A continuous casting billet quality traceability method based on a decision tree and a knowledge graph

Abstract

A continuous casting billet quality traceability method based on decision tree and knowledge graph belongs to the technical field of steel product quality traceability. The process parameters involved in the steel production process are summarized by a decision tree model, a steel quality defect knowledge graph is constructed based on a Neo4j graph database, and a model for quality traceability based on the steel quality defect knowledge graph is designed. Through decision tree analysis of steel quality defect data, it helps to find the association rules in steel quality data and the association between steel defects, reveals the potential factors and influencing factors that lead to quality problems, and helps to improve steel quality control and production efficiency. Through graphical display of complex association relationships, an intuitive expression method is provided at the semantic level, and reasoning ability is provided. The construction of a steel quality defect knowledge graph makes the traceability process more reliable and interpretable, and the use of graph computing technology completes the quality traceability on the steel quality defect knowledge graph.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of steel product quality traceability, and particularly provides a continuous casting billet quality traceability method based on a decision tree and a knowledge graph, aiming at a continuous casting billet with a steel grade of SPHC12. BACKGROUND

[0002] Improving the quality of steel products and reducing the yield of defective products have become important goals pursued by steel enterprises in China.

[0003] Quality defect traceability, i.e., finding the root cause of quality defects. According to the type of steel quality defects, by analyzing the parameters in the steel process production process, the abnormal parameters related to the occurrence of this type of quality defects are explored, and then the product production process is improved to ensure the quality of the product. Chinese steel enterprises generally adopt traditional quality detection methods in product quality control, such as the central ledger mode, which is recorded and saved by each market participant, which is an information isolation mode. If the quality of steel products is a problem, it is difficult to determine the cause of the quality problem of steel products. With the proposal of "intelligent manufacturing", how to use artificial intelligence technology to promote the improvement of the steel production quality system has become a very urgent demand.

[0004] Decision tree is a process of discovering, extracting and analyzing valuable information and knowledge hidden in large-scale data sets. A large amount of data generated in the steel production process contains key quality information. Through decision tree analysis of steel quality defect data, it can help to discover the association rules in steel quality data and the correlation between steel defects, and reveal the potential factors and influencing factors that lead to quality problems, which can help to improve steel quality control and production efficiency. Knowledge graph is an innovative knowledge representation method, which visually displays complex correlation through graphics, provides an intuitive expression method at the semantic level, and has reasoning ability, providing a more comprehensive and intuitive perspective in complex knowledge correlation, making the traceability process more reliable and interpretable. SUMMARY

[0005] The purpose of the present application is to provide a continuous casting billet quality traceability method based on a decision tree and a knowledge graph, which provides an innovative solution for steel quality defect traceability by using decision tree and knowledge graph technology. The application of this solution can improve the accuracy and efficiency of steel product quality traceability, help enterprises to find potential problems and take corresponding measures, and further promote the quality improvement and development of the steel industry.

[0006] The present application uses a decision tree model to induce the rules of steel product defects, and then constructs a steel quality defect knowledge graph based on a Neo4j graph database, as shown in the flowchart Figure 1The decision tree is a learning model based on tree chart to make decisions, which can process both classification tasks and regression tasks. The construction of the decision tree is based on the greedy strategy, that is, the optimal feature is selected for splitting each time to minimize the classification error or regression error. The decision tree model mines potential rules and patterns from a large amount of data, and reveals the internal relationship between quality defects.

[0007] The implementation steps of the present application are as follows:

[0008] 1. The rules of the influence degree of the process parameters involved in the steel production process, such as the heat flow density of the loose side of the crystallizer wide face, the argon sealing pressure of the long nozzle of the ladle, the superheat degree and the heat flow density of the fixed side of the crystallizer wide face, etc. on the steel product defects such as slag inclusion and skin curling are induced by using the decision tree model, which provides available knowledge for constructing the steel quality defect knowledge graph.

[0009] 2. Based on the Neo4j graph database, the steel quality defect knowledge graph is constructed. Through the analysis of the rules of the steel quality data mining and induction by using the decision tree technology, the knowledge modeling of the steel defect data is carried out, and based on the Neo4j graph database, the steel quality defect knowledge graph is constructed, which is a structured steel quality defect knowledge network. The scattered information is integrated together to form a more complete and comprehensive steel quality knowledge system, which is intuitive and easy to understand. At the same time, it provides auxiliary decision support function for quality traceability. The steel quality defect knowledge graph is constructed by using the top-down and bottom-up combination method and adopts the seven-step method to construct the ontology. In the steel quality defect knowledge graph, the concepts are events and defects. The attributes of the event are name and event importance; the attribute of the defect is name. The relationship between the event and the event is related, and the attribute of the relationship is related degree; the relationship between the event and the defect is that the path leads to, and the attribute of the path leading to is weight.

