A Method and System for Equipment Fault Diagnosis Based on Knowledge Graph and Large Language Model

Abstract

This invention relates to the field of equipment monitoring and fault diagnosis technology, and discloses a method and system for equipment fault diagnosis based on knowledge graphs and large language models. The method includes: acquiring multiple types of data sources for the target equipment, preprocessing and extracting features to obtain a structured dataset; extracting knowledge triples from the dataset based on a natural language processing model and constructing an equipment health status ontology; constructing evaluation functions for each fault phenomenon to quantify the adaptability of monitoring methods, forming fault phenomenon-monitoring method mapping triples; merging the knowledge and mapping triples and storing them in a graph database to form a knowledge graph and optimizing it; using a pre-trained semantic embedding model to encode triples and construct a retrieval index library, receiving user query codes and retrieving and constructing a semantic subgraph; combining the semantic subgraph and the query into prompt words and inputting them into a large language model to generate recommendation results, updating the knowledge graph or index library based on user feedback. This invention improves the accuracy and adaptability of fault diagnosis and reduces operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to the field of equipment monitoring and fault diagnosis technology, specifically to a method and system for equipment fault diagnosis based on knowledge graphs and large language models. Background Technology

[0002] With the rapid development of industrial internet, Internet of Things and big data technologies, equipment, systems and processes in various industries are becoming increasingly complex, and fault diagnosis and real-time monitoring have become key links to ensure production safety and improve operational efficiency.

[0003] Traditional methods heavily rely on human experience and static rule bases, making it difficult to effectively handle multi-source, heterogeneous, and massive real-time data. This results in slow response and poor adaptability in complex industrial scenarios. Secondly, fault information is often scattered across different systems and documents, lacking a unified semantic association and integration mechanism, forming "data silos" that limit the comprehensiveness and consistency of diagnostic knowledge. Furthermore, most existing systems lack flexibility in recommending monitoring methods, often employing fixed thresholds or single methods, failing to dynamically select the optimal monitoring strategy based on equipment type, operating status, and fault characteristics, thus reducing fault identification accuracy and early warning capabilities. Finally, traditional diagnostic models have poor interpretability, making it difficult for maintenance personnel to understand the decision-making basis. This not only affects the efficiency of fault handling but also hinders continuous system optimization and the establishment of trust. These shortcomings severely restrict the development and application effectiveness of intelligent operation and maintenance. Summary of the Invention

[0004] In view of this, in order to address the technical shortcomings of traditional methods such as knowledge isolation, poor diagnostic adaptability, single monitoring strategy, and weak decision interpretability, this invention provides a device fault diagnosis method and system based on knowledge graphs and large language models. This is not only an effective way to solve the bottlenecks of traditional fault diagnosis and monitoring, but also an important means to achieve accurate, efficient, and intelligent maintenance management in the era of industrial intelligence and digital twins.

[0005] In a first aspect, the present invention provides a device fault diagnosis method based on knowledge graphs and large language models, comprising:

[0006] Acquire multiple types of data sources from the target equipment, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals, and equipment design documents, and perform preprocessing and feature extraction to obtain a structured dataset;

[0007] Based on a natural language processing model, knowledge triples in the form of head entity-relation-tail entity are extracted from the structured dataset, and a device health status ontology is constructed based on the knowledge triples.

[0008] For each fault phenomenon of an entity in the knowledge triple, the corresponding equipment type, operating conditions, historical fault records and related parameter characteristics are comprehensively analyzed. An evaluation function is constructed to quantify the adaptability of different monitoring methods to it, forming a fault phenomenon-monitoring method mapping relationship, and generating a fault monitoring mapping triple in the form of fault phenomenon-adapted monitoring method-adaptability score.

[0009] Knowledge triples and fault monitoring mapping triples are merged and stored in a graph database to form a device fault knowledge graph, and the graph structure is optimized by entity alignment and disambiguation algorithms.

