A steel defect tracing method based on Bayesian network

Abstract

A steel defect tracing method based on Bayesian network belongs to the field of industrial design technology. The method uses Bayesian network as the base model to trace steel defects more accurately. The main steps are: 1) perform anomaly detection on each sample parameter in the data set to obtain the abnormal label of the feature or parameter; 2) give the initial weight matrix of the Bayesian network; 3) adjust the weight matrix according to the sample abnormality; 4) obtain the Bayesian network structure according to the adjusted weight matrix, so as to determine the importance of each parameter in the process of steel product defect generation. In view of the difficulties such as many defects in the steel production process, complex characteristics of measured data in steel production, and difficulty in analyzing the causes of steel defects, the present invention can fully explore the relationship between data features, and can easily obtain the corresponding relationship between sample defects and parameter abnormalities without relying on prior knowledge. The overall algorithm process is easy to understand and has strong explanatory power.

Description

Technical Field

[0001] The invention relates to the technical field of industrial design and provides a steel defect tracing method based on a Bayesian network. Background Art

[0002] In the steel production process, the production of defective products has brought a lot of pressure on economic costs, labor costs and time costs. Therefore, in addition to finding the defective products themselves, it is also particularly important to trace the causes of defective products. The steel production process is long and complicated. The production of defective products is often the result of the joint action of many factors. In addition to judging based on the workers' production experience, auxiliary technical means are also needed to further reduce the generation of defects. The method we adopted is to use the Bayesian network set-related algorithm to generate a directed graph to represent the dependency relationship between defects and the factors that cause defects. At this stage, the Bayesian method has the common prime Bayesian network, the Bayesian network based on the fault tree, and the distributed Bayesian network. The common defect tracing method also uses neural networks for tracing. As Li Lijuan et al. described in the article "Method for tracing defects in mechanical processing based on Bayesian networks", the above methods are difficult to deal with the problem of network structure learning in the absence of prior knowledge, and there are many restrictions when using them, and the interpretability is poor.

[0003] The Bayesian network is essentially a directed acyclic graph, which uses a directed acyclic graph to characterize the dependency between attributes and uses conditional probability distribution to describe the strength of the dependency between different attributes. Specifically, a Bayesian network B consists of two parts: structure G and parameter V, that is, B =<G,V> The network structure G is a directed acyclic graph, in which each node corresponds to an attribute. If two attributes have a direct dependency, they are connected by an edge; the parameter V quantitatively describes this dependency. Assuming that the attribute x i The set of parent nodes in G is π i , then it contains the conditional probability of each attribute

[0004] In view of the particularity of defect tracing, we proposed a defect tracing method based on Bayesian network that does not rely on prior knowledge and has strong interpretability. Based on the properties of Bayesian network, we can obtain the dependency relationship between steel defects and various parameters through the network structure. Different parameters are given different attention according to their importance, so as to minimize the production of defective products at the lowest cost. Summary of the invention

[0005] The present invention mainly solves the problem of tracing the source of defective steel products in the production process to reduce production costs. A steel defect tracing method based on a Bayesian network is proposed. The Bayesian network weight matrix is ​​trained through existing steel measured data to update the Bayesian network and achieve better tracing results.

[0006] The technical solution adopted by the present invention is:

[0007] A steel defect tracing method based on Bayesian network includes the following steps:

[0008] The first step is to determine the parameters of the samples in the data set and give them labels.

[0009] For the measured data set obtained during the steel production process Where m means that there are m samples in the data set (one sample is a steel product), and n means that each sample has n features (or parameters. When constructing the Bayesian network later, each parameter is a node). Through the existing anomaly detection method, all parameters of all samples are assigned abnormal or normal labels, where the abnormal parameter label is 1 and the normal parameter label is 0.