[0010] 3. A model for quality traceability based on the steel quality defect knowledge graph is designed as Figure 2The method is based on the path ranking algorithm. In the steel quality traceability, the steel quality defect knowledge graph contains quality event nodes and defect nodes related to steel quality. The quality event nodes include the low heat flux density on the loose side of the wide surface of the crystallizer, the small argon sealing pressure of the long nozzle of the ladle, the small superheat, and the large heat flux density on the fixed side of the wide surface of the crystallizer. The defect nodes have two instances of slag inclusion defects and skin defects. Through the quality traceability model, the importance of the quality event path in the steel quality defect knowledge graph can be evaluated, and the quality event path is connected by the quality events that cause defects. Thus, the key path and related nodes in the quality traceability process can be determined, such as whether the path from the quality event of the low heat flux density on the loose side of the wide surface of the crystallizer to the quality event of the small superheat or the path from the quality event of the low heat flux density on the loose side of the wide surface of the crystallizer to the quality event of the large left taper of the narrow surface of the crystallizer causes the slag inclusion defect. The quality traceability model can calculate the scores of different paths according to specific evaluation criteria and find the path that causes the quality defect, such as the quality event path of the low heat flux density on the loose side of the wide surface of the crystallizer, the small argon sealing pressure of the long nozzle of the ladle, the small superheat, the small cooling water volume in the Z1 area, the large temperature difference between the inlet and outlet water on the left side of the narrow surface of the crystallizer, and the small temperature difference between the inlet and outlet water on the fixed side of the wide surface of the crystallizer, which causes the skin defect and is visualized to complete the quality defect traceability. The algorithm of the quality traceability model is as follows:

[0011] (1) Determine the defect type: determine the type of defect according to the quality diagnosis report.

[0012] (2) Determine the relationship path set: find all paths that cause the defect to occur on the steel quality defect knowledge graph through the Cypher language.

[0013] (3) Calculate the probability of the path causing the defect: compare the input data with the events in the path for each relationship path, calculate the probability of the path causing the defect, and the calculation method is as follows:

[0014] The correlation importance of the events in the defect path causing the defect is G = {g i ∈ [0, 1 | i = 1, …, n}, and Given the defective steel I that needs to be traced, the event set N that satisfies the path event in the parameter data can be obtained, so the probability of the defective steel I tracing the defect path is:

[0015]

[0016] In formula 1, [· represents the Iverson bracket. If the condition in the square brackets is satisfied, it is 1, otherwise it is 0.

[0017] (4) Path ranking: according to the score of the path, the path is ranked to determine the importance and relevance of the path, and the path with higher probability score is considered to be more relevant to quality traceability.

[0018] (5) Application of path ranking results: based on the results of path ranking, the top three paths with the highest probability are selected as the quality traceability results for display. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 Flow chart of constructing quality defect knowledge graph.

[0020] Figure 2 Quality defect traceability model diagram. DETAILED DESCRIPTION

[0021] 1. A decision tree model and a Neo4j graph database are used to construct a steel quality defect knowledge graph. The construction of the decision tree is based on a greedy strategy, that is, the optimal feature is selected for splitting at each time to minimize the classification error or regression error. The tree structure includes three types of nodes: root nodes, internal nodes, and leaf nodes. The first node of the tree is the root node, which is the first feature judgment selected by the model. The node contains all the sample data to be predicted. From the root node, the nodes that are split down are internal nodes, each of which corresponds to an attribute test. The leaf node corresponds to the decision and output result. Generally speaking, as the division process continues, we hope that the sample contained in the branch node of the decision tree belongs to the same category as much as possible, that is, the "purity" of the node is higher and higher. Currently, there are three commonly used decision tree algorithms: ID3, C4.5, and CART. Since the CART decision tree algorithm has the following advantages: the CART decision tree algorithm can process data sets including numerical and categorical features, and does not require pre-processing of features; the CART decision tree algorithm can automatically handle missing values without manually filling in missing values; the decision tree generated by the CART decision tree algorithm can be easily interpreted and visualized, which helps people understand the decision-making process of the model; the CART decision tree algorithm is not sensitive to outliers and can handle data sets with more noise and outliers. Therefore, the CART decision tree algorithm is selected to induce the rules of steel defects.

