A method and apparatus for metallurgical incident causality analysis

Abstract

The present application relates to the technical field of metallurgical production safety management, and provides a metallurgical accident causality analysis method and device. The present application performs data preprocessing on unstructured accident reports of the metallurgical industry to generate standardized key-value pairs of search phrases and semantic paragraphs of each entity according to the entity types of metallurgical accidents; realizes the correspondence of standard entities to nodes in a graph database through clustering; generates a risk graph according to nodes marked as high-frequency hidden dangers or historical accident-prone nodes to aggregate into risk communities; and traces directed relationships from risk sources within the risk communities to restore cross-process causality chains to output pre-warning information, solving the problem that the prior art metallurgical accident reports are mostly unstructured documents, rely on manual interpretation, lack a standardized and intelligent analysis system, and lead to repeated accidents.

Description

Technical Field

[0001] This invention relates to the field of metallurgical production safety management technology, and in particular to a method and apparatus for analyzing the cause of metallurgical accidents. Background Technology

[0002] In the field of safety management in metallurgical enterprises and industrial production, after a safety accident occurs, it is necessary to analyze the cause of the accident and transform it into prevention and control measures to prevent similar or related safety accidents from happening again.

[0003] However, current accident reports in metallurgical enterprises are mostly archived and stored as unstructured documents (such as text reports, scanned copies, minutes, etc.), which suffers from problems such as data dispersion, low degree of structure, weak knowledge correlation, and low retrieval efficiency. Moreover, existing accident analysis technologies mostly focus on qualitative analysis of causes, relying heavily on manual interpretation of unstructured accident documents. They have not formed a standardized and intelligent analysis technology system, lack the ability to trace loopholes in accident accountability, lack further investigation of accident hazards, and do not link accident hazards, accountability loopholes, and related accountability and penalties for the accident. Therefore, it is impossible to prevent the recurrence of accidents.

[0004] Therefore, overcoming the shortcomings of the existing technology is an urgent problem to be solved in this technical field. Summary of the Invention

[0005] The technical problem this invention aims to solve is that existing metallurgical accident reports are mostly unstructured documents that rely on manual interpretation and lack a standardized and intelligent analysis system. They cannot organize the development process from hidden dangers to accidents and then to penalties, leading to the frequent recurrence of accidents.

[0006] Firstly, a method for causal analysis of metallurgical accidents is provided, including: Data preprocessing is performed on unstructured accident reports in the metallurgical industry to generate standardized text; The standardized text is processed according to the metallurgical process steps. Based on the predefined entity types of metallurgical accidents, the search phrases of each entity are generated as keys, and the corresponding semantic paragraphs are used as values ​​to obtain standardized key-value pairs, which are then stored in the graph database. Determine the semantic similarity between the entities, construct multiple clusters according to the entity types, and obtain the standard entity of each cluster. One standard entity corresponds to one node in the graph database. Based on the nodes marked as high-frequency hidden dangers or historical accidents in the graph database, a risk graph is generated; the risk status of the nodes in the risk graph is iteratively updated until convergence, and they are aggregated into risk communities; within the risk community, the directed relationship is traced back from the risk source to restore the causal chain across processes, so as to output early warning information.

[0007] Furthermore, generating a risk map based on nodes marked as high-frequency hidden dangers or historically frequent accidents in the graph database includes: Subgraphs that meet the conditions for safe production are extracted from the graph database, and a graph model for risk inference is constructed using node sets and directed edges with semantic orientation. For nodes in the graph database that are pre-marked as high-frequency hidden dangers or have a high incidence of historical accidents, add initial risk labels to obtain the risk sources; For each edge between nodes in the graph database, a semantic label is preset for each edge; the weight corresponding to the semantic label is determined according to the mapping rules and stored as an attribute of the edge to obtain a risk graph.

[0008] Furthermore, the iterative update of the risk state of the nodes in the risk graph until convergence is achieved, and they are aggregated into risk communities; within the risk community, directed relationships are traced back from the risk source to reconstruct the causal chain across processes, so as to output pre-warning information including: The risk concentration value of each node is iteratively updated according to the weight of the edge until convergence. After convergence, nodes with the same or similar risk labels within a preset range are aggregated into risk communities. Within the risk community, starting from the node that is the source of risk, the causal semantic backtracking is performed along the directed edge based on multiple constraints to the node of the potential accident, thus obtaining the causal chain, and outputting the causal chain as early warning information.

[0009] Furthermore, the step of iteratively updating the risk concentration value of each node according to the weight of the edge until convergence includes: The following dynamic weighted propagation formula for risk concentration is used for iterative updates: ; in ， Indicates pointing to a node The set of predecessor nodes; For nodes To the node The weight, For the number of iterations, For nodes V k To the node V i The weights; During the iteration process, the following activation function is used to perform a non-linear mapping on the updated risk concentration values: ; in, For nodes The current risk label, For nodes The risk concentration value.

[0010] Furthermore, the expression for the multiple constraints is: ; in, This represents the optimal causal path selected from the candidate causal evolution paths; Indicates the first in the path A directed edge; Indicates the first A directed edge The weights; This indicates the number of directed edges in the path; Indicates the first in the path One node; Indicates a high-risk community; Indicates the first A directed edge The semantic type.

[0011] Furthermore, the step of tracing back directed relationships from the risk source within the risk community to reconstruct the causal chain across processes also includes: When restoring the causal chain, only directed edges that satisfy the preset semantic type and whose weight is greater than or equal to the weight threshold of weak causal relationship are added to the search tree.

[0012] Furthermore, the step of processing the standardized text according to the metallurgical process steps, generating retrieval phrases for each entity as keys based on predefined entity types of metallurgical accidents, and using the corresponding semantic paragraphs as values ​​to obtain standardized key-value pairs, and storing them in the graph database includes: The standardized text is semantically sliced ​​according to the entire process sequence of metallurgical processes to obtain multiple small text blocks; The instance relationships of entities are extracted from the small text blocks by a prompt-driven large language model to obtain triples, which consist of entity A, the relationship between entity A and entity B, and entity B. Determine the retrieval identifier for the entity type, including the entity, and obtain a semantic summary of the standardized text corresponding to the entity; The search identifier is used as the value, and the corresponding semantic summary is used as the key to generate a standardized key-value pair for the entity; The triples are transformed into a topological structure consisting of node A, edges, and node B; the normalized key values ​​corresponding to the entities represented by the nodes are added to the nodes for storage in the graph database.

