Coal slime flotation control method based on ensemble learning and distributed hybrid bayesian network

Abstract

The application discloses a slime flotation control method based on ensemble learning and distributed hybrid Bayesian network, and comprises the following steps: step one, offline modeling of the distributed hybrid Bayesian network based on the ensemble learning; S11, dividing the slime flotation process and determining Bayesian network nodes; S12, determining N corresponding hybrid Bayesian network structures; S13, determining parameters of the N hybrid Bayesian network by using a maximum likelihood estimation method; S14, completing the distributed hybrid Bayesian network modeling based on the ensemble learning; step two, controlling online safe operation of the slime flotation process; S21, judging whether an abnormal working condition occurs based on current working condition data information, and respectively making safe control decisions by using N hybrid Bayesian network control models; S22, integrating the N decision results to make a final safe control strategy of the current working condition; and S23, re-executing S21 when the abnormal working condition is not eliminated. The method can make more reasonable and reliable safe control decisions.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of safe operation control of process industry, and particularly relates to a slime flotation control method based on integrated learning and distributed hybrid Bayesian network. BACKGROUND

[0002] Slime flotation is an important coal processing technology, which combines heavy medium coal preparation technology and flotation technology, and efficiently separates impurities from coal through physical and chemical methods, improves the grade and quality of coal, removes harmful elements, improves combustion efficiency, and effectively reduces environmental pollution. At the same time, the slime flotation process can also recycle coal gangue and other valuable ores, realizing sustainable utilization of resources. However, due to the interference existing in the actual production environment and the uncertainty of raw coal quality, the product coal quality of the slime flotation process may be reduced, and even abnormal working conditions may occur. Therefore, it is necessary to design a timely and reliable safe operation control and product quality control method for the slime flotation process.

[0003] As an important probabilistic graphical model, Bayesian network is widely used in fault diagnosis, reliability evaluation and prediction. In the production environment of the slime flotation process, there are many uncertain factors, such as sensor errors and raw coal fluctuations, which bring certain difficulties to the operators when making control decisions. In order to effectively solve this problem, Bayesian network is widely used in modeling and processing these uncertainties. Through probabilistic reasoning, Bayesian network can accurately estimate and infer uncertain factors, so as to help operators make more reasonable control decisions. In addition, Bayesian network can effectively combine expert knowledge and data information, use directed edges to represent the dependence relationship between process variables in the slime flotation process, use conditional probability to describe the strength of the dependence relationship, and graphically represent the interaction between variables, improving the understandability of control decisions. Therefore, it is necessary to provide a method for safe operation and product quality control of the slime flotation process using Bayesian network. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a slime flotation control method based on integrated learning and distributed hybrid Bayesian network. The method has simple implementation process and low implementation cost, can effectively combine the integrated learning strategy to significantly improve the generalization ability of the distributed hybrid Bayesian network, and can establish multiple control models and weight and fuse the decision results of each model to make more reasonable and reliable safe control decisions.

[0005] In order to achieve the above object, the present application provides a slime flotation control method based on integrated learning and distributed hybrid Bayesian network, comprising a slime flotation industrial system, wherein the slime flotation industrial system comprises a primary sizing screen, a secondary sizing screen, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a medium removal screen one, a medium removal screen two, a qualified medium barrel, a circulating pump, a slime barrel, an arc screen, a thickener, an ore pulp pretreater, a flotation cell and a controller.