[0010] A pre-trained semantic embedding model is used to encode triples in the knowledge graph to generate normalized vectors and build a retrieval index library; user natural language queries are received and encoded into semantic vectors, and relevant knowledge fragments are retrieved through similarity calculation to build a semantic subgraph that matches the query;

[0011] The semantic subgraph is transformed into a structured context and combined with the user query to form prompt words input into a large language model. This generates recommended results that include fault causes, impact analysis, warning levels, and maintenance suggestions. At the same time, the knowledge graph or semantic index is updated based on user feedback.

[0012] The equipment fault diagnosis method based on knowledge graphs and large language models provided in this invention solves the core problems of traditional fault diagnosis, such as data silos, scattered diagnostic results, and poor interpretability, by integrating multi-source data processing, knowledge extraction, fit calculation, graph construction, semantic retrieval, and LLM recommendation throughout the entire process. Preprocessing of multiple data sources ensures data quality; knowledge triples and ontology construction realize the structuring of fault knowledge; fit calculation generates accurate fault phenomenon-monitoring method mappings; graph database storage and optimization improve knowledge query efficiency; semantic embedding and subgraph construction ensure retrieval accuracy; LLM-generated recommendation results cover all dimensions of fault information; and a user feedback update mechanism enables dynamic iteration of the knowledge graph. The overall method balances diagnostic accuracy, interpretability, and adaptability, significantly improving equipment fault diagnosis efficiency, reducing operation and maintenance costs, and providing a complete technical solution for intelligent maintenance of industrial equipment.

[0013] In one optional implementation, the expression for the evaluation function is:

[0014]

[0015] in: This indicates the degree of suitability of monitoring method m on the target device, and is obtained by quantitative scoring based on the matching matrix between device type and monitoring method;

[0016] This represents the diagnostic effectiveness of monitoring method m in historical fault cases of the target equipment, calculated comprehensively based on the detection rate, false alarm rate, and detection delay rate of historical fault cases;

[0017] This represents the sensitivity of monitoring method m to key features of the fault phenomenon, obtained from the sensitivity mapping table of monitoring methods to fault features;

[0018] This indicates the reliability of monitoring method m under the corresponding fault phenomenon, and is obtained based on whether the monitoring method meets the safety integrity level requirements.

[0019] The weight parameters for each indicator are defined, and the sum of the weights of each indicator is 1. The weight parameters are determined based on the fault risk level of the target equipment and industry operation and maintenance standards.

[0020] This invention provides a scientific and reproducible computational framework for quantifying the fit between fault phenomena and monitoring methods by clearly defining the expression of the evaluation function and the definitions of each index. This addresses the problem of traditional fit determination relying on human experience and being highly subjective. It transforms fit calculation from a "qualitative judgment" to a "quantitative analysis," significantly improving the objectivity and accuracy of monitoring method selection. This avoids diagnostic misses and false alarms caused by inappropriate method fit, and provides high-quality prior knowledge support for subsequent knowledge graph construction and LLM recommendation, further ensuring the overall effectiveness of fault diagnosis.

[0021] In one optional implementation, the pre-trained semantic embedding model is a SentenceTransformer model, which processes triples in the knowledge graph. The expression for encoding and generating the embedding vector is:

[0022] ,

[0023] in, Concatenate text using triples. It is a d-dimensional embedding vector;

[0024] The embedded vectors are normalized, and a retrieval index is constructed based on the normalized vectors.

[0025] This invention employs the SentenceTransformer model as a pre-trained semantic embedding model. Through standardized triple encoding and normalization, it eliminates the interference of vector magnitude differences on retrieval results. The retrieval index built upon this model can achieve efficient and accurate semantic matching. When receiving user natural language queries, it can quickly retrieve relevant knowledge fragments, avoiding retrieval biases caused by semantic ambiguity. This provides an accurate data source for subsequent semantic subgraph construction, ensuring the contextual integrity of LLM input and ultimately improving the accuracy and reliability of fault diagnosis recommendation results.

