean value of the collector XGBoost reagent addition model (3.43) > the SHAP mean value of the frother XGBoost reagent addition model (3.01), so the collector addition amount is selected as the primary dependent variable, and the frother addition amount is selected as the secondary dependent variable;

[0072] From the training data set of the embodiment, the collector addition amount and the frother addition amount that meet the production requirements of the coal preparation plant are extracted, and the corresponding dry coal slime content m(g) and reagent ratio k are used to fit the relationship between the two using the method of S43 in the research content, and the relationship formula is k = (3.54m(g)) / 1000-0.21;

[0073] S5: In this embodiment, considering the unstable factors of industrial flotation feed conditions, when the variable characteristics are less than the conventional range, there may be a condition of insufficient reagent only relying on the machine learning model, therefore, the SHAP mean value is needed to segment the boundary value of the added reagent amount, Figure 4 The dependence_plot graph of the pulp concentration contribution to the collector XGBoost model in the embodiment is shown;

[0074] As can be seen from the graph, the change of concentration value is positively correlated with the output of the collector XGBoost model, in this case, the corresponding clean coal yield and combustible recovery rate curves are also plotted, the results show that when the pulp concentration ρ>60g / L, the XGBoost model needs to add reagent, which is basically stable, and meets the production requirements of the coal preparation plant;

[0075] Similarly, the lower limit of other variables in this embodiment is in turn the feed flow rate Q>51m 3 / h, tailings ash content Ad>65%, and air supply A>89m 3 / h;

[0076] In summary, this embodiment describes in detail the application of the segmented interpretable coal flotation intelligent reagent addition method proposed in the present application to the specific implementation of the production data of the coal preparation plant. For different coal preparation plants, the service life of the flotation machine and the flotation process are not the same, therefore, the embodiments described in the present application are illustrative, not limiting, and obviously, those skilled in the art can make various modifications and changes to the present application without departing from the spirit and scope of the present application, and any other embodiments derived from the technical solutions of the present application by those skilled in the art also fall within the protection scope of the present application.

Claims

1. A segmented and interpretable intelligent dosing method for coal flotation, characterized in that, Includes the following steps: S1: Capture characteristic data and flotation indicators of the coal flotation reagent addition process, and establish a coal slime flotation reagent addition prediction dataset; S2: Design and build an XGBoost flotation dosing regression prediction model and calculate the training error; S3: The SHAP algorithm is used to explain the contribution and importance of flotation process characteristic variables to the amount of flotation reagents added, and the most important characteristics for predicting the amount of reagents added are calculated respectively. S4: Select one flotation reagent as the main dependent variable based on the SHAP mean and add it according to the predicted value of the XGBoost model in step S2. The other reagent is used as the secondary dependent variable and the reagent ratio is predicted according to the dry coal slime content. S5: Use the SHAP mean to segment the boundary values ​​of the added drug dosage; S6: Combining the above steps S1 to S5, set the dosing rules; Step S3 above includes: S31: Taking the i-th group of training data as an example, each group corresponds to a floating feature variable. The prediction of the SHAP value of the i-th training sample can be expressed as: in in the formula It is the kth feature. Z represents the average prediction of the samples, S is the sample set excluding the k-th feature, and G represents the prediction model. S This represents a model trained using only set S; S32: Each flotation feature variable in the training data has a corresponding SHAP value. The mean SHAP value of the samples is selected to represent the importance of the feature for predicting the amount of flotation reagent added, and finally the importance ranking of the flotation process features is obtained. S33: Select the variable feature with the largest SHAP mean as the most important feature for drug dosage prediction; S51: Based on the sorting of the flotation variable characteristics in step S2, draw the SHAP dependence_plot and the relationship graphs between yield, combustible gas recovery rate and flotation variables in sequence; S52: Find the concentration value ρ in the SHAP dependence_plot where the predicted dosage of the corresponding variable causes the flotation index to fail to meet the requirements of the coal preparation plant. l Flow value Q l Inflation volume A l and tailings ash Ad l The minimum dosage segment point is set as ρ. l Q l A l and Ad l The corresponding collector dosage is the minimum dosage Y. b .