[0006] The primary sizing screen is used to separate raw coal into coarse coal and clean coal by screening; the feed inlet of the secondary sizing screen is connected with the top discharge outlet of the primary sizing screen, which is used to remove slime impurities in the coal powder by screening; the feed end of the conveyor is connected with the top discharge outlet of the secondary sizing screen, which is used to output the screened coal powder to the feed inlet of the mixing barrel; the feed inlet of the mixing barrel is connected with the discharge outlet of the conveyor, which is used to mix the clean coal with the dense medium suspension to form a mixture; the inlet end of the pressure pump is connected with the discharge outlet of the mixing barrel through a pipeline, which is used to output the mixture to the dense medium cyclone; the feed inlet of the dense medium cyclone is connected with the outlet end of the pressure pump, which is used to separate low-density material and high-density material by using the density difference of the material to obtain overflow and underflow products; the feed inlet of the first desliming screen is connected with the high-density discharge outlet of the dense medium cyclone, which is used to perform dewatering and medium removal on the high-density material, and the high-density discharge outlet thereof discharges gangue; the feed inlet of the second desliming screen is connected with the low-density discharge outlet of the dense medium cyclone, which is used to perform dewatering and medium removal on the low-density material, and the high-density discharge outlet thereof discharges clean coal; the feed inlet of the qualified medium barrel is connected with the low-density discharge outlet of the first desliming screen, the low-density discharge outlet of the second desliming screen, the magnetic medium filling pipeline and the circulating water filling pipeline, respectively, which is used to complete the preparation of the dense medium suspension; the inlet end of the circulating pump is connected with the discharge outlet of the qualified medium barrel through a pipeline, and the outlet end thereof is connected with the feed inlet of the mixing barrel through a pipeline, which is used to deliver the dense medium suspension to the mixing barrel; the feed inlet of the slime barrel is connected with the high-density discharge outlet of the second desliming screen, which is used to preliminarily process the clean coal; the feed inlet of the arc screen is connected with the discharge outlet of the slime barrel, which is used to control the particle size of the subsequent floated slime by screening and separation, and the top discharge outlet thereof discharges coarse particle clean coal; the feed inlet of the thickener is connected with the bottom discharge outlet of the arc screen, which is used to increase the concentration of the slime and reduce the moisture content of the slime, so that the pulp density is kept within the range of stable operation; the feed inlet of the pulp pretreater is connected with the circulating water filling pipeline, the flotation reagent filling pipeline and the high-density discharge outlet of the thickener, respectively, which is used to fully mix the flotation reagent with the pulp to make the final preparation for the slime froth flotation; the feed inlet of the flotation tank is connected with the air filling pipeline and the discharge outlet of the pulp pretreater, respectively, which is used to obtain the final clean coal in the overflow product of the flotation tank by using the difference in the hydrophilicity of different solid particles in the pulp, and the top discharge outlet thereof outputs clean coal, and the bottom discharge outlet thereof outputs impurities; the controller is connected with the primary sizing screen, the secondary sizing screen, the conveyor, the mixing barrel, the pressure pump, the dense medium cyclone, the first desliming screen, the second desliming screen, the qualified medium barrel, the circulating pump, the slime barrel, the arc screen, the thickener, the pulp pretreater and the flotation tank, respectively, which is used to control the components;

[0007] The slime flotation control method comprises the following steps:

[0008] Step one: offline modeling of distributed hybrid Bayesian network based on ensemble learning;

[0009] S11: divide the slime flotation process and determine the Bayesian network nodes;

[0010] According to the actual situation of the slime flotation industrial system, combined with operation experience, the slime flotation process is divided into multiple local modules, and the control variables, process variables and target variables of each local module are determined, and then each variable is mapped to the corresponding Bayesian network node, and the level state of each node is further divided;

[0011] S12: stratified sampling of the training set N times to obtain sub-training sets D1, D2...D N , determine the N corresponding hybrid Bayesian network structure;

[0012] Collect multiple sets of abnormal working condition data information using a plurality of sensors installed at a plurality of sampling points in the slime flotation industrial system; use the first Bayesian network model to formulate control decisions for multiple sets of abnormal working conditions, and calculate the control decision reasoning accuracy according to formula (1); when the control decision reasoning accuracy reaches 98%, it is considered that the Bayesian network model can provide accurate safety control decisions, and there is no need to integrate subsequent Bayesian network models, otherwise continue to integrate the decision results of the second Bayesian network model, further judge the reasoning accuracy of the integrated control decisions, and so on, until the control decision reasoning accuracy of the integrated Nth Bayesian network model is greater than or equal to 98%, the number of models N in the ensemble learning strategy is determined;

[0013]

[0014] In the formula, TP represents effective safety control decisions that can eliminate abnormal working conditions; TN represents invalid safety control decisions that cannot eliminate abnormal working conditions;

[0015] S13: determine the parameters of the N hybrid Bayesian networks using the maximum likelihood estimation method;

[0016] S14: complete the modeling of distributed hybrid Bayesian networks based on ensemble learning;

[0017] Step two: control of the slime flotation process online safety operation based on distributed hybrid Bayesian network modeling;

