Online updating method of Bayesian network model for safety control of heavy medium coal preparation process based on incremental learning

Abstract

The present invention discloses an online updating method for a Bayesian network model of safety control of a heavy medium coal preparation process based on incremental learning, and belongs to the technical field of intelligent control of industrial safe operation. The main steps of the present invention are as follows: First, the phenomenon variables of abnormal working conditions are input as evidence information into the safety control Bayesian network model for reasoning. Secondly, a corresponding safety control decision is made according to the principle of maximum a posteriori probability, and whether the safety control decision is reasonable is judged based on the knowledge of domain experts. If it is reasonable, the control decision is implemented. If it is unreasonable, it continues to judge whether a new abnormal working condition occurs, and structural update learning and parameter update learning are performed according to the judgment conditions. Finally, the safety control decision is retrieved using the updated safety control Bayesian network model. The method for updating the safety control model of the heavy medium coal preparation process proposed by the present invention can improve the ability of the model to adapt to changes in the coal preparation environment and provide safe and reliable control decisions for operators.

Description

Technical Field

[0001] The invention relates to an online updating method of a Bayesian network model for safety control of a heavy medium coal preparation process based on incremental learning, and belongs to the technical field of intelligent control of industrial safety operation. Background Art

[0002] At present, when abnormal conditions occur in the heavy medium coal preparation process, the operator's experience and knowledge are mainly used to analyze the cause and location of the abnormal conditions, and to give corresponding operating variable values ​​to restore the production process to a stable operating state. However, this operation method is affected by the operator's own experience and knowledge limitations, and often cannot give reliable safety control decisions in time, causing certain losses to production.

[0003] Now, with the development of big data technology, on the one hand, due to the improvement of sensor accuracy and the application of big data technology, the coal preparation process has accumulated a large amount of industrial data. These data contain rich status information, which can help operators improve the quality of product coal. On the other hand, a large amount of expert knowledge is also contained in the process of process mechanism, material flow and equipment maintenance. This knowledge can guide operators to perform corresponding production operations and facilitate operators to understand. As an efficient visual graph model, Bayesian network has significant advantages in describing the causal relationship between complex variables, and can solve various problems under uncertain conditions based on Bayesian reasoning. Therefore, the use of Bayesian network can effectively combine the characteristics of abnormal working condition data information and expert knowledge, and help operators to obtain safe and reliable control decisions through Bayesian reasoning when facing abnormal working conditions. It has become an effective means.

[0004] However, in the actual production process, coal preparation processes are interrelated, raw coal properties change frequently, and the coal preparation production environment is complex and changeable. Therefore, new abnormal conditions are likely to occur. Faced with new abnormal conditions, the original safety control Bayesian model may not be able to adapt to environmental changes. If the safety control decisions provided by the original model are still used, not only will the abnormal conditions not be eliminated, but the lives of operators may even be threatened. In addition, some structures in the original model may still be able to represent some new data information and adapt to environmental changes. In view of this, it is necessary to find a suitable method that can effectively identify the parts of the existing safety control Bayesian model that are not adapted to environmental changes, and to perform structural update learning on them so that the model has the ability to adapt to environmental changes, thereby obtaining safe and reliable control decisions. Summary of the invention

[0005] In order to overcome the shortcomings of the above-mentioned prior art, the present invention provides an online updating method of the Bayesian network model for safety control of heavy medium coal preparation process based on incremental learning, which can improve the model's ability to adapt to changes in the coal preparation environment and provide operators with safe and reliable control decisions.