[0013] Further, determining the semantic similarity between the entities, constructing multiple clusters based on the entity types, and obtaining standard entities for each cluster, wherein a standard entity corresponds to a node in the graph database includes: Calculate the semantic similarity between entities in the graph database; Set an asymmetric similarity threshold based on the entity type in the metallurgical emergency scenario; Entities whose semantic similarity reaches the corresponding asymmetric similarity threshold are grouped into the same cluster; In each cluster, standard entities are selected and the effective attributes of other entities are merged. Redundant nodes are deleted to achieve unique node updates.

[0014] Secondly, a metallurgical accident causality analysis device is provided, the metallurgical accident causality analysis device comprising: a processor and a memory for storing processor-executable instructions; The processor is configured to execute the metallurgical accident causal analysis method.

[0015] Thirdly, a non-volatile computer storage medium is provided, the computer storage medium storing computer-executable instructions, which are executed by one or more processors to perform the metallurgical accident causal analysis method described in the first aspect.

[0016] Fourthly, a computer program product containing instructions is provided that, when executed on a computer or processor, causes the computer or processor to perform the metallurgical accident causal analysis method as described in the first aspect.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention preprocesses unstructured accident reports to generate standardized text and standardized key-value pairs, storing them in a graph database. This avoids simply creating unstructured accident reports and solves the problems of scattered, weakly correlated, and difficult-to-retrieve accident report data in existing technologies. It provides a standardized and intelligent analysis system that preprocesses unstructured accident reports sequentially, extracts entities, clusters the data stored in the graph database, and iteratively simulates risk propagation. This automatically reconstructs the causal chains across processes from unstructured accident reports, outputting pre-emptive warning information containing the corresponding causal chains. No manual intervention is required, avoiding reliance on manual interpretation, and achieving a leap from post-event tracing to pre-event warning, effectively preventing recurring accidents. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of a metallurgical accident causal analysis method provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating step 20 provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating step 30 provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating step 40 provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating another step 40 provided in an embodiment of the present invention; Figure 6 This is an example of redundancy in the attributes of a knowledge graph node in the prior art, provided by an embodiment of the present invention. Figure 6 This is a box diagram illustrating a specific example of a scoring comparison between a prior art solution and a solution from an embodiment of the present invention. Figure 7 This is a schematic diagram illustrating a specific example of an accident chain retrieval accuracy comparison provided by an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating a specific example of a knowledge graph provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating a specific example of a cross-process splash warning provided by an embodiment of the present invention; Figure 9 This is a schematic diagram of a metallurgical accident causality analysis device provided in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0021] Unless the context otherwise requires, throughout the specification and claims, the term "comprising" is interpreted as openly inclusive, meaning "including, but not limited to." In the description of the specification, terms such as "one embodiment," "some embodiments," "exemplary embodiment," "example," "specific example," or "some examples" are intended to indicate that a particular feature, structure, material, or characteristic associated with that embodiment or example is included in at least one embodiment or example of this disclosure. The illustrative representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics mentioned may be included in any suitable manner in any one or more embodiments or examples; that is, although they may be incorporated into embodiments or examples using the above terms for reasons such as order and position, it does not limit them to be incorporated in combination by a single embodiment or example.

[0022] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of embodiments of this disclosure, unless otherwise stated, "a plurality of" means two or more. Furthermore, for example, the description may use the prefix "A" or "B" to describe the same type of nouns as two independent entities. In this case, the corresponding features defined with "A" and "B" are used only to distinguish between similar entities and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features.

[0023] In describing some embodiments, the terms "coupled," "coupled," and "connected," and their derivative expressions, may be used. For example, the term "connected" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact with each other. Similarly, the term "coupled" may be used in describing some embodiments to indicate that two or more components have direct physical or electrical contact. However, the terms "connected" or "coupled" may also refer to two or more components that do not have direct contact with each other but still cooperate or interact with each other, such as "optical coupling," "wireless connection," etc. The embodiments disclosed herein are not necessarily limited to the scope of this invention.

[0024] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0025] Existing technologies include some solutions that utilize knowledge graphs to achieve knowledge integration and intelligent applications. For example, there are technical concepts that display accident knowledge in the form of visual graphics and tables, and a basic accident knowledge visualization framework has been constructed. However, this solution only focuses on the visual presentation of knowledge and does not realize the automated extraction, semantic modeling, and correlation analysis of accident knowledge. The application of technology is relatively superficial and cannot fully leverage the intelligent advantages of knowledge graphs. The data value needs to be further explored.

[0026] Current knowledge graph and retrieval-augmented generation (RAG) solutions for the field of metallurgical accidents suffer from systemic bottlenecks in underlying modeling, industry adaptation, and business empowerment. Specific shortcomings are as follows: (1) There are defects in the underlying data modeling and entity recognition, and the data modeling accuracy is insufficient and misaligned with the actual scenario. On the one hand, the existing modeling scheme is relatively crude, adopting a general entity definition and failing to build a dedicated key-value (KV) correspondence table for metallurgical accident reports. This results in a low matching degree between the triples extracted based on the knowledge graph and the actual metallurgical accident scenario, which seriously affects the accuracy of knowledge representation. On the other hand, conventional techniques cannot identify the phenomenon of different names for the same entity in industrial data (for example, identifying the First Company Research Institute and the Research Institute as multiple isolated entities), resulting in a large amount of redundancy in the nodes within the knowledge base, which seriously undermines the statistical value.

[0027] (2) The analysis is only point-to-point based on the causes of accidents, lacking in-depth adaptation to the entire metallurgical process. The accident logic is not linked to the process flow and the accident causal chain for chain-to-chain analysis. First, existing research mostly draws on solutions from other fields such as construction and transportation, and does not design exclusive modeling solutions for the unique process flow of the metallurgical industry (for example, the steps in a process flow are: pelletizing, ironmaking, steelmaking, refining, continuous casting and rolling). This results in the inability of existing technology to provide risk warnings between the steps in the process flow, and the phenomenon of knowledge silos is prominent. Second, the technical system is fragmented. It does not establish a sequential connection between accidents and corresponding hidden dangers and their direct causes, indirect causes, emergency response, and the division of responsibilities and penalties after the accident through accident analysis. Therefore, it cannot provide intelligent causal analysis and accurate decision support, and cannot provide in-depth causal data support for safety performance evaluation and other content.

[0028] (3) The depth of the hidden accident chain is insufficient, making it impossible to provide early warning for accidents without a clear development path. General solutions often treat risk sources as isolated nodes for point-to-point static queries, failing to discover the latent long-distance correlations between nodes through the complex industrial mechanism topology. For example, general solutions cannot deduce from the abnormal state of a certain raw material in the sintering process (i.e., the indirect cause) across physical distances to a certain equipment hazard in the refining process (i.e., the direct cause), resulting in the truncation of the latent long-distance correlation. It is impossible to find the cause and establish the corresponding correlation relationship, construct the causal chain across processes, and thus fail to identify hidden risks.