[0018] S21: collect current working condition data information using a plurality of sensors installed at a plurality of sampling points in the slime flotation industrial system, and determine whether an abnormal working condition occurs based on the current working condition data information; if S 38.5%, it is determined that an abnormal working condition occurs, and N hybrid Bayesian network control models established in the offline modeling process are used to formulate safety control decisions respectively;

[0019] S22: integrate N decision results to make the final safety control strategy for the current working condition;

[0020] S22-1: calculate the BIC score between the current working condition data and the i (i£N)th hybrid Bayesian network control model,

[0021] to evaluate the fitting degree of the current working condition and each model;

[0022] S22-2: assign a corresponding weight to each control decision according to the fitting degree of each model;

[0023] S22-3: weight and integrate all decisions according to formula (2) to form the final safety control decision Adjustment for the current working condition

[0024]

[0025] wherein, a i represents the weight of the i th model to provide control decisions, wherein, represents the BIC score function of the i th model and the current working condition data, which is used to represent the fitting degree of the i th model and the current working condition data; Adjustment i represents the safety control decision made by the i th model;

[0026] S23: predict the overflow ash content after adjustment. If S < 8.5%, it proves that the abnormal working condition is eliminated, the current safety control decision is executed, and the current normal working condition mode is maintained. Otherwise, S21 is re-executed, and the current working condition data information Bayesian network inference process is continued.

[0027] Aiming at the problems of product quality decline and abnormal working conditions caused by working condition fluctuation in the actual production of coal slime flotation process, the present application proposes a coal slime flotation process safety control method based on integrated learning and distributed hybrid Bayesian network. Bayesian network is widely used in the safe operation control of complex industrial processes due to its strong uncertainty reasoning ability. However, the fluctuation of raw materials and the change of operating parameters in the running environment of industrial process may lead to insufficient generalization ability of Bayesian network control model, which cannot provide accurate safety operation control decision. At the same time, coal slime flotation is a complex industrial process with high degree of scale and multi-variable interaction, and directly establishing a Bayesian network model for the whole process may reduce the accuracy of the model. In order to solve this challenge, the present application provides an effective solution through distributed modeling. By decomposing the coal slime flotation process into multiple local modules, each local module is modeled according to its own needs and characteristics, and each local module is adjusted and optimized according to the actual situation, reducing the complexity of calculation and reasoning. When abnormal conditions occur, these models generate corresponding safety control decisions. Then, the BIC score function is used to measure the fitting degree of the current working condition and each model, and the weighted fusion of the decision results of each model is used to form a final safety control decision. The widely used discrete Bayesian network model needs to discretize the data during modeling, resulting in a large amount of information loss. Therefore, in the reasoning control decision process, this model can only provide a rough adjustment direction, and cannot provide accurate adjustment strategy for the operator, nor can it further improve the product quality of coal slime flotation process. In order to solve this problem, Bayesian network improves the reasoning accuracy by introducing continuous nodes.

[0028] The method has simple implementation process and low implementation cost, which can effectively combine the integrated learning strategy to significantly improve the generalization ability of the distributed hybrid Bayesian network, and can establish more reasonable and reliable safety control decisions by establishing multiple control models and weighted fusion of the decision results of each model. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 is the flowchart of the present application;

[0030] Figure 2 is the control decision reasoning accuracy corresponding to different model quantities of the present application;

[0031] Figure 3 is the three hybrid Bayesian network structures of the present application, wherein left: structure 1; middle: structure 2; right: structure 3;

[0032] Figure 4 is the conditional probability distribution of part of the continuous nodes in model 1 in the present application;

[0033] Figure 5is the contribution graph of four abnormal working conditions of the application;

[0034] Figure 6 is the overflow ash change curve of working condition 1 of the application;

[0035] Figure 7 is the structural schematic diagram of the coal slime flotation industrial system in the application. DETAILED DESCRIPTION

[0036] The application will be further described below with reference to the drawings.