[0006] The present invention is implemented by the following technical scheme: an online updating method of a Bayesian network model for safety control of a heavy medium coal preparation process based on incremental learning, comprising the following steps:

[0007] Step 1: Collect data information of abnormal operating condition variables in the heavy medium coal preparation process, and use the discretized data as evidence to input into the existing safety control Bayesian network model for reasoning, and use formula (1) to obtain the posterior probability of the operating variable X related to the abnormal phenomenon:

[0008]

[0009] Among them, x is the value of variable X, e is the value of variable E;

[0010] Step 2: According to the results calculated in step 1, make corresponding control decisions according to the principle of maximizing the posterior probability;

[0011] Step 3: Based on the implementation effect of the control decision in step 2 in the heavy medium coal preparation process, determine whether the safety control decision is reasonable; if it is unreasonable, proceed to step 4, otherwise proceed to step 9;

[0012] Step 4: Use equations (2)-(3) to determine whether a new abnormal condition occurs. If no new abnormal condition occurs, it indicates that the structure of the existing safety control Bayesian network model does not need to be changed, and only the parameters of the original model need to be updated, that is, proceed to step 7. If a new abnormal condition occurs, it indicates that both the structure and parameters of the existing safety control Bayesian network model need to be updated, that is, proceed to step 5.

[0013]

[0014]

[0015] Among them, D old and D new Represent historical data and new sampled data respectively, M old and M new Represent the sizes of the old and new datasets respectively; n is the number of variables, r i is the variable X i The number of states, q i is the variable X i The parent node Pa(X i ) state combination number, m ijk is the variable X i is the number of samples in state k and its parent node is in state j, is the likelihood conditional probability, m is the sample size;

[0016] If ΔScore BIC >ξ, it indicates that the current model only needs to update the model parameters; if ΔScore BIC <ξ, it indicates that the current model needs to perform model structure update learning and parameter update learning, and ξ is the set threshold;

[0017] Step 5: Update the Bayesian network structure using the Bayesian network structure update method based on incremental learning, and then proceed to step 7;

[0018] Step 5.1: Use equations (4)-(5) to identify the partial structure of the old model that is not adapted to the environmental changes;

[0019]

[0020]

[0021] Among them, Pa(X i ) represents the variable X i The parent node of i ,Pa(X i )> represents node X i 、its parent node Pa(X i ) and the local structure formed by the edges between them; when ΔScore(<X i ,Pa(X i )>|D)>δ, indicating that node X i The local structure nearby does not need to be updated, where δ is the set threshold; when ΔScore(<X i ,Pa(X i )>|D)<δ, indicating that node X i The nearby local structures require structural renewal;

[0022] Step 5.2: Identify the local structure G that adapts to environmental changes according to step 5.1 0 and its corresponding node set V 1 ={X 1 ,..,X m}, and the remaining nodes are recorded as V 2 ={X m+1 ,...,X n};

[0023] Step 5.3: Use formula (6) to calculate the set V 1 and V 2 The average mutual information of 2 The variables in are sorted from large to small, and the sorted set is recorded as V 2 * ;

[0024]

[0025] Among them, x i and x j The variable set V 1 and V 2 The value of

[0026] Step 5.4: Set the learning step size t and determine V 2 * Is it empty? If it is empty, the structure update is completed and the structure G is returned. 0 ; If it is not empty, proceed to the next step;

[0027] Step 5.5: Take out set V 2 * The first t variables in are put into the set V 1 In, and according to the set V 1 The variables in the dataset D i ;

[0028] Step 5.6: Use the BIC scoring function shown in formula (3) as the scoring standard, greedy search as the search strategy, G 0 is the initial search seed structure, and the structure G of this iteration is obtained i ;

[0029] Step 5.7: G 0 Replace with G i , and return to step 5.4;

[0030] Step 6: The structure update of the safety control Bayesian network model is completed, and it replaces the old model structure, and then proceeds to step 7;

[0031] Step 7: Use formula (7) to learn the parameter update of the safety control Bayesian network model;

[0032]

[0033] Among them, m old ijk is the variable X in the old data i is the number of samples in state k and its parent node is in state j, m new ijk is the variable X in the new data i is the number of samples in state k and its parent node is in state j; when ΔScore BIC When ξ<ξ, the old structural model does not need to be changed, so the new parameters are obtained by integrating the new data information with the original data information; otherwise, the new data information is used to obtain the Bayesian network parameters;

[0034] Step 8: Replace the old model with the updated safety control Bayesian network model, return to step 1, and re-reason to obtain the safety control decision;

[0035] Step 9: Implement security control decisions.

