A Coal Slime Flotation Process Control Method Based on Distributed Pyramid Dynamic Bayesian Network

Abstract

A control method for coal slime flotation process based on a distributed pyramidal dynamic Bayesian network is proposed. The method divides the coal slime flotation process into three sub-processes, determines global and local quality index variables, and establishes a global dynamic Bayesian network. Key variables in each sub-process are identified, their correlations are analyzed, and process delays in each process are determined. The number of time slices in the first-layer dynamic Bayesian network is determined. Structure learning is performed to establish the network structure of the first layer of the model. Identical nodes in adjacent time slices are merged into one node to construct the second-layer dynamic Bayesian network structure. Bayesian network parameters are learned. Actual operating condition data is collected, and the sub-process causing the anomaly is identified when an anomaly occurs. The root cause of the anomaly is determined using a local pyramidal dynamic Bayesian network. Based on the cause of the anomaly, a control decision scheme is formulated to eliminate the anomaly. This method can promptly eliminate anomalies.

Description

Technical Field

[0001] This invention belongs to the field of safety control technology for complex industrial processes, specifically a control method for coal slime flotation process based on a distributed pyramid dynamic Bayesian network. Background Technology

[0002] Coal slime flotation, as a typical complex industrial process, consists of multiple coupled sub-processes, such as raw coal pretreatment, separation, centrifugal separation, and froth flotation. Each sub-process contains numerous variables, and these variables have complex interactions and coupling relationships. With the rapid development of modern industry, the scale of coal slime flotation processes is constantly expanding, which not only increases the difficulty of process control but also makes the system's dynamism, time delay, and uncertainty more significant. The expansion of the scale of coal slime flotation processes brings a series of problems, including increased equipment maintenance costs, increased operational complexity, enhanced sensitivity to raw material fluctuations, and higher requirements for adaptability to environmental changes. Meanwhile, the current low level of automation in coal slime flotation processes leads to frequent occurrences of abnormal operating conditions. If these abnormal conditions are not eliminated in a timely manner, it will result in decreased production efficiency, unstable product quality, and may even lead to safety accidents. Furthermore, insufficient automation also leads to over-reliance on the subjective experience of operators when dealing with abnormal conditions, increasing the risk of human error and thus affecting the stability and reliability of the entire coal slime flotation process. Therefore, taking effective measures to improve the automation level of the coal slime flotation process is crucial to ensuring its efficient, stable and safe operation while scaling up.

[0003] The increasing scale of coal slime flotation processes makes centralized modeling exceptionally difficult. In large-scale flotation plants, the dynamic behavior of the system is influenced by numerous variables, including the properties of the raw coal, the operating conditions of the flotation machine, and environmental factors. The interactions and changes of these variables increase the complexity of modeling. Centralized modeling requires consideration of the global characteristics of the entire system, including the interactions and information transmission between various subsystems. However, since the difficulty of data collection, processing, and analysis increases significantly with the scale of the system, this requirement is often difficult to achieve in practice. Furthermore, centralized models often require substantial computational resources and time, which is impractical for real-time control and decision support.

[0004] A Bayesian network is a probabilistic model used to represent conditional dependencies between variables. It consists of a set of random variables and conditional probability distributions defined on these variables, typically represented as a directed acyclic graph. Dynamic Bayesian networks are an extension of Bayesian networks; they model time-series data by replicating the network structure at the time level, thus capturing the dynamic characteristics of variables over time. To reduce model complexity, dynamic Bayesian networks are usually required to satisfy the first-order Markov property, meaning the state of a variable at the current moment depends only on the state of the variable at the previous moment. However, in the coal slime flotation process, due to the significant time delays between variables, the state of a variable at the current moment depends not only on the state of the variable at the previous moment but also on the state of the variable at one or several past time moments. This time delay characteristic violates the first-order Markov assumption, making traditional dynamic Bayesian networks difficult to directly apply to the coal slime flotation process. Therefore, it is necessary to provide a dynamic Bayesian network model that not only satisfies the first-order Markov assumption, but also takes into account the dynamic characteristics and time delay in the coal slime flotation process. This model can accurately identify the causes of abnormal operating conditions in the coal slime flotation process, and based on this diagnostic information, formulate scientific control strategies to eliminate abnormalities in a timely manner and ensure the safety and stability of the coal slime flotation process. Summary of the Invention

[0005] To address the problems existing in the prior art, this invention provides a control method for coal slime flotation process based on a distributed pyramidal dynamic Bayesian network. This method has a simple modeling process and high reliability. It can accurately identify the causes of abnormal operating conditions in the coal slime flotation process and, based on this, quickly and accurately formulate scientific control strategies, thereby eliminating abnormal operating conditions in a timely manner and effectively ensuring the safety and stability of the coal slime flotation process.

[0006] To achieve the above objectives, the present invention provides a coal slime flotation process control method based on a distributed pyramidal dynamic Bayesian network, including a coal slime flotation industrial system, wherein the coal slime flotation industrial system includes a raw coal processing subsystem, a heavy media coal preparation subsystem, and a concentration flotation system;

[0007] The coal slime flotation process control method includes the following steps:

[0008] Step 1: Divide the coal slime flotation process into three sub-processes, determine the global quality index variables and local quality index variables, and establish a global dynamic Bayesian network;

[0009] S11: Through in-depth analysis of historical data of the coal slime flotation process, combined with rich process knowledge and the actual situation of the coal slime flotation industrial system, the coal slime flotation process is divided into three sub-processes: raw coal processing, heavy media coal preparation and concentration flotation, using a distributed approach. The three sub-processes correspond to the raw coal processing subsystem, heavy media coal preparation subsystem and concentration flotation subsystem in the coal slime flotation industrial system, respectively.

[0010] S12: The overflow ash content S of the flotation cell is determined as the global quality index variable, and a quality index expectation value is set for each subprocess as its local quality index variable. The local quality index variables of the three subprocesses are S1, S2, and S3, respectively.

[0011] S13: Establish a global Bayesian network based on the causal relationship between global and local quality indicator variables;

[0012] S14: Connect the sub-processes between different time slices to construct a global dynamic Bayesian network;

[0013] Step 2: Identify the key variables in the three sub-processes, collect process data, analyze the correlation between variables, and determine the process delay T in each process;

[0014] S21: Based on knowledge of coal slime flotation technology, the key variables in the three sub-processes are determined according to their ability to reflect product quality and ease of measurement, and they are divided into operational variables and measurement variables. The changes in product indicators are reflected by the measurement variables in each sub-process, and the product quality is improved by adjusting the operational variables.