[0037] As Figure 7As shown, the present application provides a slime flotation control method based on integrated learning and distributed hybrid Bayesian network, including a slime flotation industrial system, the slime flotation industrial system includes a primary classifier, a secondary classifier, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a medium screen one, a medium screen two, a qualified medium barrel, a circulating pump, a slime barrel, an arc screen, a thickener, a slurry pretreater, a flotation tank and a controller; the primary classifier is used for separating raw coal into coarse coal and clean coal by screening; the feed inlet of the secondary classifier is connected with the top discharge outlet of the primary classifier, and is used for removing slime impurities in the coal powder by screening; the feed end of the conveyor is connected with the top discharge outlet of the secondary classifier, and is used for outputting the screened coal powder to the feed inlet of the mixing barrel; the feed inlet of the mixing barrel is connected with the discharge outlet of the conveyor, and is used for mixing the clean coal and the heavy medium suspension to form a mixed material; the inlet end of the pressure pump is connected with the discharge outlet of the mixing barrel through a pipeline, and is used for outputting the mixed material to the heavy medium cyclone; the feed inlet of the heavy medium cyclone is connected with the outlet end of the pressure pump, and is used for separating low-density material and high-density material to obtain overflow and underflow products by utilizing the density difference of the materials; the feed inlet of the medium screen one is connected with the high-density discharge outlet of the heavy medium cyclone, and is used for carrying out dewatering and medium removing operation on the high-density material, and the high-density discharge outlet of the medium screen one discharges gangue; the feed inlet of the medium screen two is connected with the low-density discharge outlet of the heavy medium cyclone, and is used for carrying out dewatering and medium removing operation on the low-density material, and the high-density discharge outlet of the medium screen two discharges clean coal; the feed inlet of the qualified medium barrel is connected with the low-density discharge outlet of the medium screen one, the low-density discharge outlet of the medium screen two, a magnetic medium filling pipeline and a circulating water filling pipeline respectively, and is used for completing the adjustment of the heavy medium suspension; the inlet end of the circulating pump is connected with the discharge outlet of the qualified medium barrel through a pipeline, and the outlet end of the circulating pump is connected with the feed inlet of the mixing barrel through a pipeline, and is used for conveying the heavy medium suspension to the mixing barrel; the feed inlet of the slime barrel is connected with the high-density discharge outlet of the medium screen two, and is used for preliminarily processing the clean coal; the feed inlet of the arc screen is connected with the discharge outlet of the slime barrel, and is used for controlling the particle size of the slime for subsequent flotation by screening and separation, and the top discharge outlet of the arc screen discharges coarse particle clean coal; the feed inlet of the thickener is connected with the bottom discharge outlet of the arc screen, and is used for improving the concentration of the slime and reducing the moisture content of the slime, so that the slurry density is kept in the range of stable operation process; the feed inlet of the slurry pretreater is connected with a circulating water filling pipeline, a flotation reagent filling pipeline and the high-density discharge outlet of the thickener respectively, and is used for fully mixing the flotation reagent with the slurry, and making the final preparation for the slime froth flotation; the feed inlet of the flotation tank is connected with an air filling pipeline and the discharge outlet of the slurry pretreater respectively, and is used for obtaining the final clean coal in the overflow product of the flotation tank by utilizing the difference of the hydrophilicity of different solid particles in the slurry, and the top discharge outlet of the flotation tank outputs the clean coal, and the bottom discharge outlet of the flotation tank outputs impurities.The controller is connected with the primary classifying screen, the secondary classifying screen, the conveyor, the mixing barrel, the pressure pump, the heavy medium cyclone, the medium removal screen one, the medium removal screen two, the qualified medium barrel, the circulating pump, the slime barrel, the arc screen, the thickener, the ore pulp pretreater and the flotation tank respectively, and is used for controlling each component;

[0038] As shown in Figure 1 To improve the generalization ability of the distributed hybrid Bayesian network, an integrated learning strategy is combined, multiple control models are established, the decision results of each model are weighted and fused, and a more reliable safety control decision is made.

[0039] Step one: offline modeling of the distributed hybrid Bayesian network based on integrated learning;

[0040] S11: divide the coal slime flotation process and determine the Bayesian network nodes;

[0041] According to the actual situation of the coal slime flotation industrial system and combined with operation experience, the coal slime flotation process is divided into multiple local modules, the control variables, process variables and target variables of each local module are determined, and then each variable is mapped to the corresponding Bayesian network node, and the level state of each node is further divided;

[0042] S12: N times stratified sampling is performed on the training set to obtain sub-training sets D1, D2...D N , and N corresponding hybrid Bayesian network structures are determined;

[0043] With the fluctuation of the working condition of the coal slime flotation process, the Bayesian network structure may change, in order to be able to construct a more adaptive Bayesian network control model for the current working condition and make a more accurate control decision, the application adopts a method based on score search to determine the Bayesian network structure.