[0026] In one optional implementation, the retrieved entities are expanded by first-order or second-order adjacency according to the causal logic chain of fault phenomenon-fault cause-related component-monitoring method, to ensure that the expanded semantic subgraph covers fault cause, adapted monitoring method and related component information.

[0027] This invention addresses the problems of "disordered expansion range" and "missing key information" in traditional semantic subgraph construction by clarifying the causal logical chain expansion rules of fault phenomenon - fault cause - related components - monitoring method. First-order or second-order adjacency expansion ensures that the subgraph is neither overly redundant and inefficient, nor does it omit core knowledge, covering information on fault causes, monitoring methods, and related components. This allows the semantic subgraph to fully present the causal relationships and solutions of the fault. This structured expansion method avoids fragmented subgraph information, providing a clear framework for subsequent transformation into a structured context. This ensures that LLM can accurately understand the fault context, generate logically coherent and comprehensive recommendation results, and help users quickly clarify the fault's path, improving the efficiency of understanding and adopting diagnostic results.

[0028] In one optional implementation, the monitoring method includes at least one of safety integrity analysis, multivariate function analysis, adaptive observer, and time series correlation method, wherein safety integrity analysis is adapted to high-risk safety components, multivariate function analysis is adapted to fluid components, adaptive observer is adapted to rotating components, and time series correlation is adapted to devices with time series correlation relationships.

[0029] This invention, through the setting of precise matching between method and scenario, avoids problems such as low monitoring efficiency and missed fault detection caused by improper method selection, greatly improves the applicability of monitoring methods and fault detection accuracy, provides a clear basis for the fitness calculation in the evaluation function, and further ensures the accuracy of the mapping relationship between "fault phenomenon and monitoring method".

[0030] In one optional implementation, updating the knowledge graph or semantic index based on user feedback includes:

[0031] Record the user's adoption tag for the recommendation result. If adopted, the weight of the corresponding fault phenomenon-monitoring method mapping relationship is increased by a preset ratio. If not adopted, the triples in the knowledge graph are added or corrected based on the fault handling information supplemented by the user. At the same time, when the user feedback data reaches the preset data volume, the encoding parameters of the pre-trained semantic embedding model are fine-tuned based on the user feedback data, and the vector data in the retrieval index is updated.

[0032] This invention addresses the problem that traditional static knowledge graphs and retrieval indexes are ill-suited to adapting to new equipment failure scenarios by constructing a user feedback-driven update mechanism. Recording user-adopted tags and adjusting mapping weights prioritizes effective knowledge; adding or correcting triples based on user-supplemented information enriches the knowledge graph's coverage and fills gaps; fine-tuning the semantic embedding model and updating the index when feedback data reaches a preset amount improves the continuous accuracy of semantic retrieval. This dynamic update mechanism continuously optimizes the knowledge graph and index as equipment operating experience accumulates, ensuring that the fault diagnosis method adapts to new equipment failure modes and operating conditions, maintains high diagnostic accuracy over the long term, extends the lifespan of the technical solution, and provides continuous support for long-term stable equipment operation and maintenance.

[0033] Secondly, this invention provides a device fault diagnosis system based on knowledge graphs and large language models, the system comprising:

[0034] The target equipment data acquisition and processing module is used to acquire multiple types of data sources from the target equipment, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals and equipment design data, and to perform preprocessing and feature extraction to obtain a structured dataset.

[0035] The knowledge triple construction module is used to extract knowledge triples in the form of head entity-relation-tail entity from the structured dataset based on the natural language processing model, and to construct a device health status ontology based on the knowledge triples.

