A metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on a knowledge graph and a GraphRAG

Abstract

The application discloses a metallurgical equipment fault reasoning and hierarchical maintenance list generation system and method based on a knowledge graph and a GraphRAG, and belongs to the technical field of intelligent operation and maintenance of metallurgical equipment. The system comprises a knowledge graph construction module, a GraphRAG reasoning module, a fault risk grading module, a maintenance list generation and optimization module, a data interaction module and a model updating module, and forms an operation and maintenance closed loop of data input-knowledge modeling-fault reasoning-risk grading-maintenance generation-model updating; GraphRAG hierarchical reasoning is adopted; a Cox proportional risk regression or an equivalent model is used to realize dynamic RPN correction according to working conditions, and the evaluation result dynamically coupled with RQR (operation behavior-product quality-system reliability) is mapped and divided into risk grades by a segmented function; and a standardized hierarchical maintenance list is output by fusing RCM and RBM multi-objective optimization.

Description

Technical Field

[0001] This invention belongs to the field of intelligent operation and maintenance technology for metallurgical equipment, and relates to a system for fault reasoning and hierarchical maintenance list generation for metallurgical equipment based on knowledge graphs and GraphRAG. Background Technology Hydrogen-based vertical shaft furnace feeding systems, rolling mills, continuous casting machines, and other metallurgical equipment are core infrastructures in steel production. Their operational status directly determines production continuity, product quality, and safety levels. Metallurgical equipment is complex in structure, operates under harsh conditions such as high temperature, high pressure, high load, and strong corrosion, and exhibits diverse and highly correlated failure modes. Traditional fault diagnosis and maintenance management methods have many limitations and cannot meet the needs of intelligent operation and maintenance.

[0002] In projects such as the construction of a fault mechanism database for metallurgical equipment and reliability research, the collaborating parties have clearly defined hydrogen metallurgical continuous reaction equipment as the verification carrier to build a complete technical system from fault database to reliability modeling and intelligent operation and maintenance decision-making. Technical indicators include: a fault mechanism database covering no fewer than 30 typical failure modes and being scalable; establishing a dynamic coupling model of operational behavior-product quality-system reliability (RQR); and integrating reliability-centered maintenance (RCM) and risk-based maintenance (RBM) to generate a graded maintenance task list. The existing product platform still needs to develop systematic invention solutions in knowledge reasoning, graph-enhanced retrieval, and automatic maintenance list generation.

[0003] In existing technologies, one type of solution is based on traditional Failure Mode and Effects Analysis (FMEA) to complete fault identification and risk assessment, with maintenance lists being manually compiled. Its main problems include: weak fault reasoning ability, making it difficult to cope with multiple fault coupling and cascading failures; the priority of the maintenance list is based only on the static Risk Priority Number (RPN), without combining real-time operating conditions, RQR assessment, and downtime window; and the knowledge is scattered in document form, making it difficult to retrieve in a unified manner with test outlines and maintenance cases.

[0004] Another approach constructs a static fault knowledge graph for fault reasoning, but suffers from problems such as lagging knowledge updates, insufficient mining of unstructured knowledge, and lack of a closed loop with maintenance list generation. Yet another approach uses general retrieval-enhanced generation (RAG) for maintenance question answering, but the general RAG does not incorporate the hierarchical structure of metallurgical equipment, lacks structured guidance and GraphRAG community summary retrieval, and the maintenance list generation is uncontrollable, making it difficult to meet the requirements of high-reliability maintenance.

[0005] In summary, the existing technologies have at least the following shortcomings: First, fault reasoning lacks the synergy between structured and precise retrieval and graph augmentation generation, resulting in insufficient accuracy and interpretability, and the mechanistic knowledge is difficult to update dynamically with new cases and test outlines; Second, risk quantification is still mainly based on static RPN, failing to incorporate the dynamic RPN and RQR dynamically coupled evaluation results into risk classification; Third, maintenance list generation lacks a hierarchical optimization that integrates reliability-centered maintenance (RCM) and risk-based maintenance (RBM), and an iterative closed loop has not yet been formed between perception, evaluation, and decision-making. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a system for fault reasoning and graded maintenance list generation of metallurgical equipment based on knowledge graphs and GraphRAG, in order to practically solve the following technical problems: how to obtain interpretable fault conclusions by adopting hierarchical reasoning with structured precise retrieval priority, graph retrieval and vector retrieval enhancement, and large model generation as a fallback, supported by multi-source mechanism knowledge of metallurgical equipment; how to jointly classify risk levels by using piecewise function mapping to combine the comprehensive risk level output by the dynamically coupled model of RPN and RQR that is dynamically corrected according to operating conditions; and how to generate and optimize graded maintenance lists by integrating RCM and RBM decision rules under the drive of risk classification results, thereby forming a closed loop of operation and maintenance of perception-evaluation-decision.