[0015] S22: Using sensor groups arranged in the raw coal processing subsystem, heavy media coal preparation subsystem, and concentration flotation subsystem, process data in the raw coal processing subprocess, heavy media coal preparation subprocess, and concentration flotation subprocess are collected in real time, and time series data are obtained.

[0016] S23: Analyze the time series data in the raw coal processing sub-process, heavy media coal preparation sub-process, and thickening flotation sub-process using the maximum information coefficient method, and determine the correlation between variables;

[0017] S24: Based on the correlation results, approximate the time delay between local quality index variables and key variables in the three sub-processes, and take the largest time delay as the process delay T of the process.

[0018] Step 3: Based on the process delay T of each subprocess, use formula (1) to determine the number of time slices of the first layer of dynamic Bayesian network in the local pyramid dynamic Bayesian network model.

[0019] W=2 (K-1) (1);

[0020] In the formula, W represents the number of time slices in the first layer of the dynamic Bayesian network of the pyramid; K = ceil(log2(T+1); ceil(·) represents the floor function;

[0021] Step 4: Perform structure learning and establish the network structure of the first layer of the pyramid dynamic Bayesian network model;

[0022] After determining the number of time slices in the first layer of the local pyramid dynamic Bayesian network in each subprocess, the structure of the first layer dynamic Bayesian network is learned using the Bayesian information criterion scoring function in formula (2).

[0023]

[0024] In the formula, D represents the given dataset; G represents the given structure; N represents the number of training samples; n is the number of nodes; I P (X i ;Pa(X i )) represents a node X in a given structure G. i The mutual information between it and its parent node; Dim[G] is the number of independent parameters of model G; H P (X i ) represents node X i Entropy;

[0025] Step 5: Using a variable-length data window, merge identical nodes in adjacent time slices of the first-layer dynamic Bayesian network into one node to construct the second-layer dynamic Bayesian network structure.

[0026] By using a variable-length data window, identical nodes in adjacent time slices in the first-layer dynamic Bayesian network are merged into one node, so that the nodes in the second-layer dynamic Bayesian network contain more dynamic data information. At the same time, by changing the size of the data window and reusing the Bayesian information criterion scoring function for structure learning, the structure of the second-layer dynamic Bayesian network of the pyramid is obtained.

[0027] Step 6: Repeat step 5 until the number of time slices in the dynamic Bayesian network is 2;

[0028] The number of time slices is reduced by merging network nodes until the number of time slices is reduced to 2, at which point the merging process stops. Ultimately, the pyramid dynamic Bayesian network has a total of 2 time slices. (K-1) ,2 (K-2) ,2 (K-3) ,...,2 1 ;

[0029] Step 7: Parameter learning for global dynamic network and local pyramid dynamic Bayesian network;

[0030] Parameter learning is performed to enable dynamic Bayesian networks to capture the dynamic relationships between variables in time series data; then, the state of variables that cause abnormal operating conditions is inferred through conditional probability distribution, and the optimal safety control decision scheme is formulated.

[0031] Step 8: Collect actual working condition data, determine whether abnormal working conditions have occurred, and identify the sub-process that caused the abnormality when abnormal working conditions occur;

[0032] The actual operating data of the global quality index variables are compared with the set safety threshold, which is set at 9.2%. When the ash content S of the flotation cell overflow is less than or equal to 9.2%, the coal slime flotation process is in a safe operating state. When the ash content S of the flotation cell overflow is greater than or equal to 9.2%, the coal slime flotation process is abnormal. The online abnormal data is input as evidence into the global dynamic Bayesian network, and the subprocess that caused the abnormality is inferred based on the global network inference principle. During the global network inference process, the abnormality probability of each subprocess in the current time and the previous time is calculated, and the subprocess with the highest abnormality probability in the current time and the previous time is found. If the subprocess with the highest abnormality probability in the previous time is adjacent to the subprocess with the highest abnormality probability in the current time, and is before the subprocess with the highest abnormality probability in the current time, it means that the abnormality of the current subprocess is caused by the abnormality of the previous subprocess, and the abnormality detection of the previous subprocess needs to be performed. Otherwise, it is considered that the cause of the abnormality is the current subprocess.

[0033] Step 9: Use a local pyramid dynamic Bayesian network to determine the root cause of the abnormal operating conditions;

[0034] After identifying the abnormal sub-process, the local online abnormal data is input as evidence into each layer of the local pyramid dynamic Bayesian network at the current time, and reasoning is performed layer by layer from top to bottom. The abnormal diagnosis results of each layer of the pyramid dynamic Bayesian network are determined according to the maximum a posteriori probability principle.

[0035] Step 10: Based on the causes of anomalies diagnosed by the local pyramid dynamic Bayesian network, formulate and implement reasonable control decision-making schemes to eliminate abnormal operating conditions;

[0036] S101: When the causes of anomalies diagnosed by each layer of the local pyramid dynamic Bayesian network are the same, it means that the time delay in the process does not affect the fault location results and decision-making. The control decision scheme based on the corresponding cause of anomaly is directly implemented. At the same time, observe whether the overflow ash S of the flotation cell returns to the safe threshold. If the overflow ash S of the flotation cell drops to the safe threshold, it means that the abnormal condition has been eliminated and the reasoning process ends. Otherwise, proceed to step eleven.

[0037] S102: When the causes of anomalies diagnosed by different layers of the local pyramid dynamic Bayesian network are not the same, control decisions shall be implemented according to the following steps:

[0038] A1: First, the adjustment scheme deduced from the first layer of the local pyramid is used for operation adjustment. If the overflow ash content S of the flotation cell returns to within the safe threshold, the reasoning process ends.

[0039] A2: If the coal slime flotation process does not return to normal operation after the first-level adjustment plan is implemented, the second-level inferred adjustment plan will be used for further operational adjustments based on the first-level operational adjustments. When implementing the second-level adjustment plan, only the newly emerging operational strategies will be adopted and executed, and variables that have already been adjusted in the first-level operational adjustments will not be adjusted again. At the same time, observe whether the overflow ash content S of the flotation cell returns to the safe threshold. If the overflow ash content S of the flotation cell drops to the safe threshold, the inference process ends.

[0040] A3: After implementing the adjustment schemes of the first and second layers, if the coal slime flotation process still has not returned to normal operation, the inference schemes of the subsequent layers of the pyramid dynamic Bayesian network will continue to be adopted and implemented. The implementation process is the same as that of the adjustment scheme of the second layer. Only the newly emerging adjustment strategies will be implemented, and the previously implemented adjustment schemes will not be repeated until the coal slime flotation process returns to normal operation, at which point the inference process ends. When the adjustment scheme inferred from the last layer is implemented, if the coal slime flotation process still has not returned to normal operation, step eleven will be executed.