[0029] To address the problems of existing technologies, this embodiment proposes a method for causal analysis of metallurgical accidents. In one embodiment, such as... Figure 1 As shown, it includes: Step 10: Perform data preprocessing on unstructured accident reports from the metallurgical industry to generate standardized text.

[0030] In this process, the content of unstructured accident reports is transformed into structured, standardized text through data preprocessing. The specific method is determined by those skilled in the art based on the specific application scenario and is not limited here. In one embodiment, a multi-source collection of full-process accident reports from the metallurgical industry can be performed through a data cleaning and preprocessing module. The reports are then normalized to obtain Portable Document Format (PDF), Word, or Optical Character Recognition (OCR) documents. The content in the documents is then subjected to layout analysis and semantic reconstruction to ultimately generate a standardized text dataset.

[0031] Step 20: Process the standardized text according to the metallurgical process steps, generate retrieval phrases for each entity as keys based on the predefined entity types of metallurgical accidents, and use the corresponding semantic paragraphs as values ​​to obtain standardized key-value pairs, and store them in the graph database.

[0032] In this approach, a search phrase for an entity is used as the key, and the corresponding semantic paragraph in the standardized text is used as the value to construct standardized key-value pairs. The entity type is determined by those skilled in the art based on the specific use case; a specific example will be provided below.

[0033] This invention provides a knowledge graph-based solution. A knowledge graph is a structured semantic network knowledge base that stores data in the form of triples. This invention uses it to integrate fragmented knowledge and achieve knowledge visualization and intelligent analysis. In one embodiment, the entity relationship extraction and graph-enhanced key-value modeling module first predefines eight types of entities to obtain the corresponding entity types. According to each stage of the metallurgical process, while taking into account the report structure, the standardized text is divided into multiple small text chunks. Then, the large language model extracts the retrieval phrases and corresponding semantic paragraphs of each entity from each small text chunk, thereby obtaining the corresponding triples and generating standardized key-value pairs, which are then injected into the Neo4j graph database, thus constructing a knowledge graph.

[0034] Step 30: Determine the semantic similarity between the entities, construct multiple clusters according to the entity types, and obtain the standard entity of each cluster. One standard entity corresponds to one node in the graph database.

[0035] The process involves clustering entities based on entity type and semantic similarity to obtain multiple clusters. Each cluster is then processed to select representative entities, resulting in a standard entity for each cluster. Each entity corresponds to a node stored in the graph database, and each standard entity also corresponds to a node in the graph database.

[0036] In one embodiment, data for each node in the graph database can be obtained through a differential threshold clustering analysis model module, metallurgical word vectors can be trained and entity vectors can be generated, clustering can be performed according to differential thresholds, and then standard entities can be updated and redundant nodes can be deleted.

[0037] Step 40: Generate a risk graph based on the nodes marked as high-frequency hidden dangers or historical accidents in the graph database; iteratively update the risk status of the nodes in the risk graph until convergence, and aggregate them into risk communities; trace back the directed relationship from the risk source within the risk community to restore the causal chain across processes, so as to output early warning information.

[0038] In one embodiment, those skilled in the art pre-mark certain nodes in a graph database as nodes with high-frequency hidden dangers or frequent historical accidents, thereby generating a risk graph characterizing the safety risks corresponding to each node, and determining the risk status of each node in the overall risk graph through iterative training. A specific example will be provided below, and will not be elaborated further here.

[0039] This invention preprocesses unstructured accident reports to generate standardized text and standardized key-value pairs, storing them in a graph database. This avoids simply creating unstructured accident reports and solves the problems of scattered, weakly correlated, and difficult-to-retrieve accident report data in existing technologies. It provides a standardized and intelligent analysis system that preprocesses unstructured accident reports sequentially, extracts entities, clusters the data stored in the graph database, and iteratively simulates risk propagation. This automatically reconstructs the causal chains across processes from unstructured accident reports, outputting pre-emptive warning information containing the corresponding causal chains. No manual intervention is required, avoiding reliance on manual interpretation, and achieving a leap from post-event tracing to pre-event warning, effectively preventing recurring accidents.

[0040] The method for causal analysis of metallurgical accidents according to embodiments of the present invention will be further described below: To address the systemic problems of traditional knowledge graph solutions in constructing accident knowledge bases in the metallurgical industry, such as low modeling accuracy, poor industry adaptability, inaccurate deduplication, and inability to provide early warnings, this invention constructs a novel metallurgical industry accident knowledge base. This base enables the extraction of cross-process causal chains from unstructured accident reports based on intelligent decision reconstruction and the output of early warning information. Specifically, this invention provides a causal analysis method for metallurgical accidents based on knowledge graphs and multidimensional semantic clustering. Using unstructured accident reports in the metallurgical industry as the processing object, the method is implemented step-by-step through data cleaning and preprocessing, entity relationship extraction and graph-enhanced key-value modeling, differential threshold clustering analysis, potential risk chain discovery, two-layer enhanced RAG retrieval, and Neo4j knowledge graph display and interaction modules. The specific operation steps for each module are explained below: In one embodiment, in order to transform messy, unstructured reports into standardized text, the data cleaning and preprocessing module first performs the following steps: Multi-source data acquisition: Collect accident investigation reports from metallurgical enterprises through system interfaces, web crawlers, or manual import, covering the entire process of pelletizing, ironmaking, steelmaking, refining, continuous casting, and rolling.

[0041] Formatting and layout processing: Normalize data from various formats such as PDF, Word, and scanned images; convert scanned images into editable text using OCR; use the mineu tool to perform layout analysis and table restoration on all text, preserving the original hierarchical structure.

[0042] Semantic Reconstruction and Quality Control: The parsed text is reconstructed into a semantic stream with Markdown hierarchical tags. Core parameters of metallurgical processes such as furnace temperature and gas pressure are specifically annotated, and garbled characters and duplicate text are removed to form a standardized text dataset.

[0043] Next, addressing the systemic problems of existing knowledge graph solutions in the construction of accident knowledge bases in the metallurgical industry, such as low modeling accuracy, poor industry adaptability, inaccurate deduplication, and inability to provide early warnings, the following solutions are provided: To construct a well-structured and semantically rich knowledge graph of metallurgical accidents, embodiments of this invention perform entity relation extraction and graph-enhanced key-value modeling. In one embodiment, such as... Figure 2 As shown, step 20 includes: Step 201: Semantically slice the standardized text according to the entire process sequence of the metallurgical process to obtain multiple small text blocks.