[0010] The second step is to obtain the Bayesian network weight matrix

[0011] First, establish the initial weight matrix W1, ···, W of the Bayesian network n ,W n+1 , where the first n matrices W1, ···, W n The size of can be n×n, but the diagonal elements must be 0 to indicate that the nodes are not connected to themselves. The n+1th matrix W n+1 The size of is n×1, which is the connection weight between each node and the defective node y, where y∈{0,1} is the product defect label of the sample (which can be regarded as the n+1th feature of the sample), y=1 indicates that the sample is a defective product, and y=0 indicates that the sample is a normal product. The initial weight matrix of the Bayesian network can be established based on empirical knowledge, or the initial zero matrix can be directly used without using empirical knowledge.

[0012] Then, according to the Booleanized data set in the first step The Bayesian network weight matrix is ​​updated by the product defect label data of each sample in the data set, and the data set D is traversed: if the sample is a defective product, the weight of its abnormal parameters in the weight matrix is ​​increased by Δw in the corresponding position between the two. + ; If it is a normal product, the weight of the connection position in the matrix corresponding to the abnormal parameter of the sample is reduced by Δw - , until the data set is completely traversed, stop updating, and get the Bayesian network weight matrix

[0013] The third step is to construct a Bayesian network based on the Bayesian network weight matrix obtained in the second step.

[0014] First, some set definitions are given and initialized. Define the current child node set N s , used to store the current child nodes, and find the parent node from the child node, and initialize the current child node set N s ={x n+1}(To simplify the code, x n+1 The actual sample defect node y); define the remaining parent node set N f , used to store the parent nodes that can be selected by the current child node set, and initialize the remaining parent node set N f ={x1,···,x n}, define the current parent node set N′ f When a node in the child node set looks for a parent node, it temporarily stores the selected parent node and uses it as the current child node set of the next layer to initialize the current parent node set. Define the adjacency matrix G = {0} (n+1)×(n+1) , responsible for recording the connection relationship between nodes and facilitating the experiment to draw the Bayesian network diagram; defining the number of parent node selections k.

[0015] Then, taking the defective node y as the current child node set of the first layer, we start to reversely search for the parent node to build the Bayesian network. i new , for the current child node set N s Traverse, N s For each node x j , in the remaining parent node set N f Select the first k largest nodes connected to it As x j The parent node is recorded in the current parent node set N′ f In the above example, G(j,[t1,···,t k ])=1, current child node set N s After the traversal is complete, the current child node set N s From the remaining parent node set N f Remove the current parent node set N′ f As the current child node set N for the next loop s , and the current parent node set N′ f Set it to an empty set and rejoin the loop until the loop is completed. Through the connection relationship between nodes recorded by the critical matrix G, find the other coordinate of the element with a value of 1 from the last layer of the matrix to determine the parent node, and recursively draw the Bayesian network.

[0016] In the fourth step, according to the Bayesian network drawn in the third step, the number of connected edges of each node and the parent-child relationship between nodes are clear at a glance: the more connected edges, the more important the node; the closer the parent-child relationship with the defect node y, the higher the node influence. Based on this, the influence of each node on the product defect can be directly obtained from the obtained Bayesian network, thereby determining the importance of each parameter in the process of steel product defect generation.

[0017] The beneficial effects of the present invention are:

[0018] The present invention provides a steel defect tracing method based on a Bayesian network. Aiming at the difficulties such as many defects in the steel production process, complex characteristics of measured data of steel production, and difficulty in analyzing the causes of steel defects, the connection relationship of various parameters in defect generation is obtained by training the Bayesian network using measured data, thereby obtaining the importance of each parameter for defect generation and determining the parameters that need to be focused on in the steel production process. The method proposed by the present invention is highly interpretable, can provide technical guidance for the steel production process, improve the quality of steel production, reduce the defect rate of steel products, and reduce costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is the fully connected graph of the Bayesian network (schematic diagram of the initial Bayesian network)

[0020] Figure 2 This is a schematic diagram of the Bayesian network training process;

[0021] Figure 3 Explain the parameters of the experimental examples;

[0022] Figure 4 Schematic diagram of weight update during training process for experimental examples;

[0023] Figure 5 Schematic diagram of the Bayesian network trained for the experimental paradigm. DETAILED DESCRIPTION

[0024] In order to make the method problems solved by the present invention, the method solutions adopted and the method effects achieved clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are only used to explain the present invention, rather than to limit the present invention. It is also necessary to explain that, for the convenience of description, only the parts related to the present invention are shown in the accompanying drawings, rather than all the contents.