[0041] Step 11: Select the sub-process with the second highest anomaly probability in the global reasoning process for reasoning, and re-execute Step 9 to continue using online anomaly data as evidence to reason out control decision schemes until the abnormal operating conditions are eliminated.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] Coal slime flotation is a crucial step in coal washing, aiming to improve coal quality and thus increase coal resource utilization efficiency through physical and chemical methods. However, the fluctuations in raw material properties, frequent operational adjustments, and complex and variable working environment conditions during coal slime flotation result in numerous dynamic characteristics and time delays. These dynamic characteristics and time delays not only increase the difficulty of process control but also lead to frequent abnormal operating conditions. Failure to take effective measures to identify the causes of these abnormalities in a timely manner and to formulate correct control decisions to eliminate them can lead to resource waste, environmental pollution, damage to production equipment, and safety accidents. Traditional dynamic Bayesian networks are a powerful tool for modeling and inferring uncertainties and dynamic characteristics in time-series data. By extending the Bayesian network along the time axis, dynamic Bayesian networks link the current state of the process with the previous state, thereby capturing the dynamic characteristics of the process. However, dynamic Bayesian networks typically assume that changes in the state of variables at the current moment depend only on the state of variables at the previous moment. This first-order Markov assumption simplifies the model's complexity in many cases, but it also limits the model's ability to capture causal relationships caused by time delays. In the coal slime flotation process, due to the existence of time delays, relying solely on traditional dynamic Bayesian networks may not accurately predict and control the delay relationships between process variables, thus affecting the efficiency and accuracy of the control strategy. To effectively address the impact of dynamic characteristics and time delays on anomaly identification in the coal slime flotation process, this invention provides a safety control method for the coal slime flotation industrial process based on a distributed pyramid dynamic Bayesian network. By establishing a dynamic Bayesian network with progressively decreasing time slices, not only can the dynamic characteristics of the coal slime flotation process be accurately captured, but the time delays in the process can also be gradually covered. More importantly, the dynamic Bayesian networks established by this method all satisfy the first-order Markov assumption, without increasing modeling complexity. The main process of anomaly diagnosis and decision-making using the distributed pyramid dynamic Bayesian network consists of four parts: the first part is data acquisition, sub-process division, variable determination, and the establishment of a global dynamic Bayesian network. Connecting the three sub-processes across different time slices to establish a global dynamic Bayesian network effectively reflects the interrelationships between them, thus facilitating the rapid resolution of information transmission anomalies. The second part involves correlation analysis of variables within the sub-processes, establishing a dynamic Bayesian network for the first layer of the pyramid. The maximum information coefficient method is used to analyze the correlation of variables in each sub-process and approximate the time delay of each variable. The maximum time delay is taken as the process delay T, which facilitates determining the number of time slices in the first layer of the pyramid dynamic Bayesian network and constructing its network structure. The third part merges identical nodes in adjacent time slices in the first-layer dynamic Bayesian network to obtain the dynamic Bayesian networks for subsequent layers.By using variable-length data windows, data from identical nodes in adjacent time slices of the first-layer dynamic Bayesian network are merged, resulting in more data information in the network nodes of the second layer. This also avoids data loss during node merging. Furthermore, structure and parameter learning are re-performed to complete the construction of the second-layer dynamic Bayesian network model, facilitating the estimation of conditional probabilities of variables in the network from historical data. The construction method for the dynamic Bayesian network models of subsequent layers of the pyramid is the same as the second layer. Modeling ends when the number of time slices in the dynamic Bayesian network is reduced to two. The fourth part is control decision reasoning to eliminate abnormal operating conditions. Online abnormal data is input as evidence into the global dynamic Bayesian network model to infer abnormal sub-processes. Then, based on the local pyramid dynamic Bayesian network of the abnormal sub-processes, control strategies for abnormal operating conditions are inferred layer by layer from top to bottom. Finally, the inferred control strategies are implemented to eliminate abnormal operating conditions.

[0044] This method has a simple modeling process and high reliability. It can accurately identify the causes of abnormal operating conditions in the coal slime flotation process, and based on this, quickly and accurately formulate scientific control strategies, thereby eliminating abnormal operating conditions in a timely manner and effectively ensuring the safety and stability of the coal slime flotation process. Attached Figure Description

[0045] Figure 1 This is a flowchart of the present invention;

[0046] Figure 2 This is a simplified principle block diagram of the coal slime flotation industrial system in this invention;

[0047] Figure 3 This is a diagram of the global dynamic Bayesian network structure in this invention;

[0048] Figure 4 This is a graph showing the maximum information coefficient result of sub-process 1 in this invention;

[0049] Figure 5 This is a graph showing the maximum information coefficient result of sub-process 2 in this invention;

[0050] Figure 6 This is a graph showing the maximum information coefficient result of sub-process 3 in this invention;

[0051] Figure 7 This is a diagram of the pyramid dynamic Bayesian network structure of sub-process 1 in this invention;

[0052] Figure 8 This is a diagram of the pyramid dynamic Bayesian network structure of subprocess 2 in this invention;

[0053] Figure 9 This is a diagram of the pyramid dynamic Bayesian network structure of subprocess 3 in this invention;

[0054] Figure 10 This is the result of global dynamic Bayesian inference in Case 1 of this invention;

[0055] Figure 11 This is the result of global dynamic Bayesian inference in Case 2 of this invention;

[0056] Figure 12 This refers to the change in the overflow ash content curve of the flotation cell during the coal slime flotation process in Case 1 of this invention.

[0057] Figure 13 This refers to the change in the overflow ash content curve of the flotation cell during the coal slime flotation process in Case 2 of this invention. Detailed Implementation

[0058] The invention will now be further described with reference to the accompanying drawings.

