Coal slime flotation safe operation control method based on distributed dynamic bayesian network

Abstract

A slime flotation safe operation control method based on a distributed dynamic Bayesian network, the slime flotation process is divided into blocks, and a global quality index is determined; the variables related to the global quality index of each module are determined; the state of the variables in the global network and the state of the variables in the local network are determined; the global Bayesian network and the local Bayesian network are established; the global Bayesian network is converted into a global dynamic Bayesian network; it is judged whether there is a loop structure in the local Bayesian network, and if there is a loop structure, the weakest causal relationship is determined; the parameters of the local dynamic Bayesian network and the global dynamic Bayesian network are learned; the establishment of the distributed dynamic Bayesian network model is completed; the inference of the control decision is carried out by using the online abnormal data; the control decision is inferred and implemented; it is judged whether the abnormal working condition is removed, and if not, the adjustment is continued. The method can effectively ensure that the product quality meets the standard, and is also conducive to ensuring the service life of the production equipment.

Description

Technical Field

[0001] This invention belongs to the field of industrial safety operation control technology, specifically relating to a method for safe operation control of coal slime flotation based on distributed dynamic Bayesian networks. Background Technology

[0002] Mineral processing is a crucial production link in mineral resource processing, directly affecting the utilization and recovery rates of mineral resources. Flotation is the most important mineral separation method in mineral processing, widely used in hydrometallurgy, coal, and chemical industries. The coal slime flotation process is complex, with strong nonlinear and uncertain relationships between variables. During coal slime flotation, drastic fluctuations in working conditions and frequent changes in raw materials often lead to various abnormal operating conditions. These abnormal conditions directly affect product quality, shorten the service life of production equipment, and in severe cases, even threaten the personal safety of operators. In actual production, eliminating abnormal operating conditions is often done manually based on experience, but this method is highly subjective. To address the shortcomings of manual handling of abnormal operating conditions, a more intelligent safety control method is urgently needed.

[0003] Conventional control methods require the construction of a system mechanistic model. However, the coal preparation process is lengthy, complex, and involves nonlinear relationships between variables, making it impossible to establish an accurate mechanistic model. To improve control performance, it is necessary to integrate expert knowledge and data information to develop intelligent industrial process safety operation control methods. Conventional methods based on intelligent decision support systems have received considerable attention. Decision support systems assist decision-makers in analyzing data and making correct decisions through model analysis. Bayesian networks are the most representative graphical models, and due to their significant advantages in representing uncertainty, they have become an important method for constructing decision support systems.

[0004] Modern industrial processes are increasingly large-scale, making it impractical to simply place all process variables into a large Bayesian network for safe control. Furthermore, while this approach is simple and easy to understand, it is difficult to implement in terms of both time and space complexity. Therefore, it cannot effectively control the safe operation of coal slime flotation processes. Thus, there is an urgent need for a more precise control method to meet the control requirements of large-scale industrial applications. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a safe operation control method for coal slime flotation based on a distributed dynamic Bayesian network. This method is simple to implement and has low implementation cost. By establishing a distributed dynamic Bayesian network, it can achieve modular control of the coal slime flotation process, effectively ensuring product quality standards are met, and at the same time, it helps to ensure the service life of production equipment.

[0006] To achieve the above objectives, this invention provides a method for safe operation control of coal slime flotation based on a distributed dynamic Bayesian network, comprising the following steps:

[0007] Step 1: Divide the coal slime flotation process into three modules: raw coal processing module, heavy media coal preparation module, and thickening flotation module, and determine the coal slime ash content as a global quality indicator.

[0008] Step 2: Identify the variables related to the global quality metrics for each module;

[0009] The variables related to the raw coal processing module and the global quality indicators were determined to be the raw coal feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash content of the flotation cell.

[0010] The variables related to the overall quality indicators of the heavy medium coal preparation module are determined to be: coal inlet flow rate of the mixed medium tank, slurry density of the mixed medium tank, medium density of the hydrocyclone, inlet pressure of the hydrocyclone, ash content of the hydrocyclone overflow, density adjustment of the qualified medium tank, medium density of the qualified medium tank, and ash content of the flotation cell overflow.

