Self-healing control method for complex industrial processes based on elastic time-varying bayesian networks

Abstract

The application discloses a complex industrial process self-recovery control method based on an elastic time-varying Bayesian network, collects field data and carries out analysis and processing; the ADF square root algorithm is used to test the stationarity of the data, if the data is non-stationary, the data is subjected to difference processing, and the data is stabilized; after the stabilized data and the original stationary data are fused, data correlation analysis is carried out; through the correlation analysis, the relationship strength and direction between multiple variables are determined; the data is input into the elastic time-varying dynamic Bayesian network model, the network structure and parameters are learned; according to the learned structure and parameters, the abnormal time slice is accurately positioned in combination with the probability density function, and the process variable causing the abnormality is determined; the abnormal data is input into the dynamic Bayesian network as evidence to infer a decision scheme capable of eliminating the abnormal working condition, and the decision scheme is converted into a control operation. The method can make effective control decisions after the abnormality of the slime flotation industrial process.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of process industry self-recovery control, and particularly relates to a complex industrial process self-recovery control method based on an elastic time-varying Bayesian network. BACKGROUND

[0002] With the continuous improvement of industrial automation and intelligence, the challenges faced by process industry process control are increasingly complex and diverse. Complex industrial processes are industrial production processes with high dynamicity, nonlinearity, multivariable, time-varying characteristics and strong coupling. The process usually involves multiple interrelated subsystems and operation steps, and various factors inside and outside the system jointly affect its running state and performance. Due to its dynamicity and time-varying nature, complex industrial processes are mostly non-stationary industrial processes, which are affected by various factors during operation, resulting in changes in system characteristics and behavior over time. These factors include but are not limited to changes in environmental conditions, aging and wear of equipment, fluctuations in raw material characteristics, adjustments in operating conditions, etc. Non-stationarity is reflected in the dynamic changes of system parameters and states, and when the system is abnormal, traditional steady-state modeling and control methods are difficult to apply.

[0003] Non-stationary industrial processes usually exhibit characteristics of time-varying parameters and structures, which lead to changes in system state and parameters, making it difficult for a single fixed model to accurately describe and judge system behavior, thereby affecting control effect and decision accuracy. In addition, the complexity and multivariate correlation characteristics of industrial processes also increase the difficulty of process control and safety management. How to achieve efficient monitoring and safe control of non-stationary industrial processes in a dynamic environment has become an important problem to be solved in the field of industrial process control. Existing safety control methods for non-stationary complex industrial processes can be divided into model-based methods, knowledge-driven methods and data-driven methods. Model-driven methods are based on physical and chemical principles, and the model has strong interpretability and is suitable for specific process. However, the model is complex to establish and requires a deep understanding of the system mechanism. For nonlinear and non-stationary processes, modeling is too difficult, and model updating and maintenance are difficult, making it difficult to adapt to rapidly changing working conditions. Commonly used model-driven methods include first-principle modeling, state-space modeling, and mechanism modeling. Knowledge-driven methods use expert knowledge and experience, and the rules are clear and the reasoning process is transparent, which is suitable for known working conditions and fault types. However, the rule base is difficult to establish and maintain, and it cannot handle new unknown faults, and it has limited ability to handle complex nonlinear relationships. Knowledge-driven methods include expert systems, fuzzy logic, and rule-based reasoning. Data-driven methods can handle large amounts of data, capture complex patterns and nonlinear relationships, and have strong adaptability, which can be applied to various process. However, it relies on a large amount of high-quality data, which is difficult to interpret and understand, and has low sensitivity to new faults and working conditions. Commonly used data-driven methods include neural networks, machine learning, support vector machines, and deep learning. Therefore, it is necessary to combine the advantages of the above methods and comprehensively use expert knowledge, data analysis and model simulation to make multi-level and multi-dimensional decisions to achieve safety control of non-stationary complex industrial processes. SUMMARY

[0004] In view of the problems existing in the prior art, the present application provides a complex industrial process self-healing control method based on an elastic time-varying Bayesian network, which has simple steps and high reliability. It can quickly make effective control decisions after an abnormality occurs in the coal slime flotation industrial process, ensuring the safe and stable operation of the industrial process and providing automatic decision support for abnormal working conditions of the flotation process.