[0059] Coal slime flotation is a highly efficient mineral processing technology primarily used to improve coal quality by separating useful components from impurities in coal slime through physical and chemical methods. This process involves several key steps, including raw coal crushing, mixing, pH adjustment, collector addition, stirring, aerated flotation, separation and concentration, dewatering, filtration, and drying. During flotation, coal particles, due to their hydrophobic properties, adhere to air bubbles and float, while hydrophilic impurities remain in the slurry, thus achieving separation. Coal slime flotation plays a crucial role in improving coal resource utilization, reducing environmental pollution, and enhancing economic benefits. With the rapid development of automation and intelligent technologies, the efficiency and quality control of the coal slime flotation process face new opportunities for improvement. Automation technology can optimize flotation effects and reduce human error by precisely controlling various parameters in the flotation process, such as slurry concentration, pH value, and reagent dosage. Intelligent technologies can analyze historical data, predict and identify anomalies in the production process, thereby achieving more precise process control and decision support. The application of these technologies can not only improve the production efficiency of coal slime flotation and reduce operating costs but also effectively reduce environmental pollution and promote the sustainable development of the coal industry. Therefore, integrating automation and intelligent technologies into the coal slime flotation process plays a crucial role in ensuring the safe and stable operation of the coal slime flotation process.

[0060] like Figure 1 As shown, this invention provides a method for controlling a coal slime flotation process based on a distributed pyramidal dynamic Bayesian network, including a coal slime flotation industrial system, such as... Figure 2As shown, the coal slime flotation industrial system includes a raw coal processing subsystem, a heavy media coal preparation subsystem, and a concentration flotation subsystem. The raw coal processing subsystem mainly removes impurities such as rocks, soil, and plant debris from the raw coal through screening, crushing, and washing steps to improve coal quality and usability. The heavy media coal preparation subsystem mainly utilizes the density difference between coal and gangue, using equipment such as heavy media hydrocyclones and separators to efficiently separate high-calorific-value coal while reducing transportation costs. The concentration flotation subsystem mainly further purifies fine-particle coal by removing sulfur and other harmful impurities through equipment such as flotation machines, agitators, and reagent addition systems to improve coal combustion efficiency and reduce environmental pollution.

[0061] The raw coal processing subsystem includes a crusher, a double-layer screen, and a single-layer screen. The crusher is used to crush the raw coal. The feed inlet of the double-layer screen is connected to the discharge outlet of the crusher, and it is used to recover and reduce the ash content of the crushed coarse coal slime by screening, and output clean coal. The feed inlet of the single-layer screen is connected to the tailings outlet of the double-layer screen, and it is used to remove impurities from the raw coal by screening.

[0062] The heavy media coal preparation subsystem includes a mixing box, a heavy media hydrocyclone, an arc screen (I), a desliming screen (I), a clean coal centrifuge, a diversion device, a magnetic separator (I), a coal slime bucket, an arc screen (II), a centrifuge (I), a desliming screen (II), a magnetic separator (II), and a centrifuge (II). The inlet of the mixing box is connected to the tail material outlet of the single-layer screen, used to mix raw coal and recovered liquid before outputting the mixture. The inlet of the heavy media hydrocyclone is connected to the outlet of the mixing box, used to separate low-density and high-density materials based on their density differences. The inlet of the arc screen (I) is connected to the low-density outlet of the heavy media hydrocyclone. This device is used to perform pre-demediation operations; the feed inlet of the first demediation screen is connected to the fixed discharge outlet of the first arc screen for secondary demediation operations; the feed inlet of the clean coal centrifuge is connected to the solid discharge outlet of the first demediation screen, and the solid discharge outlet of the clean coal centrifuge discharges clean coal, which is used to separate clean coal and liquid; the feed inlet of the diversion device is connected to the liquid discharge outlet of the first arc screen, and the concentrate discharge outlet of the diversion device is connected to the feed inlet of the mixing box, which is used to separate the top flow containing concentrate and the bottom flow containing tailings; the feed inlet of the first magnetic separator is connected to the liquid discharge outlet of the first demediation screen and the tailings discharge outlet of the diversion device. The low-density material outlet of magnetic separator one is connected to the inlet of the mixing tank. Magnetic separator one is used to remove iron impurities and separate low-density and high-density materials. The inlet of the coal slime bucket is connected to the liquid outlet of the clean coal centrifuge and the high-density material outlet of magnetic separator one, used to ensure thorough mixing of the coal slime. The inlet of the arc-shaped screen two is connected to the outlet of the coal slime bucket, used for desliming. The inlet of centrifuge one is connected to the solid outlet of arc-shaped screen two, the liquid outlet of centrifuge one is connected to the inlet of the coal slime bucket, and the solid outlet of centrifuge one discharges clean coal. Centrifuge one is used for... The system achieves the separation of clean coal and liquid. The inlet of the second desliming screen is connected to the high-density outlet of the heavy medium cyclone separator, and the liquid outlet of the second desliming screen is connected to the inlet of the mixing tank. The second desliming screen is used for desliming operations. The inlet of the second magnetic separator is connected to the solid outlet of the second desliming screen, and the low-density material outlet of the second magnetic separator is connected to the inlet of the mixing tank. The second magnetic separator is used to remove iron impurities and separate low-density and high-density materials. The inlet of the second centrifuge is connected to the liquid outlet of the second desliming screen, and the fixed outlet of the second centrifuge discharges middlings. The second centrifuge is used for separating middlings from liquid.

[0063] The thickening and flotation subsystem includes a coal slime thickener, a slurry preprocessor, a flotation cell, a tailings thickener I, a clean coal filter, a drying device, a tailings thickener II, a thickener III, and a filter press. The feed inlet of the coal slime thickener is connected to the liquid outlet of the arc screen I, used to increase the concentration of the coal slime and reduce its moisture content. The feed inlet of the slurry preprocessor is connected to the solid outlet of the coal slime thickener, used to pre-treat the slurry before flotation to improve flotation performance. The feed inlet of the flotation cell is connected to the outlet of the slurry preprocessor, used to separate coal and impurities. The feed inlet of the tailings thickener I is connected to the tailings discharge outlet of the flotation cell; the solid outlet of the tailings thickener I discharges tailings, and the liquid outlet of the tailings thickener I discharges circulating water; the tailings thickener I is used to increase the concentration of tailings and reduce its moisture content. The feed inlet of the clean coal filter is connected to the concentrate discharge outlet of the flotation cell, and the clean coal passes through... The liquid outlet of the filter is connected to the inlet of the slurry preprocessor. The clean coal filter is used to increase the concentration of clean coal and reduce its moisture content. The drying equipment is connected to the solid outlet of the clean coal filter and is used to dry the clean coal and output it. The inlet of the tailings thickener II is connected to the high-density outlet of the magnetic separator II and the liquid outlet of the centrifuge II. The liquid outlet of the tailings thickener II discharges circulating water. The tailings thickener II is used to increase the concentration of tailings and reduce its moisture content. The inlet of the thickener III is connected to the liquid outlet of the coal slime thickener and the fixed outlet of the tailings thickener II. The liquid outlet of the thickener III discharges circulating water. The thickener III is used to separate fixed impurities from liquid. The inlet of the filter press is connected to the solid outlet of the thickener III. The fixed outlet of the filter press discharges the filter cake, and the liquid outlet of the filter press outputs circulating water. The filter press is used for solid-liquid separation.

