Self-healing control method based on cross-space-time stable causal dynamic Bayesian network

Abstract

The invention discloses a self-healing control method based on a cross-space-time stable causal dynamic Bayesian network. Time features extracted from industrial process data by an LSTM algorithm and space features extracted based on full-flow physical unit distribution knowledge are fused in an interpolation mode; constructing a knowledge and data dual-driven cross-space-time causal dynamic Bayesian network model by combining stable learning and DBN based on the time features and the space features; discriminating a false causal relationship by using a cross-spatio-temporal causal dynamic Bayesian network model and combining spatio-temporal features; a sample weight is adjusted by adopting a sample reweighting technology; using the selected features and sample weights to train a cross-spatio-temporal causal dynamic Bayesian network model to obtain a cross-spatio-temporal stable causal DBN model; on-line abnormal data variables are used as evidences to be input into the cross-space-time stability causal DBN model, time slices where anomalies occur and variables causing abnormal working conditions are determined, and a control scheme is deduced; and implementing the control scheme. The method can quickly and effectively eliminate sudden abnormal working conditions in the process.

Description

Technical Field

[0001] The present invention belongs to the technical field of industrial process self-healing control, and specifically is a self-healing control method based on a cross-temporal and spatial stable causal dynamic Bayesian network. Background Art

[0002] Heavy medium coal preparation is a commonly used coal washing method. It uses heavy medium suspension to separate coal and gangue to improve coal quality and reduce impurity content. The basic principle of heavy medium coal preparation is to use the difference in density between coal and gangue. Through the action of heavy medium suspension, the coal with lower density floats on the surface of the suspension, while the gangue with higher density sinks to the bottom of the suspension, thereby achieving the separation of coal and gangue. In the heavy medium coal preparation process, the raw coal is first crushed and screened to remove large pieces of coal and gangue to ensure that the particle size of the selected coal is uniform. The pretreated coal enters the heavy medium separator. In the heavy medium suspension, the coal with lower density will float on the surface, while the gangue with higher density will sink to the bottom. The clean coal floating on the surface of the suspension and the gangue sinking to the bottom are dehydrated by dewatering screens respectively, and the clean coal and gangue are obtained after removing the suspension on the surface. Finally, the magnetite powder in the suspension is recovered by a magnetic separator, and then added back into the suspension for recycling. With the advancement of science and technology and the improvement of environmental protection requirements, the development trend of heavy medium coal preparation is moving towards automation and intelligence. The use of automation and safety control technology can effectively improve the sorting efficiency and ensure the safe and stable operation of the industrial process.

[0003] BN is a directed acyclic graph model that often reasons about uncertain knowledge in the form of solving probabilities, using network nodes and directed edges between nodes to represent random variables and their conditional dependencies. Given any BN structure, such as variable X→Y, the dependency and strength between X and Y are represented by a conditional probability table (CPT), and nodes X (parent node) to Y (child node) are connected by directed arcs. DBN is an extended form of BN used to model processes that change over time. Unlike traditional static BN, DBN introduces a time dimension, allowing the state of variables to change over time, and can model and reason about time series data. In DBN, the network structure can change over time, and the dependencies between variables can change over time. Summary of the invention

[0004] In view of the problems existing in the above-mentioned prior art, the present invention provides a self-healing control method based on a cross-temporal and spatial stable causal dynamic Bayesian network. The implementation process of the method is simple and the implementation cost is low. It can effectively solve the problem of false causality and can quickly and effectively eliminate sudden abnormal conditions in the process.