[0005] In order to achieve the above object, the application provides a complex industrial process self-recovery control method based on elastic time-varying Bayesian network, including a heavy medium coal preparation system, the heavy medium coal preparation system comprising a conveyor, a heavy medium cyclone, an arc-shaped screen A, a clean coal medium screen, an arc-shaped screen B, a gangue medium screen, a dilution medium barrel, a dilution medium pump, a dilution medium magnetic separator, a water adding pump, a medium adding pump, a medium adding magnetic separator, a qualified medium barrel and a pressure pump; the feed inlet of the heavy medium cyclone is connected with the discharge end of the conveyor, the feed inlet of the arc-shaped screen A is connected with the overflow outlet of the heavy medium cyclone, the feed inlet of the clean coal medium screen is connected with the discharge outlet of the arc-shaped screen A, the upper outlet of the clean coal medium screen is used for outputting clean coal, and the lower outlet of the clean coal medium screen is respectively connected with a first liquid discharge pipeline and a second liquid discharge pipeline; the feed inlet of the arc-shaped screen B is connected with the underflow outlet of the heavy medium cyclone, the feed inlet of the gangue medium screen is connected with the discharge outlet of the arc-shaped screen B, the upper outlet of the gangue medium screen is used for outputting coal gangue, and the lower outlet of the gangue medium screen is respectively connected with a third liquid discharge pipeline and a fourth liquid discharge pipeline; the feed inlets at the top of the dilution medium barrel are respectively connected with the outlet ends of the first liquid discharge pipeline and the third liquid discharge pipeline; the inlet end of the dilution pump is connected with the dilution medium outlet at the bottom of the dilution medium barrel; the inlet end of the dilution medium magnetic separator is connected with the outlet end of the dilution pump; the water inlet end of the water adding pump is connected with a water source; the inlet end of the medium adding pump is connected with a medium source; the inlet end of the medium adding magnetic separator is connected with the outlet end of the medium adding pump; the feed inlets at the top of the qualified medium barrel are respectively connected with the outlet ends of the second liquid discharge pipeline, the fourth liquid discharge pipeline, the dilution medium magnetic separator, the water adding pump and the medium adding magnetic separator; the inlet end of the pressure pump is connected with the outlet end at the bottom of the qualified medium barrel, and the outlet end of the pressure pump is connected with the feed inlet of the heavy medium cyclone.

[0006] The method comprises the following steps:

[0007] Step one: establishing a heavy medium coal preparation system and determining variables;

[0008] The heavy medium coal preparation system is established by using a conveyor, a heavy medium cyclone, an arc-shaped screen A, a clean coal medium screen, an arc-shaped screen B, a gangue medium screen, a dilution medium barrel, a dilution medium pump, a dilution medium magnetic separator, a water adding pump, a medium adding pump, a medium adding magnetic separator, a qualified medium barrel and a pressure pump;

[0009] According to the operation characteristics of the heavy medium coal preparation system, the measurement variables and the operation variables in the heavy medium coal preparation system are determined; the measurement variables are six, and are in turn double-layer screen discharge flow, single-layer screen discharge flow, mixed medium barrel slurry density, cyclone medium density, qualified medium barrel medium density and cyclone overflow ash content; the operation variables are two, and are in turn raw coal bunker coal inflow and cyclone medium inflow pressure;

[0010] Step two: collecting coal preparation process data and using ADF square root algorithm to test the stationarity of the industrial process;

[0011] S21: In the process of running the dense medium coal preparation system, the sensor group arranged in the dense medium coal preparation system is used to collect coal preparation process data and output to the controller, and the controller processes the coal preparation process data according to the internal clock module to obtain time series data;

[0012] S22: The ADF square root algorithm is used to test the stationarity of the time series data;

[0013] S221: According to the characteristics of the time series, an ADF test model with constant term and time trend is selected, and the first-order difference Δy of the time series is obtained by using formula (1) t ;

[0014]

[0015] In the formula, Δy t represents y t -y t-1 ; y t-1 is the first-order lag time series value; μ is the constant term; λt is the time trend term; α is the most important coefficient in the ADF test, which is used to judge the existence of the unit root; β i is the coefficient of the difference term, which is used to eliminate the high-order autocorrelation in the sequence; ∈ t is the white noise error term; and p is the lag order;

[0016] S222: The parameters in the ADF test model are estimated by using the least squares method;

[0017] S223: The ADF statistic is calculated, which is in the form of the estimated α divided by its standard error;

[0018] S224: The hypothesis test is constructed, the null hypothesis H0 is that the unit root exists, at this time, the data is non-stationary, the alternative hypothesis H1 is that the unit root does not exist, at this time, the data is stationary;

[0019] The calculated ADF statistic is compared with the corresponding critical value; the critical value comes from the Dickey-Fuller distribution table; if the ADF statistic is less than the critical value, the null hypothesis is rejected, and it is considered that the time series is stationary; if the ADF statistic is greater than the critical value, the null hypothesis cannot be rejected, and it is considered that the time series is non-stationary;

[0020] Step three: stationary treatment of non-stationary data;

[0021] For non-stationary data, the trend and periodic components in the time series are eliminated by using the difference method until the data reaches the stationary state; in this process, the unit root test or the observation of the time series diagram of the differentiated data is used to check whether the differentiated data reaches the stationary state;

[0022] Step four: structure learning is performed to establish an elastic time-varying dynamic Bayesian network model;

[0023] S41: the data after the stationary treatment is fused with the original stationary data to obtain fused data, the fused data is analyzed for correlation to determine the relationship strength and direction between multiple variables, and the relationship between the variables is displayed in the form of a graph;

[0024] S42: the key variables affecting the coal separation effect are taken as nodes in the elastic time-varying dynamic Bayesian network, and an elastic time-varying dynamic Bayesian network model reflecting the causal relationship between industrial process variables is constructed to capture the dependence relationship changes of the dense medium coal separation system at different time periods;