[0044] Before the establishment of the hybrid Bayesian network, the number N of models in the ensemble learning strategy needs to be determined; a plurality of sensors installed at a plurality of sampling points in the coal slime flotation industrial system are used to collect a plurality of sets of abnormal working condition data information, as an optimization, 500 sets of abnormal working condition data information are collected, and then whether the control decision reasoning accuracy meets the predetermined threshold is evaluated to determine the number N of models required; specifically, a plurality of sets of abnormal working condition data are formulated into control decisions by using a first Bayesian network model, and the control decision reasoning accuracy is calculated according to formula (1); when the control decision reasoning accuracy reaches 98%, it is considered that the Bayesian network model can provide sufficiently accurate safety control decisions, and there is no need to integrate subsequent Bayesian network models, otherwise the decision results of a second Bayesian network model are integrated, and the integrated control decision reasoning accuracy is further judged, and so on, until the control decision reasoning accuracy formulated by the Nth integrated Bayesian network model is greater than or equal to 98%, and the number N of models in the ensemble learning strategy is determined;

[0045]

[0046] In the formula, TP represents an effective safety control decision that can eliminate abnormal working conditions; TN represents an invalid safety control decision that cannot eliminate abnormal working conditions;

[0047] S13: Determine the parameters of the N hybrid Bayesian networks by using the maximum likelihood estimation method;

[0048] S14: Complete the distributed hybrid Bayesian network modeling based on ensemble learning;

[0049] Step two: control the online safe operation of the coal slime flotation process based on the distributed hybrid Bayesian network modeling;

[0050] S21: Collect current working condition data information by using a plurality of sensors installed at a plurality of sampling points in the coal slime flotation industrial system, and determine whether an abnormal working condition occurs based on the current working condition data information; if S 38.5%, it is determined that an abnormal working condition occurs, and N hybrid Bayesian network control models established in the offline modeling process are used to formulate safety control decisions respectively;

[0051] S22: Integrate N decision results, and use the weight method as an integration strategy of the control decisions to formulate the final safety control strategy of the current working condition;

[0052] S22-1: Calculate the BIC score between the current working condition data and the i(i￡N)th hybrid Bayesian network control model,

[0053] to evaluate the fitting degree of the current working condition and each model;

[0054] S22-2: According to the fitting degree of each model, each control decision is assigned a corresponding weight;

[0055] S22-3: All decisions are fused by weighting according to formula (2) to form the final safety control decision Adjustment for the current working condition

[0056]

[0057] wherein, α i represents the weight of the i-th model to provide control decisions, wherein, represents the BIC score function of the i-th model with the current working condition data, used to indicate the fitting degree of the i-th model with the current working condition data; Adjustment i represents the safety control decision made by the i-th model;

[0058] S23: Predict the overflow ash content after adjustment. If S < 8.5%, it proves that the abnormal working condition is eliminated, the current safety control decision is executed, and the current normal working condition mode is maintained. Otherwise, S21 is re-executed, and the current working condition data information Bayesian network inference process is continued to be used.

[0059] Embodiment:

[0060] Coal resources are the backbone of China's economic and social development, providing key impetus and support for sustainable and rapid economic growth. China has abundant coal reserves, with production ranking first in the world and occupying an important proportion in the energy structure. However, raw coal contains a large amount of impurities, and direct combustion releases harmful gases, causing serious damage to the ecological environment. Therefore, in order to improve the utilization efficiency of coal and reduce the harm of combustion to the environment, coal preparation technology is a crucial link in the process of coal utilization. As an important coal preparation technology, slime flotation process is widely used due to its fine particle size and high separation precision. The slime flotation process first uses the Archimedes principle to wash and select the coarse coal with larger particle size, and then recovers the fine coal after washing and selection for froth flotation. Finally, the overflow from the flotation tank is scraped out by the scraper device, and the product coal is dried after dehydration. The gangue that cannot float is left in the flotation tank bottom as tailings. However, the actual slime flotation production environment is poor, and the properties of raw coal change frequently, which may cause the coal preparation process to deviate from the set optimal operating point, leading to a decrease in product coal quality, even abnormal working conditions, not only affecting production performance and causing huge economic losses, but also posing a serious threat to the service life of equipment and the personal safety of operators. When the slime flotation process deviates from the optimal operating point, resulting in a decrease in product quality or even abnormal working conditions, operators mainly rely on their knowledge and experience to determine the cause of the deviation from the optimal operating point, and then develop appropriate safety control decisions and quality control decisions. However, this knowledge-driven method is subjective and the decisions are not precise enough, which cannot ensure that the slime flotation process can return to the optimal operating point, and even may deepen the abnormal working conditions. In addition, existing control methods for the slime flotation process mainly focus on model-driven and data-driven methods. The slime flotation process includes raw coal treatment, heavy medium separation, and froth flotation, which are complex and difficult to establish an accurate mechanism model. Therefore, it is difficult for model-driven methods to effectively develop appropriate control decisions. With the increasing maturity of big data technology, the slime flotation process has accumulated a large amount of working condition data in actual operation, which can be used to establish a control model. However, the output of data-driven methods is less interpretable, which is not conducive to the understanding of operators. Therefore, it is necessary to design a more intelligent safety and quality integrated control method for the slime flotation process.