[0005] In order to achieve the above-mentioned purpose, the present invention provides a self-healing control method based on a cross-time and space stable causal dynamic Bayesian network, including a heavy medium coal preparation system, the heavy medium coal preparation system including a main grading screen, a secondary grading screen, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a first de-medium screen, a second de-medium screen, a magnetic separator, a magnetic medium storage bin, a qualified medium barrel, a circulation pump and a controller; the main grading screen is used to separate the raw coal into coarse coal and clean coal by screening; the feed port of the secondary grading screen is connected to the top discharge port of the main grading screen, and is used to remove coal slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the secondary The top discharge port of the grading screen is connected to output the screened clean coal to the feed port of the mixing barrel; the feed port of the mixing barrel is connected to the discharge port of the conveyor to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge port of the mixing barrel through a pipeline to output the mixture to the heavy medium cyclone; the feed port of the heavy medium cyclone is connected to the outlet end of the pressure pump to utilize the density difference of the materials to separate the low-density materials from the high-density materials to obtain overflow and underflow products; the feed port of the de-mediating screen is connected to the high-density discharge port of the heavy medium cyclone to separate the high-density materials from the low-density materials. The high-density discharge port discharges gangue from the high-density discharge port; the feed port of the de-medium screen 2 is connected to the low-density discharge port of the heavy medium cyclone for de-watering and de-mediuming the low-density material, and the high-density discharge port discharges clean coal; the feed port of the magnetic separator is respectively connected to the low-density discharge port of the de-medium screen 1 and the low-density discharge port of the de-medium screen 2 for removing iron impurities and separating the low-density material from the high-density material, the high-density discharge port outputs coal slime, and the low-density discharge port outputs dilute medium; the magnetic medium storage bin is used to store and quantitatively output magnetic media; the feed ports of the qualified medium barrels are respectively connected to the low-density discharge ports of the de-medium screen 1 and the low-density discharge ports of the de-medium screen 2 for removing iron impurities and separating the low-density material from the high-density material. The discharge port of the magnetic medium storage bin is connected to the circulating water filling pipeline to complete the preparation of the heavy medium suspension; the inlet end of the circulating pump is respectively connected to the discharge port of the qualified medium barrel and the water replenishment filling pipeline, and its outlet end is connected to the feed port of the mixing barrel to transport the heavy medium suspension to the mixing barrel; the controller is respectively connected to the main grading screen, the secondary grading screen, the conveyor, the mixing barrel, the pressure pump, the heavy medium cyclone, the first de-medium screen, the second de-medium screen, the magnetic separator, the magnetic medium storage bin, the qualified medium barrel, the control valve on the circulating water filling pipeline, the water replenishment valve on the water replenishment pipeline and the circulating pump to control the actions of each component;

[0006] The control method comprises the following steps:

[0007] Step 1: Full process data collection: Use a number of sensors installed at key monitoring nodes in the heavy medium coal preparation system to collect industrial process data in the coal slime flotation process in real time;

[0008] Step 2: Preprocess the data and generate spatiotemporal features. Use interpolation to fuse the temporal features extracted from the industrial process data by the LSTM algorithm and the spatial features extracted based on the knowledge of the physical unit distribution in the entire process. Map the temporal features and spatial features to the same dimension according to formula (1) to generate unified spatiotemporal features. In the process of extracting sub-unit features, the local information in each sub-unit is considered as a whole. Not only the local correlation within a single sub-unit is considered, but also the spatial features between different sub-units are extracted to better capture the spatial correlation between the operation sub-units.

[0009] Z(t i ,s j )=(1-α)T(t i )+αS(s j ) (1);

[0010] In the formula, α is a weight parameter used to control the degree of fusion of temporal features and spatial features, Z(t i ,s j ) is a comprehensive feature at a specific time t and spatial location s;

[0011] Step 3: Based on temporal and spatial features, we combine stable learning with DBN to build a knowledge- and data-driven cross-temporal and spatial causal dynamic Bayesian network model;

[0012] Step 4: Use the cross-spatial causal dynamic Bayesian network model combined with spatiotemporal features to identify false causal relationships. By finding features that maintain causal consistency under different data distributions, use DBN to capture causal relationships in time evolution and select causal relationships that remain stable in all units. At the same time, introduce lag nodes to learn the DBN structure in different time periods to capture causal relationships in spatiotemporal evolution and screen out causal relationships that remain stable in each spatiotemporal unit.

[0013] Step 5: Use sample reweighting technology to adjust sample weights, optimize sample distribution, reduce the impact of confounding factors and selection bias on causal relationships, and make the data better reflect the true causal relationship;

[0014] Step 6: Use the selected features and sample weights to train the cross-temporal causal dynamic Bayesian network model to obtain a cross-temporal stable causal DBN model;

[0015] Step 7: Input the online abnormal data variables as evidence into the cross-temporal and spatial stable causal DBN model, use the cross-temporal and spatial stable causal DBN model to perform causal inference, identify the true causal relationship between the variables, determine the time slice where the abnormality occurs and the variables that cause the abnormal condition, and infer the control plan;