[0007] To achieve the above objectives, the present invention provides the following technical solution: In the first aspect, the present invention provides a metallurgical equipment fault reasoning and graded maintenance list generation system based on knowledge graph and GraphRAG, including a knowledge graph construction module, a GraphRAG reasoning module, a fault risk grading module, a maintenance list generation and optimization module, a data interaction module, and a model update module, forming a complete operation and maintenance closed loop of data input-knowledge modeling-fault reasoning-risk grading-maintenance generation-model update. The knowledge graph construction module is connected to the data interaction module and is used to construct and maintain a metallurgical equipment fault knowledge graph, define knowledge ontology, integrate multi-source knowledge, realize knowledge storage and management, and provide knowledge support for fault reasoning. The data interaction module is connected to the knowledge graph construction module, the GraphRAG inference module, and the model update module, respectively, and is used to collect, transmit, and preprocess multi-source data to realize the association mapping between data and knowledge graph; The GraphRAG reasoning module is connected to the knowledge graph construction module, the data interaction module, and the fault risk classification module. It is used to combine the structured knowledge of the knowledge graph with the retrieval enhancement capabilities of GraphRAG to realize fault reasoning and cause analysis, and output an interpretable fault reasoning report. The fault risk classification module is connected to the GraphRAG inference module and the maintenance list generation and optimization module. It is used to, based on the fault mode and component association information output by the GraphRAG inference module, combine the dynamic RPN value calculated by the Cox proportional hazards regression model or the equivalent dynamic correction model, and the operational behavior index R1, product quality index Q, system reliability index R2 and comprehensive risk level L output by the RQR dynamic coupling model. The comprehensive risk level L and the dynamic RPN value are jointly classified into fault risk levels through piecewise function mapping, and the faults are prioritized. The maintenance list generation and optimization module is connected to the fault risk classification module, the data interaction module, and the model update module. It is used to generate a basic maintenance list based on the fault reasoning results, the dynamic RPN value, the evaluation results of the RQR dynamic coupling model, and the decision rules that integrate reliability-centered maintenance RCM and risk-based maintenance RBM. It is then optimized by combining downtime window, maintenance resource availability, and product quality constraints to output a standardized graded maintenance list. The model update module is connected to the knowledge graph construction module, the GraphRAG inference module, and the maintenance list generation and optimization module, respectively, and is used to realize the dynamic updating and optimization of the knowledge graph, the inference model, and the maintenance optimization model.

[0008] Preferably, the knowledge graph construction module includes a knowledge ontology definition unit, a multi-source knowledge fusion unit, and a knowledge graph storage and management unit; the knowledge ontology definition unit defines core entities and their relationships based on a four-level hierarchical structure of metallurgical equipment system-subsystem-equipment-component; the multi-source knowledge fusion unit integrates structured data, semi-structured data, and unstructured data to construct a fault mechanism library, which covers no less than 30 typical failure modes and supports field expansion; the knowledge graph storage and management unit uses a graph database and / or a relational database for storage, supporting incremental updates and knowledge source traceability.

[0009] Preferably, the GraphRAG inference module includes a structured query unit, a Graph retrieval unit, a RAG enhanced inference unit, and an inference result verification unit. The structured query unit prioritizes structured precise queries for statistical, field-based, and association-based questions and directly returns verifiable results. The Graph retrieval unit supports GraphRAG global retrieval and GraphRAG local retrieval. The global retrieval generates multi-level summaries of the knowledge graph community during the indexing stage and uses a Map-Reduce strategy for parallel retrieval and merging during the query stage. The local retrieval expands a k-hop subgraph using entities in the user's question as seeds and retrieves associated text blocks, where k is 1 to 3, preferably 2. When the time taken for a single graph retrieval request exceeds a threshold T1 or the number of consecutive failures reaches N, it is automatically downgraded to vector retrieval question answering, where T1 is 30 to 120 seconds, preferably 60 seconds, and N is 1 to 3 times, preferably 2 times. If vector retrieval still fails, a preset maintenance prompt is returned.

[0010] Preferably, the fault risk classification module includes a dynamic RPN calculation unit, an RQR status input unit, a risk level classification unit, and a risk priority ranking unit; the dynamic RPN calculation unit uses a Cox proportional hazards regression model to represent instantaneous fault risk as... , dynamic When the sample size is insufficient to calibrate the Cox parameter, an equivalent dynamic correction model in linear or log-linear form is adopted; the quantization mapping of the RQR dynamic coupling model is as follows: R1(t)=Σiw i ·x i ′(t), Q(t)=γ1·M(t)+γ2·P(t)+γ3·C(t), R2(t)=α·R1(t)+β·Q(t)+ε·R1(t)·Q(t)+R 20 .

[0011] Preferably, the maintenance list generation and optimization module includes a basic maintenance list generation unit, a maintenance list hierarchical optimization unit, and a maintenance list standardization output unit; the maintenance list hierarchical optimization unit adopts a genetic algorithm or a multi-objective optimization algorithm, with the goal of minimizing the total system risk and the maintenance cost, and optimizes the execution order of maintenance tasks in combination with maintenance resource availability, downtime window and product quality constraints.

[0012] Preferably, the model update module includes a knowledge graph update unit, an inference model optimization unit, and a maintenance optimization model update unit. It uses an incremental update mechanism to update the knowledge graph, optimizes the graph retrieval algorithm, GraphRAG index, and large language model prompt words based on inference verification feedback, and optimizes the maintenance list generation algorithm and constraints based on the maintenance task execution results.

[0013] Secondly, this invention provides a method for fault reasoning and graded maintenance list generation of metallurgical equipment based on knowledge graph and GraphRAG, applied to the system described in the first aspect, comprising six steps: knowledge graph and mechanism library initialization, data acquisition and preprocessing, hierarchical fault reasoning, fault risk grading, maintenance list generation and optimization, and dynamic model updating; wherein, the fault risk grading step is to jointly divide the risk level by the comprehensive risk level and the dynamic RPN through piecewise function mapping, and to prioritize the faults.

[0014] Thirdly, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the computer program, when executed by the processor, implements the steps of the method described in the second aspect.