[0061] The data of the present example is from a coal slime flotation process full-process simulation platform (registration number: 2020SR1160052), in order to improve the explainability of control decision, further considering the process variables involved in the coal slime flotation process, through a more clear graph network structure, explaining the influence relationship between the control variables and the target variables. Based on the above established modeling method, in order to provide more timely and accurate safety control decision, first of all, the number N of models in the ensemble learning strategy needs to be determined. In the present embodiment, 500 groups of abnormal working condition data are selected, and the safety control decision is made by using the established hybrid Bayesian network. The reasoning accuracy of the control decision corresponding to different number of models in the ensemble learning is as shown in Figure 2 The results show that when the number N of Bayesian network models is 3, the reasoning accuracy of the control decision reaches 98.7%, which meets the set accuracy threshold requirement, and with the continuous increase of the number of models, the reasoning accuracy of the control decision is not greatly improved. Therefore, the number N of models in the ensemble learning strategy is set to 3.

[0062] The structure of the three hybrid Bayesian network models established by the present application is as shown in Figure 3 The physical meaning of each node is as shown in Table 1, the conditional probability distribution table of the discrete node is as shown in Table 2, and the conditional probability distribution of the continuous node is as shown in Figure 4 The initial value of the control variable is a discrete node, according to the expert knowledge and the influence of each control variable on the target variable, it can be divided into five levels: 1, low level, 2, lower level, 3, middle level, 4, higher level, and 5, high level. In order to ensure the reasoning accuracy of the control decision, the control variable adjustment value, the process variable and the target variable are continuous nodes.

[0063] Table 1: Bayesian network nodes and physical meaning

[0064]

[0065]

[0066] Table 2: Conditional probability distribution table of nodes A, B, C, H, I, M, N and O in model 1

[0067]

[0068] Randomly list four kinds of abnormal working conditions that may occur, the corresponding control variable initial value and target variable are as shown in Table 3, and then the safety control decision is made for the four kinds of abnormal working conditions by using the online application strategy. First of all, the abnormal module is determined by using the contribution graph algorithm, and the contribution graph of the four kinds of abnormal working conditions is as shown in Figure 5The abnormal module is the flotation module, for example, in the working condition 1. Next, the BIC score of the current working condition data and the i (i≤N)th Bayesian network model is calculated, the fitting degree of the current working condition and each model is further obtained, and each model is assigned a corresponding weight of control decision according to the fitting degree. The corresponding weight of each model under the four abnormal working conditions is shown in Table 4.

[0069] Table 3: Four possible abnormal working conditions

[0070]

[0071] Table 4: The corresponding weight of each model under the four abnormal working conditions

[0072]

[0073] The initial value of the control variable of all modules and the overflow ash safety threshold S = 8.5% are taken as evidence information, and each model uses Bayesian network inference to obtain the corresponding safety control decision, that is, the control variable adjustment value of the abnormal module. Combined with the weight of each model in Table 4, the final safety control decision is made according to formula (7), and the result is shown in Table 5. For example, in the working condition 1, the physical meaning of the control decision represents that the air supply of the flotation tank is increased by 0.067 m 3 / m 2 *h, the motor stirring speed is increased by 0.752 rad / min, and the reagent addition amount is increased by 1.889 kg / t. The same is true for other working conditions.