[0016] Among them, the cross-temporal and spatial stable causal DBN model obtains the probability of each variable according to the following process, and determines the variable that causes the abnormal working condition according to the probability;

[0017] S71: For a given set of variables Z = (x1, x2, x3, ..., x n ), define a DBN(Z 1 ,Z→), where Z 1 Denotes the BN at the initial moment, and defines P(Z 1 ) is the probability at the initial moment, as shown in Formula 2;

[0018]

[0019] S72: Combining formula (3) and formula (4), obtain the conditional distribution of the DBN initial time and adjacent time slices;

[0020]

[0021] S73: At each time step, the state of the variable is inferred from the state of the previous time step and the model parameters, so that the node with time step t depends not only on the parent node of the current time step It also depends on the state at the previous moment and its parent node

[0022] For a given series of variables Z = (X1, X2, X3, ..., X n ), calculate the joint probability distribution P(Z) from t = 1 to T according to formula (5) 1:T );

[0023]

[0024] Step 8: The controller controls the heavy medium coal preparation system based on the inferred control scheme, and at the same time, monitors the abnormal operating conditions. If the abnormal operating conditions are eliminated, the normal working mode is entered. If the abnormal operating conditions are not eliminated, step 7 is re-executed, and online data is continued to be input as evidence information guided by knowledge to infer the process control scheme until the abnormal conditions are eliminated and the normal working mode is entered.

[0025] Modern industrial processes have obvious spatiotemporal characteristics, and the interactions of interrelated industrial units are complex, facing the challenges of high dynamics and strong correlation in industrial processes. Existing safe operation control methods based on BN often use static Bayes to model dynamic industrial processes, or only perform causal modeling from the time dimension, which will inevitably lead to causal false correlations between variables, making it difficult to accurately locate abnormal conditions in plant-level processes, and ultimately affect the decision-making effect. Therefore, the present application aims to solve the problem of causal false correlations between variables in the coal preparation process and construct a knowledge- and data-driven cross-temporal and spatial stable causal dynamic Bayesian network model. In order to ensure the safe and stable operation of the coal preparation process, the present application proposes an industrial process safety control method based on a cross-temporal and spatial stable causal dynamic Bayesian network. Compared with the prior art, the present invention proposes a self-healing control method based on a cross-temporal and spatial stable causal dynamic Bayesian network, which is knowledge-oriented and can accurately reveal the causal relationship between and within subunits using spatiotemporal information. First, a cross-temporal and spatial causal dynamic Bayesian network model is constructed using the temporal and spatial characteristics of the industrial process, and the temporal and spatial characteristics extracted by the LSTM algorithm and the plant-level process physical unit distribution knowledge are fused by interpolation. In the process of sub-unit feature extraction, the local information in each industrial unit is considered as a whole, not only focusing on the local correlation within a single unit, but also extracting the spatial features between different units to better capture the spatial correlation between operating units. However, it was found in the experiment that with the change of data distribution, the correlation between certain features and quality variables is not a true causal relationship. This false correlation leads to the low accuracy of the inferred decision-making scheme, and the abnormal working conditions cannot be effectively eliminated, which in turn affects the generalization ability of the model in practical applications. Therefore, this application combines stable learning (SL) with DBN to effectively remove the causal false correlation. By finding features that maintain causal consistency under different data distributions, DBN is used to capture the causal relationship in time evolution, and the causal relationship that remains stable in all units is selected. The DBN structure under different time periods is learned by introducing lag nodes, and sample reweighting is used to reduce the impact of confounding factors and selection bias on the model. By adjusting the sample weights, the data better reflects the true causal relationship and is close to the distribution under ideal conditions. Finally, the abnormal working condition data is input into the stable causal DBN model as evidence to infer the decision-making scheme that can eliminate the abnormal working conditions, and the decision-making scheme is converted into a robust control operation. The method has a simple implementation process and low implementation cost. It can effectively solve the problem of false causality and can quickly and effectively eliminate sudden abnormal conditions in the process. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a flow chart of the present invention;

[0027] Figure 2is a schematic diagram of a heavy medium coal preparation process in the present invention;

[0028] Figure 3 is the Granger causality diagram of the present invention;

[0029] Figure 4 It is the DBN causal relationship diagram of quality variables in the present invention;

[0030] Figure 5 It is the DBN structure diagram of all variables in the present invention;