[0015] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in the second aspect.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention employs a GraphRAG hierarchical reasoning mechanism that prioritizes structured and precise retrieval, enhances graph and vector retrieval, and provides a fallback through large model generation. This mechanism can handle scenarios involving multiple fault coupling and cascading failures, with an expected reasoning accuracy of over 90% and a diagnostic time that is more than 50% shorter than the traditional FMEA method. This invention uses Cox proportional hazards regression or equivalent models to dynamically correct static RPN according to operating conditions, and incorporates the dynamic coupling assessment results of operating behavior-product quality-system reliability (RQR) into risk classification, so that the risk level changes dynamically and accurately with real-time operating conditions, providing a reliable basis for maintenance decisions. This invention integrates RCM functional importance analysis and RBM risk quantification results, and combines downtime window, spare parts resources and product quality constraints for multi-objective optimization. The output graded maintenance list takes into account safety risks, downtime losses and quality losses, and is expected to reduce operation and maintenance costs by 15% to 25% and reduce unplanned downtime by more than 50%. This invention supports rapid adaptation to new working conditions through dynamic knowledge reuse and an expandable mechanism library. The model continuously iterates based on maintenance feedback and is adaptable to various metallurgical equipment such as hydrogen-based vertical shaft furnaces and rolling mills.

[0017] The above-mentioned technical effects are the results that the present invention has achieved after 12 months of trial operation verification in the hydrogen-based vertical shaft furnace feeding system and rolling mill embodiment, or the effects that are expected to be achieved according to the engineering implementation plan.

[0018] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0019] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 A system structure block diagram provided for embodiments of the present invention; Figure 2 A flowchart of the method provided in an embodiment of the present invention; The steps are as follows: S101, Initialization of knowledge graph and mechanism base; S102, Data acquisition and preprocessing; S103, Hierarchical fault reasoning; S104, Fault risk classification; S105, Maintenance list generation and optimization; S106, Dynamic model update. Figure 3 This is a schematic diagram of the hierarchical structure of a knowledge graph. Figure 4 This is a schematic diagram of the GraphRAG hierarchical inference process; Among them, 41 is the structured query unit; 42 is the graph retrieval unit; 43 is the global retrieval branch; 44 is the local retrieval branch; 45 is the RAG enhanced reasoning unit; 46 is the timeout degradation judgment; 47 is the vector retrieval degradation; and 48 is the reasoning result verification unit. Figure 5 A schematic diagram of the processing flow for the fault risk classification and maintenance list generation and optimization module; Among them, 51 is the dynamic RPN calculation unit; 52 is the RQR status input unit; 53 is the risk level classification unit; 54 is the basic generation unit of the maintenance list generation and optimization module; 55 is the maintenance list hierarchical optimization unit; and 56 is the standardized output unit. Figure 6 This is a diagram showing the comparison of fault reasoning accuracy (including measured and expected data from examples). Figure 7 This is a diagram comparing operation and maintenance costs (including actual and expected data from specific examples). Figure 8 This is a comparative diagram of unplanned downtime (including measured and expected data from the examples). Detailed Implementation

[0020] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0021] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings and examples. The examples are used to explain the present invention and do not limit the scope of protection; any aspects not described in detail can be addressed using methods known in the art.

[0022] This invention can be implemented in an intelligent knowledge platform software system. The platform's backend provides a project library, BOM, FMEA, test outline parsing, document and triplet management, knowledge Q&A, knowledge graph visualization, equipment telemetry, and GraphRAG indexing / query interfaces, and reserves HTTP interfaces for reliability prediction, RQR dynamic assessment, and maintenance strategy optimization services. The platform's implemented or planned capabilities include: GraphRAG workspace indexing and querying, structured question-and-answer direct response, vector retrieval and gold standard answers, asynchronous test outline parsing, and equipment operation curve display. The reliability prediction model, RQR service, and RCM / RBM maintenance strategy optimization service will be completed during the integration phase according to the engineering cooperation agreement. The technical effects of this invention are described based on the expected indicators after integration.

[0023] Example 1 Please see Figure 1 The metallurgical equipment fault reasoning and graded maintenance list generation system based on knowledge graph and GraphRAG in this embodiment includes a knowledge graph construction module 1, a data interaction module 2, a GraphRAG reasoning module 3, a fault risk grading module 4, a maintenance list generation and optimization module 5, and a model update module 6. The six modules work together to complete the perception-evaluation-decision closed loop.

[0024] I. Knowledge Graph Construction Module 1 (1) Definition of knowledge ontology Taking the hydrogen-based vertical shaft furnace feeding system as an example, a knowledge ontology is defined based on a four-level structure of system-subsystem-equipment-component, defining fault modes, fault causes, fault effects, maintenance measures, RPN parameters, component attributes and relationships.

[0025] The RPN parameter is defined as follows: Severity S indicates the degree of impact of the failure mode on personnel safety, regulatory compliance, and system functionality (1-10). Occurrence frequency O indicates the estimated frequency (1-10) of the failure cause or mode occurring during the life cycle. The detection difficulty D indicates the ease (1-10) of detecting the failure mode before or during operation. Static RPN = S × O × D.

[0026] Please see Figure 3 , Figure 3 The left side shows the four-level equipment structure hierarchy of system-subsystem-equipment-component, while the right side shows entities such as failure modes, failure causes, failure effects, maintenance measures, RPN parameters, and component attributes, as well as their existence, initiation, cause, and corresponding relationships.

[0027] In other embodiments, the knowledge graph, based on the four-level structure of system-subsystem-equipment-component, can be expanded to five or six levels according to the complexity of the equipment. Engineering-device mapping nodes and test-stress mapping nodes can be added to support the mounting of stress-failure mapping and test item-stress mapping analysis results in the reliability enhancement test outline. In this embodiment, the four-level structure is the main framework, and the expansion of levels and mapping nodes are optional enhancements, without changing the main process of fault reasoning and risk classification.

[0028] (2) Multi-source knowledge fusion Enter the FMEA table, BOM, test outline analysis sub-table, maintenance cases, and common failure data; Unstructured knowledge is extracted from documents using BERT combined with a dictionary for the metallurgical equipment field. Build a failure mechanism library that covers no fewer than 30 typical failure modes, with an extensible field structure.

