An industrial alarm network construction method based on improved bayesian network structure

Abstract

The application discloses an industrial alarm network construction method based on an improved Bayesian network structure and relates to the technical field of industrial alarm network construction; wherein, S1: selecting TE process alarm data as a data set, a data sampling interval being 3 minutes, and dividing the data set into a training set and a test set; the TE process alarm data is divided in the following manner: 100 alarm data in the first 5 hours are taken as the training set, and 860 alarm data from the 5th hour to the 48th hour are taken as the test set; S2: sorting the data set extracted by S1; if there is an interval phenomenon in the statistical data, a mean value method is used to perform value supplementing on the discontinuous place; the industrial alarm network learned by the method is closer to the original network, because the EEMI considers the information entropy of itself when calculating the correlation degree between nodes, is more advantageous in accuracy, and the accuracy of the VLL directional rule is higher.

Description

Technical Field

[0001] This invention relates to the field of industrial alarm network construction technology, specifically to a method for constructing an industrial alarm network based on an improved Bayesian network structure. Background Technology

[0002] Modern process industries refer to industrial production sectors characterized by long-term, continuous production processes that require strict process control and safety assurance. They are characterized by large-scale, complex industrial production systems and massive production volumes. Process industries are an important component of the national economy, playing a vital role in promoting economic growth and enhancing national economic strength. In recent years, the country has actively advocated the deep integration of industrialization and informatization to develop process industries in a more efficient and intelligent manner.

[0003] However, due to the extreme complexity of production systems in process industries, with numerous interconnected and highly coupled production equipment, any equipment or unit malfunction can be transmitted between different systems through material and information flows, making the production process susceptible to internal and external disturbances. Even minor malfunctions can trigger a chain reaction, evolving and affecting the entire production process, thus causing unpredictable impacts on production safety, stability, and product quality.

[0004] Bayesian network structures are probabilistic graphical models that can describe causal relationships between variables and can be used to analyze and predict the probability of events. In the traditional Bayesian network learning process, mutual information is affected by the information entropy of the random variables themselves. When the information entropy difference between variables is large, using mutual information to determine whether a relationship exists between nodes is inaccurate. To address this problem, the inventors propose an industrial alarm network construction method based on an improved Bayesian network structure. Summary of the Invention

[0005] To address the issue of low similarity between industrial alarm networks and original networks, the present invention aims to provide a method for constructing industrial alarm networks based on an improved Bayesian network structure.

[0006] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: a method for constructing an industrial alarm network based on an improved Bayesian network structure, comprising the following steps:

[0007] S1. Select the TE chemical process dataset and divide the dataset into training and test sets;

[0008] S2. Data preprocessing: Extract the TE chemical process dataset. If there are gaps in the statistical data, use the mean method to fill in the gaps. At the same time, set the threshold according to the industrial 3δ rule.

[0009] S3, based on the mutual information of the lack of improvement, the calculation of mutual information between nodes, the concept of de-entropy mutual information is introduced, and the de-entropy mutual information and de-entropy conditional mutual information are obtained, and then the undirected network graph is generated;

[0010] S4, the generated undirected network graph is used to obtain the improved Bayesian network structure (EEMI&VLL) by VLL orientation rule;

[0011] S5, using Asia, Car, Child, Alarm four standard network data sets to experiment EEMI&VLL, and comparing the experimental effect evaluation index.

[0012] Preferably, the TE process alarm data is divided, specifically: the alarm data of the first 5 hours is taken as the training set, and the alarm data from the 5th hour to the 48th hour is taken as the test set.

[0013] Preferably, the process of data preprocessing in S2 specifically includes:

[0014] A: extract the TE chemical process data set, and sort the data in time sequence;

[0015] B: the data is counted according to time interval T, wherein T=3min;

[0016] C: if the statistical data appears discontinuous, the mean method is used for value filling, and the specific formula is as follows:

[0017]

[0018] Wherein, S t represents the alarm data at time t, S t+1 , S t-1 respectively represent the alarm data at t+1, t-1 time;

[0019] D: at the same time, according to the industrial 3δ rule, the continuous data is converted into 0, 1 discrete sequence according to the high and low threshold.