[0031] Figure 6 It is a curve diagram of the simulation results of the present invention. DETAILED DESCRIPTION

[0032] The present invention provides a self-healing control method based on a cross-temporal and spatial stable causal dynamic Bayesian network, which mainly includes the following three main parts: the first part is data preprocessing; the time and space features extracted by the LSTM algorithm and the distribution knowledge of physical units in the whole process are fused by interpolation; the time features and the space features are mapped to the same dimension to generate unified time and space features; the second part is the construction of a cross-temporal and spatial stable causal DBN; the key features that maintain causal consistency under different time and space data distributions are found, and the sample reweighting method is used to reduce the influence of confounding factors and selection bias on the model, so as to ensure that the DBN can capture the real causal relationship that evolves over time in each time slice; the third part is decision reasoning and control scheme implementation; the abnormal data is input as evidence into the cross-temporal and spatial stable causal DBN model to infer a decision scheme that can eliminate abnormal working conditions; and the decision scheme is converted into a robust control operation and applied to the coal preparation process.

[0033] The present invention will be further described below in conjunction with the accompanying drawings.

[0034] like Figures 1 to 6As shown, the present invention provides a self-healing control method based on a cross-time and space stable causal dynamic Bayesian network, including a heavy medium coal preparation system, the heavy medium coal preparation system including a main grading screen, a secondary grading screen, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a first de-medium screen, a second de-medium screen, a magnetic separator, a magnetic medium storage bin, a qualified medium barrel, a circulation pump and a controller; the main grading screen is used to separate raw coal into coarse coal and clean coal by screening; the feed port of the secondary grading screen is connected to the top discharge port of the main grading screen, and is used to remove coal slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the secondary grading screen The top discharge port is connected to output the screened clean coal to the feed port of the mixing barrel; the feed port of the mixing barrel is connected to the discharge port of the conveyor to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge port of the mixing barrel through a pipeline to output the mixture to the heavy medium cyclone; the feed port of the heavy medium cyclone is connected to the outlet end of the pressure pump to utilize the density difference of the materials to separate the low-density materials from the high-density materials to obtain overflow and underflow products; the feed port of the de-mediating screen is connected to the high-density discharge port of the heavy medium cyclone to separate the high-density materials Dehydration and demediation are carried out, and gangue is discharged from its high-density discharge port; the feed port of the demediation screen No. 2 is connected to the low-density discharge port of the heavy medium cyclone, which is used to dehydrate and demediate the low-density material, and the high-density discharge port discharges clean coal; the feed port of the magnetic separator is respectively connected to the low-density discharge port of the demediation screen No. 1 and the low-density discharge port of the demediation screen No. 2, which is used to remove iron impurities and separate the low-density material from the high-density material, and the high-density discharge port outputs coal slime, and the low-density discharge port outputs dilute medium; the magnetic medium storage bin is used to store and quantitatively output magnetic media; the feed ports of the qualified medium barrels are respectively connected to the magnetic separators. The discharge port of the qualified medium storage bin is connected to the circulating water filling pipeline to complete the preparation of the heavy medium suspension; the inlet end of the circulating pump is respectively connected to the discharge port of the qualified medium barrel and the water replenishment filling pipeline, and its outlet end is connected to the feed port of the mixing barrel to transport the heavy medium suspension to the mixing barrel; the controller is respectively connected to the main grading screen, the secondary grading screen, the conveyor, the mixing barrel, the pressure pump, the heavy medium cyclone, the first de-medium screen, the second de-medium screen, the magnetic separator, the magnetic medium storage bin, the qualified medium barrel, the control valve on the circulating water filling pipeline, the water replenishment valve on the water replenishment pipeline and the circulating pump to control the actions of each component;

[0035] The control method comprises the following steps:

[0036] Step 1: Full process data collection: Use a number of sensors installed at key monitoring nodes in the heavy medium coal preparation system to collect industrial process data in the coal slime flotation process in real time;

[0037] Step 2: Preprocess the data and generate spatiotemporal features. Use interpolation to fuse the temporal features extracted from the industrial process data by the LSTM algorithm and the spatial features extracted based on the knowledge of the physical unit distribution in the entire process. Map the temporal features and spatial features to the same dimension according to formula (1) to generate unified spatiotemporal features. In the process of extracting sub-unit features, the local information in each sub-unit is considered as a whole. Not only the local correlation within a single sub-unit is considered, but also the spatial features between different sub-units are extracted to better capture the spatial correlation between the operation sub-units.