[0025] Figure 1 This is a schematic diagram of the fully connected Bayesian network used in this article, which is also the initial Bayesian network structure diagram. Figure 1Assuming there are four process parameters x1, x2, x3, x4 and a defect label y, the Bayesian network structure trained by the Bayesian network must be a subgraph of the fully connected graph. Therefore, this is taken as an example as the initial structure of the Bayesian network. Starting from the defect label node y, we reversely push back the parent node of the current node to complete the construction of the Bayesian network. Figure 2 The present invention provides a flowchart for implementing a method for tracing the source of steel defects based on a Bayesian network. The present invention is applicable to the field of process design of steel engineering. It trains the measured data obtained from the steel product production process to obtain the importance of various parameters in the product production process to the generation of defects, and provides an implementation case: The implementation case uses measured data from the steel production process. Taking the tracing of protective slag defects in the continuous casting process as an example, the protective slag process sensor measured data set D is used. The data set size is 15631×15, that is, the data set contains 15631 samples, and the sample features are 15. The sample features of this data set include product defect labels, that is, each sample contains 14 process parameters x1,···,x 14 With 1 product defect label parameter x 15 =y∈{0,1}, where x1 to x 14 The corresponding parameter is a continuous parameter (i.e. the parameter value is continuous), x 15 The corresponding parameter is a Boolean parameter (that is, the value can only be 0 or 1). Figure 3 The following are the definitions of the parameters in the embodiments of the present invention. Figure 2 As shown, a steel defect tracing method based on a Bayesian network provided by an embodiment of the present invention includes:

[0026] The first step is to determine the parameters of the samples in the protective slag data set and assign labels. This embodiment uses the protective slag data set of the continuous casting process of steel production. According to the scale of the data set (15,631 samples, sample features are 15), this embodiment uses the box plot algorithm to perform anomaly detection on each parameter in the data set. Taking parameter x2 (minimum TD molten steel weight) as an example, the following operations are performed on each parameter: the parameter x2 values ​​of the 15,631 samples are arranged in ascending order from small to large and recorded as x sort , calculate the first quartile and the third quartile Get the interquartile range IQR = Q3-Q1 = 90, calculate the outlier cutoff point The sample feature values ​​whose parameter x2 values ​​are within the interval [453,813] are marked as normal, and the Boolean data is 0. The sample feature values ​​whose parameter x2 values ​​are outside the interval are marked as abnormal, and the Boolean data is 1. After all parameters are detected for abnormality, the abnormality of each parameter of each sample can be obtained. For example, the fourth sample has abnormal parameters x2, x7, and x9, and the Boolean value is 1.

[0027] The second step is to obtain the Bayesian network weight matrix trained with the protective slag dataset. Without using prior knowledge, this embodiment uses the initial weights W1, ···, W 14 ,W 15 Set to zero matrix, let the weighted intensity Δw + =0.1, weight reduction intensity Δw - = 0.01, extract the sample x that has not been extracted from the data set D i , find the sample x i n equals 1 i parameters, denoted as Build All Include All A directed graph with a sequential structure of nodes and y as the final child node, totaling n i ! The weight matrix is ​​updated according to the product defect label of the sample: if the sample x i For defective products (i.e. ), then n i ! The directed edges of the directed graph are in W1,···,W 14 The corresponding weight in plus Δw + =0.1; if sample x i For normal products (i.e. ), then n i ! The directed edges of the directed graph are in W1,···,W 14 ,W 15 The corresponding weight in minus Δw - =0.01. Take the sample x in the fourth steel product, the protective slag data set 4 For example, sample x 4 Among the 14 parameters of Build includes node And the directed graph with y as the final child node is as follows Figure 4 As shown in the network diagram outlined by the solid line, there are 6 subgraphs containing three nodes and product defect nodes. 4 For normal products, Then the directed edges of the six subgraphs (i.e. Figure 4 The weight of all solid lines is reduced by 0.01. After all samples are traversed, the weight matrix Update completed.