[0011] The variables related to the global quality indicators of the thickening and flotation module were determined to be thickener underflow rate, thickener media density, slurry preprocessor media density, flotation cell stirring speed, and flotation cell overflow ash content.

[0012] Step 3: Determine the state of variables in the global network and the state of variables in the local network:

[0013] S31: It is determined that the main variables in the global network are feature variables and quality variables, and the quality variables are mainly in two states: normal and abnormal.

[0014] S32: It is determined that the variables in the local network are mainly process variables and quality variables, and the process variable states are mainly three states: normal, low outlier, and high outlier.

[0015] Step 4: Establish global Bayesian networks and local Bayesian networks;

[0016] A global Bayesian network is established based on the relationship between the raw coal processing module, the heavy media coal preparation module, the concentration flotation module, and the global quality indicators.

[0017] Based on the relationship between the raw coal processing module, the heavy media coal preparation module, and the concentration flotation module, a local Bayesian network is established among the three modules.

[0018] A local Bayesian network for the raw coal processing module is established based on the relationship between the raw coal silo feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash from the flotation cell.

[0019] A local Bayesian network for the heavy media coal preparation module is established based on the relationship between the coal flow rate in the mixing medium tank, the slurry density in the mixing medium tank, the medium density in the hydrocyclone, the inlet pressure of the hydrocyclone, the overflow ash content of the hydrocyclone, the density adjustment of the qualified medium tank, the medium density of the qualified medium tank, and the overflow ash content of the flotation cell.

[0020] A local Bayesian network for the thickening and flotation module is established based on the relationship between the underflow rate of the thickener, the density of the thickener medium, the density of the slurry preprocessor medium, the stirring speed of the flotation cell, and the ash content of the flotation cell overflow.

[0021] Step 5: Based on the correlation between the characteristic variables in the raw coal processing module, the characteristic variables in the heavy media coal preparation module, and the characteristic variables in the concentration flotation module, the global Bayesian network is transformed into a global dynamic Bayesian network;

[0022] Step 6: Determine if there is a loop structure in the local Bayesian network. If there is a loop structure, proceed to Step 7; otherwise, proceed to Step 8.

[0023] Step 7: Use the transfer entropy of formula (1) to determine the weakest causal relationship of the local loop-loop Bayesian network, and place the parent node in the weakest relationship in the previous time slice and the child node in the current time slice to transform this causal relationship into a local dynamic Bayesian network.

[0024]

[0025] In the formula, T(X) i+1 |X i ,pa G (X)) is the variable X to variable pa G The transit entropy of (X), X and pa G (X) represents the directed edge connecting the child node and the parent node, pa G (X) is the parent node of variable X, x i and pa G (X) j Representing X and pa G (X) in the data batches i and j, x i+1 This indicates that X represents data in the next batch in the future;

[0026] Step 8: Using the collected coal slime flotation process data, the parameters of the local dynamic Bayesian network and the global dynamic Bayesian network are learned by formula (2), and the maximum likelihood estimate of the parameters is obtained by formula (3).

[0027] θ * =arg max θ L(θ|D) (2);

[0028]

[0029] In the formula, θ represents a given parameter, and the sample dataset D contains M samples D={D1,D2,...,D...} M}, where D m ={d m1 ,d m2 ,...,d mn m=1,2,...,M,θ * To utilize the parameters estimated by maximum likelihood; L(θ|D) represents the likelihood function of θ, which is expressed in its logarithmic form. m ijk This indicates that sample data D satisfies X i =k and pa(X) i The number of samples j; This is the maximum likelihood estimate;

[0030] Step 9: Complete the establishment of the distributed dynamic Bayesian network model;

[0031] Step 10: First, use the sensor array deployed in the coal slime flotation industrial system to collect data of the coal slime flotation process online. Then, preprocess the collected data and discretize the online abnormal data as evidence information to input into the distributed dynamic Bayesian network for control decision reasoning.