[0020] Preferably, the improvement process in S3 specifically includes:

[0021] A: the mutual information of node X and node Y is divided by the information entropy of the two nodes respectively to eliminate the influence of the information entropy on mutual information, and two ratios are obtained, η1 and η2, and the specific formula is as follows:

[0022]

[0023] Wherein η1, η2 represents the ratio of mutual information in X and Y node information entropy, I (X, Y) represents the mutual information of node X and node Y, H (X) represents the entropy of X node, H (Y) represents the entropy of Y node, the de-entropy mutual information between two nodes is shown in the following formula:

[0024] EEMI XY = p1η1+ p2η2

[0025] Wherein, EEM XY represents the de-entropy mutual information between X and Y nodes, and the definition formula of P1 and P2 is shown in the following formula:

[0026]

[0027]

[0028] B: According to the above theory, the conditional mutual information of node X and node Y is divided by the conditional information entropy of each node respectively to eliminate the influence of conditional information entropy on conditional mutual information, and two ratios are obtained, which are denoted as η3 and η4, and the specific formula is shown in the following formula:

[0029]

[0030] Wherein η3 and η4 represent the ratio of conditional mutual information in X and Y node conditional information entropy, I (X, Y|Z) represents the conditional mutual information of X and Y given the random variable Z, H (X|Z) represents the conditional entropy of X given Z, H (Y|Z) represents the conditional entropy of Y given Z, and the conditional de-entropy mutual information between two nodes is shown in the following formula:

[0031] EECM XY|Z = P3η3+ P4η4

[0032] Wherein, EECM XY|Z represents the conditional de-entropy mutual information between X and Y nodes given Z, and the definition formula of P3 and P4 is shown in the following formula:

[0033]

[0034]

[0035] Preferably, the discrimination function is constructed by using the difference between the de-entropy conditional mutual information and the de-entropy mutual information, the VLL structure in the maximum support tree is verified and oriented by the discrimination function, and the formula of the discrimination function is shown in the following formula:

[0036] Δ = EECMI XY|K (X i ,X j )-EEMI XY(X i ,X j )

[0037] wherein EECMI XY|K represents the conditional entropy mutual information between X and Y nodes under given condition X K represents the conditional entropy mutual information between X and Y nodes under given condition X XY represents the conditional entropy mutual information between X and Y nodes under given condition X K represents the difference between the conditional entropy mutual information and the entropy mutual information between X and Y nodes under given condition X

[0038] Preferably, after the orientation, the improved Bayesian network structure (EEMI&VLL) is obtained, and the EEMI&VLL is experimented by using four standard network data sets of Asia, Car, Child and Alarm, and the experimental effect evaluation indexes are compared.

[0039] Compared with the prior art, the application has the beneficial effects that:

[0040] In the application, the industrial alarm network learned by the method is closer to the original network, because the EEMI considers the information entropy of itself when calculating the correlation degree between nodes, has more advantages in accuracy, and the VLL orientation rule is more accurate. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description only show some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0042] Fig. 1 The flow chart of the industrial alarm network construction based on the improved Bayesian network structure is provided.

[0043] Fig. 2 The information intersection graph among the three variables is provided.

[0044] Fig. 3 The VLL orientation rule graph is provided. DETAILED DESCRIPTION

[0045] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0046] Embodiment: As shown in the figure, the application provides an industrial alarm network construction method based on improved Bayesian network structure, comprising the following steps: Figs. 1-3

[0047] S1, select TE process alarm data as a data set, the data sampling interval is 3 minutes, and the data set is divided into a training set and a test set. The TE process alarm data is divided into: 100 alarm data in the first 5 hours are taken as the training set, and 860 alarm data from the 5th hour to the 48th hour are taken as the test set;

[0048] S2, sort the data set extracted in S1, if there is a gap in the statistical data, adopt the mean method to supplement the value at the discontinuous place, and the specific formula is as follows:

[0049]

[0050] Wherein, S t represents the alarm data at time t, S t+1 , S t-1 respectively represent the alarm data at t+1, t-1, and the threshold is set according to the industrial 3δ rule, the continuous data is converted to 1 according to the high and low threshold, and the data in the threshold is converted to 0. The discrete sequence is processed, the data scale is reduced under the premise of ensuring the state unchanged, and the training time is reduced;

[0051] S3, the proportion of the mutual information I (X, Y) of the variable X and Y is greater than the proportion of the mutual information I (Y, Z) of the variable Y and Z, but in the figure, I (Y, Z) > I (X, Y), through the related theory of mutual information, we think that the actual correlation between X and Z is higher than that between X and Y, which is exactly the opposite of the actual situation. This is because in the process of learning the traditional Bayesian network structure, the mutual information is affected by the information entropy of the random variable itself. When the information entropy between variables is very different, it is not accurate to use mutual information to judge whether there is a relationship between nodes. Therefore, the application improves the calculation of mutual information between nodes based on the deficiency of mutual information, and introduces the concept of de-entropy mutual information. The specific improvement is as follows:

[0052] According to the above problems, the mutual information of node X and node Y is divided by the information entropy of the two nodes respectively to eliminate the influence of the information entropy of itself on the mutual information, and two ratios are obtained, η1 and η2, and the specific formula is as follows:

[0053]

[0054] ​Wherein, η1, η2 represent the ratio of mutual information in X and Y node information entropy, I (X, Y) represent the mutual information of node X and node Y, H (X) represent the entropy of X node, H (Y) represent the entropy of Y node, it can be seen that the smaller the information entropy of node, the higher the corresponding mutual information ratio, the de-entropy mutual information between two nodes is as shown in the following formula:

[0055] EEMI XY = p1η1 + p2η2

[0056] Wherein, EEM XY The de-entropy mutual information between X and Y nodes, the definition of P1 and P2 is as shown in the following formula:

[0057]

[0058]

[0059] According to the above theory, the conditional mutual information of node X and node Y is divided by the conditional information entropy of each node respectively to eliminate the influence of conditional information entropy on conditional mutual information, two ratios are η3 and η4, the specific formula is as shown in the following formula:

[0060]

[0061] Wherein, η3 and η4 represent the ratio of conditional mutual information in X and Y node conditional information entropy, I (X, Y|Z) represent the conditional mutual information of X and Y given the random variable Z, H (X|Z) represent the conditional entropy of X given Z, H (Y|Z) represent the conditional entropy of Y given Z, the conditional de-entropy mutual information between two nodes is as shown in the following formula:

[0062] EECM XY|Z = P3η3 + P4η4

[0063] Wherein, EECM XY|Z The conditional de-entropy mutual information between X and Y nodes given Z, the definition formula of P3 and P4 is as shown in the following formula:

[0064]

[0065]

[0066] The assignment method is a parameter self-adjusting mechanism, which can independently assign a larger weight to a larger mutual information ratio, which has practical significance, the mutual information ratio of the node with smaller information entropy can reflect the mutual information ratio of the node, the above content can be obtained, the application considers the information entropy of itself when calculating the correlation degree between nodes, which is more advantageous in accuracy, and the advantage is more obvious when the number of variables is larger;

[0067] S4, according to S3, the de-entropy mutual information and de-entropy conditional mutual information are obtained, and then the undirected network graph is obtained, the generated undirected network graph is oriented by VLL orientation rule, the idea is to use the difference between de-entropy conditional mutual information and de-entropy mutual information to construct a discrimination function, all VLL structures in the maximum support tree are checked and oriented through the discrimination function, the formula of the discrimination function is as follows:

[0068] Δ = EECMI XY|K (X i ,X j )-EEMI XY (X i ,X j )

[0069] Wherein, EECMI XY|K represents the conditional de-entropy mutual information between X and Y nodes under the given condition X K , EEM XY represents the de-entropy mutual information between X and Y nodes, Δ represents the difference between the conditional de-entropy mutual information and the de-entropy mutual information between X and Y nodes under the given condition X K , if Δ is greater than 0, it indicates that the dependence between node X k and node X i is enhanced when the state of node X j is known, that is, node X i and node X j form a VLL structure at node X k , so the undirected edge between the three nodes is oriented as X i →X k , X i →X k , otherwise, the three nodes do not constitute a VLL structure, and no orientation is performed, the specific formula of VLL orientation rule is as follows:

[0070]

[0071] Wherein, e ik =1 represents the directed edge X i →X k , e jk =1 represents the directed edge X j →X k , e ik =0 represents that there is no directed edge X i →X k , e jk =0 represents that there is no directed edge X j →X k , and then the improved Bayesian network structure (EEMI&VLL) is obtained;

[0072] S5, the improved Bayesian network structure (EEMI&VLL) obtained through S4, in order to verify the feasibility of the industrial alarm network structure obtained by EEMI&VLL, comparative experiments are carried out on EEMI&VLL and traditional MI&CRAR using four standard network data sets of Asia, Car, Child and Alarm, in this experiment, data sets with sample sizes of 1000, 5000 and 10000 are generated respectively, 100 experiments are independently carried out for each group of data, and the average value is taken as the final result, and the experimental results are shown in the chart:

[0073]

[0074] Wherein T represents that there is a correctly directed edge between variable X i and variable X j , F represents that there is a incorrectly directed edge between variable X i and variable X j , acc: correct orientation rate, the specific formula is as follows:

[0075]

[0076] It can be seen that the industrial alarm network learned by EEMI&VLL is closer to the original network, which is because EEMI considers the information entropy when calculating the correlation degree between nodes, and has more advantages in accuracy, and the accuracy of VLL orientation rule is higher, and when the number of variables is larger, the advantage is more obvious.