[0044] To address the issues of low accuracy and poor adaptability in extracting core elements of metallurgical accidents due to the generic entity definitions and lack of a dedicated modeling system tailored to specific metallurgical accident scenarios in traditional knowledge graphs, this invention predefines core entity types. In one specific instance, these include: accident, company name, accident location, direct cause, emergency response error type, indirect cause, company penalty, and personnel penalty. An augmented key-value (KV) modeling table is also created: the original accident report is segmented into multiple small text chunks according to metallurgical process stages (e.g., pelletizing, ironmaking, steelmaking, refining, continuous casting, and rolling), improving text processing efficiency while ensuring the scenario-based relevance of entity extraction. The specific method of semantic slicing is determined by those skilled in the art based on the specific use case.

[0045] Step 202: Extract instance relationships of entities from the small text block using a prompt-driven large language model to obtain triples, wherein the triples consist of entity A, the relationship between entity A and entity B, and entity B.

[0046] For each extracted entity, a standardized key-value pair is generated. The key of the standardized key-value pair is a retrieval identifier, which is a natural language phrase generated according to rules combining entity types and core features. The value of the standardized key-value pair is a semantic summary paragraph of the original text fragment. The key and value pairs are stored as attributes and corresponding triples for the respective entities in a graph database. The specific method for extracting instance relationships of entities from small text blocks using a prompt-driven large language model is determined by those skilled in the art based on the specific use case and is not limited here.

[0047] Step 203: Determine the retrieval identifier of the entity type including the entity, and obtain the semantic summary of the standardized text corresponding to the entity.

[0048] The specific methods for converting or processing the entity type of an entity and using it as a retrieval identifier, obtaining the standardized text corresponding to the entity, and obtaining the semantic summary of the corresponding standardized text are determined by those skilled in the art based on the specific use case. In one embodiment, the standardized text is processed by a natural language model according to prompts containing entity types.

[0049] Step 204: Use the search identifier as the value and the corresponding semantic summary as the key to generate a standardized key-value pair for the entity.

[0050] In one embodiment, a prompt-driven large language model can be used to extract entities and relationships; at the same time, standardized key-value pairs are generated for each entity to realize the process of steps 202 to 204 above: where the key is a search identifier in a unified format and the value is a semantic paragraph summarizing the relevant text fragments, which is used to assist in the authenticity verification of subsequent entity extraction and improve the accuracy of retrieval.

[0051] Step 205: Transform the triples into a topological structure, which consists of node A, edges, and node B; add the normalized key value corresponding to the entity represented by the node to the node for storage in the graph database.

[0052] In one embodiment, the generated KV data (i.e., normalized key-value pairs) is injected as attributes into nodes of Neo4j (i.e., a graph database). The key of the normalized key-value pair serves as the node's main label or index, and the value corresponding to that key serves as the detailed attribute description of the node, supporting subsequent semantic verification. A specific example is provided in this embodiment, as shown in Table 1 below: Table 1. A standardized key-value pair and specific examples of its business scenarios.

[0053] In one embodiment, a core entity specific to metallurgical accidents is defined through an entity relation extraction and graph-enhanced key-value modeling module, clarifying the basic relationships between entities (e.g., causing, occurring at, and punishing). Then, text semantic slicing is performed: standardized text is segmented into independent semantic chunks according to metallurgical process steps and accident report structure (e.g., accident overview, cause analysis, handling process, and punishment results). Triple extraction is achieved through LLM, using customized prompts for each text chunk based on metallurgical equipment, processes, and safety regulations to guide LLM in extracting triple data chunks composed of entity-relationship-entity pairs. Finally, key-value generation and Neo4j injection are implemented, generating standardized key-value pairs for the extracted entities or relations. The key is determined based on entity type and core features, while the value is a semantic summary paragraph of the original text. The key-value data is injected into Neo4j, with the key serving as the node's main label or retrieval index, and the value serving as the node's detailed attributes. Simultaneously, the triples are transformed into a "node-edge-node" topological structure in Neo4j, consisting of nodes and adjacent edges.

[0054] Existing Knowledge Graph-Retrieval-augmented Generation (KG-RAG) technologies suffer from significant node redundancy, compromising data uniqueness and statistical value. Furthermore, traditional clustering methods, employing a single similarity threshold, fail to address the inaccurate deduplication and poor industry adaptability issues arising from the existence of synonyms and ambiguities in the metallurgical field. For example, "continuous casting workshop" and "continuous casting section" are concrete examples of synonyms; another example is the ambiguity of the term "hazard," which can refer to both equipment hazards and management hazards in different reports.

[0055] To eliminate node redundancy caused by duplicate names and ensure data uniqueness, this invention also provides a differentiated threshold clustering analysis scheme. In one embodiment, such as... Figure 3 As shown, step 30 includes: Step 301: Calculate the semantic similarity between entities in the graph database.

[0056] To address the aforementioned issues, this invention provides a differential threshold clustering analysis model module to achieve accurate deduplication and aggregation, as detailed below: First, in one embodiment, a cosine similarity algorithm is used to calculate the semantic similarity between entities.

[0057] Step 302: Set the asymmetric similarity threshold according to the entity type of the metallurgical emergency scenario.

[0058] Then, based on the entity types in the metallurgical emergency scenario, an asymmetric similarity threshold is set, and a quantitative standard is defined. In one embodiment: for entities with strict qualitative requirements that cannot be mistakenly merged (e.g., accident type, direct cause, and emergency response error type), a high similarity threshold (0.90 to 0.95) is set to prevent different entities from being mistakenly merged; for entities with a large number of abbreviations and full names, and common names and standard names mixed together (e.g., accident location, company name, and equipment name), a moderately lenient similarity threshold (e.g., a similarity threshold of 0.70 to 0.80) is set to facilitate the automatic aggregation of similar entities in the future.

[0059] Step 303: Group entities whose semantic similarity reaches the corresponding asymmetric similarity threshold into the same cluster.

[0060] The specific method for grouping entities into the same cluster shall be determined by those skilled in the art based on the specific application scenario, and is not limited here.

[0061] Step 304: Select standard entities in each cluster and merge the valid attributes of other entities, and delete redundant nodes to achieve unique node updates.

[0062] Finally, standardized update rules for clustered entities are established, retaining the most frequent and semantically complete entities in the cluster as standard entities, and integrating the effective information of other entities to ensure the integrity and uniqueness of the deduplicated data.