[0032] Step 11: First, input the abnormal quality indicators as evidence information into the global network, and infer the abnormal probability of the raw coal processing module, the abnormal probability of the heavy medium coal preparation module, and the abnormal probability of the concentration flotation module. Then, determine the module that has an abnormality based on the abnormal probability and obtain the adjustment decision for the module. Input the normal quality indicators and the state of some process variables as evidence information into the local network, and infer the adjustment decision for the variables.

[0033] Step 12: Implement the obtained adjustment decision in the coal slime flotation control system and determine whether the abnormal operating condition has been removed. If it has been removed, enter the normal operating state. If it has not been removed, select the module with the second highest abnormal probability during global network inference for adjustment and implement the adjusted decision in the coal slime flotation control system.

[0034] Step Thirteen: Determine whether the abnormal operating condition has been removed. If it has been removed, proceed to normal operating condition. If it has not been removed, continue to adjust according to Step Twelve until the abnormal operating condition is removed.

[0035] Step Fourteen: Set the update time and use formula (4) to update the parameters of the distributed dynamic Bayesian network according to the update time to ensure the reliability of the model;

[0036]

[0037] In the formula, For variable X in historical data i Let k be the number of samples whose parent node is in state j. For the new data, variable X i The number of samples in state k where its parent node is in state j;

[0038] Step 15: In the coal slime flotation control system, steps 10 to 14 are executed cyclically to control the actual coal slime flotation production process.

[0039] In this invention, the entire industrial process is first divided into three sub-modules. Then, based on the relevant variables of each module and global quality indicators, a global Bayesian network and a local Bayesian network are obtained. The global Bayesian network is then transformed into a global dynamic Bayesian network. Next, for local Bayesian networks with loops, the weakest causal relationship is used to transform them into local dynamic Bayesian networks. Finally, maximum likelihood estimation is used for parameter learning to obtain a distributed dynamic Bayesian network. After the distributed dynamic Bayesian network is established, outlier data is used for control decision reasoning, and control decisions are implemented to effectively eliminate abnormal operating conditions. This method effectively ensures product quality standards are met and also helps ensure the service life of production equipment. To improve the effectiveness of distributed modeling in the coal slime flotation process, this invention addresses two existing problems in the coal slime flotation process: first, the problem of anomaly transmission between modules cannot be ignored after modularizing the entire process; and second, the problem of loop structures appearing in the Bayesian network due to media backflow and feedback in the local network. To address these issues, this application creatively proposes a network security operation control method based on distributed dynamic Bayesian networks. This method establishes the global network within a distributed Bayesian network as a dynamic Bayesian network model, thereby achieving modular control of the coal slime flotation process. Since modules in different time slices are interconnected within the dynamic Bayesian network model, it effectively solves the two problems mentioned above. During the establishment of the local Bayesian network, this application utilizes transfer entropy to identify the weakest causal relationship in the loop, placing the parent node of the weakest relationship in the previous time slice and the child node in the current time slice. The network security operation control method based on distributed dynamic Bayesian networks proposed in this invention is simple to implement and has low implementation costs, making it particularly suitable for use in coal slime flotation processes. Verification results show that this method can effectively eliminate abnormal operating conditions in the coal slime flotation process, providing a reliable technical basis for operators' safety control decisions. Attached Figure Description

[0040] Figure 1 This invention describes the modeling process and online control strategy for distributed dynamic Bayesian networks.

[0041] Figure 2 This is an example diagram of the global Bayesian network in this invention;

[0042] Figure 3 This is an example diagram of a local Bayesian network in this invention;

[0043] Figure 4 This is a flow chart of the coal slime flotation process in this invention;

[0044] Figure 5 This is a static Bayesian network structure diagram of the coal slime flotation process in this invention;

[0045] Figure 6 This is the global dynamic Bayesian network in this invention;

[0046] Figure 7 This is a diagram of the dynamic Bayesian network structure for coal slime flotation in this invention.