[0036] The beneficial effect of the present invention is that the present invention enables the established safety control model to have the ability to adapt to environmental changes, thereby ensuring the decision-making performance of the model under the new environment. Combined with the idea of ​​incremental learning, the data information under new abnormal working conditions and part of the structural information in the old model are used to update the parts of the original safety control Bayesian network model that are not adapted to environmental changes, so that the model has the ability to adapt to environmental changes and ensure the decision-making performance of the model under the new environment.

[0037] The present invention first improves the model's ability to adapt to environmental changes and ensures the model's decision-making performance in the face of new environments. Secondly, combined with the idea of ​​incremental learning, the data information under new abnormal working conditions and part of the structural information in the old model are used to update the parts of the original safety control Bayesian network model that are not adapted to environmental changes, thereby improving the efficiency of model updating. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention is further described below with reference to the accompanying drawings and embodiments.

[0039] Figure 1 An online updating strategy for the Bayesian network model of safety control in heavy medium coal preparation process based on incremental learning is proposed.

[0040] Figure 2 Simplified process flow chart for heavy medium coal preparation.

[0041] Figure 3 It is an existing security control Bayesian network model.

[0042] Figure 4 It is the difference in BIC scores between the new and old data sets on the old structure.

[0043] Figure 5 is the score difference between each variable and the old data at different sampling points.

[0044] Figure 6 The Bayesian network structure for security control after the update. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings of the specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and does not serve as any limitation on the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without carrying out creative work are within the scope of protection of the present invention.

[0046] Technologies, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, the technologies, methods, and equipment should be considered part of the specification.

[0047] refer to Figure 1-Figure 6 The present invention proposes an online updating method for the Bayesian network model of safety control of heavy medium coal preparation process based on incremental learning. First, the phenomenon variables of abnormal working conditions are input into the safety control Bayesian network model as evidence information for reasoning. Secondly, the corresponding safety control decision is made according to the principle of maximum a posteriori probability, and whether the safety control decision is reasonable is judged based on the knowledge of domain experts. If it is reasonable, the control decision is implemented. If it is unreasonable, it continues to judge whether a new abnormal working condition occurs, and the structure update learning and parameter update learning are performed according to the judgment conditions. Finally, the safety control decision is retrieved using the updated safety control Bayesian network model.

[0048] like Figure 1-Figure 6 The embodiment shown in this example proposes an online updating method of the Bayesian network model for safety control of heavy medium coal preparation process based on incremental learning.

[0049] Step 1: Collect data information of abnormal operating conditions in the heavy medium coal preparation process, and use the discretized data as evidence to input into the existing safety control Bayesian network model for reasoning. The existing safety control model is as follows: Figure 1 As shown, the physical meaning of the nodes of the Bayesian network is shown in Table 1. Formula (1) is used to obtain the posterior probability of the operating variable X related to the abnormal phenomenon.

[0050]

[0051] Here, x is the value of the variable X, and e is the value of the variable E.

[0052] Table 1 The physical meaning of Bayesian network nodes and their classification levels

[0053] node Physical meaning node Physical meaning A Dilute medium pump speed F Suspension density B Medium pump speed G The cyclone inlet pressure is too low C Water pump speed H The diameter of the cyclone bottom flow port becomes larger D Pressure pump speed I Coal content in gangue E Qualified medium barrel liquid level J pressure

[0054] Step 2: According to the results calculated in step 1, make corresponding control decisions according to the principle of maximizing the posterior probability;

[0055] Step 3: Assume that the operating environment of the coal preparation process changes at this time, and a new abnormal condition appears: the cyclone inlet pressure is too high. When the cyclone inlet pressure is too high, the denser substances in the mixture will cause great wear on the inner wall of the cyclone, resulting in an increase in the diameter of the cyclone bottom flow port. At this time, the safety control decision made by the existing safety control Bayesian network model is unreasonable, so go to step 4.

[0056] Step 4: Use equations (2)-(3) to determine whether a new abnormal condition has occurred. The difference in BIC scores between the new and old data sets on the old structure is as follows: Figure 2 As shown in Figure 2, the threshold ξ is set to -0.083. It can be seen that the existing safety control Bayesian network model structure cannot adapt to environmental changes and needs to be updated, that is, go to step 5.