[0063] In one embodiment, in each cluster, the entity with the highest frequency and the most complete semantics is retained as the standard entity. The valid attribute information of other entities in the cluster is merged into the standard entity. Then, other redundant nodes are deleted, and entity mapping rules are generated for subsequent retrieval and reasoning.

[0064] In one embodiment, an industry-specific word vector model is trained using a differentiated threshold clustering analysis model module: a customized word vector model for the metallurgical industry is trained based on corpora specific to metallurgical safety management, accident handling, and process technology. Entity vectorization and index construction involve loading the trained word vector model, converting all entity names in the graph into normalized semantic vectors, and constructing a cosine similarity index using FAISS technology. Differential similarity thresholds are set: a high similarity threshold of 0.90 to 0.95 is set for accident type, direct cause, and emergency response error type; a relaxed similarity threshold of 0.70 to 0.80 is set for accident location, company name, and equipment name. Entity clustering calculation involves using the cosine similarity algorithm to calculate the semantic similarity between all entities based on the cosine similarity index, grouping entities that reach the corresponding threshold into the same cluster. Ultimately, a unified update of the graph nodes is achieved. In each cluster, the entity with the highest frequency and the most complete semantics is selected as the standard entity, and the effective attributes of other entities in the cluster are integrated. The standard entity is then synchronized to Neo4j, redundant nodes are deleted, and entity mapping rules are generated.

[0065] To uncover hidden risk chains across processes and achieve early warning, embodiments of this invention provide a potential risk mining scheme based on directed graph search. In one embodiment, such as... Figure 4 As shown, in step 40, generating a risk map based on nodes marked as high-frequency hidden dangers or historically frequent accidents in the graph database includes: Step 401a: Extract subgraphs that meet the conditions for safe production from the graph database, and construct a graph model for risk inference using node sets and directed edges with semantic orientation.

[0066] The specific methods for determining the safety production conditions and constructing the graph model are determined by those skilled in the art based on the specific application scenario. The safety production conditions are used to filter out the nodes corresponding to entities that are strongly related to safety production in the graph database, thereby obtaining the corresponding subgraphs.

[0067] Step 402a: Add initial risk labels to the nodes in the graph database that are pre-marked as high-frequency hidden dangers or historically frequent accidents to obtain the risk sources.

[0068] The risk labels are subsequently updated iteratively based on the initial risk labels defined here.

[0069] Then, risk graph projection and seed injection are performed. In one embodiment, the system first extracts subgraphs strongly related to safe production from the Neo4j knowledge graph and constructs a risk projection graph model G=(V,E). Here, V is a set of nodes (e.g., equipment, process links, hidden dangers, etc.), and E is a directed edge with semantic reference (e.g., edge causes, edge triggers, edge belongs to). The system automatically identifies nodes in the knowledge base marked as high-frequency hidden dangers or frequently occurring historical accidents and injects initial risk labels into them. As a source of risk spread.

[0070] Step 403a: For the edges between nodes in the graph database, a semantic label is preset for each edge; the weight corresponding to the semantic label is determined according to the mapping rule and stored as the attribute of the edge to obtain the risk graph.

[0071] In one embodiment, the preset semantic labels may include direct causes and indirect causes, etc. The mapping rules between the semantic labels, the semantic labels of each edge and the weights of that edge, and the specific method for constructing the risk graph are determined by those skilled in the art based on the specific use case.

[0072] The potential risk chain discovery module based on directed graph search in this invention addresses the shortcomings of traditional knowledge graphs, which can only achieve post-event tracing, cannot discover potential risks across processes in the entire metallurgical industry, and lack pre-event warning capabilities. It performs risk evolution deduction and combines directed graph path search to achieve causal chain mining. In one embodiment, such as... Figure 5 As shown, in step 40, the iterative update of the risk status of the nodes in the risk graph until convergence is achieved, and they are aggregated into risk communities; within the risk community, directed relationships are traced back from the risk source to restore the causal chain across processes, so as to output pre-warning information, including: Step 401b: Iteratively update the risk concentration value of each node according to the weight of the edge until convergence, and aggregate nodes with the same or similar risk labels within a preset range after convergence into a risk community.

[0073] The preset range is determined by those skilled in the art based on the specific application scenario, and is not limited here.

[0074] In one embodiment, nodes with the same risk label after convergence are aggregated into a risk community. After aggregating the converged nodes into risk communities, the method further includes: detecting node pairs that are within the same risk community, have a physical distance of 3 or more in the metallurgical process flow, and belong to different process stages; identifying the association between node pairs as a cross-process implicit logical association; and using this as a basis for systemic risk warning.

[0075] Step 402b: Within the risk community, starting from the node that is the source of risk, backtrack along the directed edge causal semantics based on multiple constraints to the node of the potential accident to obtain the causal chain, and output the causal chain as early warning information.

[0076] In one embodiment, within the risk community, node pairs that are physically distant by ≥3 and cross processes are detected (e.g., node pairs consisting of sintering risks and finished product warehouse risks).

[0077] After multiple iterations and convergence, risk community discovery and implicit logic mining aggregate nodes with the same or highly similar risk labels into risk communities. Cross-process association discovery: The system can detect nodes that are in the same risk community but are physically far apart in the metallurgical process (association path length ≥ 3 and belonging to different process stages) (e.g., lubrication risks in the sintering stage and hoisting risks in the finished product warehouse), and discover implicit logical associations across physical space to identify potential systemic risks.

[0078] In one embodiment, a breadth-first search can be used to trace back from the source of the hazard to the node of the potential accident in the risk graph (i.e., a directed graph), thereby reconstructing the specific causal chain (e.g., from the source of the hazard to the intermediate anomaly and then to the potential accident) and providing it to the user as a key checklist.

[0079] In one embodiment, a potential risk chain discovery module using directed graph search can be used to project the risk graph and inject seeds. After weighted propagation of the risk graph, risk communities are obtained and implicit associations are mined, ultimately restoring the causal chain for visualization.

[0080] In one embodiment, a two-layer retrieval enhanced RAG model module can perform precise matching of nodes and key-value attributes at the lower layer; and perform global keyword retrieval and aggregation at the higher layer, concatenating the values ​​of standardized key-value pairs with the user's question, and finally generating the answer through a large language model.