[0047] Figure 8 This is the module anomaly probability of global network inference in this invention;

[0048] Figure 9 This is a graph showing the change in the coal mine position after an abnormal working condition occurs, as described in this invention. Detailed Implementation

[0049] The present invention will be further described below.

[0050] like Figure 1 As shown, this invention provides a safe operation control method for coal slime flotation based on a distributed dynamic Bayesian network, comprising the following steps:

[0051] Step 1: Simplified full process diagram of coal slime flotation as follows Figure 4 As shown, the coal slime flotation process includes raw coal screening, coal-medium mixing, heavy media preparation and recovery, hydrocyclone separation, coal slime demediuming, coal slime thickening, flotation pretreatment, and froth flotation. The entire process includes equipment such as coal slime bins, single-layer and double-layer screens, coal slime mixing tanks, hydrocyclones, magnetic separators, coal slime bins, thickeners, slurry pretreatment units, and flotation cells. Based on the entire process flow or on process knowledge accumulated through long-term practice, the coal slime flotation process is divided into three sub-modules: raw coal processing, heavy media coal preparation, and thickening flotation. The ash content of the coal slime is determined as a global quality indicator. Therefore, the overall process operation and adjustment problem can be approximated as the operation and adjustment problem of these three modules.

[0052] Step 2: The coal slime flotation process involves 10 measurement variables and 6 operation variables, of which the 10 measurement variables are shown in Table 1 and the 6 operation variables are shown in Table 2.

[0053] Table 1. Measured variables in the coal slime flotation process

[0054]

[0055]

[0056] Table 2 Manipulated Variables in Coal Slime Flotation Process

[0057]

[0058] The variables related to the global quality indicators for each module were identified and assigned. The results of the variable assignment are shown in Table 3.

[0059] Table 3 Submodule Variable Allocation Results

[0060]

[0061] The variables related to the raw coal processing module and the global quality indicators were determined to be the raw coal feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash content of the flotation cell.

[0062] The variables related to the overall quality indicators of the heavy medium coal preparation module are determined to be: coal inlet flow rate of the mixed medium tank, slurry density of the mixed medium tank, medium density of the hydrocyclone, inlet pressure of the hydrocyclone, ash content of the hydrocyclone overflow, density adjustment of the qualified medium tank, medium density of the qualified medium tank, and ash content of the flotation cell overflow.

[0063] The variables related to the global quality indicators of the thickening and flotation module were determined to be thickener underflow rate, thickener media density, slurry preprocessor media density, flotation cell stirring speed, and flotation cell overflow ash content.

[0064] Step 3: Determine the state of variables in the global network and the state of variables in the local network:

[0065] S31: It is determined that the main variables in the global network are feature variables and quality variables, and the quality variables are mainly in two states: normal and abnormal. The normal state of the quality variable is represented by 1, and the abnormal state of the quality variable is represented by 2.

[0066] S32: It is determined that the variables in the local network are mainly process variables and quality variables, and the process variable states are mainly three states: normal, low outlier, and high outlier. The normal state of the process variable is represented by 1, the low outlier state of the process variable is represented by 2, and the high outlier state of the process variable is represented by 3.

[0067] Step 4: Establish global Bayesian networks and local Bayesian networks;

[0068] like Figure 2 As shown, a global Bayesian network is established based on the relationship between the raw coal processing module, the heavy media coal preparation module, the concentration flotation module, and the global quality indicators; in this process, expert knowledge can be used to establish a local Bayesian network.

[0069] like Figure 3 As shown, a local Bayesian network is established between the three modules based on the relationship between the raw coal processing module, the heavy media coal preparation module, and the concentration flotation module. In this process, expert knowledge can be used to establish the local Bayesian network.

[0070] like Figure 5 As shown, a local Bayesian network for the raw coal processing module is established based on the relationship between the raw coal feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash from the flotation cell. In this process, expert knowledge can be used to establish the local Bayesian network.