[0074] In order to ensure the safe and stable operation of the coal slime flotation process, the safety control decision obtained needs to be verified next, to predict whether the overflow ash after executing the decision meets the requirements. If S≥8.5%, the decision is not executed, the working condition information is updated, and the Bayesian network inference process is continued; otherwise, the decision is executed. The initial value of the control variable of all modules and the control variable adjustment value of the abnormal module are taken as evidence information, and the adjusted overflow ash is obtained by inference, as shown in Table 6. The results show that the overflow ash of all abnormal working conditions is reduced below the safety threshold, indicating that the method can effectively eliminate the abnormal working conditions in the coal slime flotation process.

[0075] Table 5: Safety control decision

[0076]

[0077] Table 6: Overflow ash after executing safety control decision

[0078]

[0079] Take working condition 1 as an example, the obtained safety control decision is applied to the slime flotation process simulation experiment platform. The overflow ash safety threshold is set to S = 8.5%, and the sampling interval is 1 second. The overflow ash change curve after executing the control decision is as shown in Figure 6 The abnormal working condition data information collected by the operator at the 943th moment is used as evidence to obtain a safety control decision by using Bayesian network inference, and the slime flotation process returns to normal working condition at the 984th sampling point after executing the control decision. Figure 6 The results show that the method proposed in the application can effectively eliminate the abnormal working condition of the slime flotation process and provide reliable safety control decisions. At the same time, Figure 1 The results show that compared with a single Bayesian network model, the control decision made by the method proposed in the application has higher accuracy.

[0080] The application proposes a slime flotation process safety operation control method based on integrated learning and distributed hybrid Bayesian network. The method uses the Bagging strategy in integrated learning to construct a group of hybrid Bayesian network models by stratified sampling of the training set. When an abnormal working condition occurs, these models generate corresponding safety control decisions respectively. Then, the BIC score function is used to measure the fitting degree of the current working condition and each model, and the weight method is used to weight and fuse the decision results of each model, thereby forming a final safety control decision. This process significantly improves the accuracy and robustness of the control decision. Finally, the effectiveness of the proposed method is verified through a slime flotation simulation experiment platform.