[0029] (3) Knowledge graph storage and management The Neo4j graph database is used to store the relationship network, and the MySQL relational database is used to store the mechanism database business tables. The two are linked by entity identifiers. It maintains and records knowledge sources and timestamps through a RESTful API, supporting incremental updates, source tracing, and integration with intelligent knowledge platform engineering systems.

[0030] II. Data Interaction Module 2 Data interaction module 2 collects sensor and SCADA data (temperature, pressure, vibration, process parameters, etc., with a sampling frequency of, for example, once per minute), equipment alarms, and inspection records; Import historical fault, repair, FMEA, and test data; Preprocessing and feature extraction, such as 3σ denoising, interpolation, and MinMax normalization, are performed and mapped to corresponding component entity nodes such as the discharge valve in the knowledge graph.

[0031] The data interaction module 2 can interact with the equipment telemetry subsystem, reliability prediction service and RQR service through a standard HTTP interface, and simultaneously receive the output of the RQR service and reliability prediction service.

[0032] III. GraphRAG Inference Module 3 Please see Figure 4 The GraphRAG inference module 3 includes a structured query unit 41, a graph retrieval unit 42, a global retrieval branch 43, a local retrieval branch 44, a RAG enhanced inference unit 45, a timeout degradation judgment 46 (threshold T1), a vector retrieval degradation 47, and an inference result verification unit 48 (threshold θ), forming a hierarchical inference mechanism that prioritizes structured precise retrieval, enhances graph retrieval / vector retrieval, and provides a fallback for large model generation.

[0033] (1) Structured query unit 41 For statistical, field-based, and project-FMEA-maintenance case related issues such as FMEA statistics, project attributes, maintenance case aggregation, and test outline status, prioritize structured precise queries and directly return verifiable results to ensure verifiability.

[0034] (2) Graph retrieval unit 42 Based on real-time anomaly data, graph retrieval algorithms, shortest path algorithms, or subgraph matching algorithms optimized by graph neural networks (GNN) are used to locate associated fault modes, components, and historical cases in the knowledge graph and identify cascading faults.

[0035] The preferred GNN is a Graph Attention Network (GAT) or a Graph Convolutional Network (GCN), which optimizes the sorting of retrieval paths by using node neighbor attention weights or convolutional aggregation. In this embodiment, GAT is preferred to highlight key related edges when multiple types of neighbors such as components, failure modes, and maintenance measures coexist.

[0036] The differences between the implementation of global search and local search in GraphRAG are as follows: Global retrieval generates multi-level summaries for knowledge graph communities during the indexing phase. During the query phase, a Map-Reduce strategy is used to first retrieve the summaries of each community in parallel and then merge them. This is suitable for problems that require cross-community integration, such as summarizing failure modes across projects and common risks of multiple equipment. Local retrieval uses entities (parts, failure modes, items) in the user's question as seeds to expand a k-hop subgraph in the graph database, with k preferably ranging from 1 to 3. It only retrieves text blocks corresponding to entity descriptions and relational edges associated with the subgraph, making it suitable for precise question answering regarding common failures and solutions for a specific component.

[0037] (3) RAG Enhanced Reasoning Unit 45 The knowledge fragments obtained from the fusion graph retrieval and unstructured knowledge such as expert experience are input into the industrial big data language model in a Prompt manner. Combined with the real-time operating status, an interpretable fault reasoning report containing fault modes, root causes, scope of impact, and reasoning basis is output.

[0038] (4) Timeout Degradation and Result Verification The HTTP timeout for graph retrieval is set to 60 seconds for a single query, with a preferred range of 30 to 120 seconds. The index creation timeout is set to 3600 seconds. When the time taken for a single query is greater than or equal to T1 or the number of consecutive failures N is greater than or equal to 2, the query is automatically downgraded to vector retrieval, with N preferably being 1 to 3 times. If the vector retrieval still fails, a preset maintenance prompt is returned to avoid service blockage.

[0039] The reasoning result verification unit 48 compares the reasoning result with actual fault records, expert judgments, or offline evaluation sets. The expected reasoning accuracy rate can reach more than 90% (θ=90%). When it is lower than the threshold, the model update module 6 is triggered.

[0040] The offline evaluation set consists of cases reviewed by experts from historical fault records and maintenance work orders. The cases are derived from the operation and maintenance data of metallurgical equipment such as hydrogen-based vertical shaft furnace feeding systems and rolling mills. The comparison baseline includes pure vector retrieval (RAG) and static knowledge graph retrieval. The evaluation index is mainly based on the accuracy of fault mode / root cause identification, supplemented by random checks on the completeness of interpretable evidence.

[0041] IV. Fault Risk Classification Module 4 Please see Figure 5 The fault risk classification module 4 includes a dynamic RPN calculation unit 51, an RQR status input unit 52, and a risk level classification unit 53, which are connected to the basic generation unit 54, the maintenance list classification optimization unit 55, and the standardized output unit 56 of the maintenance list generation and optimization module 5.

[0042] (1) Dynamic RPN calculation Combining the initial values ​​of S, O, and D with the operating status and failure trends, Cox proportional hazards regression or an equivalent model is used for dynamic correction. The Cox proportional hazards regression model represents the instantaneous failure risk as:

[0043] in, Let β be the baseline risk function, where ΔS(t), ΔO(t), and ΔD(t) are the deviations from the design baseline based on real-time operating conditions, and β is the risk factor. S β O β D The regression coefficients are: dynamic severity S′(t), dynamic occurrence frequency O′(t), and dynamic detection difficulty D′(t), which are obtained by mapping the baselines S0, O0, and D0 to the aforementioned deviations. .

[0044] When the sample size is insufficient to calibrate the Cox parameter, an equivalent dynamic correction model is used. This model performs the same dynamic correction based on operating conditions as the Cox proportional hazards regression model, and adopts a linear or log-linear form. , Where, δ S δ O δ D α represents the normalized operational deviation. S α O α D This is an empirical coefficient.