[0071] like Figure 5 As shown, a local Bayesian network for the heavy media coal preparation module is established based on the relationships between the coal flow rate in the mixing medium tank, the slurry density in the mixing medium tank, the medium density in the hydrocyclone, the inlet pressure of the hydrocyclone, the overflow ash content of the hydrocyclone, the density adjustment of the qualified medium tank, the medium density of the qualified medium tank, and the overflow ash content of the flotation cell. In this process, expert knowledge can be used to establish the local Bayesian network.

[0072] like Figure 5 As shown, a local Bayesian network for the thickening and flotation module is established based on the relationship between the underflow rate of the thickener, the density of the thickener medium, the density of the slurry preprocessor medium, the stirring speed of the flotation cell, and the ash content of the flotation cell overflow. In this process, expert knowledge can be used to establish the local Bayesian network.

[0073] Step 5: Based on the correlations between the characteristic variables in the raw coal processing module, the characteristic variables in the heavy media coal preparation module, and the characteristic variables in the concentration and flotation module, transform the global Bayesian network into a global dynamic Bayesian network, such as... Figure 6 As shown, in Figure 6 In the table, Sub1, Sub2, and Sub3 represent the characteristic variables in the raw coal processing module, the heavy media coal preparation module, and the concentration flotation module, respectively, and S represents the global quality variable.

[0074] Step Six: Determine if a loop structure exists in the local Bayesian network. If a loop structure exists, proceed to Step Seven; otherwise, proceed to Step Eight. Figure 5 As can be seen, the local Bayesian network of the raw coal processing module has no loop structure, while the local Bayesian network of the heavy media coal preparation module and the local Bayesian network of the concentration flotation module both have loop structures.

[0075] Step 7: Use the transfer entropy of formula (1) to determine the weakest causal relationship of the local loop-loop Bayesian network, and place the parent node in the weakest relationship in the previous time slice and the child node in the current time slice to transform this causal relationship into a local dynamic Bayesian network.

[0076]

[0077] In the formula, T(X) i+1 |X i ,pa G (X)) is the variable X to variable pa G The transit entropy of (X), X and pa G (X) represents the directed edge connecting the child node and the parent node, pa G (X) is the parent node of variable X, x i and pa G (X) jRepresenting X and pa G (X) in the data batches i and j, x i+1 This indicates that X represents data in the next batch in the future;

[0078] The calculated values ​​of the loop structures in the local Bayesian networks of the heavy media coal preparation module and the thickening flotation module are shown in Table 4. The established local dynamic Bayesian networks are as follows: Figure 7 As shown;

[0079] Table 4. Entropy values ​​between variables in the loop structure.

[0080]

[0081] In Table 4, the K->E in the local Bayesian network of the heavy media coal preparation module and the N->M in the local Bayesian network of the concentration flotation module are transformed into the connection relationship between time slices.

[0082] Step 8: Using the collected coal slime flotation process data, the parameters of the local dynamic Bayesian network and the global dynamic Bayesian network are learned by formula (2). The maximum likelihood estimation method in formula (2) learns the parameters by the likelihood between the data and the parameters. The parameters mainly include three types of parameters: prior probability, conditional probability and transitive probability between time slices. The maximum likelihood estimate of the parameters is obtained by formula (3).

[0083] θ * =arg max θ L(θ|D) (2);

[0084]

[0085] In the formula, θ represents a given parameter, and the sample dataset D contains M samples D={D1,D2,...,D...} M}, where D m ={d m1 ,d m2 ,...,d mn m=1,2,...,M,θ * To utilize the parameters after maximum likelihood estimation; L(θ|D) represents the likelihood function of θ, specifically in the form P(D|θ), which is expressed in its logarithmic form. m ijk This indicates that sample data D satisfies X i =k and pa(X) i The number of samples j; This is the maximum likelihood estimate;

[0086] To further illustrate this, some conditional probabilities are shown in Tables 5 to 14;

[0087] Table 5: Prior Probability Table for Node A

[0088]

[0089] Table 6 Conditional Probability Table for Node S* in Module 1

[0090]

[0091] Table 7. Prior probability table for nodes D, G, and I.