[0022] The CART decision tree algorithm uses the "Gini index" to select the division attribute, and the purity of the data set D can be represented by the Gini value as follows:

[0023]

[0024] Gini(D) reflects the probability of inconsistency of the class labels of two samples randomly selected from the data set D. Therefore, the smaller Gini(D) is, the higher the purity of the data set D is. The Gini index of attribute a is defined as:

[0025]

[0026] 2. Knowledge graphs model entities and concepts in the real world and their relationships, storing entities and their relationships in the form of structured triples. Triples typically describe a fact in a particular domain, consisting of a head entity, a tail entity, and a relationship describing the connection between the two entities. Knowledge graphs are considered as data graphs that aim to accumulate and transfer real-world knowledge, with nodes representing entities of interest and edges representing relationships between these entities. Knowledge graphs describe various entities in the real world, which can be concrete things (such as people, places, or objects) or abstract concepts (such as events or keywords). Entities can be identified by unique identifiers and can have attributes and relationships. Attributes describe the characteristics or properties of an entity and provide additional information about the entity, helping us better understand the entity. Attributes can be simple values (such as age, address), or complex structures (such as hierarchies, lists). Relationships define the connections and interactions between entities and describe the relevance, dependency, and constraints between entities. Relationships can be unidirectional or bidirectional, and can have different attributes and characteristics.

[0027] Learn the rules of steel defects through decision trees, based on the Neo4j graph database, using a top-down and bottom-up combined knowledge modeling method. Top-down modeling and bottom-up modeling are two common knowledge modeling approaches that play different roles and applications in the construction of knowledge graphs. Top-down modeling, also known as concept-level modeling or ontology-driven modeling, refers to defining the data schema, i.e., ontology, when constructing a knowledge graph. Generally, it is compiled by domain experts manually, starting from the topmost concept and then gradually refining it to form a well-structured hierarchical structure. Bottom-up modeling is based on actual data and inductively forms the bottom-level concepts for specific entities, then gradually abstracts them to form upper-level concepts. It is a data-driven modeling approach that constructs a knowledge graph by analyzing and mining the potential patterns and relationships in the data. The bottom-up approach is more commonly used for ontology construction in open-domain knowledge graphs, as the open world is too complex to be fully considered using the top-down approach. Moreover, as the world changes, the corresponding concepts continue to grow, and the bottom-up approach can meet the needs of growing concepts. Combining top-down modeling and bottom-up modeling can fully utilize the normative nature of ontology definition and the real-time nature of data-driven modeling.

[0028] Learn the rules of steel defects through decision trees, based on the Neo4j graph database, using a seven-step ontology construction method to build a knowledge graph ontology, which mainly includes the following seven steps:

[0029] (1) Determine the purpose and scope of the ontology: clarify the application scenario and purpose of the ontology, determine the coverage and required knowledge of the ontology.

[0030] (2) Collect domain knowledge: collect domain knowledge related to the ontology, including relevant literature, standards, dictionaries, and expert knowledge, etc.

[0031] (3) Analyze knowledge: analyze and organize the collected knowledge, extract key information such as concepts, attributes, and relationships, and determine classes and attributes in the ontology.

[0032] (4) Design ontology structure: design the hierarchical structure, attributes, and relationships of the ontology, and determine the hierarchical relationships and attribute relationships between concepts in the ontology.

[0033] (5) Establish ontology: use ontology editing tools to formalize the designed ontology structure, including concepts, attributes, relationships, constraints, etc.

[0034] (6) Ontology verification: verify whether the ontology conforms to the description of domain knowledge and semantic logic, including ontology consistency, completeness, reusability, extensibility, etc.

[0035] (7) Ontology maintenance: with the continuous updating of domain knowledge and the changes of ontology use, the ontology needs to be continuously maintained and updated, including adding, deleting, modifying concepts, attributes, and relationships, etc.

[0036] In the constructed steel quality defect knowledge graph, the ontology has events and defects. The attributes of the event are name and event importance; the attribute of the defect is name. The relationship between events and events is related, and the attribute is related degree; the relationship between events and defects is that the path leads to.

[0037] 3. Quality traceability method based on quality defect knowledge graph

[0038] Quality traceability on the steel defect knowledge graph can be realized with the help of graph computing technology. Graph computing is a data processing and analysis method based on graph structure, which can effectively process large-scale graph data and perform complex graph analysis and computation. Graph computing is an important tool for quality traceability on the steel defect knowledge graph. By using graph computing technology, the nodes and edges in the graph can be analyzed and calculated to reveal the correlation, path and characteristics of steel production events and quality defects, providing strong support for quality traceability. The core of graph computing is graph algorithms. Common graph algorithms include graph analysis algorithms, community discovery algorithms, recommendation algorithms and structure prediction algorithms. Among them, graph analysis algorithms are based on graph theory and use iterative traversal operations on graph data vertices or edges to analyze and mine the overall characteristics of the whole graph or the local characteristics of part of the graph structure. There are many graph analysis algorithms, and common graph analysis algorithms include path finding algorithms, graph embedding algorithms and similarity calculation algorithms. Among them, path finding algorithms are used to find paths between two nodes that meet certain conditions in a graph. In a knowledge graph, path search is performed to find the most relevant path from the starting node to the target node. This helps to track the defect propagation path in the steel production process, thereby achieving quality traceability.