[0038] Z(t i ,s j )=(1-α)T(t i )+αS(s j ) (1);

[0039] In the formula, α is a weight parameter used to control the degree of fusion of temporal features and spatial features, Z(t i ,s j ) is a comprehensive feature at a specific time t and spatial location s;

[0040] Step 3: Based on temporal and spatial features, we combine stable learning (SL) with DBN to build a knowledge- and data-driven cross-temporal and spatial causal dynamic Bayesian network model;

[0041] Step 4: Use the cross-spatial causal dynamic Bayesian network model combined with spatiotemporal features to identify false causal relationships. By finding features that maintain causal consistency under different data distributions, use DBN to capture causal relationships in time evolution and select causal relationships that remain stable in all units. At the same time, introduce lag nodes to learn the DBN structure in different time periods to capture causal relationships in spatiotemporal evolution and screen out causal relationships that remain stable in each spatiotemporal unit.

[0042] Step 5: Use sample reweighting technology to adjust sample weights, optimize sample distribution, reduce the impact of confounding factors and selection bias on causal relationships, and make the data better reflect the true causal relationship;

[0043] Step 6: Use the selected features and sample weights to train the cross-temporal causal dynamic Bayesian network model to obtain a cross-temporal stable causal DBN model;

[0044] Step 7: Input the online abnormal data variables as evidence into the cross-temporal and spatial stable causal DBN model, use the cross-temporal and spatial stable causal DBN model to perform causal inference, identify the true causal relationship between the variables, determine the time slice where the abnormality occurs and the variables that cause the abnormal condition, and infer the control plan;

[0045] DBN is an extended form of BN, which is used to model processes that change over time. Unlike traditional static BN, DBN introduces the time dimension, allowing the state of variables to change over time, and can model and reason about time series data; in DBN, the network structure can change over time, and the dependencies between variables can change over time. In order to simulate the time evolution relationship, the continuous time slice is divided into a series of discrete time slices;

[0046] Among them, the cross-temporal and spatial stable causal DBN model obtains the probability of each variable according to the following process, and determines the variable that causes the abnormal working condition according to the probability;

[0047] S71: For a given set of variables Z = (x1, x2, x3, ..., x n ), define a DBN(Z 1 ,Z→), where Z 1 Denotes the BN at the initial moment, and defines P(Z 1 ) is the probability at the initial moment, as shown in Formula 2;

[0048]

[0049] S72: Combining formula (3) and formula (4), obtain the conditional distribution of the DBN initial time and adjacent time slices;

[0050]

[0051] S73: At each time step, the state of the variable is inferred from the state of the previous time step and the model parameters, so that the node with time step t depends not only on the parent node of the current time step It also depends on the state at the previous moment and its parent node

[0052] For a given series of variables Z = (X1, X2, X3, ..., X n ), calculate the joint probability distribution P(Z) from t = 1 to T according to formula (5) 1:T );

[0053]

[0054] Step 8: The controller controls the heavy medium coal preparation system based on the inferred control scheme, and at the same time, monitors the abnormal operating conditions. If the abnormal operating conditions are eliminated, the normal working mode is entered. If the abnormal operating conditions are not eliminated, step 7 is re-executed, and online data is continued to be input as evidence information guided by knowledge to infer the process control scheme until the abnormal conditions are eliminated and the normal working mode is entered.

[0055] The technical solution of the present invention is described in detail below in conjunction with embodiments, and its feasibility is verified.

[0056] Example:

[0057] The schematic diagram of the heavy medium coal preparation unit process is as follows: Figure 2 As shown in Table 1, the heavy medium coal preparation process mainly includes 6 measurement variables and 2 operation variables. The detailed variable allocation is shown in Table 1 and Table 2.