[0092]

[0093] Table 8. Transition Probability Table for Module 2

[0094]

[0095] Table 9. Prior probabilities of nodes L and P

[0096]

[0097] Table 10 Transition Probability Table for Module 3

[0098]

[0099]

[0100] Table 11 Prior probabilities of feature variables of global network modules

[0101]

[0102] Table 12 Transition Probability Table between Module 1 and Module 2

[0103]

[0104] Table 13 Transition Probabilities Between Module 2 and Module 3

[0105]

[0106] Table 14 Transition Probabilities Between Module 3 and Module 1

[0107]

[0108] Step 9: Complete the establishment of the distributed dynamic Bayesian network model;

[0109] Step 10: First, use the sensor array deployed in the coal slime flotation industrial system to collect data of the coal slime flotation process online. Then, preprocess the collected data and discretize the online abnormal data as evidence information to input into the distributed dynamic Bayesian network for control decision reasoning.

[0110] Step 11: First, input the abnormal quality indicators as evidence into the global network, and then infer the probability of anomalies in the raw coal processing module, the heavy media coal preparation module, and the concentration flotation module. Specifically, the anomaly probabilities are as follows: Figure 8 As shown, the module that has an anomaly is then identified based on the anomaly probability, and an adjustment decision is made for the module.

[0111] The global network inference rules are as follows: The result of global network inference includes the probability of anomaly in the previous time slice and the probability of anomaly in the current time slice. If the probability of anomaly in the current time slice is the highest, it is necessary to consider whether the probability of the previous module in the previous time slice is the highest. If so, the previous module is operated on; otherwise, the current module is operated on.

[0112] To test the effectiveness of the established model, four cases were selected for study, as shown in Table 15. The normal quality indicators and the state of some process variables were used as evidence information and input into the local network to infer the adjustment decision of the variables.

[0113] The local network inference rule is as follows: Input the state of the current time slice variable as evidence information, and infer the adjustment of the operation variable in the previous time slice. The specific operation adjustments are shown in Table 16.

[0114] Table 15 Evidence Information from Cases

[0115]

[0116] Table 16 Adjustment Strategies for Distributed Dynamic Bayesian Network Cases

[0117]

[0118] Step 12: Implement the obtained adjustment decision in the coal slime flotation control system and determine whether the abnormal operating condition has been removed. If it has been removed, enter the normal operating state. If it has not been removed, select the module with the second highest abnormal probability during global network inference for adjustment and implement the adjusted decision in the coal slime flotation control system.

[0119] Step Thirteen: Determine whether the abnormal operating condition has been removed. If it has been removed, proceed to normal operating condition. If it has not been removed, continue to adjust according to Step Twelve until the abnormal operating condition is removed.

[0120] Step Fourteen: Set the update time and use formula (4) to update the parameters of the distributed dynamic Bayesian network according to the update time to ensure the reliability of the model. The parameter update method of the dynamic Bayesian network is mainly completed by integrating new data and historical data information.

[0121]

[0122] In the formula, For variable X in historical data i Let k be the number of samples whose parent node is in state j. For the new data, variable X i The number of samples in state k where its parent node is in state j;

[0123] Step 15: In the coal slime flotation control system, steps 10 to 14 are executed cyclically to control the actual coal slime flotation production process.

[0124] To verify the effectiveness of the adjustment algorithm mentioned in this application, four selected cases were simulated on the constructed simulation platform, and the control decisions provided by the model were applied to observe whether the anomalies in the quality indicators could be eliminated. The network parameters corresponding to the four cases are all different. Figure 9 As shown, a quality index threshold of 9.5% is set. Values ​​exceeding this threshold are used as evidence input into the established model for inference and control decision-making. The four control curves in the figure utilize the distributed dynamic Bayesian network model proposed in this application for control implemented after inference. It is evident that the control effect is accurate and timely. Therefore, the control decisions provided by the distributed dynamic Bayesian network proposed in this application are not only accurate inferences but also timely.