Claims

1. A coal slime flotation control method based on ensemble learning and distributed hybrid Bayesian networks, comprising a coal slime flotation industrial system, wherein the coal slime flotation industrial system includes a main classifying screen, a secondary classifying screen, a conveyor, a mixing tank, a pressure pump, a heavy medium cyclone separator, a first desliming screen, a second desliming screen, a qualified medium tank, a circulating pump, a coal slime tank, an arc screen, a thickener, a slurry preprocessor, a flotation cell, and a controller; The main grading screen is used to separate raw coal into coarse coal and clean coal by screening. The feed inlet of the secondary grading screen is connected to the top discharge outlet of the main grading screen, and is used to remove coal slurry impurities from the coal powder by screening. The feed end of the conveyor is connected to the top discharge outlet of the secondary grading screen, and is used to output the screened coal powder to the feed inlet of the mixing tank. The feed inlet of the mixing tank is connected to the discharge outlet of the conveyor, and is used to mix the clean coal with the heavy medium suspension to form a mixture. The inlet of the pressure pump is connected to the discharge outlet of the mixing tank through a pipeline, and is used to output the mixture to the heavy medium hydrocyclone. The feed inlet of the heavy medium hydrocyclone is connected to the outlet of the pressure pump, and is used to utilize the density difference of the materials to... Low-density and high-density materials are separated to obtain overflow and underflow products. The inlet of the first desliming screen is connected to the high-density outlet of the heavy medium cyclone separator for dewatering and desliming the high-density material, with gangue discharged from its high-density outlet. The inlet of the second desliming screen is connected to the low-density outlet of the heavy medium cyclone separator for dewatering and desliming the low-density material, with clean coal discharged from its high-density outlet. The inlet of the qualified medium tank is connected to the low-density outlets of the first and second desliming screens, the magnetic medium injection pipeline, and the circulating water injection pipeline, respectively, for preparing the heavy medium suspension. The inlet of the circulating pump is connected to the outlet of the qualified medium tank via a pipeline. The outlet of the slurry pre-processor is connected to the inlet of the mixing tank via a pipeline, for conveying the heavy medium suspension into the mixing tank; the inlet of the coal slime tank is connected to the high-density outlet of the desliming screen, for preliminary treatment of the clean coal; the inlet of the arc screen is connected to the outlet of the coal slime tank, for controlling the particle size of the coal slime in subsequent flotation by screening and separation, with coarse clean coal discharged from its top outlet; the inlet of the thickener is connected to the bottom outlet of the arc screen, for increasing the concentration of the coal slime and reducing its moisture content, so that the slurry density is maintained within the range of stable process operation; the inlet of the slurry pre-processor is connected to the circulating water injection pipeline, the flotation reagent injection pipeline, and the thickener respectively. The high-density discharge port of the machine is connected to ensure thorough mixing of flotation reagents and slurry, preparing for the final stage of coal slime froth flotation. The inlet of the flotation cell is connected to the air injection pipeline and the outlet of the slurry preprocessor, respectively, to obtain the final clean coal from the overflow product of the flotation cell by utilizing the different hydrophilicities of different solid particles in the slurry. The top outlet outputs clean coal, and the bottom outlet outputs impurities. The controller is connected to the main classifying screen, secondary classifying screen, conveyor, mixing tank, pressure pump, heavy medium hydrocyclone, desliming screen one, desliming screen two, qualified medium tank, circulating pump, coal slime tank, arc screen, thickener, slurry preprocessor, and flotation cell, respectively, for controlling each component. Its features are, The coal slime flotation control method includes the following steps: Step 1: Offline modeling using distributed hybrid Bayesian networks based on ensemble learning; S11: Divide the coal slime flotation process and determine the Bayesian network nodes; Based on the actual situation of the coal slime flotation industrial system and combined with operational experience, the coal slime flotation process is divided into multiple local modules, and the control variables, process variables and target variables of each local module are determined. Then, each variable is mapped to the corresponding Bayesian network node, and the level state of each node is further divided. S12: Process the training set Sub-stratified sampling to obtain a sub-training set ,Sure A corresponding hybrid Bayesian network structure; Multiple sets of abnormal operating condition data are collected using several sensors installed at several sampling points in the coal slime flotation industrial system. A first Bayesian network model is used to formulate control decisions for these abnormal operating conditions, and the inference accuracy of the control decisions is calculated according to formula (1). When the inference accuracy reaches 98%, it is considered that the Bayesian network model can provide sufficiently accurate safety control decisions, and there is no need to integrate subsequent Bayesian network models. Otherwise, the decision results of the second Bayesian network model are integrated, and the inference accuracy of the integrated control decisions is further judged. This process is repeated until the first integrated Bayesian network model is reached. When the control decision reasoning accuracy of a Bayesian network model is greater than or equal to 98%, the number of models in the ensemble learning strategy is determined. ; （1）； In the formula, This represents effective safety control decisions that can eliminate abnormal operating conditions; This represents ineffective safety control decisions that cannot eliminate abnormal operating conditions; S13: Determine using the maximum likelihood estimation method Parameters of a hybrid Bayesian network; S14: Complete the modeling of a distributed hybrid Bayesian network based on ensemble learning; Step 2: Controlling the online safe operation of the coal slime flotation process based on distributed hybrid Bayesian network modeling; S21: Collect current operating condition data using several sensors installed at several sampling points in the coal slime flotation industrial system, and determine whether an abnormal operating condition has occurred based on the current operating condition data; if If an abnormal operating condition is detected, the data established during the offline modeling process will be used. Each hybrid Bayesian network control model formulates security control decisions; S22: Integration Based on the decision results, formulate the final safety control strategy for the current operating conditions; S22-1: Calculate the current operating condition data and the... The BIC score among the hybrid Bayesian network control models is used to evaluate the fit between the current operating condition and each model. S22-2: Assign corresponding weights to each control decision based on the fit of each model; S22-3: Based on formula (2), all decisions are weighted and integrated to form the final safety control decision for the current working condition. ; （2）； In the formula, Representing the Each model provides weights for control decisions. ,in, Representing the The BIC scoring function of the model and the current operating condition data is used to represent the BIC scoring function of the model. The degree of fit between the model and the current operating condition data; Representing the Security control decisions are formulated based on a model; S23: Predict the adjusted overflow ash content, if If the abnormal operating condition is proven to be eliminated, the current safety control decision is executed to maintain the current normal operating condition mode; otherwise, S21 is re-executed to continue the Bayesian network inference process using the current operating condition data.

Images