[0045] Initial values ​​S0, O0, and D0 are taken from the static evaluation results of FMEA and used as the baseline input for the Cox or equivalent dynamic correction model; the baseline risk function h0(t) is obtained by Kaplan-Meier estimation or parametric distribution fitting of historical failure interval data, and the regression coefficient β... S β O β D Calibrated using partial likelihood estimation from historical operating condition-fault samples; real-time operating condition deviations ΔS(t), ΔO(t), ΔD(t) or normalized deviation δ S δ O δ D The data is calculated from the telemetry and alarm characteristics input by the data interaction module, thus realizing the connection between "FMEA static initial value → dynamic correction of operating conditions → dynamic RPN output".

[0046] (2) RQR status input The RQR status input unit 52 receives the operational behavior index R1, product quality index Q, system reliability index R2, and comprehensive risk level L output by the RQR chain, and incorporates them into the hierarchical decision-making process.

[0047] The quantification formula for the dynamic coupling model of operational behavior-product quality-system reliability (RQR) is: R1(t)=Σiw i ·x i ′(t) Where, x i ′(t) is the normalized value of the i-th operational telemetry feature, w i As weight; Q(t)=γ1·M(t)+γ2·P(t)+γ3·C(t) Where M(t), P(t), and C(t) are the normalized deviations of metallization rate, output, and key quality indicators, respectively, and γ1, γ2, and γ3 are the weights. R2(t)=α·R1(t)+β·Q(t)+ε·R1(t)·Q(t)+R20 Wherein, α, β, ε, R 20 These are calibration coefficients.

[0048] The overall risk level L is determined by the dynamic coupling model based on R²(t). The fault risk level is jointly classified by the dynamic RPN value and the overall risk level L through a piecewise function mapping. For example, when the dynamic RPN ≥ 100 and the overall risk level L is higher than a set threshold, it is classified as plant-level risk; when 50 ≤ RPN < 100, it is classified as equipment-level risk; and when RPN < 50, it is classified as general-level risk. If L is an inverse indicator in the partner's RQR model (the lower the value, the higher the risk), then the original text is valid, and this item does not need to be modified.

[0049] The RQR service outputs R1, Q, R2, and L to this system via an HTTP interface for use by the fault risk classification module 4 and the maintenance list generation and optimization module 5. The output frequency of the RQR service is aligned with the calculation cycle of the system's dynamic RPN, preferably once per minute, consistent with the SCADA / telemetry acquisition cycle; alternatively, it can be configured as an event-triggered mode, that is, the risk level recalculation of this system is triggered synchronously when the RQR service output is updated; one of the above two synchronization methods is selected and the configuration is fixed during system deployment.

[0050] (3) Risk Priority Ranking Based on the comprehensive impact range, the speed of failure development, and the impact on product quality such as direct reduced iron (DRI), failures of the same level are prioritized.

[0051] V. Repair List Generation and Optimization Module 5 (1) Generation of basic maintenance list Based on the relationship between fault modes and maintenance measures in the knowledge graph, the basic generation unit 54 generates a basic maintenance list that includes task types such as inspection, adjustment, replacement, and test verification, by combining the fault reasoning results and the RCM functional importance and RBM risk quantification results.

[0052] (2) Optimization of maintenance list classification The maintenance list hierarchical optimization unit 55 takes minimizing the total system risk and the maintenance cost as the objective function, and combines downtime window, spare parts and product quality constraints to determine the execution order of maintenance tasks using a genetic algorithm or a multi-objective optimization algorithm.

[0053] The parameterized form of multi-objective optimization is, for example: min F = ω1·ΣRPN residual +ω2·C total , where ΣRPN residual C represents the total residual risk of the system after maintenance. totalThe total maintenance cost is represented by ω1 and ω2, which are weighting coefficients, and ω1 + ω2 = 1. The weights are determined based on the enterprise's safety management preferences; in this example, ω1 = 0.6 and ω2 = 0.4 are preferred (emphasizing residual risk control while also considering cost). Constraints include maintenance resource availability constraints, downtime window duration constraints, and constraints on preventing the deterioration of key quality indicators. The genetic algorithm encoding uses chromosomes arranged in the task execution order. After selection, crossover, and mutation iterations until convergence or reaching a set number of generations, the optimal or near-optimal execution order is output.

[0054] Within the same downtime window, maintenance tasks with the same or similar risk levels and adjacent maintenance locations are combined into a single batch for execution, in order to make full use of the downtime window and reduce losses from repeated start-ups and shutdowns.

[0055] (3) Standardized output The standardized output unit 56 outputs a tiered maintenance list including task ID, fault mode, risk level, maintenance task, priority, execution order, required resources, estimated duration, and completion standards. It also supports integration with enterprise operation and maintenance management systems to generate a final maintenance plan through maintenance strategy optimization services.

[0056] VI. Model Update Module 6 The model update module 6 adopts an incremental update mechanism, updating the knowledge graph based on newly added fault cases, maintenance records, expert experience, and test outline analysis results, and undergoing manual review and confirmation; it optimizes the graph retrieval algorithm, GraphRAG index, and large language model prompt words based on inference verification feedback and newly added data, and adapts to new equipment types and new working conditions through transfer learning; it optimizes the maintenance list generation algorithm and constraints based on maintenance task execution results (such as maintenance time deviation).

[0057] Please see Figure 2 The method flow of this embodiment includes: S101, knowledge graph and mechanism library initialization; S102, data acquisition and preprocessing; S103, hierarchical fault reasoning; S104, fault risk classification; S105, maintenance list generation and optimization; S106, model dynamic update.