[0025] S43: the fused data is used to train the elastic time-varying dynamic Bayesian network model to mine the causal relationship behind the data and determine the dynamic dependence relationship between the nodes;

[0026] Step five: elastic time-varying dynamic Bayesian network parameter learning;

[0027] Given the conditional probability table of each node, the elastic time-varying dynamic Bayesian network is subjected to parameter learning to obtain a non-stationary elastic time-varying dynamic Bayesian network model, which provides a basis for reasoning control decision schemes; when an abnormal working condition occurs in the coal separation process, the probability density function is used to reason out the abnormal time slice; then, the conditional probability distribution is used to determine the variables causing the abnormal working condition, infer the state of each variable at the current time, and reason out the optimal safe control decision scheme;

[0028] Step six: determining the time slice where the abnormality occurs and the variables causing the abnormal working condition;

[0029] The online abnormal data phenomenon variables are input into the non-stationary elastic time-varying dynamic Bayesian network model, which accurately locates the time slice where the abnormality occurs in the dense medium coal separation process according to the learned structure and parameters, and determines the process variables causing the abnormality;

[0030] Step seven: reasoning out the control decision scheme;

[0031] The expected quality indicators and variable state values are input into the local Bayesian network structure for reasoning to obtain the posterior probability of the operating node, and according to the principle of maximum posterior probability, the adjustment scheme of the dense medium coal separation process decision, i.e., the control decision, is obtained;

[0032] Step eight: implementing the control decision;

[0033] The controller adjusts the operation parameters of the dense medium coal preparation system according to the obtained control decision, and if the abnormal working condition is eliminated, enters the normal working mode, and if the abnormal working condition is not eliminated, re-collects the online abnormal data, and re-executes step seven, continues to infer the control decision scheme by using the online abnormal data phenomenon variables as evidence information until the abnormal working condition is eliminated.

[0034] In the coal preparation process, the characteristics of raw materials, the performance of equipment and environmental conditions can all affect the entire production process, and the system's dependencies and parameters can change over time, making the entire process non-stationary. Simple self-healing control methods based on data or expert knowledge alone are difficult to adapt to highly dynamic, multi-variable, time-varying, and strongly coupled non-stationary complex industrial production processes. The elastic time-varying dynamic Bayesian network model is based on a vector autoregressive model and can be used to describe the time dynamics between multiple variables. To address the self-healing control problem of non-stationary complex industrial processes under abnormal operating conditions, the present invention provides a self-healing control decision model based on an elastic time-varying dynamic Bayesian network. The elastic time-varying dynamic Bayesian network decision model combines data-driven and mechanism-based knowledge modeling, and is a model and data jointly driven modeling method. In the case of abundant data, the model can automatically adjust the parameters by learning from the data; in the case of scarce data, prior knowledge can be used to supplement it to ensure the effectiveness of the model. And it can be updated in real time and adapt to the changes of non-stationary processes. By dynamically adjusting the model parameters and structure, the state changes of the system at different time periods can be accurately reflected, ensuring accurate description of the system state and prediction of the system behavior. The elastic time-varying dynamic Bayesian network can handle complex dependencies between multiple variables, which is particularly important in dealing with the multi-variable coupling problem in complex industrial processes, as multiple process variables are usually interdependent and correlated. The model can capture long-term trend changes and sudden abnormal conditions in industrial processes when dealing with time-varying parameters and structures, making the model more flexible and accurate. Dynamic Bayesian networks have the advantage of handling uncertainty, and can quantify and manage uncertainty through probability distributions. In the elastic time-varying dynamic Bayesian network model, uncertainty is described by conditional probability distributions, which can effectively deal with randomness and noise in industrial processes. To achieve abnormal detection and decision reasoning for non-stationary complex industrial processes, the main process of the elastic time-varying dynamic Bayesian network model is divided into three parts: the first part is data preprocessing. The collected field data is analyzed and processed, and the ADF square root algorithm is used to test the stationarity of the data. If the data is non-stationary, the data is differentiated to stabilize the data. In the coal preparation process, understanding the stationarity of the data is crucial to establishing an effective control strategy and improving the separation efficiency. The trend in non-stationary data can interfere with the decision-making ability of the model and easily lead to overfitting. Converting a non-stationary data process to a stationary data process has significant benefits for control decisions. This not only improves the accuracy and robustness of the prediction model, but also simplifies the calculation and analysis process. Through stabilization, the data's interpretability and operability are enhanced, and the sensitivity and reliability of the control strategy are improved. Thus, through the data stabilization technique, not only the effectiveness of the control strategy is ensured, but also the separation efficiency is improved.The data after the smoothing treatment is fused with the original stationary data. After data fusion, correlation analysis is carried out, and through the correlation analysis, the relationship and dependency between different variables are evaluated, which can provide intuitive basis for control decision. The second part is the structure learning and parameter learning of the elastic time-varying dynamic Bayesian network. A network structure reflecting the causal relationship between the variables of the industrial process is constructed, and the change of the dependency relationship of the system in different time periods is captured. According to the learned structure and parameters, the abnormal time slice is accurately located, and the process variable causing the abnormality is determined. The third part is decision reasoning. The abnormal data is input into the elastic time-varying dynamic Bayesian network as evidence to reason out the decision scheme that can eliminate the abnormal working condition, and the decision scheme is converted into a robust control operation and applied to the coal preparation process.