[0064] The coal slime flotation process control method includes the following steps:

[0065] Step 1: Divide the coal slime flotation process into three sub-processes, determine the global quality index variables and local quality index variables, and establish a global dynamic Bayesian network;

[0066] S11: Through in-depth analysis of historical data of the coal slime flotation process, combined with rich process knowledge and the actual situation of the coal slime flotation industrial system, the coal slime flotation process is divided into three sub-processes: raw coal processing, heavy media coal preparation and concentration flotation, using a distributed approach. The three sub-processes correspond to the raw coal processing subsystem, heavy media coal preparation subsystem and concentration flotation subsystem in the coal slime flotation industrial system, respectively.

[0067] S12: The overflow ash content S of the flotation cell is determined as the global quality index variable, and a quality index expectation value is set for each subprocess as its local quality index variable. The local quality index variables of the three subprocesses are S1, S2, and S3, respectively.

[0068] S13: Establish a global Bayesian network based on the causal relationship between global and local quality indicator variables;

[0069] S14: Considering the issue of information transmission anomalies between sub-processes, each sub-process is connected between different time slices to construct a global dynamic Bayesian network; the structure of the global dynamic Bayesian network constructed based on the data information of global quality index variables and local index variables is as follows. Figure 3 As shown.

[0070] Step 2: Identify the key variables in the three sub-processes, collect process data, analyze the correlation between variables, and determine the process delay T in each process;

[0071] The Maximum Information Coefficient (MIC) is a method for measuring the correlation between two variables, applicable to both linear and nonlinear variables. The principle of MIC is based on the concept of mutual information. It obtains the mutual information value by gridding data points in a two-dimensional space and calculating the joint probability distribution under different grid divisions. MIC seeks the grid division that maximizes this mutual information value and normalizes the mutual information value to fall within the range of 0 to 1. The closer to 1, the stronger the correlation; conversely, the closer to 0, the more independent the two variables. The maximum value of the normalized mutual information is the MIC. The correlation analysis results are as follows... Figures 4 to 6 As shown.

[0072] S21: Based on knowledge of coal slime flotation technology, and taking into account the ability to reflect product quality and ease of measurement, key variables in the three sub-processes are determined and divided into operational variables and measurement variables; there are 6 operational variables and 8 measurement variables; the measurement variables in each sub-process reflect the changes in product indicators, and the product quality is improved by adjusting the operational variables.

[0073] The states of the operational and measured variables are further divided into 1, 2, and 3, representing that the corresponding variable values ​​are within the normal range, below the normal range, and above the normal range, respectively. For the global quality indicator variable S and the local quality indicator variables S1, S2, and S3, they only need to measure whether any abnormalities occur in the entire coal slime flotation process and its corresponding sub-processes. Therefore, these four variables only need to be divided into two states, 1 and 2, representing normal and abnormal, respectively. The node symbols and corresponding state divisions for the measured and operational variables are shown in Tables 1 and 2.

[0074] Table 1: Measurement Variables in the Coal Slurry Flotation Process

[0075]

[0076]

[0077] Table 2: Operating Variables in Coal Slime Flotation Process

[0078]

[0079]

[0080] S22: Using sensor groups arranged in the raw coal processing subsystem, heavy media coal preparation subsystem, and concentration flotation subsystem, process data in the raw coal processing subprocess, heavy media coal preparation subprocess, and concentration flotation subprocess are collected in real time, and time series data are obtained.

[0081] S23: Analyze the time series data in the raw coal processing sub-process, heavy media coal preparation sub-process, and thickening flotation sub-process using the maximum information coefficient method, and determine the correlation between variables, such as... Figures 4 to 6 As shown;

[0082] S24: Based on the correlation results, the time delay of local quality index variables and key variables in the three sub-processes is approximately estimated, and the maximum time delay is taken as the process delay T of the process; the variables involved in each sub-process, the time delay of the variables, and the process delay are shown in Table 3.

[0083] Table 3: Subprocess Variable Allocation and Time Delay

[0084] Subprocess Variable time delay Process delay Raw coal processing <![CDATA[A(t-2),B(t-9),C(t-5),S1(t-9)]]> T＝9 Heavy media coal preparation <![CDATA[D(t-3),E(t-1),F(t-4),G(t-14),H(t-12),I(t-9),S2(t-5)]]> T＝14 Concentration Flotation <![CDATA[J(t-13),K(t-4),L(t-8),M(t-10),S3(t-7)]]> T＝13

[0085] Step 3: Based on the process delay T of each subprocess, use formula (1) to determine the number of time slices of the first layer of dynamic Bayesian network in the local pyramid dynamic Bayesian network model.

[0086] W=2 (K-1) (1);

[0087] In the formula, W represents the number of time slices in the first layer of the dynamic Bayesian network of the pyramid; K = ceil(log2(T+1); ceil(·) represents the floor function; the purpose of using T+1 and the floor function is to ensure that the number of time slices can cover the time delay in the industrial process.

[0088] The parameter N is calculated to be equal to 4 from the process delay T of each subprocess. Therefore, the number of time slices in the first layer of the local pyramid dynamic Bayesian network for each subprocess is 2. N-1 =8.

[0089] Step 4: Perform structure learning and establish the network structure of the first layer of the pyramid dynamic Bayesian network model;

[0090] After determining the number of time slices in the first layer of the local pyramid dynamic Bayesian network in each subprocess, the structure of the first layer dynamic Bayesian network is learned using the Bayesian information criterion scoring function in formula (2).

[0091]

[0092] In the formula, D represents the given dataset; G represents the given structure; N represents the number of training samples; n is the number of nodes; I P (X i ;Pa(X i )) represents a node X in a given structure G. i The mutual information between it and its parent node; Dim[G] is the number of independent parameters of model G; H P (X i ) represents node X i Entropy;

[0093] Step 5: Using a variable-length data window, merge identical nodes in adjacent time slices of the first-layer dynamic Bayesian network into one node to construct the second-layer dynamic Bayesian network structure.

[0094] By using a variable-length data window, identical nodes in adjacent time slices in the first-layer dynamic Bayesian network are merged into one node, so that the nodes in the second-layer dynamic Bayesian network contain more dynamic data information. At the same time, the data window size is changed, and the Bayesian information criterion scoring function is reused for structure learning to obtain the structure of the second-layer dynamic Bayesian network of the pyramid. Using a variable-length data window can avoid the loss of node data information during the node merging process.