[0058] Example 2 This embodiment uses the fault reasoning and maintenance list generation of a hydrogen-based vertical shaft furnace charging system as an example for explanation. The hydrogen-based vertical shaft furnace charging system is a core piece of equipment in steel production, operating under harsh conditions and prone to faults such as bridging, vibration, and leakage. Traditional maintenance methods rely on manual FMEA analysis, resulting in low fault diagnosis efficiency and a lack of specificity in the maintenance list, often leading to unplanned downtime. The implementation process of this embodiment is as follows: (1) Knowledge graph initialization Based on the four-level structure of system-subsystem-equipment-component definition of the hydrogen-based vertical shaft furnace feeding system, a knowledge ontology is constructed by integrating nearly 3 years of fault records, FMEA reports, and expert experience. It includes 48 typical fault modes, more than 120 fault causes, and more than 80 maintenance measures.

[0059] (2) Data acquisition and preprocessing The system collects data in real time, such as the vibration and temperature of the discharge valve, the load of the reducer, and the pressure of the cooling pump. After preprocessing, the data is associated with the corresponding components in the knowledge graph. Historical fault data is also imported for model training.

[0060] (3) GraphRAG Fault Reasoning When the vibration amplitude of the discharge valve exceeds the standard by 25% and the temperature rises by 20°C, the Graph retrieval unit finds the associated fault mode as excessive vibration and bearing wear in the knowledge graph. RAG enhanced inference, combined with expert experience, determines the root cause to be insufficient bearing lubrication leading to wear. The fault impact is the potential for the discharge valve to seize, causing production interruption. In this embodiment, the inference accuracy increased from 78% to 93% over 12 months.

[0061] (4) Fault risk classification The dynamic RPN is calculated to be 132, classified as a plant-level risk, and set to the highest priority.

[0062] (5) Maintenance list generation and optimization Generate a basic maintenance list, optimize the execution order by combining downtime window (1 hour) and maintenance cost (≤5000 yuan), and output a standardized hierarchical maintenance list that clarifies maintenance tasks, priorities, required resources and completion standards.

[0063] (6) Model update The fault case and maintenance records were added to the knowledge graph to optimize the reasoning model and maintenance optimization model.

[0064] In this embodiment, the diagnostic time is reduced from approximately 2 hours to approximately 30 minutes. Comparative data by period is shown in Table 1. Traditional FMEA uses statistical values ​​from the control group during the same period, while this invention uses statistical interpolation for platform trial operation and acceptance targets. Table 1

[0065] 12-month summary: Accuracy rate approximately 93%; unplanned downtime reduced by approximately 60%; maintenance costs reduced by approximately 20%; production continuity improved by approximately 25%.

[0066] Example 3 This embodiment uses rolling mill fault reasoning and maintenance list generation as an example for explanation. Rolling mills are key equipment in metallurgical production, subjected to high loads and vibrations, and are prone to faults such as gear wear, bearing failure, and excessive vibration. Traditional maintenance methods rely on manual experience, resulting in chaotic prioritization of maintenance lists and high maintenance costs. The implementation process of this embodiment is as follows: (1) Knowledge graph initialization Based on the knowledge ontology of the system-subsystem-equipment-component structure of the rolling mill, and integrating rolling mill fault records, maintenance logs, and expert experience, a fault knowledge graph is constructed, which includes 52 typical fault modes, more than 130 fault causes, and more than 90 maintenance measures.

[0067] (2) Data acquisition and preprocessing Data such as mill bearing temperature, gear vibration, and load are collected in real time by sensors, pre-processed, and then linked to a knowledge graph; historical fault data is imported to complete model training.

[0068] (3) GraphRAG Fault Reasoning When the vibration amplitude of the mill gears exceeds the standard by 30% and the load fluctuation is abnormal, the Graph retrieval unit finds the associated fault mode as gear wear and uneven load. RAG enhanced inference determines that the root cause is insufficient gear lubrication and uneven load distribution. The fault effect is further gear wear, leading to mill shutdown. In this embodiment, the inference accuracy increased from 76% to 91% over 12 months.

[0069] (4) Fault risk classification The dynamic RPN is calculated to be 118, classified as a plant-level risk, and set to the highest priority.

[0070] (5) Maintenance list generation and optimization Generate a basic maintenance list, optimize the execution order by combining downtime window (2 hours) and maintenance cost (≤8000 yuan), and output a tiered maintenance list.

[0071] (6) Model update Update the knowledge graph and model to improve the accuracy of subsequent reasoning and maintenance optimization. Comparison data by period is shown in Table 2. Table 2

[0072] 12-month summary: Accuracy rate approximately 91%; maintenance costs reduced by approximately 22%; downtime reduced by approximately 55%.

[0073] As can be seen from the above embodiments 2 and 3, the system and method of the present invention can achieve accurate fault reasoning and generation of graded maintenance lists in metallurgical equipment such as hydrogen-based vertical furnace feeding systems and rolling mills, effectively solving the defects of traditional operation and maintenance methods, improving operation and maintenance efficiency and equipment reliability, reducing operation and maintenance costs, and having broad engineering application value.

[0074] Please see Figures 6-8 , Figures 6-8 The horizontal axis represents the operating cycle (months), and the vertical axis represents the fault reasoning accuracy (%), relative maintenance cost index, and unplanned downtime (h), respectively. The three curves correspond to the traditional FMEA, the hydrogen-based vertical furnace feeding system example (the system of this invention), and the rolling mill example (the system of this invention), respectively, reflecting the trends of improved accuracy, reduced costs, and shorter downtime under different operating cycles: (1) Accuracy improvement curve of fault reasoning ( Figure 6 The accuracy of the method of this invention continues to improve with the running cycle, and stabilizes at over 90% after 9 months, which is significantly higher than the traditional FMEA method (around 70%). (2) Operation and maintenance cost reduction curve ( Figure 7 The operation and maintenance cost of the method of this invention is continuously optimized with system iteration, and is reduced by more than 20% compared with the traditional method at the end of a 12-month cycle. (3) Curve for reduction in unplanned downtime ( Figure 8 The unplanned downtime of the method of this invention decreases continuously with the operating cycle, and is reduced by more than 50% compared with the traditional method at the end of the 12-month cycle, effectively improving production continuity.