2. The segmented, interpretable intelligent dosing method for coal flotation as described in claim 1, characterized in that, Step S1 above includes: S11: Use instruments to capture characteristic variables of the coal flotation reagent addition process, and use combustible recovery rate and clean coal yield as indicators to mark the indicator parameters when the combustible recovery rate and clean coal yield meet the requirements of the concentrator. S12: Remove outlier data and use a normalization method to place the values ​​of each variable's features within the interval [0,1] based on the maximum and minimum values ​​of individual variables. The normalization formula is as follows: v(x)=[a(x)-mina(x)] / [maxa(x)-mina(x)] Where a(x) is a variable characteristic of the flotation process or tailings ash content, and v(x) is a model input characteristic; S13: Divide the data into training and validation sets in an 80:20 ratio to establish a coal slime flotation reagent prediction dataset.

3. The segmented, interpretable intelligent dosing method for coal flotation as described in claim 2, characterized in that, In step S11 above, the flotation feed concentration ρ, feed flow rate Q, and aeration volume A during the industrial flotation process are collected using a control module and sensors, and the ash content Ad of the flotation tailings is monitored using a tailings visual ash analyzer as four characteristic variables. When collecting sample data, the ash content of the clean coal in the flotation machine, the amount of collector added, and the amount of frother added are also collected simultaneously.

4. The segmented, interpretable intelligent dosing method for coal flotation as described in claim 1, characterized in that, Step S2 above includes: S21: The flotation dosing prediction dataset is read in using the 10-fold cross-validation method; S22: The XGBoost model in step S2 is a number ensemble model, internally composed of multiple regression trees. To optimize the loss function, the objective function for further training is: in In the formula g i h is the first derivative of the loss function. i Let be the second derivative of the loss function, f represent the flotation dosage that minimizes the loss function, Ω be the regularization term, γ be the penalty parameter for the leaf nodes, and T be the number of leaf nodes. S23: Grid search optimization of hyperparameters to be optimized in XGBoost, including the number of weak agent learners n_estimators, the maximum tree depth max_depth, the penalty coefficient gamma, and the learning rate learning_rate; S24: Encode a set of flotation characteristic variables and calculate the flotation reagent dosage, optimize the root mean square error (RMSE) of the loss function, and continue to optimize the dosing model based on the direction of decreasing loss function. The formula for calculating RMSE is as follows: Where Y obj Y represents the actual amount of flotation reagent added. model is the amount of flotation reagent to be added as predicted by the model, and n is the number of variable samples.

5. The segmented, interpretable intelligent dosing method for coal flotation as described in claim 1, characterized in that, Step S4 above includes: S41: Train the flotation dosing model using the method in step S2, and then calculate the average SHAP (collector) S using the method in S3. b and SHAP (foaming agent) q ; S42: If Sb>Sq, select the amount of collector added as the main variable and the amount of foaming agent added as the secondary variable. The main variable is directly predicted using the XGBoost model, and the secondary variable = main variable / k, where k is the agent ratio; otherwise, the amount of foaming agent added is the main variable and the amount of collector added is the secondary variable. S43: The optimal reagent ratio is determined by the real-time dry coal slime content of the flotation system, i.e., the amount of collector added / the amount of frother added. The formula for calculating the dry coal slime content is as follows: m(g)=Q(l)·ρ(l) Where Q is the feed slurry flow rate and ρ is the feed slurry concentration; then, under the premise that the flotation index meets the production requirements of the coal preparation plant, the relationship between the flotation reagent ratio and the dry coal slime content is fitted by SPSS in combination with the flotation experiment as k=f(m(g)), where k is the reagent ratio.

6. The segmented, interpretable intelligent dosing method for coal flotation as described in claim 1, characterized in that, Based on the above steps S1 to S5, set the dosing rules: Read the float variables ρ, Q, A, and Ad; Then calculate the dosage of the collector Y as Y model And the dosage of foaming agent P is Y model / k; if concentration value ρ < ρ l or traffic value Q l or tailings ash Ad <Ad l or inflation volume A l ;​​ then Y=Y b ,andP=Y b / k; else Y=Y model andP=Y model / k。

Images