[0028] The third step is to construct a Bayesian network based on the Bayesian network weight matrix obtained in the second step. Establish and initialize the current child node set N s ={x 15}, the remaining parent node set N f={x1,···,x 14}, adjacency matrix G = {0} 15×15 , the current parent node set Determine the parent node selection number k = [4,3,3,2], first in Find the node x associated with the product defect 15 Connect the top 4 nodes x2, x7, x9, x 13 Record the current parent node set N′ f In, that is, N′ f ={x2,x7,x9,x 13}, and record G(15,[2,7,9,13])=1, and then update the current parent node set N′ f is the current child node set, that is, N s ={x2,x7,x9,x 13}, and set the current parent node set to an empty set, that is, Continue from N s Start from the remaining parent node set N f ={x1,···,x 14}, find the next layer of parent node set, x7 finds the top 3 parent nodes with the largest connection weight, which are x2, x3, x8 and record them in the current parent node set N′ f In, that is, N′ f ={x2,x3,x8}, and remember G(7,[2,3,8])=1, x9 finds the parent nodes with the largest connection weights, which are x2,x3,x 7, x8 and record it in the current parent node set N′ f In, that is, N′ f ={x2,x3,x7,x8}, and record G(9,[2,3,7,8])=1, x 13 Find the top three parent nodes with the largest connection weights, x4, x5, and x6, and record them in the current parent node set N′ f In, that is, N′ f ={x2,x3,x4,x5,x6,x7,x8}, and record G(13,[4,5,6])=1, the current child node set N s After the traversal is completed, the current child node set N s ={x2,x7,x9,x 13}From the remaining parent node set N f Eliminate, the current child node set N s Update to N′ f , and set the current parent node set to an empty set to store the parent node of the next layer, that is Continue from N s Start from the remaining parent node set N fContinue to search for the next layer of parent nodes in the loop until the number of layers of the Bayesian network structure is reached or the termination condition is reached. The parent-child relationship between nodes is stored in the adjacency matrix G. From the last row of the adjacency matrix G (i.e., row 15), find the ordinate (2, 7, 9, 13) of the point with a value of 1, which is the node x 15 The parent node x2,x7,x9,x 13 , draw from x2,x7,x9,x 13 Start pointing to node x 15 Then continue to search for points x2, x7, x9, x13 in the 2nd, 7th, 9th, and 13th order of the adjacency matrix G. 13 The parent nodes of the tree are drawn to point to their child nodes respectively, until the search is completed and the Bayesian network is constructed. Figure 5 This is the Bayesian network structure finally obtained by training in this embodiment.

[0029] The fourth step is to obtain the Bayesian network structure Figure 5 , it is concise and clear to see the parent-child relationship between nodes and the relationship between parameters and product defects: when the sample is a defective product, parameters x2, x7, x9, x 13 The possibility of abnormality is extremely high. When parameter x9 is abnormal, parameters x2, x3, x 7, x8 is also related to its abnormality, etc. It can be concluded that to reduce the defect rate of steel products, when adding protective slag in the continuous casting process, we should focus on parameters such as tdweightmin, castspeedavg, castspeedmin, liquidphasetemp, and try to control these parameters without abnormalities; secondly, focus on parameters such as tdweightavg, castspeedmax, tdsteeltempmax, etc., so as to achieve the minimum cost while reducing the defect rate of products as much as possible.