[0077] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application belong to the scope of the claims of the present application and their equivalent technologies, the present application also intends to include these modifications and variations.

Claims

1. An industrial alarm network construction method based on an improved Bayesian network structure, characterized by, The method comprises the following steps: S1, selecting a TE chemical process data set, and dividing the data set into a training set and a test set; S2, data preprocessing: extracting the TE chemical process data set, if the statistical data has discontinuous phenomena, using the mean value method for interpolation, and setting the threshold value according to the industrial 3δ rule; S3, based on the mutual information, the calculation of mutual information between nodes is improved, the concept of de-entropy mutual information is introduced, the de-entropy mutual information and the de-entropy conditional mutual information are obtained, and a non-directional network graph is generated; The improved process specifically comprises: A: the mutual information of nodes X and Y is divided by the information entropy of the two nodes respectively to eliminate the influence of the information entropy on the mutual information, and two ratios η1 and η2 are obtained, and the specific formula is as follows: Wherein η1, η2 represents the ratio of mutual information in X and Y node information entropy, I(X, Y) represents the mutual information of nodes X and Y, H(X) represents the entropy of X node, H(Y) represents the entropy of Y node, and the de-entropy mutual information formula between the two nodes is as follows: EEMI XY = p1η1 + p2η2 where EEMI XY represents the de-entropy mutual information between X and Y nodes, and the definition formulas of P1 and P2 are shown as follows: B: according to the above theory, the conditional mutual information of nodes X and Y is divided by the conditional information entropy of the nodes respectively to eliminate the influence of the conditional information entropy on the conditional mutual information, and two ratios η3 and η4 are obtained, and the specific formula is as follows: Wherein η3 and η4 represent the ratio of conditional mutual information in X and Y node conditional information entropy, I(X, Y|Z) represents the conditional mutual information of X and Y given the random variable Z, H(X|Z) represents the conditional entropy of X given Z, H(Y|Z) represents the conditional entropy of Y given Z, and the conditional de-entropy mutual information formula between the two nodes is as follows: EECMI XY|Z = P3η3+ P4η4 where EECMI XY|Z denotes the conditional mutual information between X and Y given Z, and P3 and P4 are defined as follows: S4, the generated non-directional network graph is used to obtain the improved Bayesian network structure (EEMI&VLL) by VLL orientation rule; S5, using Asia, Car, Child, Alarm four standard network data sets to experiment EEMI&VLL, and comparing the experimental effect evaluation index.

2. The industrial alarm network construction method based on the improved Bayesian network structure of claim 1, wherein, The TE process alarm data is divided into: the alarm data of the first 5 hours is taken as the training set, and the alarm data from the 5th hour to the 48th hour is taken as the test set.

3. The industrial alarm network construction method based on the improved Bayesian network structure of claim 1, wherein, The process of data preprocessing in S2 specifically comprises: A: extracting the TE chemical process data set, and sorting the data in time sequence; B: the data is counted according to the time interval T, wherein T=3min; C: if the statistical data has discontinuous phenomena, the mean value method is used for interpolation, and the specific formula is as follows: wherein S t represents alarm data at time t; S t+1 , S t-1 represent alarm data at times t+1, t-1, respectively; D: at the same time, the threshold value is set according to the industrial 3δ rule, and the continuous data is converted into 0, 1 discrete sequence according to the high and low threshold value.

4. The industrial alarm network construction method based on the improved Bayesian network structure of claim 1, wherein, The difference between the de-entropy conditional mutual information and the de-entropy mutual information is used to construct a discrimination function, all VLL structures in the maximum support tree are tested and oriented through the discrimination function, and the formula of the discrimination function is as follows: Δ = EECMI XY|K (X i ,X j ) - EEMI XY (X i ,X j ) where EECMI XY|K denotes the conditional mutual information between the nodes X K denotes the conditional mutual information between the nodes X XY denotes the conditional mutual information between the nodes X K denotes the difference between the conditional mutual information and the mutual information between the nodes X 5. The industrial alarm network construction method based on the improved Bayesian network structure of claim 1, wherein, After the orientation, the improved Bayesian network structure (EEMI&VLL) is obtained. The four standard network data sets of Asia, Car, Child and Alarm are used to experiment EEMI&VLL, and the experimental results are compared and evaluated.

Images