[0075] It should be noted that the present invention can be implemented by hardware, software, or a combination of both. Computer devices and computer-readable storage media are both used to implement the above-described systems and methods.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A system for fault reasoning and hierarchical maintenance list generation of metallurgical equipment based on knowledge graphs and GraphRAG, characterized in that, It includes a knowledge graph construction module, a GraphRAG reasoning module, a fault risk classification module, a maintenance list generation and optimization module, a data interaction module, and a model update module; the modules work together to form a complete operation and maintenance closed loop of data input - knowledge modeling - fault reasoning - risk classification - maintenance generation - model update; The knowledge graph construction module is connected to the data interaction module and is used to construct and maintain a metallurgical equipment fault knowledge graph, define knowledge ontology, integrate multi-source knowledge, realize knowledge storage and management, and provide knowledge support for fault reasoning. The data interaction module is connected to the knowledge graph construction module, the GraphRAG inference module, and the model update module, respectively, and is used to collect, transmit, and preprocess multi-source data, including real-time running data and historical data, to realize the association mapping between data and the knowledge graph. The GraphRAG reasoning module is connected to the knowledge graph construction module, the data interaction module, and the fault risk classification module. It is used to combine the structured knowledge of the knowledge graph with the retrieval enhancement capabilities of GraphRAG to realize fault reasoning and cause analysis, and output a fault reasoning report. The fault risk classification module, connected to the GraphRAG inference module and the maintenance list generation and optimization module, is used to, based on the fault mode and component association information output by the GraphRAG inference module, combine the dynamic RPN value with the operational behavior index R1, product quality index Q, system reliability index R2, and comprehensive risk level L output by the received RQR dynamic coupling model. The comprehensive risk level L and the dynamic RPN value are then used to jointly classify the fault risk level through a piecewise function mapping, and the faults are prioritized. The dynamic RPN value is obtained by the fault risk classification module using a Cox proportional hazards regression model or an equivalent dynamic correction model to dynamically adjust the static RPN according to the operating conditions. The maintenance list generation and optimization module is connected to the fault risk classification module, the data interaction module, and the model update module. It is used to generate a basic maintenance list based on the fault reasoning results, the dynamic RPN value, the evaluation results of the RQR dynamic coupling model, and the decision rules that integrate reliability-centered maintenance (RCM) and risk-based maintenance (RBM). It is then optimized in combination with engineering constraints to output a standardized graded maintenance list. The model update module is connected to the knowledge graph construction module, the GraphRAG inference module, and the maintenance list generation and optimization module, respectively, and is used to realize the dynamic updating and optimization of the knowledge graph, the inference model, and the maintenance optimization model.

2. The metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on knowledge graph and GraphRAG as described in claim 1, characterized in that, The knowledge graph construction module includes a knowledge ontology definition unit, a multi-source knowledge fusion unit, and a knowledge graph storage and management unit. The knowledge ontology definition unit is used to define fault modes, fault causes, fault effects, maintenance measures, RPN parameters and component attributes as core entities, as well as the relationships between entities, based on the four-level hierarchical structure of metallurgical equipment system-subsystem-equipment-component. The RPN parameters include: Severity S indicates the degree of impact of the failure mode on personnel safety, regulatory compliance, and system functionality, and ranges from 1 to 10. Occurrence frequency O represents the estimated frequency of failure causes or modes occurring during the equipment's life cycle, with a value ranging from 1 to 10. The detection difficulty D indicates the ease or difficulty of detecting the failure mode before or during operation, and its value ranges from 1 to 10. Static Risk Priority Number RPN = S × O × D; The multi-source knowledge fusion unit is used to collect and fuse structured data, semi-structured data and unstructured data, extract entities and relationships through natural language processing technology, and write them into the fault mechanism library after data cleaning and normalization. The fault mechanism library covers no less than 30 typical failure modes and supports field expansion. The structured data includes equipment fault records, FMEA analysis reports, product BOM, reliability enhancement test outline analysis results, maintenance logs, and component parameter tables. The semi-structured data includes operation and maintenance manuals and technical specifications; the unstructured data includes fault diagnosis reports, expert experience texts, and on-site inspection records. The knowledge graph storage and management unit uses a graph database and / or a relational database to store knowledge graph and mechanism library data, provides CRUD interfaces, and supports incremental updates and knowledge source tracing.

3. The metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on knowledge graph and GraphRAG as described in claim 1, characterized in that, The GraphRAG inference module includes a structured query unit, a Graph retrieval unit, a RAG enhanced inference unit, and an inference result verification unit. The structured query unit is used to prioritize and perform structured, precise queries on statistical, field-related, and project-FMEA-maintenance case related issues, and directly return verifiable results. The Graph retrieval unit employs graph retrieval algorithms, shortest path algorithms, or subgraph matching algorithms optimized by graph neural networks. Based on real-time equipment operation data, it performs structured retrieval in the knowledge graph to locate associated fault modes, components, and historical cases, and identify potential cascading faults. The RAG enhanced reasoning unit integrates the knowledge fragments retrieved from the Graph with unstructured knowledge, inputs them into the industrial big language model in a Prompt manner, and outputs an interpretable fault reasoning report containing fault modes, root causes, scope of impact, and reasoning basis in combination with the real-time operating status. The reasoning result verification unit compares the reasoning result with actual fault records, expert judgments, or offline evaluation sets, and feeds back to the model update module when the accuracy is lower than a set threshold.