[0125] In this invention, the entire industrial process is first divided into three sub-modules. Then, based on the relevant variables of each module and global quality indicators, a global Bayesian network and local Bayesian networks are obtained. The global Bayesian network is then transformed into a global dynamic Bayesian network. Next, for local Bayesian networks containing loops, the weakest causal relationship is used to transform them into local dynamic Bayesian networks. Finally, maximum likelihood estimation is used to learn the parameters, resulting in a distributed dynamic Bayesian network. After the distributed dynamic Bayesian network is established, outlier data is used to infer control decisions and implement those decisions to effectively eliminate abnormal operating conditions. This method effectively ensures product quality standards are met and also helps ensure the service life of production equipment.

[0126] Bayesian networks, as powerful reasoning and knowledge representation tools, are widely used in safety control and decision-making. The coal slime flotation process, due to its numerous variables, leads to complex modeling. Therefore, modularization of the entire process and the establishment of a distributed Bayesian network model from global to local are proposed. However, the distributed Bayesian network established after modularizing the long process suffers from two problems: first, the problem of abnormal transmission between modules cannot be ignored after modularization; second, the existence of medium backflow and feedback in the local network leads to loop structures in the Bayesian network. To improve the effectiveness of distributed modeling in the coal slime flotation process, this invention creatively proposes a network security operation control method based on distributed dynamic Bayesian networks, addressing these two problems. This method establishes the global network in the distributed Bayesian network as a dynamic Bayesian network model, thereby realizing modular control of the coal slime flotation process. Because the modules in different time slices are connected in the dynamic Bayesian network model, it can effectively solve the above two problems. In the process of establishing the local Bayesian network, this application uses transfer entropy to identify the weakest causal relationship in the loop, placing the parent node of the weakest relationship in the previous time slice and the child node in the current time slice. The distributed dynamic Bayesian network security operation control method proposed in this invention is simple to implement and has low implementation cost. It is especially suitable for use in the coal slime flotation process. Verification results show that the method can effectively eliminate abnormal operating conditions in the coal slime flotation process and provide a reliable technical basis for operators' safety control decisions.