[0081] In one embodiment, iteratively updating the risk concentration value of each node according to the weight of the edge until convergence includes: To simulate the physical mechanism of risk transmission in metallurgical processes, this embodiment of the invention defines the metallurgical process topology graph as a directed graph G=(V,E,W) in the dynamic propagation mechanism of risk concentration by introducing edge weights. In the next iteration, the node The risk state concentration value is set to The system determines the risk label of neighbor node vj and the weight of the associated edges. The risk concentration is dynamically weighted and updated iteratively according to the following risk concentration dynamic weighted propagation formula: ; The number of predecessor nodes of the pointer node is limited; in one embodiment, the number of predecessor nodes of the pointer node ranges from 10 to 20.

[0082] in ， Indicates pointing to a node The set of predecessor nodes; For nodes To the node The weight of the direct cause of the accident (e.g., assigning a strong causal relationship to the direct cause of the accident). =0.8, weak causal association, indirect cause assignment. =0.4); For the number of iterations, For nodes V k To the node V i Weights; denominator This is a normalization term to ensure the conservation of risk concentration propagation within the topological space.

[0083] Furthermore, to accelerate convergence and simulate a surge in risk threshold, an activation function is introduced to perform a non-linear mapping on the updated risk concentration: In one embodiment, during the iteration process, the following activation function is used to perform a non-linear mapping on the updated risk concentration value: ; in, For nodes The current risk label, For nodes The risk concentration value.

[0084] This invention ensures that risk labels will preferentially spread along paths with stronger causal certainty through this mechanism, simulating the concentration propagation of risk within the topological space.

[0085] In one embodiment, the step of tracing back directed relationships from the risk source within the risk community to reconstruct the causal chain across processes further includes: when reconstructing the causal chain, only directed edges that satisfy a preset semantic type and whose weight is greater than or equal to the weight threshold of a weak causal relationship are added to the search tree. During queue expansion, each hop only adds directed edges that satisfy the aforementioned core semantics (Type) and whose weight... Edges are added to the search tree. Through this optimization equation, the system can filter out redundant and noisy paths in the graph, and rigorously transform the abstract topological network into a continuous, interpretable, shortest rigid causal chain. This shortest rigid causal chain is used to clearly describe the complete evolutionary path from the source of the hidden danger, through the causal relationship to a series of intermediate abnormal states, and finally through the triggering relationship to the potential accident; thus providing expert investigation guidance with strong mathematical logic support.

[0086] Based on the identification of risk communities, this embodiment of the invention also performs backtracking and deduction in the directed graph to further determine the specific evolutionary path: Let the mined risk communities be subgraphs, and define the initial hazard source node as... Potential accident nodes are The system searches for a set of causal evolutionary paths within this subgraph. At that time, the following multiple constraint equations must be satisfied: In one embodiment, the expression for the multiple constraints is: ; in, This represents the optimal causal path selected from the candidate causal evolution paths; Indicates the first in the path A directed edge; Indicates the first A directed edge The weights; This indicates the number of directed edges in the path; Indicates the first in the path One node; Indicates a high-risk community; Indicates the first A directed edge The semantic type. Set of causal evolutionary paths. It is the set of all that meet the requirements. This is the optimal path selected from among them.

[0087] In one embodiment, the potential risk chain discovery module via directed graph search can be integrated with the clustering analysis, RAG retrieval, and risk mining modules in Neo4j, with relevant results presented in real-time on a visual interface. Risk graph projection and seed injection are performed: extracting strongly related subgraphs for safety production from the Neo4j graph to construct a risk projection graph model G=(V,E) (V being a set of nodes such as equipment nodes, process nodes, and hidden danger nodes; E being directed edges such as edge-caused and edge-triggered edges); injecting initial risk labels Lrisk into high-frequency hidden danger nodes and historically accident-prone nodes in the knowledge graph. Weighted LPA risk propagation settings are then implemented, assigning weight factors based on the semantic type of edges: Whigh=0.8 for strong causal relationships and Wmid=0.4 for weak relationships; iterative propagation of risk labels is initiated, calculating the node risk state by weighting the risk labels of neighboring nodes and the weights of associated edges until the iteration converges. Risk community and implicit association mining involves aggregating nodes with the same or highly similar risk labels into risk communities after convergence. In one embodiment, node pairs within a risk community with path lengths greater than or equal to a preset value and belonging to different process stages are detected to mine implicit logical associations across processes. The preset value is determined by those skilled in the art based on the specific application scenario; in an optional embodiment, the preset value can be 3. Causal chain reconstruction involves using a breadth-first search (BFS) algorithm within the risk community, starting from the initial risk source node, and backtracking along directed edges to extract quantifiable transmission paths. The topological path is then transformed into a causal chain from the potential accident hazard to intermediate transmission nodes to the potential accident.

[0088] In one embodiment, the method further includes a two-layer retrieval enhancement generation step: for specific entity detail queries, precise matching of nodes and edges in a graph database is used to extract value information from corresponding key-value pairs as low-level retrieval results; for abstract topic queries, global keyword retrieval in a graph database is used to aggregate value information from key-value pairs of multiple related entities and relationships as high-level retrieval results; the low-level or high-level retrieval results are structurally concatenated with the user's query question and input into a large language model to generate the final answer. Low-level retrieval: For specific entities (e.g., the direct cause of an accident), precise matching of nodes using Neo4j is used to extract their key-value attributes. High-level retrieval: For abstract topics (e.g., common hidden dangers in continuous casting processes), global keyword retrieval of the graph is performed to aggregate information from multiple nodes. The values ​​in the retrieved standardized key-value pairs (i.e., the semantic paragraphs corresponding to the entities) are concatenated with the user's question and input into an LLM to generate the final answer, ensuring that the answer is fact-based and traceable.

[0089] In one embodiment, a two-layer enhanced RAG model module can perform low-level precise retrieval, i.e., for specific entity detail queries, extracting and summarizing corresponding key-value (KV) attribute information through precise matching of Neo4j graph nodes and edges. It can also perform high-level global retrieval, i.e., for abstract topic queries, aggregating KV attribute information of multiple related entities and relationships through global graph keyword retrieval. Finally, it enhances answer generation by structurally concatenating the Value portion of the KV attributes obtained from the two-layer retrieval with the user query structure, inputting it into the LLM, and generating the answer from the LLM.

[0090] This invention elevates accident analysis from merely post-incident tracing in existing technologies to proactive early warning. Verified by multiple batches of metallurgical industry accident reports (i.e., accident reports covering the entire process of ironmaking, steelmaking, and continuous casting), this module possesses the following technical effects: First, it enables quantitative prediction, identifying cross-process linkage risks that are difficult to identify manually, with a significantly improved accuracy compared to traditional knowledge graphs. It also allows for precise investigation, providing safety inspectors with a logically supported checklist of key areas for inspection. The knowledge graph database's visualization interface highlights risk propagation paths, allowing safety managers to intuitively observe the evolution of potential hazards, thus improving the efficiency and targeting of safety inspections. Finally, this invention enables the discovery of systemic hazard evolution paths at the industry level, providing scientific data support for the formulation and revision of safety standards in the metallurgical industry and the optimization of safety control systems.