[0039] Path finding algorithms include graph traversal algorithms, path ranking algorithms (PRA), shortest path algorithms and bidirectional search algorithms. Different path finding algorithms are used for different path finding tasks and application fields. We need to choose the appropriate algorithm based on the steel quality defect traceability problem and the characteristics of the data set. Path ranking algorithm is a graph traversal-based algorithm used to sort paths in a graph. It usually evaluates and sorts paths based on certain criteria to determine the optimal or most suitable path. The specific implementation steps are as follows:

[0040] (1) Define a set of Cypher statements to automatically and batch query the paths that cause the occurrence of the skin defect: MATCH p = (start:Event)-[*]->(end:Defect) WHERE start.name = $start_name AND end.name = $end_name RETURN [node in nodes(p) | node.name] as path. This method can obtain all paths that cause the defect, a total of 23 defect cause paths, path 1 to path 23.

[0041] (2) Calculate the probability of each defect cause path leading to the warping defect by the quality traceability algorithm model, the probability of path 1 is 90.5%, the probability of path 2 is 81.6%, the probability of path 3 is 78.2%, the probability of path 4 is 73.6%, the probability of path 5 is 72.3%, the probability of path 6 is 70.1%, the probability of path 7 is 68.5%, the probability of path 8 is 68.3%, the probability of path 9 is 66.8%, the probability of path 10 is 64.6%, the probability of path 11 is 66.3%, the probability of path 12 is 64.4%, the probability of path 13 is 61.4%, the probability of path 14 is 60.7%, the probability of path 15 is 57.4%, the probability of path 16 is 50.1%, the probability of path 17 is 48.5%, the probability of path 18 is 45.1%, the probability of path 19 is 44.1%, the probability of path 20 is 42.1%, the probability of path 21 is 39.7%, the probability of path 22 is 37.5%, the probability of path 23 is 37.3%.

[0042] (3) Select the top three paths with the largest probability, i.e. path 1, path 2 and path 3, as the quality traceability result for visual display, and complete quality traceability.

Claims

1. A continuous casting billet quality traceability method based on a decision tree and a knowledge graph, characterized in that, Comprise the following steps: (1) using decision tree model induction involved in the process of steel production process parameters, including crystallizer wide loose side heat flux density, long nozzle argon sealing pressure, superheat and crystallizer wide fixed side heat flux density; (2) based on Neo4j graph database to build steel quality defect knowledge graph, through the analysis of the use of decision tree technology to steel quality data mining and induction rules, steel defect data knowledge modeling, steel quality defect knowledge graph using top-down and bottom-up combination of methods and seven-step method to build ontology; In the steel quality defect knowledge graph, the concept has event and defect; The attributes of event are name and event importance; The attribute of defect is name; The relationship between event and event is related, and the attribute of related is related degree; The relationship between event and defect is path leading, and the attribute of path leading is weight; (3) design a model based on steel quality defect knowledge graph for quality traceability, in the steel quality traceability, the steel quality defect knowledge graph contains quality event nodes and defect nodes related to steel quality, through the quality traceability model, the importance of quality event path in the steel quality defect knowledge graph is evaluated, and the quality event path is connected by the quality event causing the defect; The quality traceability model calculates the scores of different paths according to specific evaluation criteria, and finds out the path leading to the quality defect; The specific implementation steps of the quality traceability model based on the steel quality defect knowledge graph are: a determine the defect type: determine the type of defect according to the quality diagnosis report; b determine the relationship path set: find all paths leading to the occurrence of the defect on the steel quality defect knowledge graph through Cypher language; c calculate the probability of path leading to defect: for each relationship path, compare the input data with the events in the path, and calculate the probability of the path leading to the defect, the calculation method is as follows: The associated importance of the events in the defect path that caused the defect to occur is G = {gi | i = 1,..., n} and i ∈ [0, 1) | i = 1,..., n}, and Given a defective steel I that needs to be traced, obtaining a set of events N that satisfy the path events in the parameter data, the probability that the defective steel I is traced to the defect path is: In formula 1, [·] represents the Iverson bracket, which is 1 if the condition in the square brackets is met, and 0 if it is not met; d path sorting: according to the score of the path, the importance and correlation of the path are determined, and the path with higher probability score is considered to be more related to the quality traceability; e apply path sorting results: based on the results of path sorting, select the top three paths with the highest probability as the quality traceability results for display. 2.The method according to claim 1, wherein, The quality event nodes in step (3) include low crystallizer wide loose side heat flux density, long nozzle argon sealing pressure, small superheat and large crystallizer wide fixed side heat flux density, and the defect nodes have two examples of slag inclusion defect and skin defect.

Images