[0095] Step 6: Repeat step 5 until the number of time slices in the dynamic Bayesian network is 2;

[0096] The time slice data is reduced by merging network nodes until the number of time slices is reduced to 2, at which point the merging process stops. The dynamic Bayesian network with 2 time slices is the simplest dynamic Bayesian network model, so the merging of network nodes stops when the number of time slices is reduced to 2.

[0097] Figure 7 , Figure 8 , Figure 9 The model is represented by a local pyramid dynamic Bayesian network structure for the raw coal processing subsystem, the heavy media coal preparation subsystem, and the concentration flotation subsystem. Because the same nodes in each layer of the network contain different data information, the dependencies between variables may differ across time slices. Dynamic Bayesian networks can capture the dynamic characteristics between variables. Furthermore, by changing the amount of data information within a node, causal relationships caused by time delays can be gradually covered, improving the model's accuracy.

[0098] Ultimately, the number of time slices in the pyramid dynamic Bayesian network was obtained as 2. (K-1) ,2 (K-2) ,2 (K-3) ,...,2 1 ;

[0099] Step 7: Parameter learning for global dynamic network and local pyramid dynamic Bayesian network;

[0100] After the model structure learning is completed, parameter learning is required. The purpose of parameter learning is mainly to estimate the conditional probability table of variables in the network from historical data. Through parameter learning, the dynamic Bayesian network captures the dynamic relationships between variables in time series data. Then, the variable states that cause abnormal operating conditions are inferred through the conditional probability distribution, and the optimal safety control decision scheme is formulated. The conditional probability distributions of global quality index variables and local quality index variables in the last layer of the pyramid structure in each sub-process are shown in Tables 4 to 7.

[0101] Table 4: Conditional probability distribution of node S1 in subprocess 1

[0102]

[0103] Table 5: Conditional probability distribution of node S2 in subprocess 2

[0104]

[0105] Table 6: Conditional probability distribution of node S3 in subprocess 3

[0106]

[0107] Table 7: Conditional Probability Distribution of Global Quality Indicator Variable S

[0108]

[0109] The table categorizes the measured variable states into 1, 2, and 3, representing that the corresponding variable's value is within the normal range, below the normal range, and above the normal range, respectively. Since the quality variables S, S1, S2, and S3 are used to measure whether abnormal operating conditions occur in the entire process or its corresponding sub-processes, their states only need to be categorized into 1 and 2, representing normal and abnormal, respectively. When an abnormal operating condition occurs in the coal slime flotation process, the online abnormal data is input into the corresponding local pyramid dynamic Bayesian network model, and the abnormal evidence state is set to 2 to infer the root cause of the abnormality. Based on this, a scientific control decision-making scheme is formulated.

[0110] Step 8: Collect actual working condition data, determine whether abnormal working conditions have occurred, and identify the sub-process that caused the abnormality when abnormal working conditions occur;

[0111] The actual operating data of the global quality index variable (flotation cell overflow ash content S) is compared with the set safety threshold, which is set at 9.2%. When the flotation cell overflow ash content S ≤ 9.2%, it indicates that the coal slime flotation process is in a safe operating state. When the flotation cell overflow ash content S > 9.2%, it indicates that the coal slime flotation process is abnormal. The online abnormal data is input as evidence into the global dynamic Bayesian network, and the subprocess that caused the abnormality is inferred based on the global network reasoning principle.

[0112] To verify the effectiveness of the distributed pyramid dynamic Bayesian network security control strategy, two cases can be simulated on a coal slime flotation process simulation platform, and abnormal data can be input into the distributed pyramid Bayesian model. First, the abnormal sub-processes are inferred from the global dynamic Bayesian network. During the global network inference process, the abnormal probability of each sub-process in the current time and the previous time is calculated, and the sub-process with the highest abnormal probability in the current time and the previous time is identified. If the sub-process with the highest abnormal probability in the previous time is adjacent to the sub-process with the highest abnormal probability in the current time, and is before the sub-process with the highest abnormal probability in the current time, it means that the abnormality of the current sub-process is caused by the abnormality of the previous sub-process, and the abnormality detection of the previous sub-process needs to be performed; otherwise, it is assumed that the cause of the abnormality is the current sub-process.

[0113] Global network inference results as follows Figure 10 and Figure 11 As shown in Case 1, the sub-process with the highest probability of anomaly at the current moment is sub-process 3, while the sub-process with the highest probability of anomaly at the previous moment was sub-process 2. Therefore, the anomaly in sub-process 3 is caused by the anomaly in sub-process 2, and it is necessary to detect the cause of the anomaly in sub-process 2 and formulate a reasonable control decision plan.

[0114] Step 9: Use a local pyramid dynamic Bayesian network to determine the root cause of the abnormal operating conditions;

[0115] After identifying the abnormal sub-process, the local online abnormal data is input as evidence into each layer of the local pyramid dynamic Bayesian network at the current time, and reasoning is performed layer by layer from top to bottom. The abnormal diagnosis results of each layer of the pyramid dynamic Bayesian network are determined according to the maximum a posteriori probability principle.

[0116] Because the same nodes in different layers contain different data information, the inferred causes of anomalies may be the same or different. When local online anomaly data is input into each layer of the pyramid dynamic Bayesian network, it is input into the network at the current moment, and the inferred control decisions are from the previous moment, with the aim of improving the model's response speed.

[0117] Step 10: Based on the causes of anomalies diagnosed by the local pyramid dynamic Bayesian network, formulate and implement reasonable control decision-making schemes to eliminate abnormal operating conditions;

[0118] S101: When the causes of anomalies diagnosed by each layer of the local pyramid dynamic Bayesian network are the same, it means that the time delay in the process does not affect the fault location results and decision-making. The control decision scheme based on the corresponding cause of anomaly is directly implemented. At the same time, observe whether the overflow ash S of the flotation cell returns to the safe threshold. If the overflow ash S of the flotation cell drops to the safe threshold, it means that the abnormal condition has been eliminated and the reasoning process ends. Otherwise, proceed to step eleven.

[0119] S102: When the causes of anomalies diagnosed by different layers of the local pyramid dynamic Bayesian network are not the same, control decisions shall be implemented according to the following steps:

[0120] A1: First, the adjustment scheme deduced from the first layer of the local pyramid is used for operation adjustment. If the overflow ash content S of the flotation cell returns to within the safe threshold, the reasoning process ends.