[0091] The technical effects of the method in the embodiments of the present invention are described below: The differential threshold clustering analysis model proposed in this invention effectively solves the inaccuracies caused by node redundancy, homonymy, and polysemy in traditional knowledge graph applications in the metallurgical industry. Table 2 below shows a comparison of the core data processed by a traditional method and the clustering model of this invention.

[0092] Table 2 Comparison of core data processed by traditional methods and the clustering model of this invention

[0093] In one embodiment, the knowledge graph database of this invention is evaluated using existing open-source projects. The core evaluation metrics (i.e., fidelity, answer relevance, and context relevance) are plotted on the vertical axis, and the scores of the traditional RAG scheme and the scheme of this invention are plotted on the horizontal axis, resulting in a... Figure 6The box plot shown is a comparison of the scores of the traditional RAG scheme and the scheme of this invention. Tests were conducted on 20 different questions, and the results show that the scores of the present invention are higher than those of existing technical solutions in all aspects. The core reason is that the present invention achieves the authenticity verification of entity extraction through clustering algorithms and graph augmentation standard key-value pair modeling, and combined with differential clustering, ensures the accuracy of the retrieval data source, solving the retrieval bias and answer distortion problems that are prone to occur in existing RAG schemes.

[0094] This invention constructs a dedicated modeling system tailored to metallurgical accident scenarios at the data modeling level. Through predefined core entities and graph-enhanced key-value (KV) modeling, it solves the problems of multi-hop logic confusion and low modeling accuracy caused by the generalization and lack of specificity in existing technologies. Existing knowledge graphs are often chaotic, with many nodes failing to form chains; they only extract entities, resulting in a disorganized knowledge graph that fails to form a logical chain from the direct cause of the accident to the accident and corresponding penalties, thus failing to support further accident chain analysis. This invention extracts the logical chain of the direct cause of the accident and the triple from the accident to the corresponding penalty based on standardized key-value pairs. Experimental tests show the accuracy of the logical chains extracted using conventional methods and the method of this invention. A comparative example is provided. Figure 7 As shown, the entity extraction accuracy of the modeling method of this invention reaches 88.5%, which is 16.7% higher than that of existing technologies. This is mainly because existing modeling methods do not specify specific entity types and multi-hop logic chains for metallurgical accident scenarios, leading to omissions of core elements in entity extraction, confusion in relationship identification (e.g., misjudging indirect causes as direct causes), and logical gaps in multi-hop reasoning, ultimately resulting in a significant decrease in modeling accuracy. In contrast, this invention predefines eight core entity categories, clarifies the association rules between entities, and combines key-value modeling authenticity verification to standardize the multi-hop reasoning logic from the accident to direct and indirect causes, as well as corresponding handling and subsequent penalties. This improves the accuracy and industry adaptability of data modeling from a deeper mechanistic perspective, solving the problem of mechanistic gaps in existing technologies for modeling accident-related logic.

[0095] Based on the clustering and modeling of the accident chains described above, this invention achieves in-depth mining of accident chains through the Neo4j potential risk chain discovery module based on directed graph search. Compared with the traditional KG-RAG scheme, which can only achieve post-event tracing based on existing accident reports, it has the following advantages: 1) Existing technical methods can only mine local, intuitive causal chains within the accident process with a path length L≤2; the method of this invention can mine implicit risk transmission chains across the entire metallurgical process, across processes, non-intuitive, and with a path length L>3. 2) Existing technical methods only perform question-and-answer based on existing facts; the method of this invention uses improved LPA to simulate risk diffusion, deducing potential accidents from local states or combinations of states that have never occurred, thus achieving early warning.

[0096] Below is a comparative example of how existing methods and the method of this invention can detect cross-process splashing early warnings from a local state. In the flow of metallurgical processes, the physical state of the previous process often determines the safety risk of the next process. Taking the feedback status (abnormality) node of the tundish temperature monitoring system in the continuous casting process as an example of a local risk source node: Multiple studies indicate that as the number of hops in a knowledge graph increases, accurately retrieving relevant paths and filtering noise becomes a core challenge of KG-RAG. Existing KG-RAG methods are limited by mining depth, typically exhibiting localized source tracing (path length = 2). The search results primarily show the association between the system anomaly and the equipment maintenance report within that process, relying solely on existing knowledge entity matching and failing to deduce across physical connections. The mined causal chain is limited to: [intermediate batch temperature control system anomaly] - intermediate batch system failure. This method cannot predict the impact on the next process (refining).

[0097] This invention employs a deep mining strategy (path length L > 4, cross-process). Its mining logic is as follows: First, through iterative risk concentration diffusion, the risk is diffused from the tundish temperature control anomaly node to the physically connected refining process node; then, directed path reconstruction is used to reconstruct the specific path along the equipment state propagation edges and triggering edges. The final hidden causal chain (i.e., the cross-process causal chain) discovered through deep mining is: [Tundish temperature control system anomaly] - [Large fluctuation in tundish molten steel temperature] - [Unstable molten steel quality in refining furnace] - [Refining temperature control overload] - [Potential splashing accident B in refining process]. For example... Figure 8 As shown, this causal chain successfully provided an early warning of the potential splashing risk caused by overload of the refining furnace temperature control system.

[0098] The foregoing embodiments provided a method for causal analysis of metallurgical accidents. In this embodiment, another device for causal analysis of metallurgical accidents will be proposed. The device includes a processor and a memory for storing processor-executable instructions. The processor is configured to execute the method for causal analysis of metallurgical accidents described in the foregoing embodiments.

[0099] like Figure 9 As shown, the metallurgical accident causality analysis device includes a processor 21 and a memory 22, wherein the processor 21 and the memory 22 can be connected by a bus or other means.

[0100] Processor 21 can be a CPU. Processor 21 can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, or combinations of the above types of chips.

[0101] The memory 22, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the metallurgical accident causal analysis method in the aforementioned embodiments. The processor executes various functional applications and training processes by running the non-transitory software programs, instructions, and modules stored in the memory.