[0058] Granger causality test method is used to conduct in-depth causal mining analysis on the coal preparation process from the time level. Granger causality test results are judged by p-value statistics. A p-value less than 0.05 indicates that the test results have a statistically significant causal relationship. Figure 3 As shown in Figure 3, Granger causality analysis reveals the dynamic interaction between each node, providing an important basis for industrial process control; the detailed analysis of the main nodes of the heavy medium coal preparation unit and their interrelationships is shown in Table 3. The p value of node A to node H is 0.7864, indicating that the effect of coal input on the cyclone overflow ash is not significant, and increasing the amount of raw coal input will not directly improve the screening efficiency; similarly, the p value of node B to node H is 0.1674, and the effect of coal input on flotation effect is not significant, further emphasizing that the role of coal input in the entire flotation process is limited. The p value of node C to node H is 0. .0004, indicating that the change of the single-layer screen discharge flow rate will directly affect the separation effect of the hydrocyclone. The p-value between node E and node H is 5.764171e-132, indicating that the pulp density has a significant effect on the ash separation effect of the hydrocyclone. In addition, the p-value between node G and node H is 0.0042, indicating that the hydrocyclone flow rate has a significant effect on the ash content, which means that the change of flow rate may affect the separation efficiency of the hydrocyclone. Therefore, in the subsequent spatiotemporal causal relationship mining, it is necessary to focus on the nodes with significant dynamic changes to ensure that the abnormal cause can be quickly inferred when an abnormality occurs in the heavy medium process.

[0059] Table 1: Measured variables of heavy medium coal preparation unit

[0060]

[0061]

[0062] Table 2: Heavy Medium Coal Preparation Unit Operating Variables

[0063]

[0064] Table 3: Ranger causality test results for heavy medium coal preparation unit

[0065]

[0066] The cross-temporal and spatial stable DBN model further enhances the dynamic adaptability of the model by capturing the causal dynamics of process variables changing over time and selecting causal relationships that remain stable in all units. In the verification of the heavy medium coal preparation unit and the coal slime flotation unit, the model introduced lag nodes. Observing the lag effect can help identify the true causal relationship, thereby providing a causal basis for the control plan. In order to reduce the adverse effects of confounding factors and selection bias on the performance of the model, the cross-temporal and spatial stable DBN model uses a sample reweighting method. By adjusting the sample weights, the data can more accurately map the true causal structure and approach the ideal distribution state. The DBN structure of the heavy medium coal preparation unit is as follows: Figure 4 and Figure 5 As shown in the figure, when a key variable or its dependency is detected to be abnormal, the model can respond immediately and take corresponding adjustment measures to ensure the stable and safe operation of the coal preparation process. Comparing the Granger causality analysis method with the cross-temporal and spatial stable DBN method proposed in this application, it is found that the causal relationship mined by Granger causality analysis is incomplete. The method proposed in this application can mine a complete causal relationship. The quality variable H of the heavy medium coal preparation unit depends on variables C and F, and has no causal relationship with other variables, while the Granger causal analysis method learns false causal relationships E and G.

[0067] After learning the causal network structure of the heavy medium coal preparation unit, parameter learning is performed, and the conditional probability table of each node is given to provide a basis for the inference control scheme. In order to ensure the safe and stable operation of the entire process, the abnormal time slice is inferred by the parameter learning method when an abnormal condition occurs in the coal preparation process. Then, the conditional probability distribution is used to determine the variables causing the abnormal condition, infer the state of each variable at the current moment, and infer the optimal safety control scheme. When an abnormal condition occurs in the heavy medium coal preparation unit, the conditional probability of the overflow ash H of the quality variable is shown in Table 4. Among them, the process variable 1 indicates normal, 2 indicates that the abnormal value is small, and 3 indicates that the abnormal value is large. The abnormal evidence state is set to 1 in the reasoning.

[0068] Table 4: Conditional probability table of overflow ash H in heavy medium coal preparation unit

[0069]

[0070] After the parameter learning is completed, the time slice of the abnormality of the heavy medium coal preparation unit is determined, and the expected quality index and the state value of the variable are input into the network structure for reasoning, and the posterior probability of the operation node is obtained. According to the principle of maximum posterior probability, the control variable adjustment scheme of the heavy medium coal preparation unit is inferred, as shown in Table 5. Six industrial process cases are selected to verify the cross-temporal and spatial stable causal dynamic Bayesian network security control scheme. Among them, "↑" indicates that the variable adjustment direction increases, "ˉ" indicates that the adjustment direction decreases, and "-" indicates unchanged. In the heavy medium coal preparation unit, the changes in key equipment such as the single-layer screen discharge flow rate and the cyclone medium density are the key factors that directly cause the abnormal changes in the overflow ash content of the cyclone. The single-layer screen discharge flow rate affects the amount and composition of the material entering the cyclone. Too high or too low will destroy the sorting environment, thereby affecting the ash content. The cyclone medium density determines the density limit of the sorting and is the key to controlling the overflow ash content. Too high or too low density will change the sorting limit and affect the ash content. When adjusting the joint equipment that causes abnormal overflow ash content, other process variables also need to be adjusted at the same time to ensure that the single-layer screen discharge flow rate and cyclone medium density are within the optimal range. For example, in Case 1, when the cyclone overflow ash content is abnormal, according to the variable adjustment plan, it is necessary to quickly reduce the discharge flow rate and cyclone medium density, and increase the slurry density of the mixed medium barrel.