[0035] The method has simple steps and high reliability, can make effective control decisions quickly after the abnormality of the slime flotation industrial process, can ensure the safe and stable operation of the industrial process, and can provide automatic decision support for the abnormal working condition of the flotation process. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a flowchart of the present application;

[0037] Figure 2 is a flowchart of the present application;

[0038] Figure 3 is an ADF non-stationary test result graph in the present application;

[0039] Figure 4 is a visual representation of the data after smoothing in the present application;

[0040] Figure 5 is a structure diagram of the elastic time-varying dynamic Bayesian network in the present application;

[0041] Figure 6 is a scatter plot of the abnormal data distribution of the heavy medium coal preparation process in the present application;

[0042] Figure 7 is a histogram of the abnormal data distribution of the heavy medium coal preparation process in the present application;

[0043] Figure 8 Heavy medium model overflow ash content curve change. DETAILED DESCRIPTION

[0044] The present application will be further described below with reference to the accompanying drawings.

[0045] Heavy medium coal preparation is a common coal washing method, in the process of heavy medium coal preparation, raw coal is first crushed and screened to remove large pieces of coal and gangue, to ensure the uniformity of the particle size of the coal for preparation. The pretreated coal enters the heavy medium separator, the heavy medium separator is filled with heavy medium suspension, in the heavy medium suspension, the coal with smaller density will float on the surface of the suspension, while the gangue with larger density will sink to the bottom of the suspension, so as to separate the coal and gangue by using the difference in density, and through the flotation of the heavy medium suspension, the purpose of improving the quality of coal and reducing the impurity content can be achieved. The clean coal floating on the surface of the suspension and the gangue sinking to the bottom are respectively subjected to dewatering treatment by the medium removal screen, and after removing the surface suspension, the clean coal and the gangue are obtained. Finally, the magnetite powder in the suspension is recovered by the magnetic separator, and then the magnetite powder is added into the suspension for recycling. With the continuous improvement of scientific and technological progress and environmental protection requirements, the development trend of heavy medium coal preparation is towards automation and intelligentization, and it is very important to adopt automatic and safety control technology to improve the separation efficiency and ensure the safe and stable operation of the industrial process.

[0046] As shown in Figures 1 to 8 The application provides a complex industrial process self-healing control method based on an elastic time-varying Bayesian network, which comprises a heavy medium coal preparation system, the heavy medium coal preparation system comprises a conveyor, a heavy medium cyclone, an arc screen A, a clean coal medium removal screen, an arc screen B, a gangue medium removal screen, a dilute medium barrel, a dilute medium pump, a dilute medium magnetic separator, a water adding pump, a medium adding pump, a medium adding magnetic separator, a qualified medium barrel and a pressure pump; the feed inlet of the heavy medium cyclone is connected with the discharge end of the conveyor, the feed inlet of the arc screen A is connected with the overflow outlet of the heavy medium cyclone, the feed inlet of the clean coal medium removal screen is connected with the discharge outlet of the arc screen A, the upper outlet of the clean coal medium removal screen is used for outputting clean coal, and the lower outlet thereof is connected with a first liquid discharge pipeline and a second liquid discharge pipeline; the feed inlet of the arc screen B is connected with the underflow outlet of the heavy medium cyclone, the feed inlet of the gangue medium removal screen is connected with the discharge outlet of the arc screen B, the upper outlet of the gangue medium removal screen is used for outputting coal gangue, and the lower outlet thereof is connected with a third liquid discharge pipeline and a fourth liquid discharge pipeline; the feed inlets at the top of the dilute medium barrel are connected with the outlet ends of the first liquid discharge pipeline and the third liquid discharge pipeline; the inlet end of the dilute pump is connected with the dilute medium outlet at the bottom of the dilute medium barrel; the inlet end of the dilute medium magnetic separator is connected with the outlet end of the dilute pump; the water inlet end of the water adding pump is connected with a water source; the inlet end of the medium adding pump is connected with a medium source; the inlet end of the medium adding magnetic separator is connected with the outlet end of the medium adding pump; the feed inlets at the top of the qualified medium barrel are connected with the outlet ends of the second liquid discharge pipeline, the fourth liquid discharge pipeline, the dilute medium magnetic separator, the water outlet end of the water adding pump and the outlet end of the medium adding magnetic separator; the inlet end of the pressure pump is connected with the outlet end at the bottom of the qualified medium barrel, and the outlet end thereof is connected with the feed inlet of the heavy medium cyclone.