[0036] The fault monitoring mapping triplet construction module is used to comprehensively analyze the corresponding equipment type, operating conditions, historical fault records and related parameter characteristics for each fault phenomenon of an entity in the knowledge triplet. It constructs an evaluation function to quantify the adaptability of different monitoring methods to the fault phenomenon, forms a "fault phenomenon - monitoring method" mapping relationship, and generates a fault monitoring mapping triplet in the form of fault phenomenon - suitable monitoring method - adaptability score.

[0037] The equipment fault knowledge graph construction module is used to merge and store knowledge triples and fault monitoring mapping triples into a graph database to form an equipment fault knowledge graph, and optimize the graph structure through entity alignment and disambiguation algorithms.

[0038] The indexing module is used to encode triples in the knowledge graph using a pre-trained semantic embedding model to generate normalized vectors and build a retrieval index library; it receives user natural language queries and encodes them into semantic vectors, retrieves relevant knowledge fragments through similarity calculation, and builds a semantic subgraph that matches the query;

[0039] The recommendation result output and update optimization module is used to transform the semantic subgraph into a structured context, combine it with the user query to input prompt words into the large language model, generate recommendation results containing fault causes, impact analysis, warning levels and maintenance suggestions, and update the knowledge graph or semantic index library based on user feedback.

[0040] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the device fault diagnosis method based on knowledge graph and large language model described in the first aspect or any corresponding embodiment.

[0041] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the device fault diagnosis method based on knowledge graphs and large language models described in the first aspect or any corresponding embodiment thereof.

[0042] Fifthly, the present invention provides a computer program product, including computer instructions for causing a computer to execute the device fault diagnosis method based on knowledge graph and large language model described in the first aspect or any corresponding embodiment. Attached Figure Description

[0043] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0044] Figure 1 This is a flowchart illustrating a device fault diagnosis method based on knowledge graphs and large language models according to an embodiment of the present invention.

[0045] Figure 2 This is a schematic diagram illustrating the extraction of key fields related to fault diagnosis and monitoring methods based on an FMEA table according to an embodiment of the present invention.

[0046] Figure 3 This is a schematic diagram of the process of constructing a knowledge graph ontology according to an embodiment of the present invention;

[0047] Figure 4 This is an example diagram illustrating the overall framework of the equipment fault diagnosis process according to an embodiment of the present invention;

[0048] Figure 5 This is an example diagram of fault diagnosis output based on an embodiment of the present invention, taking the cause of high temperature of turbine thrust bearing as an example and inputting it into the model;

[0049] Figure 6 This is an example diagram of the output results in the input model, using the high temperature monitoring technology of steam turbine thrust bearings as an example according to an embodiment of the present invention.

[0050] Figure 7 This is a structural block diagram of a device fault diagnosis system based on knowledge graphs and large language models according to an embodiment of the present invention;

[0051] Figure 8 A schematic diagram of the hardware structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] In the field of industrial equipment fault diagnosis and monitoring, knowledge graphs, with their semantic network characteristics, have become a key tool for breaking through traditional technical bottlenecks. They can abstract information such as equipment, components, processes, and historical fault records into a "node-edge" graph structure, using entity tables and relational tables to intuitively present the overall picture of the equipment system and the connections between its various parts, improving diagnostic efficiency and laying the foundation for automated diagnosis.

[0054] Meanwhile, knowledge graphs possess powerful semantic reasoning capabilities, which can break through the limitations of traditional reliance on expert experience and preset rules. By reasoning through node relationships, they can uncover hidden fault modes and anomalies behind the data, demonstrating self-learning and adaptive advantages, and adapting to complex and ever-changing fault scenarios. They can also integrate domain knowledge and professional terminology from multiple subsystems to build comprehensive diagnostic models, solving the problem of insufficient coverage of a single knowledge base.