[0071] Table 5: Control variable adjustment strategy for heavy medium coal preparation unit

[0072]

[0073]

[0074] In order to verify the effectiveness of the cross-temporal and spatial stable DBN modeling method and the abnormal operating condition adjustment strategy in Table 5, the above six cases were simulated on the coal preparation simulation platform. When an abnormal operating condition occurs, observe whether the cyclone overflow ash content can be restored to the normal range. According to the guidance of the on-site operators, the threshold of the fine ore position of the heavy medium unit is set to 10.8%. After exceeding the threshold, the abnormal data is input as evidence into the established cross-temporal and spatial stable causal DBN model, and the decision plan is inferred and converted into a control operation. Figure 6 It can be seen that in the study of each case, when an abnormality occurs, the decision-making plan inferred by the method of this application can quickly eliminate the abnormality and ensure that the heavy medium coal preparation unit can operate safely and stably.

[0075] The present invention proposes a self-healing control method based on a cross-space stable causal dynamic Bayesian network. This method combines stable learning to mine the causal consistency characteristics under different space-time data distributions, and uses the dynamic Bayesian network to introduce lag nodes to capture the causal relationship in the space-time evolution, and screen out the causal relationship that remains stable in each space-time unit. In addition, the sample reweighting technology is used to reduce the impact of confounding factors and selection bias on the causal relationship. By adjusting the sample weights, the space-time data characteristics can better reflect the causal characteristics under ideal conditions. The method of this application has been verified in 6 working condition cases of the coal preparation process and compared with the traditional Granger causal analysis method. Experimental results show that the proposed method has significant advantages in eliminating false causal relationships and handling sudden abnormal conditions.