[0047] Comprising the following steps:

[0048] Step one: Establish a dense medium coal preparation system and determine the variables;

[0049] A dense medium coal preparation system is established using a conveyor, a dense medium cyclone, an arc screen A, a clean coal medium removal screen, an arc screen B, a gangue medium removal screen, a dilute medium barrel, a dilute medium pump, a dilute medium magnetic separator, a water addition pump, a medium addition pump, a medium addition magnetic separator, a qualified medium barrel, and a pressure pump;

[0050] As in the analysis cases of Table 1 and Table 2, the measurement variables and operation variables in the dense medium coal preparation system are determined according to the operation characteristics of the dense medium coal preparation system; the measurement variables are six, in order: double-layer screen discharge flow, single-layer screen discharge flow, mixed medium barrel slurry density, cyclone medium density, qualified medium barrel medium density, and cyclone overflow ash content; the operation variables are two, in order: raw coal bin coal inflow and cyclone medium inflow pressure;

[0051] Table 1: Measurement variables in the dense medium coal preparation process

[0052]

[0053]

[0054] Table 2: Operation variables in the dense medium coal preparation process

[0055]

[0056]

[0057] Step two: Collect coal preparation process data and use ADF square root algorithm to test the stationarity of the industrial process;

[0058] ADF (Augmented Dickey-Fuller) test is a statistical test method used to detect whether there is a unit root in time series data. The existence of a unit root indicates that the time series data is non-stationary, i.e. the mean, variance and autocorrelation structure change over time. The stationarity of time series is a basic assumption of many statistical and econometric models, so ADF test is of great significance in time series analysis. ADF test judges the stationarity of time series by testing the unit root in the model. It is an extension of the Dickey-Fuller test, considering more complex cases, including high-order autocorrelation structure in time series. ADF test eliminates high-order autocorrelation in the sequence by introducing lagged difference terms, thereby improving the reliability of the test.

[0059] S21: In the process of running the dense medium coal preparation system, the sensor group arranged in the dense medium coal preparation system is used to collect the coal preparation process data and output to the controller, and the controller processes the coal preparation process data according to the internal clock module to obtain time series data;

[0060] S22: The ADF square root algorithm is used to test the stationarity of the time series data;

[0061] S221: According to the characteristics of the time series, an ADF test model with constant term and time trend is selected, and the first order difference Δy of the time series is obtained by using formula (1) t ;

[0062]

[0063] In the formula, Δy t represents y t -y t-1 ; y t-1 is the first order lag time series value; μ is the constant term; λt is the time trend term; α is the most important coefficient in the ADF test, which is used to judge the existence of the unit root; β i is the coefficient of the difference term, which is used to eliminate the high order autocorrelation in the sequence; ∈ t is the white noise error term; p is the lag order;

[0064] S222: The parameters in the ADF test model are estimated by using the least square method;

[0065] S223: The ADF statistic is calculated, which is in the form of the estimated α divided by its standard error;

[0066] S224: The hypothesis test is constructed, the null hypothesis H0 is that the unit root exists, at this time, the data is non-stationary, the alternative hypothesis H1 is that the unit root does not exist, at this time, the data is stationary;

[0067] The calculated ADF statistic is compared with the corresponding critical value; the critical value comes from the Dickey-Fuller distribution table; if the ADF statistic is less than the critical value, the null hypothesis is rejected, and it is considered that the time series is stationary; if the ADF statistic is greater than the critical value, the null hypothesis cannot be rejected, and it is considered that the time series is non-stationary;

[0068] A time series is non-stationary if its statistical properties change over time. The presence of a unit root is one of the main reasons for the non-stationarity of a time series. The ADF test method judges whether a time series contains a unit root by constructing a hypothesis test. The null hypothesis H0 is that a unit root exists, and the data is non-stationary. The alternative hypothesis H1 is that a unit root does not exist, and the data is stationary. In the coal preparation process, understanding the stationarity of the data is crucial for establishing effective control strategies and improving separation efficiency. In the coal preparation process, key variables such as density and flow rate may fluctuate over time, which can lead to unstable process control and reduced separation efficiency. Experiments have shown that key variables in the coal preparation process exhibit non-stationarity. Automatic control systems based on traditional control methods may not be able to effectively adapt to these changes, resulting in a decline in control performance.

[0069] As shown in Figure 3 , the present application determines non-stationary process variables by performing ADF tests on time series data in the heavy medium coal preparation process, thereby providing a reliable basis for further data processing and control strategy adjustment. In the heavy medium coal preparation process, raw coal bin coal intake (A), double-layer screen discharge flow (B), single-layer screen discharge flow (C), cyclone medium density (F), cyclone medium pressure (G), and cyclone overflow ash content (H) are non-stationary variables. In the slime concentration flotation process, the thickener underflow flow (L), the thickener medium density (M), the slurry pre-treater medium density (N), and the flotation tank stirring speed (P) are non-stationary variables.