[0055] Combining knowledge graphs with machine learning techniques such as deep learning can further improve the automation and accuracy of diagnostics. Machine learning models, trained with the structured information of knowledge graphs, can identify and predict fault patterns. When faced with new faults, the reasoning capabilities of knowledge graphs can quickly generate reliable diagnostic suggestions, shortening fault recovery time and reducing maintenance costs. Furthermore, the interpretability of knowledge graphs is superior to traditional black-box models. By graphically displaying diagnostic logic, they provide support for fault prevention and equipment improvement, enhancing system reliability. Therefore, fault diagnosis and monitoring methods based on knowledge graphs and large language models are not only an effective way to overcome the bottlenecks of traditional fault diagnosis and monitoring, but also an important means to achieve precise, efficient, and intelligent maintenance management in the era of industrial intelligence and digital twins.

[0056] This embodiment provides a device fault diagnosis method based on knowledge graphs and large language models. Figure 1 This is a flowchart of a device fault diagnosis method based on knowledge graphs and large language models according to an embodiment of the present invention. The specific steps are as follows:

[0057] Step S1: Obtain multiple data sources from the target device, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals, and equipment design data, and perform preprocessing and feature extraction to obtain a structured dataset.

[0058] Specifically, this invention uses historical fault data of power equipment as a basis, selects the turbine body and its drainage system as the research object, and conducts method verification and application research. Real-time equipment status data are collected from various sensors (such as temperature, vibration, pressure, current, etc.). At the same time, collect historical operating data of the equipment. Maintenance records, FMEA (Failure Mode and Effects Analysis) forms Design materials The above data preprocessing yields a cleaned structured dataset. This involves organizing and standardizing the information in the FMEA table, extracting key fields related to fault diagnosis and monitoring methods, including system name, equipment name, fault phenomenon, fault cause, cause classification, related components, and hazard level, etc. Figure 2 As shown, this information constitutes the core triplet structure of the domain knowledge graph, which helps to comprehensively express the causal relationships and semantic connections between various turbine faults.

[0059] The preprocessing process in this embodiment of the invention includes: handling outliers in real-time monitoring data (e.g., a sudden jump in thrust bearing temperature to 150°C, which is determined to be sensor interference and replaced by the sliding average of 5 adjacent data points); standardization: normalizing and mapping real-time parameters of different dimensions to the [0,1] interval; removing redundant information and standardizing technical terms in FMEA documents and maintenance manuals; further extracting data features (statistically analyzing vibration fluctuation amplitude and historical frequency of fault occurrence, and extracting "fault mode - fault cause" feature pairs from FMEA documents using NLP technology); and finally forming an "equipment dataset" that includes: a structured feature table (e.g., "timestamp - thrust bearing temperature - temperature change rate - vibration fluctuation amplitude") and a text feature table (e.g., "fault phenomenon - associated cause - troubleshooting index"), with a unified format that can be directly used for subsequent knowledge triple extraction. A unified dataset format is the foundation for subsequent semantic encoding and subgraph expansion. Preprocessed features can ensure the semantic matching degree between user queries and knowledge graph retrieval, avoid the bias of recommendation results caused by chaotic data format and missing features, and ultimately improve the accuracy and practicality of fault diagnosis and monitoring method recommendations.

[0060] Step S2: Based on the natural language processing model, extract knowledge triples in the form of head entity-relation-tail entity from the structured dataset, and construct the device health status ontology based on the knowledge triples.

[0061] This invention requires the use of a Natural Language Processing (NLP) model to extract knowledge triples in the form of "head entity (h) - relation (r) - tail entity (t)" from a device dataset. Entities are the specific representations of fault knowledge, such as device name text, fault mode description text, and fault manifestation description text. Relationships describe the association between two entities, such as device name—fault mode, fault mode—fault impact, etc. The entity-relation-entity triples in structured fault knowledge are relatively clear and can be directly extracted; however, the triples in unstructured knowledge are more ambiguous and require text processing methods to extract entities and relations. Specifically,

[0062] 1. Entity Extraction Technology: Considering the need to extract entities from structured, semi-structured, and unstructured text for device health status knowledge extraction, this study focuses on entity extraction technology. Specifically, a portion of the original corpus is split into training and testing sets, and both sets are manually labeled. The labeled corpus is used to train the entity extraction model, thereby performing entity extraction. The test set corpus is used to test the performance of the entity extraction model. The tested entity extraction model is then used to extract entities from the remaining original corpus to obtain the device health status entity knowledge.