[0057]

[0058]

[0059] Among them, D old and D new Represent historical data and new sampled data respectively, M old and M new Represents the size of the old and new datasets respectively. n is the number of variables, r i is the variable X i The number of states, q i is the variable X i The parent node Pa(X i ) state combination number, m ijk is the variable X i is the number of samples in state k and its parent node is in state j, is the likelihood conditional probability, and m is the sample size.

[0060] If ΔScore BIC >ξ, it indicates that the current model only needs to update the model parameters. BIC <ξ, it indicates that the current model needs to perform model structure update learning and parameter update learning, and ξ is the set threshold.

[0061] Step 5: Use the Bayesian network structure updating method based on incremental learning to update the Bayesian network structure, and then proceed to step 7.

[0062] Step 5.1: Use equations (4)-(5) to identify the parts of the old model that are not adapted to the environmental changes. Figure 5is the score difference between each variable and the old data at different sampling points. The target node score change threshold is set to δ = -0.035. The white column in the figure represents ΔScore (<X i ,Pa(X i )>|D) node variables whose values ​​are greater than the threshold δ. Figure 5 It can be seen that the nodes in the existing safety control Bayesian network model that are not adapted to environmental changes are E, H and I.

[0063]

[0064]

[0065] Among them, Pa(X i ) represents the variable X i The parent node of i ,Pa(X i )> represents node X i 、its parent node Pa(X i ) and the local structure formed by the edges between them. i ,Pa(X i )>|D)>δ, indicating that node X i The local structure nearby does not need to be updated, where δ is the set threshold. i ,Pa(X i )>|D)<δ, indicating that node X i The nearby local structures require structural updates.

[0066] Step 5.2: Combined with the existing safety control model, the local structure extracted in step 5.1 that can identify and adapt to environmental changes is Its corresponding node set V 1 ={A,D,F,G,J}, and the remaining nodes are recorded as V 2 ={B,C,E,H,I}.

[0067] Step 5.3: Use formula (6) to calculate the set V 1 and V 2 The average mutual information of 2 The variables in are sorted from large to small, and the sorted set is recorded as V 2 * ={E,H,I,C,B}.

[0068]

[0069] Among them, x i and x j The variable set V1 and V 2 The value of .

[0070] Step 5.4: Set the learning step length t = 3 and determine V 2 * Is it empty? If it is empty, the structure update is completed and the structure G is returned. 0 If it is not empty, proceed to the next step.

[0071] Step 5.5: Take out set V 2 * The first t variables in are put into the set V 1 In, and according to the set V 1 The variables in the dataset D i .

[0072] Step 5.6: Use the BIC scoring function shown in formula (3) as the scoring standard, greedy search as the search strategy, G 0 is the initial search seed structure, and the structure G of this iteration is obtained i .

[0073] Step 5.7: G 0 Replace with G i , and return to step 5.4

[0074] Step 6: The structure of the safety control Bayesian network model is updated. The updated safety control Bayesian network structure is as follows: Figure 6 As shown. Replace the old model structure with it and go to step 7.

[0075] Step 7: Use formula (7) to learn the parameter update of the safety control Bayesian network model.

[0076]

[0077] Among them, m old ijk is the variable X in the old data i is the number of samples in state k and its parent node is in state j, m new ijk is the variable X in the new data i is the number of samples in state k and its parent node is in state j. BIC When ξ<ξ, the old structural model does not need to be changed, so the new parameters are obtained by integrating the new data information with the original data information. Otherwise, the new data information is used to obtain the Bayesian network parameters.

[0078] Step 8: Replace the old model with the updated safety control Bayesian network model, return to step 1, and re-infer to obtain the safety control decision.

[0079] Step 9: Implement security control decisions.