[0121] A2: If the coal slime flotation process does not return to normal operation after the first-level adjustment plan is implemented, the second-level inferred adjustment plan will be used for further operational adjustments based on the first-level operational adjustments. When implementing the second-level adjustment plan, only the newly emerging operational strategies will be adopted and executed, and variables that have already been adjusted in the first-level operational adjustments will not be adjusted again. At the same time, observe whether the overflow ash content S of the flotation cell returns to the safe threshold. If the overflow ash content S of the flotation cell drops to the safe threshold, the inference process ends.

[0122] A3: After implementing the adjustment schemes of the first and second layers, if the coal slime flotation process still has not returned to normal operation, the inference schemes of subsequent layers of the pyramid dynamic Bayesian network will continue to be adopted and implemented. The implementation process is the same as that of the adjustment scheme of the second layer. Only the newly emerging adjustment strategies will be implemented, and the previously implemented adjustment schemes will not be repeated until the coal slime flotation process returns to normal operation, at which point the inference process ends. When the adjustment scheme inferred from the last layer (2-time-slice dynamic Bayesian network) is implemented, if the coal slime flotation process still has not returned to normal operation, step eleven will be executed.

[0123] The outlier data from the two cases were input into a distributed pyramidal dynamic Bayesian network, and the resulting control decisions are shown in Table 8. In the table, "↑" indicates an increase in the direction of variable adjustment, "ˉ" indicates a decrease in the direction of adjustment, and "-" indicates no change.

[0124] Table 8: Control Decisions for the Two Cases

[0125]

[0126]

[0127] As shown in Table 8, in Case 1, subprocess 2 is adjusted. The first-level reasoning result is to increase the coal inflow rate into the mixing medium tank, while keeping everything else unchanged. The second-level reasoning result is to increase the coal inflow rate into the mixing medium tank and decrease the opening of the qualified medium tank density adjustment valve, while keeping everything else unchanged. The third-level reasoning result is to increase the coal inflow rate into the mixing medium tank, decrease the opening of the qualified medium tank density adjustment valve, and increase the opening of the hydrocyclone pressure regulating valve. Since the reasoning results at each level in Case 1 are different, they need to be adopted layer by layer. In Case 2, subprocess 3 is adjusted, and the corresponding operation is to reduce the underflow rate of the thickener and increase the stirring speed of the flotation cell.

[0128] Step 11: Select the sub-process with the second highest anomaly probability in the global reasoning process for reasoning, and re-execute Step 9 to continue using online anomaly data as evidence to reason out control decision schemes until the abnormal operating conditions are eliminated.

[0129] Furthermore, the method in this application can be verified by observing the change in the overflow ash S of the flotation cell after the control decision is implemented:

[0130] To verify the effectiveness of the distributed pyramid dynamic Bayesian network modeling method and case adjustment strategy, the effects of control decisions on two cases can be observed on a coal slime flotation process simulation platform. After an abnormal operating condition occurs, it is observed whether the overflow ash content of the flotation cell can recover to the normal range. Figure 12 and Figure 13 As shown, a safety threshold of ash content in the flotation cell overflow is set at 9.2%. Once this threshold is exceeded, it is used as evidence input into the established distributed pyramid dynamic Bayesian network model to deduce a control decision scheme. Figure 12 and Figure 13 As can be seen in each case study, when anomalies occur, the model proposed in this invention can not only deduce control decision schemes layer by layer and quickly eliminate abnormal operating conditions, but also take into account the impact of dynamic characteristics and time delays in industrial processes on the formulation of control decisions, thereby improving the accuracy of decisions.

[0131] This invention addresses the problem of frequent abnormal operating conditions caused by the dynamic characteristics and time delays in the coal slime flotation process by proposing a safety control method for the coal slime flotation industrial process based on a distributed pyramid dynamic Bayesian network. This method utilizes a distributed strategy to refine the coal slime flotation process into multiple sub-processes. By analyzing historical and real-time abnormal data, combined with process knowledge, it establishes causal relationships between global and local quality indicator variables, constructing a global dynamic Bayesian network. Simultaneously, by learning from historical data of the sub-processes, it determines the structural relationships between local quality indicator variables and related variables within each sub-process, establishing a local pyramid dynamic Bayesian network. When an anomaly occurs, online abnormal data is input into the global Bayesian network to infer the sub-process where the anomaly occurred. The local pyramid dynamic Bayesian network is then used to determine the root cause of the anomaly, formulate and implement reasonable control decisions, and eliminate the abnormal operating condition. By collecting actual operating condition data and combining it with the inference mechanism of the dynamic Bayesian network, this method can quickly respond to and handle abnormal situations in the coal slime flotation process, ensuring the safety and stability of the process. Implementation and verification in the coal slime flotation process demonstrate the effectiveness of the proposed method.