[0102] The memory 22 may include a program storage area and a training storage area. The program storage area may store the operating system and applications required for at least one function; the training storage area may store training data created by the processor. Furthermore, the memory may include high-speed random access memory and non-transitory memory, such as at least one disk storage device, flash memory, or other non-transitory solid-state storage device. In some embodiments, the memory 22 may optionally include memory remotely located relative to the processor, which can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. The one or more modules stored in the memory 22, when executed by the processor 21, perform the metallurgical accident causal analysis method as shown in the embodiments of the present invention. Specific details of the metallurgical accident causal analysis method can be understood by referring to the corresponding descriptions and effects in the embodiments of the present invention, and will not be repeated here.

[0103] This embodiment also provides a computer storage medium storing a computer program that can be executed by a processor to perform the metallurgical accident causal analysis method described in the foregoing embodiments.

[0104] The computer storage medium stores computer-executable instructions, which can execute the metallurgical accident causal analysis method in any of the above method embodiments. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk drive (HDD), or solid-state drive (SSD), etc.; the storage medium may also include combinations of the above types of memory.

[0105] The specific steps of the metallurgical accident causal analysis method are described in the foregoing embodiments and will not be repeated in this embodiment.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for causal analysis of metallurgical accidents, characterized in that, include: Data preprocessing is performed on unstructured accident reports in the metallurgical industry to generate standardized text; The standardized text is processed according to the metallurgical process steps. Based on the predefined entity types of metallurgical accidents, the search phrases of each entity are generated as keys, and the corresponding semantic paragraphs are used as values ​​to obtain standardized key-value pairs, which are then stored in the graph database. Determine the semantic similarity between the entities, construct multiple clusters according to the entity types, and obtain the standard entity of each cluster. One standard entity corresponds to one node in the graph database. A risk map is generated based on the nodes marked as high-frequency hidden dangers or historically frequent accidents in the graph database; Iteratively update the risk states of the nodes in the risk graph until convergence, and aggregate them into risk communities; Within the risk community, the directed relationships are traced back from the risk source to restore the causal chain across processes, so as to output early warning information.

2. The method for causal analysis of metallurgical accidents according to claim 1, characterized in that, The process of generating a risk map based on nodes marked as high-frequency hazards or historically frequent accidents in the graph database includes: Subgraphs that meet the conditions for safe production are extracted from the graph database, and a graph model for risk inference is constructed using node sets and directed edges with semantic orientation. For nodes in the graph database that are pre-marked as high-frequency hidden dangers or have a high incidence of historical accidents, add initial risk labels to obtain the risk sources; For each edge between nodes in the graph database, a semantic label is preset for each edge; the weight corresponding to the semantic label is determined according to the mapping rules and stored as an attribute of the edge to obtain a risk graph.

3. The method for causal analysis of metallurgical accidents according to claim 2, characterized in that, The iterative update of the risk state of the nodes in the risk graph until convergence is achieved, and they are aggregated into risk communities; Within the aforementioned risk community, directed relationships are traced back from the risk source to reconstruct the causal chain across processes, in order to output pre-emptive warning information, including: The risk concentration value of each node is iteratively updated according to the weight of the edge until convergence. After convergence, nodes with the same or similar risk labels within a preset range are aggregated into risk communities. Within the risk community, starting from the node that is the source of risk, the causal semantic backtracking is performed along the directed edge based on multiple constraints to the node of the potential accident, thus obtaining the causal chain, and outputting the causal chain as early warning information.

4. The method for causal analysis of metallurgical accidents according to claim 3, characterized in that, The step of iteratively updating the risk concentration value of each node according to the weight of the edge until convergence includes: The following dynamic weighted propagation formula for risk concentration is used for iterative updates: ； in ， Indicates pointing to a node The set of predecessor nodes; For nodes To the node The weight, For the number of iterations, For nodes V k To the node V i The weights; During the iteration process, the following activation function is used to perform a non-linear mapping on the updated risk concentration values: ； in, For nodes The current risk label, For nodes The risk concentration value.

5. The method for causal analysis of metallurgical accidents according to claim 4, characterized in that, The expression for the multiple constraints is: ； in, This represents the optimal causal path selected from the candidate causal evolution paths; Indicates the first in the path A directed edge; Indicates the first A directed edge The weights; This indicates the number of directed edges in the path; Indicates the first in the path One node; Indicates a high-risk community; Indicates the first A directed edge The semantic type.

6. The method for causal analysis of metallurgical accidents according to claim 2, characterized in that, The step of tracing back directed relationships from risk sources within the risk community to reconstruct causal chains across processes also includes: When restoring the causal chain, only directed edges that satisfy the preset semantic type and whose weight is greater than or equal to the weight threshold of weak causal relationship are added to the search tree.

7. The method for causal analysis of metallurgical accidents according to claim 1, characterized in that, The process of processing the standardized text according to the metallurgical process steps, generating search phrases for each entity as keys based on predefined entity types of metallurgical accidents, and using corresponding semantic paragraphs as values ​​to obtain standardized key-value pairs, and storing them in a graph database includes: The standardized text is semantically sliced ​​according to the entire process sequence of metallurgical processes to obtain multiple small text blocks; The instance relationships of entities are extracted from the small text blocks by a prompt-driven large language model to obtain triples, which consist of entity A, the relationship between entity A and entity B, and entity B. Determine the retrieval identifier for the entity type, including the entity, and obtain a semantic summary of the standardized text corresponding to the entity; The search identifier is used as the value, and the corresponding semantic summary is used as the key to generate a standardized key-value pair for the entity; The triples are transformed into a topological structure consisting of node A, edges, and node B; the normalized key values ​​corresponding to the entities represented by the nodes are added to the nodes for storage in the graph database.

8. The method for causal analysis of metallurgical accidents according to claim 7, characterized in that, The process of determining the semantic similarity between the entities, constructing multiple clusters based on the entity types, and obtaining standard entities for each cluster, wherein a standard entity corresponds to a node in the graph database, includes: Calculate the semantic similarity between entities in the graph database; Set an asymmetric similarity threshold based on the entity type in the metallurgical emergency scenario; Entities whose semantic similarity reaches the corresponding asymmetric similarity threshold are grouped into the same cluster; In each cluster, standard entities are selected and the effective attributes of other entities are merged. Redundant nodes are deleted to achieve unique node updates.

9. A device for analyzing the cause and effect of metallurgical accidents, characterized in that, The metallurgical accident causality analysis device includes: a processor and a memory for storing processor-executable instructions; The processor is configured to execute the metallurgical accident causal analysis method according to any one of claims 1 to 8.

10. A non-volatile computer storage medium, characterized in that, The computer storage medium stores computer-executable instructions, which are executed by one or more processors to perform the metallurgical accident causal analysis method according to any one of claims 1 to 8.

Images