[0070] Step three: stationary treatment of non-stationary data;

[0071] Difference is a method of calculating the difference between adjacent data points in a time series, used to eliminate trends and periodic components in a time series, thereby making the data more stationary. For non-stationary data, use the difference method to eliminate trends and periodic components in the time series until the data reaches a stationary state. In this process, check whether the differentiated data has reached a stationary state by unit root test or observation of the time series plot of the differentiated data. First-order difference can remove linear trends in time series by subtracting the value of the previous time point. If the sequence is a random walk, the data is still not stationary, and it may need to be differentiated twice or higher. If the data is still not stationary after second-order differentiation, it can continue to be differentiated three times or higher until the data reaches a stationary state.

[0072] Trends in non-stationary data can interfere with the decision-making ability of the model and easily lead to overfitting. Converting non-stationary data processes into stationary data processes has significant benefits for control decisions. This not only improves the accuracy and robustness of the prediction model, but also simplifies the calculation and analysis process. Through the stationary treatment, the interpretability and operability of the data are enhanced, and the sensitivity and reliability of the control strategy are improved. After non-stationary testing using the above method, the present application converts the non-stationary data process into a stationary data process using the data difference method. On the basis of stationary data, any abnormal point deviating from stationarity is easier to detect, which helps to discover and handle abnormal situations in the coal preparation process in a timely manner, and can learn more robust elastic time-varying dynamic Bayesian network structure and parameters. Figure 4 is a visual representation of the data after stationary.

[0073] Step four: structure learning, establish elastic time-varying dynamic Bayesian network model;

[0074] Due to the complexity of the production environment, the dense medium coal preparation process has the characteristics of multivariable, time-varying and strong coupling of non-stationary complex industrial processes. The elastic time-varying dynamic Bayesian network model can adapt to the time-varying parameters and structure, which is crucial for handling the non-stationary characteristics in the coal preparation process; According to the model, the causal relationship between variables is captured and these relationships are displayed in the form of a graph, thereby providing an intuitive basis for control decisions. After determining the key variables that affect the coal preparation effect, these variables are regarded as nodes in the network; After the non-stationary data is stationary, the elastic time-varying dynamic Bayesian network model is trained through data to mine the causal relationship behind the data and determine the dynamic dependency relationship between nodes. With the passage of time, the model can update these dependency relationships in real time, reflecting the real-time state changes in the coal preparation process. When an abnormality is detected in a key variable or its dependency relationship, appropriate adjustments can be made in a timely manner to ensure the safe and stable operation of the coal preparation process. According to the network structure Figure 5 The dynamic changes of the dense medium coal preparation process and the causal relationship between variables can be clearly seen. With the passage of time, the structure and edges of the graph are updated in real time.

[0075] S41: fuse the stationary data after stationary treatment with the original stationary data to obtain fusion data, analyze the correlation of the fusion data, determine the relationship strength and direction between multiple variables, and display the relationship between variables in the form of a graph;

[0076] S42: take the key variables affecting the coal preparation effect as nodes in the elastic time-varying dynamic Bayesian network, and construct an elastic time-varying dynamic Bayesian network model reflecting the causal relationship between industrial process variables to capture the dependency relationship changes of the dense medium coal preparation system at different time periods;

[0077] S43: training the elastic time-varying dynamic Bayesian network model with the fusion data, mining the causal relationship behind the data, determining the dynamic dependency relationship between the nodes, and providing an intuitive basis for control decision-making;

[0078] Step five: elastic time-varying dynamic Bayesian network parameter learning;

[0079] After learning the elastic time-varying dynamic Bayesian network structure of the heavy medium coal preparation process, parameter learning is performed; given the conditional probability table of each node, the elastic time-varying dynamic Bayesian network performs parameter learning to obtain a non-stationary elastic time-varying dynamic Bayesian network model, which provides a basis for reasoning control decision-making scheme; Table 3 is the conditional probability table of the quality variable overflow ash H when the heavy medium coal preparation process is in an abnormal working condition. Process variable 1 represents normal, 2 represents an abnormal value that is too small, and 3 represents an abnormal value that is too large. In reasoning, the abnormal evidence state is set to 2. When the coal preparation process is in an abnormal working condition, the abnormal time slice is inferred using the probability density function, and the scatter plot and histogram of the abnormal data distribution are plotted, as shown in Figure 6 and Figure 7 Then, the variable causing the abnormal working condition is determined using the conditional probability distribution, the state of each variable at the current time is inferred, and the optimal safety control decision-making scheme is inferred;

[0080] Table 3: Conditional probability table of overflow ash H in heavy medium coal preparation process

[0081]

[0082]

[0083] Step six: determining the time slice in which the abnormality occurs and the variable causing the abnormal working condition;

[0084] The online abnormal data phenomenon variable is input into the non-stationary elastic time-varying dynamic Bayesian network model, and the non-stationary elastic time-varying dynamic Bayesian network model accurately locates the time slice in which the abnormality occurs in the heavy medium coal preparation process and determines the process variable causing the abnormality;

[0085] Step seven: reasoning out the control decision-making scheme;

[0086] The expected quality index and the state value of the variable are input into the local Bayesian network structure for reasoning, and the posterior probability of the operation node is obtained. According to the principle of maximum posterior probability, the adjustment scheme of the heavy medium coal preparation process decision-making, i.e., the control decision-making, is obtained. As shown in Table 4, four process cases are selected for verification. Among them, “↑” represents the adjustment direction of the variable increasing, “ˉ” represents the adjustment direction of the variable decreasing, and “-” represents no change.