Claims

1. A self-healing control method based on a cross-time and space stable causal dynamic Bayesian network, comprising a heavy medium coal preparation system, the heavy medium coal preparation system comprising a main grading screen, a secondary grading screen, a conveyor, a mixing barrel, a pressure pump, a heavy medium cyclone, a first de-medium screen, a second de-medium screen, a magnetic separator, a magnetic medium storage bin, a qualified medium barrel, a circulation pump and a controller; the main grading screen is used to separate raw coal into coarse coal and clean coal by screening; the feed port of the secondary grading screen is connected to the top discharge port of the main grading screen, and is used to remove coal slime impurities in the clean coal by screening; the feed end of the conveyor is connected to the top discharge port of the secondary grading screen The feed port of the mixing barrel is connected to the discharge port of the conveyor to mix the clean coal with the heavy medium suspension to form a mixture; the inlet end of the pressure pump is connected to the discharge port of the mixing barrel through a pipeline to output the mixture to the heavy medium cyclone; the feed port of the heavy medium cyclone is connected to the outlet end of the pressure pump to utilize the density difference of the materials to separate the low-density materials from the high-density materials to obtain overflow and underflow products; the feed port of the de-mediating screen is connected to the high-density discharge port of the heavy medium cyclone to de-mediate the high-density materials. The dewatering operation is carried out, and its high-density discharge port discharges gangue; the feed port of the dewatering screen 2 is connected to the low-density discharge port of the heavy medium cyclone, which is used to dewater and de-mediumize the low-density material, and its high-density discharge port discharges clean coal; the feed port of the magnetic separator is respectively connected to the low-density discharge port of the dewatering screen 1 and the low-density discharge port of the dewatering screen 2, which is used to remove iron impurities and separate the low-density material from the high-density material, and its high-density discharge port outputs coal slime, and its low-density discharge port outputs dilute medium; the magnetic medium storage bin is used to store and quantitatively output magnetic media; the feed ports of the qualified medium barrels are respectively connected to the magnetic medium The discharge port of the qualified medium storage bin is connected with the circulating water filling pipeline to complete the preparation of the heavy medium suspension; the inlet end of the circulating pump is respectively connected with the discharge port of the qualified medium barrel and the water replenishment filling pipeline, and the outlet end is connected with the feed port of the mixing barrel to transport the heavy medium suspension to the mixing barrel; the controller is respectively connected with the main grading screen, the secondary grading screen, the conveyor, the mixing barrel, the pressure pump, the heavy medium cyclone, the first de-medium screen, the second de-medium screen, the magnetic separator, the magnetic medium storage bin, the qualified medium barrel, the control valve on the circulating water filling pipeline, the water replenishment valve on the water replenishment pipeline and the circulating pump to control the actions of each component; It is characterized in that The following steps are involved: Step 1: Full process data collection: Use a number of sensors installed at key monitoring nodes in the heavy medium coal preparation system to collect industrial process data in the coal slime flotation process in real time; Step 2: Preprocess the data and generate spatiotemporal features; The temporal features extracted from the industrial process data by the LSTM algorithm and the spatial features extracted based on the knowledge of the distribution of the physical units in the whole process are fused by interpolation. The temporal features and spatial features are mapped to the same dimension according to formula (1) to generate unified spatiotemporal features. In the process of extracting sub-unit features, the local information in each sub-unit is considered as a whole. Not only the local correlation within a single sub-unit is paid attention to, but also the spatial features between different sub-units are extracted to better capture the spatial correlation between the operation sub-units. Z(t i ,s j )=(1-α)T(t i )+αS(s j ) (1); In the formula, α is a weight parameter used to control the degree of fusion of temporal features and spatial features, Z(t i ,s j ) is a comprehensive feature at a specific time t and spatial location s; Step 3: Based on temporal and spatial features, we combine stable learning with DBN to build a knowledge- and data-driven cross-temporal and spatial causal dynamic Bayesian network model; Step 4: Use the cross-spatial causal dynamic Bayesian network model combined with spatiotemporal features to identify false causal relationships. By finding features that maintain causal consistency under different data distributions, use DBN to capture causal relationships in time evolution and select causal relationships that remain stable in all units. At the same time, introduce lag nodes to learn the DBN structure in different time periods to capture causal relationships in spatiotemporal evolution and screen out causal relationships that remain stable in each spatiotemporal unit. Step 5: Use sample reweighting technology to adjust sample weights, optimize sample distribution, reduce the impact of confounding factors and selection bias on causal relationships, and make the data better reflect the true causal relationship; Step 6: Use the selected features and sample weights to train the cross-temporal causal dynamic Bayesian network model to obtain a cross-temporal stable causal DBN model; Step 7: Input the online abnormal data variables as evidence into the cross-temporal and spatial stable causal DBN model, use the cross-temporal and spatial stable causal DBN model to perform causal inference, identify the true causal relationship between the variables, determine the time slice where the abnormality occurs and the variables that cause the abnormal condition, and infer the control plan; Among them, the cross-temporal and spatial stable causal DBN model obtains the probability of each variable according to the following process, and determines the variable that causes the abnormal working condition according to the probability; S71: For a given set of variables Z = (x1, x2, x3, ..., x n ), define a DBN(Z 1 ,Z→), where Z 1 Denotes the BN at the initial moment, and defines P(Z 1 ) is the probability at the initial moment, as shown in Formula 2; S72: Combining formula (3) and formula (4), obtain the conditional distribution of the DBN initial time and adjacent time slices; S73: At each time step, the state of the variable is inferred from the state of the previous time step and the model parameters, so that the node with time step t depends not only on the parent node of the current time step It also depends on the state at the previous moment and its parent node For a given series of variables Z = (X1, X2, X3, ..., X n ), calculate the joint probability distribution P(Z) from t = 1 to T according to formula (5) 1:T ); Step 8: The controller controls the heavy medium coal preparation system based on the inferred control scheme, and at the same time, monitors the abnormal operating conditions. If the abnormal operating conditions are eliminated, the normal working mode is entered. If the abnormal operating conditions are not eliminated, step 7 is re-executed, and online data is continued to be input as evidence information guided by knowledge to infer the process control scheme until the abnormal conditions are eliminated and the normal working mode is entered.

Images