[0030] Finally, it should be noted that the above embodiments are only used to illustrate the method scheme of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, ordinary method personnel in the field should understand that modifying the method schemes recorded in the aforementioned embodiments, or equivalently replacing some or all of the method features therein, does not cause the essence of the corresponding method scheme to deviate from the scope of the method schemes of the embodiments of the present invention.

Claims

1. A steel defect tracing method based on Bayesian network, characterized in that: It includes the following steps: The first step is to determine the parameters of the samples in the data set and give them labels; For the measured data set obtained during the steel production process Where m means that there are m samples in the data set, and one sample is a steel product. n means that each sample has n features. When constructing the Bayesian network, each feature is a node. Through the existing anomaly detection method, all parameters of all samples are assigned abnormal or normal labels, where the parameter abnormal label is 1 and the parameter normal label is 0. The second step is to obtain the Bayesian network weight matrix 2.1) Establish the initial weight matrix W1, ···, W of the Bayesian network n ,W n+1 , where the first n matrices W1, ···, W n The size of W is n×n, but the diagonal elements must be 0 to indicate that the node is not connected to itself. The n+1th matrix W n+1 The size of is n×1, which is the connection weight between each node and the defective node y, where y∈{0,1} is the product defect label of the sample, y=1 indicates that the sample is a defective product, and y=0 indicates that the sample is a normal product; the initial weight matrix of the Bayesian network is established according to empirical knowledge, or the initial zero matrix is ​​directly adopted without using empirical knowledge; 2.2) According to the data set Booleanized in the first step The Bayesian network weight matrix is ​​updated by the product defect label data of each sample in the data set, and the data set D is traversed: if the sample is a defective product, the weight of its abnormal parameters in the weight matrix is ​​increased by Δw in the corresponding position between the two. + ; If it is a normal product, the weight of the connection position in the matrix corresponding to the abnormal parameter of the sample is reduced by Δw - , until the data set is completely traversed, stop updating, and get the Bayesian network weight matrix The third step is to construct a Bayesian network based on the Bayesian network weight matrix obtained in the second step; 3.1) Given a set definition and initialize it; Define the current child node set N s , used to store the current child nodes, and find the parent node from the child node, and initialize the current child node set N s ={x n+1 }, to simplify the code, x n+1 The actual sample defect node y; define the remaining parent node set N f , used to store the parent nodes that can be selected by the current child node set, and initialize the remaining parent node set N f ={x1,···,x n }, define the current parent node set N f ′ When a node in the child node set looks for a parent node, it temporarily stores the selected parent node and uses it as the current child node set of the next layer to initialize the current parent node set. Define the adjacency matrix G = {0} (n+1)×(n+1) , responsible for recording the connection relationship between nodes and drawing the Bayesian network diagram; defining the number of parent node selections k; 3.2) Taking the defective node y as the current child node set of the first layer, start to reversely search for the parent node to build the Bayesian network; Through the weight matrix W i new , for the current child node set N s Traverse, N s For each node x j , in the remaining parent node set N f Select the first k largest nodes connected to it As x j The parent node is recorded in the current parent node set N f ′ In the above example, G(j,[t1,···,t k ])=1, current child node set N s After the traversal is complete, the current child node set N s From the remaining parent node set N f Remove the current parent node set N f ′ As the current child node set N for the next loop s , and the current parent node set N f ′ Set it to an empty set and rejoin the loop until the loop is completed; through the connection relationship between nodes recorded by the critical matrix G, find the other coordinate of the element with a value of 1 from the last layer of the matrix to determine the parent node, and recursively draw the Bayesian network; The fourth step is to determine the number of edges connecting each node and the parent-child relationship between nodes based on the Bayesian network drawn in the third step; The closer the parent-child relationship with the defect node y, the higher the node influence; accordingly, the influence of each node on the product defect can be directly obtained from the obtained Bayesian network, thereby determining the importance of each parameter in the process of steel product defect generation.

Images