[0087] Table 4: Control variable adjustment strategy of heavy medium coal preparation process

[0088]

[0089]

[0090] Step eight: implement the control decision;

[0091] The controller adjusts the operating parameters of the dense medium coal preparation system according to the obtained control decision, and if the abnormal condition is eliminated, it enters the normal working mode, and if the abnormal condition is not eliminated, it re-collects the online abnormal data and re-executes step seven, continues to use the online abnormal data phenomenon variables as evidence information to reason out the control decision scheme until the abnormal condition is eliminated.

[0092] Further, the method in the present application can be verified by observing the change of the overflow ash content after the implementation of the control decision:

[0093] In order to verify the effectiveness of the elastic time-varying dynamic Bayesian network modeling method and the case adjustment strategy, four cases can be simulated on the dense medium coal preparation process simulation platform. After the abnormal condition occurs, it is observed whether the cyclone overflow ash content can return to the normal range. As shown in Figure 8 The threshold of the dense medium process fine coal position is set to 0.11, and after exceeding the threshold, it is input as evidence into the established elastic time-varying dynamic Bayesian network model to reason out the control decision scheme. From Figure 8 It can be seen that in the research of each case, when the abnormality occurs, the decision scheme reasoned out by applying the model of the present application can quickly eliminate the abnormality and ensure that the non-stationary dense medium coal preparation process can run safely and stably.

[0094] The present invention is directed to the problem of frequent abnormal conditions caused by non-stationary and time-varying characteristics in dense medium coal preparation process. A self-healing control decision method for non-stationary complex industrial processes based on elastic time-varying dynamic Bayesian network is proposed. The model is based on vector autoregressive model to describe the time dynamics between multiple variables. By dynamically adjusting the model structure and parameters, the system changes are reflected in real time, ensuring the model accurately describes the system behavior. By collecting data from coal preparation plants, the ADF square root algorithm is used to test the stationarity of the data. For non-stationary data, the difference method is used to make it stationary, and the stationary data and the original stationary data are fused for correlation analysis. Determine the strength and direction of the relationship between variables, understand the relationship between variables, and make more accurate decisions when abnormalities occur. After data preprocessing, the data is input into the elastic time-varying dynamic Bayesian network model to learn the network structure and parameters. According to the learned structure and parameters, combined with the probability density function, the abnormal time slice is accurately located, and the process variables causing the abnormality are determined. The abnormal data is input into the network model as evidence, and the decision scheme that can eliminate the abnormal condition is inferred, and it is converted into a robust control operation. Finally, the effectiveness of the proposed method is verified in the dense medium coal preparation process, and the application potential of the elastic time-varying dynamic Bayesian network model in non-stationary industrial processes is proved.