[0063] 2. Relation extraction technology, based on entity extraction from text, also needs to extract the semantic relationships between two or more entities to express the main meaning of the text along with the entities. Specifically, entity extraction is performed on the original corpus, and a portion of the extracted original corpus is split into training and testing sets. The labeled corpus is used to train the entity extraction model. The test set corpus is used to test the performance of the relation extraction model. The tested relation extraction model is then used to extract relationships from the remaining original corpus to obtain the device health status relation knowledge.

[0064] Knowledge graphs can be constructed using either a top-down or bottom-up approach. Considering that equipment-related knowledge may be structured, semi-structured, or unstructured, a top-down approach will be adopted. This involves extracting resource patterns from equipment health status knowledge, selecting new patterns with higher confidence levels, and adding them to the knowledge base after manual review. The ontology construction process for the equipment health status knowledge graph is as follows: Figure 3 As shown. Combining expert knowledge and based on the results of knowledge element analysis, entity, relation, and attribute types are defined to complete the ontology graph definition. Knowledge coverage analysis is performed on the ontology graph, and the ontology is revised based on the analysis results to obtain the device health status knowledge graph ontology.

[0065] Step S3: For each fault phenomenon of an entity in the knowledge triple, comprehensively analyze its corresponding equipment type, operating conditions, historical fault records and related parameter characteristics, construct an evaluation function to quantify the adaptability of different monitoring methods to it, form a fault phenomenon-monitoring method mapping relationship, and generate a fault monitoring mapping triple in the form of fault phenomenon-adapted monitoring method-adaptability score.

[0066] The monitoring methods involved in these embodiments of the invention include at least one of security integrity analysis, multivariate function analysis, adaptive observers, and time-series correlation methods. Specifically,

[0067] 1. Based on Safety Integrity Analysis, the Safety Integrity Level (SIL) refers to the level of reliability required to ensure that risks are reduced to a tolerable level. It is used to quantify functional safety levels and to assess and reduce potential risks in industrial systems. SIL assessment mainly involves project management (FSM) for the safety equipment development process, the safety and reliability of the development process, hardware / software, and includes basic safety characteristics. It is mainly applied in high-risk industries such as process industries, power systems, and railway systems. The Safety Integrity Analysis method is a systematic approach to assessing equipment risks and determining necessary safety performance requirements.

[0068] The core principle of Safety Integrity (SIL) is to quantify the safety of equipment and systems into different safety levels by analyzing and assessing potential risks. By conducting SIL assessments on equipment, it can be ensured that the equipment operates within a predetermined safety level, thereby reducing the occurrence of failures and accidents. SIL assessment includes the identification and assessment of potential risks, as well as the analysis of existing protective measures. The SIL level is obtained by calculating the Risk Reduction Factor (RRF), which is used to assess the ability of a safety-related system to reduce risk. The RRF is calculated as follows:

[0069] in, Indicates the current risk level. This indicates the risk tolerance standard (i.e., the acceptable level of risk).

[0070] 2. Equipment condition monitoring technology based on multivariate function analysis is an advanced analytical method in the field of industrial intelligent maintenance. Its basic principle lies in utilizing multivariate function theory in mathematics to analyze multi-source, multi-dimensional data on equipment operation, uncovering the complex correlation between equipment status and operating parameters, and overcoming the limitations of single-parameter monitoring. This technology is implemented in four steps: First, classify equipment parameters, use mathematical functions to analyze multi-dimensional parameter correlations, and screen for strongly linearly correlated parameter sequences; second, time-calibrate key parameter sequences to achieve data point matching and determine stable mathematical relationships between parameters, i.e., a multivariate function-analytical health baseline; third, construct the baseline and perform a goodness-of-fit test; finally, estimate the baseline interval coefficients through parameter intervals, generate upper and lower limit thresholds, and complete the construction of the health baseline. The baseline quality also needs to be evaluated using multi-criteria anomaly detection and a normal data test set.