4. The metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on knowledge graph and GraphRAG as described in claim 3, characterized in that, The Graph search unit supports both GraphRAG global search and GraphRAG local search. The GraphRAG global retrieval generates multi-level summaries of communities in the knowledge graph during the indexing stage, and adopts a Map-Reduce strategy to first retrieve the summaries of each community in parallel and then merge the results during the query stage. It is suitable for mechanism induction problems that are cross-project and cross-equipment. The GraphRAG local search uses entities in the user's question as seeds to expand k-hop subgraphs in the graph database, where k ranges from 1 to 3. It only retrieves text blocks corresponding to entity descriptions and relation edges associated with the subgraph, making it suitable for precise question and answering of specific components, failure modes, or maintenance cases. When a single graph retrieval request takes longer than a threshold T1 or the number of consecutive failures reaches N, it is automatically downgraded to vector retrieval question answering; T1 is 30 to 120 seconds, and N is 1 to 3 times; if the vector retrieval still fails, a preset maintenance prompt is returned to avoid service blockage.

5. The metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on knowledge graph and GraphRAG as described in claim 1, characterized in that, The fault risk classification module includes a dynamic RPN calculation unit, an RQR status input unit, a risk level classification unit, and a risk priority sorting unit. The dynamic RPN calculation unit adopts the Cox proportional hazards regression model and uses an equivalent dynamic correction model when the sample is insufficient to calibrate the Cox parameter. The Cox proportional hazards regression model represents transient failure risk as... ， in, Let β be the baseline risk function, where ΔS(t), ΔO(t), and ΔD(t) are the deviations from the design baseline based on real-time operating conditions, and β is the risk factor. S β O β D The regression coefficients are: dynamic severity S′(t), dynamic occurrence frequency O′(t), and dynamic detection difficulty D′(t), which are obtained by mapping the baselines S0, O0, and D0 to the aforementioned deviations. ; The equivalent dynamic correction model performs the same dynamic correction based on operating conditions as the Cox proportional hazards regression model, using... The linear or logarithmic linear form, where δ S δ O δ D α represents the normalized operational deviation. S α O α D This is an empirical coefficient; The RQR status input unit is used to receive the operation behavior index R1, the product quality index Q, the system reliability index R2, and the comprehensive risk level L output by the dynamic coupling model of operation behavior-product quality-system reliability, and transmit them to the risk level division unit as input for fault risk level division; wherein, the comprehensive risk level L is determined by the dynamic coupling model based on the system reliability index R2. The quantization mapping of the dynamic coupling model is as follows: R1(t)=Σiw i ·x i ′(t), x i ′(t) is the normalized value of the i-th operational telemetry feature, w i As weight; Q(t) = γ1·M(t) + γ2·P(t) + γ3·C(t), where M(t), P(t), and C(t) are the normalized deviations of metallization rate, output, and key quality indicators, respectively, and γ1, γ2, and γ3 are the weights. R2(t)=α·R1(t)+β·Q(t)+ε·R1(t)·Q(t)+R 20 , α, β, ε, R 20 are calibration coefficients; The risk level classification unit classifies the fault into plant-level, equipment-level, and general-level risk levels based on the dynamic RPN value, the comprehensive risk level L, and the enterprise safety management specifications through a piecewise function mapping. The risk priority ranking unit prioritizes faults of the same level by considering the scope of their impact, their development speed, and their impact on product quality.

6. The metallurgical equipment fault reasoning and hierarchical maintenance list generation system based on knowledge graph and GraphRAG as described in claim 1, characterized in that, The maintenance list generation and optimization module includes a basic maintenance list generation unit, a maintenance list hierarchical optimization unit, and a maintenance list standardization output unit. The basic maintenance list generation unit generates a basic maintenance list that includes inspection, adjustment, replacement, and test verification task types based on the relationship between failure modes and maintenance measures in the knowledge graph, combined with failure reasoning results and the importance of RCM functions and the risk quantification results of RBM. The hierarchical optimization unit of the maintenance list adopts a genetic algorithm or a multi-objective optimization algorithm to optimize the order of maintenance tasks by minimizing the total system risk and the maintenance cost, and combines maintenance resource availability, downtime window and product quality constraints. The standardized maintenance list output unit outputs a tiered maintenance list that includes task ID, fault mode, risk level, maintenance task, priority, execution order, required resources, estimated duration, and completion criteria, and supports integration with enterprise operation and maintenance management systems.

7. A method for fault reasoning and hierarchical maintenance list generation of metallurgical equipment based on knowledge graph and GraphRAG, characterized in that, The system applied to any one of claims 1 to 6 includes the following steps: Step 1, Knowledge Graph and Mechanism Base Initialization: Define the knowledge ontology of metallurgical equipment, integrate FMEA, BOM, test outline, maintenance cases and common failure data, and construct a fault mechanism base and knowledge graph; Step 2, Data Acquisition and Preprocessing: Collect real-time and historical data of metallurgical equipment, clean, normalize, extract features, and map them to entities in the knowledge graph; Step 3, Hierarchical Fault Reasoning: Prioritize structured queries; if not satisfied, perform Graph retrieval and RAG-enhanced reasoning, output a fault reasoning report and verify it; Step 4, Fault Risk Classification: Combining the evaluation results of the dynamic RPN and RQR dynamic coupling model, the risk level is jointly classified by the comprehensive risk level and the dynamic RPN through piecewise function mapping, and the faults are prioritized. Step 5, Maintenance List Generation and Optimization: Based on the dynamic RPN and RQR evaluation results, integrate RCM and RBM rules to generate and optimize the hierarchical maintenance list; Step 6, Dynamic Model Update: Update the knowledge graph and optimize the reasoning model and maintenance optimization model.

8. A computer device, characterized in that, It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in claim 7.

9. A computer-readable storage medium storing a computer program that, when executed by a processor, causes the processor to perform the steps of the method as claimed in claim 7.

Images