Claims

1. A method for controlling the coal slime flotation process based on a distributed pyramidal dynamic Bayesian network, comprising a coal slime flotation industrial system, wherein the coal slime flotation industrial system includes a raw coal processing subsystem, a heavy media coal preparation subsystem, and a concentration flotation subsystem; Its features are, The coal slime flotation process control method includes the following steps: Step 1: Divide the coal slime flotation process into three sub-processes, determine the global quality index variables and local quality index variables, and establish a global dynamic Bayesian network; S11: Through in-depth analysis of historical data of the coal slime flotation process, combined with rich process knowledge and the actual situation of the coal slime flotation industrial system, the coal slime flotation process is divided into three sub-processes: raw coal processing, heavy media coal preparation and concentration flotation, using a distributed approach. The three sub-processes correspond to the raw coal processing subsystem, heavy media coal preparation subsystem and concentration flotation subsystem in the coal slime flotation industrial system, respectively. S12: The overflow ash content S of the flotation cell is determined as the global quality index variable, and a quality index expectation value is set for each subprocess as its local quality index variable. The local quality index variables of the three subprocesses are S1, S2, and S3, respectively. S13: Establish a global Bayesian network based on the causal relationship between global and local quality indicator variables; S14: Connect the sub-processes between different time slices to construct a global dynamic Bayesian network; Step 2: Identify the key variables in the three sub-processes, collect process data, analyze the correlation between variables, and determine the process delay T in each process; S21: Based on knowledge of coal slime flotation technology, the key variables in the three sub-processes are determined according to their ability to reflect product quality and ease of measurement, and they are divided into operational variables and measurement variables. The changes in product indicators are reflected by the measurement variables in each sub-process, and the product quality is improved by adjusting the operational variables. S22: Using sensor groups arranged in the raw coal processing subsystem, heavy media coal preparation subsystem, and concentration flotation subsystem, process data in the raw coal processing subprocess, heavy media coal preparation subprocess, and concentration flotation subprocess are collected in real time, and time series data are obtained. S23: Analyze the time series data in the raw coal processing sub-process, heavy media coal preparation sub-process, and thickening flotation sub-process using the maximum information coefficient method, and determine the correlation between variables; S24: Based on the correlation results, approximate the time delay between local quality index variables and key variables in the three sub-processes, and take the largest time delay as the process delay T of the process. Step 3: Based on the process delay T of each sub-process, use formula (1) to determine the number of time slices of the first layer of dynamic Bayesian network in the local pyramid dynamic Bayesian network model. W＝2 (K-1) (1)； In the formula, W represents the number of time slices in the first layer of the dynamic Bayesian network of the pyramid; K = ceil(log2(T+1); ceil(·) represents the floor function; Step 4: Perform structure learning and establish the network structure of the first layer of the pyramid dynamic Bayesian network model; After determining the number of time slices in the first layer of the local pyramid dynamic Bayesian network in each subprocess, the structure of the first layer dynamic Bayesian network is learned using the Bayesian information criterion scoring function in formula (2). In the formula, D represents the given dataset; G represents the given structure; N represents the number of training samples; n is the number of nodes; I P (X i ;Pa(X i )) represents a node X in a given structure G. i The mutual information between it and its parent node; Dim[G] is the number of independent parameters of model G; H P (X i ) represents node X i Entropy; Step 5: Using a variable-length data window, merge identical nodes in adjacent time slices of the first-layer dynamic Bayesian network into one node to construct the second-layer dynamic Bayesian network structure. By using a variable-length data window, identical nodes in adjacent time slices in the first-layer dynamic Bayesian network are merged into one node, so that the nodes in the second-layer dynamic Bayesian network contain more dynamic data information. At the same time, by changing the size of the data window and reusing the Bayesian information criterion scoring function for structure learning, the structure of the second-layer dynamic Bayesian network of the pyramid is obtained. Step 6: Repeat step 5 until the number of time slices in the dynamic Bayesian network is 2; The number of time slices is reduced by merging network nodes until the number of time slices is reduced to 2, at which point the merging process stops. Ultimately, the pyramid dynamic Bayesian network has a total of 2 time slices. (K-1) ,2 (K-2) ,2 (K-3) ,...,2 1 ; Step 7: Parameter learning for global dynamic network and local pyramid dynamic Bayesian network; Parameter learning is performed to enable dynamic Bayesian networks to capture the dynamic relationships between variables in time series data; then, the state of variables that cause abnormal operating conditions is inferred through conditional probability distribution, and the optimal safety control decision scheme is formulated. Step 8: Collect actual working condition data, determine whether abnormal working conditions have occurred, and identify the sub-process that caused the abnormality when abnormal working conditions occur; The actual operating data of the global quality index variables are compared with the set safety threshold, which is set at 9.2%. When the ash content S of the flotation cell overflow is less than or equal to 9.2%, the coal slime flotation process is in a safe operating state. When the ash content S of the flotation cell overflow is greater than or equal to 9.2%, the coal slime flotation process is abnormal. The online abnormal data is input as evidence into the global dynamic Bayesian network, and the subprocess that caused the abnormality is inferred based on the global network inference principle. During the global network inference process, the abnormality probability of each subprocess in the current time and the previous time is calculated, and the subprocess with the highest abnormality probability in the current time and the previous time is found. If the subprocess with the highest abnormality probability in the previous time is adjacent to the subprocess with the highest abnormality probability in the current time, and is before the subprocess with the highest abnormality probability in the current time, it means that the abnormality of the current subprocess is caused by the abnormality of the previous subprocess, and the abnormality detection of the previous subprocess needs to be performed. Otherwise, it is considered that the cause of the abnormality is the current subprocess. Step 9: Use a local pyramid dynamic Bayesian network to determine the root cause of the abnormal operating conditions; After identifying the abnormal sub-process, the local online abnormal data is input as evidence into each layer of the local pyramid dynamic Bayesian network at the current time, and reasoning is performed layer by layer from top to bottom. The abnormal diagnosis results of each layer of the pyramid dynamic Bayesian network are determined according to the maximum a posteriori probability principle. Step 10: Based on the causes of anomalies diagnosed by the local pyramid dynamic Bayesian network, formulate and implement reasonable control decision-making schemes to eliminate abnormal operating conditions; S101: When the causes of anomalies diagnosed by each layer of the local pyramid dynamic Bayesian network are the same, it means that the time delay in the process does not affect the fault location results and decision-making. The control decision scheme based on the corresponding cause of anomaly is directly implemented. At the same time, observe whether the overflow ash S of the flotation cell returns to the safe threshold. If the overflow ash S of the flotation cell drops to the safe threshold, it means that the abnormal condition has been eliminated and the reasoning process ends. Otherwise, proceed to step eleven. S102: When the causes of anomalies diagnosed by different layers of the local pyramid dynamic Bayesian network are not the same, control decisions shall be implemented according to the following steps: A1: First, the adjustment scheme deduced from the first layer of the local pyramid is used for operation adjustment. If the overflow ash content S of the flotation cell returns to within the safe threshold, the reasoning process ends. A2: If the coal slime flotation process does not return to normal operation after the first-level adjustment plan is implemented, the second-level inferred adjustment plan will be used for further operational adjustments based on the first-level operational adjustments. When implementing the second-level adjustment plan, only the newly emerging operational strategies will be adopted and executed, and variables that have already been adjusted in the first-level operational adjustments will not be adjusted again. At the same time, observe whether the overflow ash content S of the flotation cell returns to the safe threshold. If the overflow ash content S of the flotation cell drops to the safe threshold, the inference process ends. A3: After implementing the adjustment schemes of the first and second layers, if the coal slime flotation process still has not returned to normal operation, the inference schemes of subsequent layers of the pyramid dynamic Bayesian network will continue to be adopted and implemented. The implementation process is the same as that of the adjustment scheme of the second layer. Only the newly emerging adjustment strategies will be implemented, and the previously implemented adjustment schemes will not be repeated until the coal slime flotation process returns to normal operation, at which point the inference process ends. When the adjustment scheme inferred from the last layer is implemented, if the coal slime flotation process still has not returned to normal operation, step eleven will be executed. Step 11: Select the sub-process with the second highest anomaly probability in the global reasoning process for reasoning, and re-execute Step 9 to continue using online anomaly data as evidence to reason out control decision schemes until the abnormal operating conditions are eliminated.

Images