Claims

1.A method for self-healing control of a complex industrial process based on an elastic time-varying Bayesian network, comprising a heavy medium coal preparation system, the heavy medium coal preparation system comprising a conveyor, a heavy medium cyclone, an arc screen A, a clean coal medium screen, an arc screen B, a gangue medium screen, a dilute medium tank, a dilute medium pump, a dilute medium magnetic separator, a water pump, a medium pump, a medium magnetic separator, a qualified medium tank and a pressure pump; the feed inlet of the heavy medium cyclone is connected with the discharge end of the conveyor, the feed inlet of the arc screen A is connected with the overflow outlet of the heavy medium cyclone, the feed inlet of the clean coal medium screen is connected with the discharge outlet of the arc screen A, the overflow outlet of the clean coal medium screen is used for outputting clean coal, and the underflow outlet of the clean coal medium screen is respectively connected with a first liquid discharge pipeline and a second liquid discharge pipeline; the feed inlet of the arc screen B is connected with the underflow outlet of the heavy medium cyclone, the feed inlet of the gangue medium screen is connected with the discharge outlet of the arc screen B, the overflow outlet of the gangue medium screen is used for outputting coal gangue, and the underflow outlet of the gangue medium screen is respectively connected with a third liquid discharge pipeline and a fourth liquid discharge pipeline; the feed inlets at the top of the dilute medium tank are respectively connected with the outlet ends of the first liquid discharge pipeline and the third liquid discharge pipeline; the inlet end of the dilute medium pump is connected with the dilute medium outlet at the bottom of the dilute medium tank; the inlet end of the dilute medium magnetic separator is connected with the outlet end of the dilute medium pump; the water inlet end of the water pump is connected with a water source; the inlet end of the medium pump is connected with a medium source; the inlet end of the medium magnetic separator is connected with the outlet end of the medium pump; the feed inlets at the top of the qualified medium tank are respectively connected with the outlet ends of the second liquid discharge pipeline, the fourth liquid discharge pipeline, the dilute medium magnetic separator, the water outlet end of the water pump and the outlet end of the medium magnetic separator; the inlet end of the pressure pump is connected with the outlet end at the bottom of the qualified medium tank, and the outlet end of the pressure pump is connected with the feed inlet of the heavy medium cyclone. characterized in that The method comprises the following steps: Step one: establishing a heavy medium coal preparation system and determining variables; a heavy medium coal preparation system is established by using a conveyor, a heavy medium cyclone, an arc screen A, a clean coal medium screen, an arc screen B, a gangue medium screen, a dilute medium tank, a dilute medium pump, a dilute medium magnetic separator, a water pump, a medium pump, a medium magnetic separator, a qualified medium tank and a pressure pump; measurement variables and operation variables in the heavy medium coal preparation system are determined according to the operation characteristics of the heavy medium coal preparation system; the measurement variables are six, and are in turn double-layer screen discharge flow, single-layer screen discharge flow, mixed medium tank slurry density, cyclone medium density, qualified medium tank medium density and cyclone overflow ash content; the operation variables are two, and are in turn raw coal bunker coal inflow and cyclone medium inflow pressure; Step two: collecting coal preparation process data and using ADF square root algorithm to test the stationarity of the industrial process; S21: in the process of running the heavy medium coal preparation system, coal preparation process data are collected by using a sensor group arranged in the heavy medium coal preparation system and output to a controller, the controller processes the coal preparation process data according to an internal clock module to obtain time series data; S22: using ADF square root algorithm to test the stationarity of the time series data; S221: According to the characteristics of the time series, select the ADF test model with constant term and time trend, and obtain the first-order difference of the time series by formula (1) ; （1）； wherein represents is a lagged first-order time series value; is a constant term; is a time trend term; is the most important coefficient in the ADF test for determining the presence of a unit root; is the coefficient of the difference term for eliminating high-order autocorrelation in the series; is a white noise error term; is the lag order; S222: using least square method to estimate parameters in the ADF test model; S223: Calculate the ADF statistic, which takes the form of the estimated divided by its standard error; S224: construct hypothesis test, null hypothesis there is a unit root, in which case the data is non-stationary, alternative hypothesis there is no unit root, in which case the data is stationary; Compare the calculated ADF statistics with the corresponding critical value; the critical value comes from the Dickey-Fuller distribution table; if the ADF statistics is less than the critical value, reject the null hypothesis, consider the time series to be stationary; if the ADF statistics is greater than the critical value, cannot reject the null hypothesis, consider the time series to be non-stationary; Step three: stabilize the non-stationary data; For non-stationary data, use the difference to eliminate the trend and periodic components in the time series until the data reaches a stationary state; in this process, check whether the differentiated data reaches a stationary state by unit root test or observing the time series graph of the differentiated data; Step four: structure learning, establish an elastic time-varying dynamic Bayesian network model; S41: fuse the stabilized data with the original stationary data to obtain fused data, analyze the correlation of the fused data, determine the relationship strength and direction between multiple variables, and display the relationship between variables in the form of a graph; S42: take the key variables affecting the coal separation effect as the nodes in the elastic time-varying dynamic Bayesian network, construct an elastic time-varying dynamic Bayesian network model reflecting the causal relationship between industrial process variables, and capture the dependence relationship changes of the dense medium coal separation system at different time periods; S43: train the elastic time-varying dynamic Bayesian network model using the fused data, mine the causal relationship behind the data, and determine the dynamic dependence relationship between nodes; Step five: elastic time-varying dynamic Bayesian network parameter learning; Given the conditional probability table of each node, the elastic time-varying dynamic Bayesian network performs parameter learning to obtain a non-stationary elastic time-varying dynamic Bayesian network model, providing a basis for reasoning control decision scheme; when an abnormal working condition occurs in the coal separation process, the probability density function is used to infer the abnormal time slice; then, the conditional probability distribution is used to determine the variables causing the abnormal working condition, infer the state of each variable at the current time, and infer the optimal safety control decision scheme; Step six: determine the time slice where the anomaly occurs and the variables causing the abnormal working condition; Input the online abnormal data phenomenon variables into the non-stationary elastic time-varying dynamic Bayesian network model, and the non-stationary elastic time-varying dynamic Bayesian network model accurately locates the time slice where the abnormality occurs in the dense medium coal separation process and determines the process variables causing the abnormality; Step seven: reason out the control decision scheme; Input the expected quality indicators and variable state values into the local Bayesian network structure for reasoning to obtain the posterior probability of the operation node, and according to the principle of maximum posterior probability, obtain the adjustment scheme of the dense medium coal separation process decision, i.e., the control decision; Step eight: implement the control decision; The controller adjusts the operating parameters of the dense medium coal separation system according to the obtained control decision, and if the abnormal working condition is eliminated, it enters the normal working mode, if the abnormal working condition is not eliminated, it collects online abnormal data again and re-executes step seven to continue reasoning the control decision scheme using online abnormal data phenomenon variables as evidence information until the abnormal working condition is eliminated.

Images