Claims

1. A method for safe operation control of coal slime flotation based on distributed dynamic Bayesian networks, characterized in that, Includes the following steps: Step 1: Divide the coal slime flotation process into three modules: raw coal processing module, heavy media coal preparation module, and thickening flotation module, and determine the coal slime ash content as a global quality indicator. Step 2: Identify the variables related to the global quality metrics for each module; The variables related to the raw coal processing module and the global quality indicators were determined to be the raw coal feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash content of the flotation cell. The variables related to the overall quality indicators of the heavy medium coal preparation module are determined to be: coal inlet flow rate of the mixed medium tank, slurry density of the mixed medium tank, medium density of the hydrocyclone, inlet pressure of the hydrocyclone, ash content of the hydrocyclone overflow, density adjustment of the qualified medium tank, medium density of the qualified medium tank, and ash content of the flotation cell overflow. The variables related to the global quality indicators of the thickening and flotation module were determined to be thickener underflow rate, thickener media density, slurry preprocessor media density, flotation cell stirring speed, and flotation cell overflow ash content. Step 3: Determine the state of variables in the global network and the state of variables in the local network: S31: It is determined that the main variables in the global network are feature variables and quality variables, and the quality variables are mainly in two states: normal and abnormal. S32: It is determined that the variables in the local network are mainly process variables and quality variables, and the process variable states are mainly three states: normal, low outlier, and high outlier. Step 4: Establish global Bayesian networks and local Bayesian networks; A global Bayesian network is established based on the relationship between the raw coal processing module, the heavy media coal preparation module, the concentration flotation module, and the global quality indicators. Based on the relationship between the raw coal processing module, the heavy media coal preparation module, and the concentration flotation module, a local Bayesian network is established among the three modules. A local Bayesian network for the raw coal processing module is established based on the relationship between the raw coal silo feed rate, the double-layer screen discharge flow rate, the single-layer screen discharge flow rate, and the overflow ash from the flotation cell. A local Bayesian network for the heavy media coal preparation module is established based on the relationship between the coal flow rate in the mixing medium tank, the slurry density in the mixing medium tank, the medium density in the hydrocyclone, the inlet pressure of the hydrocyclone, the overflow ash content of the hydrocyclone, the density adjustment of the qualified medium tank, the medium density of the qualified medium tank, and the overflow ash content of the flotation cell. A local Bayesian network for the thickening and flotation module is established based on the relationship between the underflow rate of the thickener, the density of the thickener medium, the density of the slurry preprocessor medium, the stirring speed of the flotation cell, and the ash content of the flotation cell overflow. Step 5: Based on the correlation between the characteristic variables in the raw coal processing module, the characteristic variables in the heavy media coal preparation module, and the characteristic variables in the concentration flotation module, the global Bayesian network is transformed into a global dynamic Bayesian network; Step 6: Determine if there is a loop structure in the local Bayesian network. If there is a loop structure, proceed to Step 7; otherwise, proceed to Step 8. Step 7: Use the transfer entropy of formula (1) to determine the weakest causal relationship of the local loop-loop Bayesian network, and place the parent node in the weakest relationship in the previous time slice and the child node in the current time slice to transform this causal relationship into a local dynamic Bayesian network. In the formula, T(X) i+1 |X i ,pa G (X)) is the variable X to variable pa G The transit entropy of (X), X and pa G (X) represents the directed edge connecting the child node and the parent node, pa G (X) is the parent node of variable X, x i and pa G (X) j Representing X and pa G (X) in the data batches i and j, x i+1 This indicates that X represents data in the next batch in the future; Step 8: Using the collected coal slime flotation process data, the parameters of the local dynamic Bayesian network and the global dynamic Bayesian network are learned by formula (2), and the maximum likelihood estimate of the parameters is obtained by formula (3). i * =argmax θ L(θ|D) (2); In the formula, θ represents a given parameter, and the sample dataset D contains M samples D={D1,D2,...,D...} M }, where D m ={d m1 ,d m2 ,...,d mn m=1,2,...,M,θ * To utilize the parameters estimated by maximum likelihood; L(θ|D) represents the likelihood function of θ, which is expressed in its logarithmic form. m ijk This indicates that sample data D satisfies X i =k and pa(X) i The number of samples j; This is the maximum likelihood estimate; Step 9: Complete the establishment of the distributed dynamic Bayesian network model; Step 10: First, use the sensor array deployed in the coal slime flotation industrial system to collect data of the coal slime flotation process online. Then, preprocess the collected data and discretize the online abnormal data as evidence information to input into the distributed dynamic Bayesian network for control decision reasoning. Step 11: First, input the abnormal quality indicators as evidence information into the global network, and infer the abnormal probability of the raw coal processing module, the abnormal probability of the heavy medium coal preparation module, and the abnormal probability of the concentration flotation module. Then, determine the module that has an abnormality based on the abnormal probability and obtain the adjustment decision for the module. Input the normal quality indicators and the state of some process variables as evidence information into the local network, and infer the adjustment decision for the variables. Step 12: Implement the obtained adjustment decision in the coal slime flotation control system and determine whether the abnormal operating condition has been removed. If it has been removed, enter the normal operating state. If it has not been removed, select the module with the second highest abnormal probability during global network inference for adjustment and implement the adjusted decision in the coal slime flotation control system. Step 13: Determine if the abnormal operating condition has been removed. If it has been removed, proceed to normal operating condition. If it has not been removed, continue to adjust according to Step 12 until the abnormal operating condition is removed. Step Fourteen: Set the update time and use formula (4) to update the parameters of the distributed dynamic Bayesian network according to the update time to ensure the reliability of the model; In the formula, For variable X in historical data i Let k be the number of samples whose parent node is in state j. For the variable X in the new data i The number of samples in state k where its parent node is in state j; Step 15: In the coal slime flotation control system, steps 10 to 14 are executed cyclically to control the actual coal slime flotation production process.

Images