[0080] The above description is only an exemplary embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. An online updating method of Bayesian network model for safety control of heavy medium coal preparation process based on incremental learning, characterized by: The steps include: Step 1: Collect data information of abnormal operating condition variables in the heavy medium coal preparation process, and use the discretized data as evidence to input into the existing safety control Bayesian network model for reasoning, and use formula (1) to obtain the posterior probability of the operating variable X related to the abnormal phenomenon: Among them, x is the value of variable X, e is the value of variable E; Step 2: Based on the results calculated in step 1, make corresponding safety control decisions according to the principle of maximum posterior probability; Step 3: Based on the implementation effect of the safety control decision in step 2 in the heavy medium coal preparation process, determine whether the safety control decision is reasonable; if it is unreasonable, proceed to step 4, otherwise proceed to step 9; Step 4: Use equations (2)-(3) to determine whether a new abnormal condition occurs. If no new abnormal condition occurs, it indicates that the structure of the existing safety control Bayesian network model does not need to be changed, and only the parameters of the original model need to be updated, that is, proceed to step 7. If a new abnormal condition occurs, it indicates that both the structure and parameters of the existing safety control Bayesian network model need to be updated, that is, proceed to step 5. Among them, D old and D new Represent historical data and new sampled data respectively, M old and M new Represent the sizes of the old and new datasets respectively; n is the number of variables, r i is the variable X i The number of states, q i is the variable X i The parent node Pa(X i ) state combination number, m ijk is the variable X i is the number of samples in state k and its parent node is in state j, is the likelihood conditional probability, m is the sample size; If ΔScore BIC >ξ, it indicates that the current model only needs to update the model parameters; if ΔScore BIC <ξ, it indicates that the current model needs to perform model structure update learning and parameter update learning, and ξ is the set threshold; Step 5: Update the Bayesian network structure using the Bayesian network structure update method based on incremental learning, and then proceed to step 7; Step 5.1: Use equations (4)-(5) to identify the partial structure of the old model that is not adapted to the environmental changes; Among them, Pa(X i ) represents the variable X i The parent node of i ,Pa(X i )> represents node X i 、its parent node Pa(X i ) and the local structure formed by the edges between them; when ΔScore(<X i ,Pa(X i )>|D)>δ, indicating that node X i The local structure nearby does not need to be updated, where δ is the set threshold; when ΔScore(<X i ,Pa(X i )>|D)<δ, indicating that node X i The nearby local structures require structural renewal; Step 5.2: According to step 5.1, identify the local structure G0 that adapts to environmental changes and its corresponding node set V1 = {X1, .., X m }, and record the remaining nodes as V2 = {X m+1 ,...,X n }; Step 5.3: Use formula (6) to calculate the average mutual information of sets V1 and V2, and sort the variables in set V2 from large to small according to the mutual information. The sorted set is recorded as V2. * ; Among them, x i and x j are the values ​​of variable sets V1 and V2 respectively; Step 5.4: Set the learning step size t and determine V2 * Is it empty? If it is empty, the structure update is completed and the structure G0 is returned; if it is not empty, proceed to the next step; Step 5.5: Take out set V2 * The first t variables in the set are put into the set V1, and the data set D is generated based on the variables in the set V1. i ; Step 5.6: Use the BIC scoring function shown in formula (3) as the scoring standard, greedy search as the search strategy, G0 as the initial search seed structure, and obtain the structure G of this iteration i ; Step 5.7: Replace G0 with G i , and return to step 5.4; Step 6: The structure update of the safety control Bayesian network model is completed, and it replaces the old model structure, and then proceeds to step 7; Step 7: Use formula (7) to learn the parameter update of the safety control Bayesian network model; Among them, m old ijk is the variable X in the old data i is the number of samples in state k and its parent node is in state j, m new ijk is the variable X in the new data i is the number of samples in state k and its parent node is in state j; when ΔScore BIC When ξ<ξ, the old structural model does not need to be changed, so the new parameters are obtained by integrating the new data information with the original data information; otherwise, the new data information is used to obtain the Bayesian network parameters; Step 8: Replace the old model with the updated safety control Bayesian network model, return to step 1, and re-reason to obtain the safety control decision; Step 9: Implement security control decisions.

Images