[0071] In system fault analysis, the primary concern is the discrepancy between the system's actual health state and its expected health state. Considering the multiple components of the system and the logical relationships between them, the fault problem can be described as follows:

[0072]

[0073] in, This is a system malfunction. and This represents the difference between the actual health status and the expected health status of subsystems e and c. A larger value generally indicates a higher degree of system failure. Furthermore, the failures of subsystems can all be calculated using corresponding linear formulas, and specific design is required based on the system's operating status. The metric is to minimize the number of faulty components. These are the corresponding weighting coefficients. The smaller the value, the smaller the number of faulty components. External environment fault characteristic matching index It is its weighting coefficient. A smaller value indicates a closer resemblance to fault characteristics calculated from external sources. Based on the aforementioned anomaly detection criteria, a health baseline is constructed using normal equipment data, and a test set is established based on a subset of normal data to assess the quality of the constructed health baseline. The anomaly detection criteria used determine whether the normal data test set falls within the normal range of the constructed health baseline. If it does, the constructed health baseline is deemed to meet the quality requirements; otherwise, it does not.

[0074] 3. Adaptive observer-based equipment condition monitoring technology is suitable for systems with high-dimensional, massive parameters and nonlinear correlations. It first categorizes equipment condition parameters under the target operating condition, aiming to construct an adaptive health baseline through deep learning and machine learning methods to achieve equipment anomaly detection and performance degradation early warning. Specifically, this includes: classifying multi-dimensional equipment condition parameters under the target operating condition, conducting correlation analysis, and selecting parameter sequences with strong linear correlations; performing time calibration on the selected key parameter sequences to ensure that the data points of the parameter sequence combinations can be matched one-to-one, thereby determining the stable mathematical relationship between multi-dimensional parameters under this monitoring state, laying the foundation for health baseline construction.

[0075] Subsequently, the focus is on constructing a health baseline based on a radial basis function (RBF) network observer. Using parameter interval estimation methods from statistics, the correlation health baseline interval and upper and lower thresholds are calculated, ultimately achieving accurate identification of abnormal equipment states. In terms of specific monitoring logic, the fault observer network is trained using the equipment's original input signal and the output signal from the previous moment to obtain the predicted output signal. The difference between the actual output value and the predicted output value is defined as the residual. An RBF neural network is trained based on normal system operation data, and an adaptive threshold is determined by combining the baseline residual and the correction coefficient obtained from simulation. Then, the difference between the residual and the adaptive threshold is normalized to the 0-1 interval using the arctangent function to obtain the system health monitoring index. Simultaneously, as equipment deteriorates, the monitoring parameters will approach or even exceed the health baseline threshold. By analyzing parameter exceedances and constructing equipment performance degradation curves, future performance degradation trends can be predicted in a timely manner, providing a scientific basis for equipment health status monitoring and maintenance decisions.

[0076] 4. Equipment status monitoring technology based on time sequence association is a key technology for dealing with massive time-series data of equipment monitoring and solving the problem of insufficient retention of time-series features in traditional algorithms, providing important support for intelligent industrial maintenance.

[0077] The current development of equipment monitoring technology has generated massive amounts of data. For equipment with temporal correlations, the typical temporal task profile is characterized by sensor parameter characteristics reaching a specific amplitude range in the parameter dimension, and by this amplitude lasting for a specific duration in the time dimension; otherwise, the system is prone to anomalies. However, traditional statistical feature vectors and multivariate function analytical algorithms are difficult to effectively retain the feature information of time-series data. Since equipment monitoring data is mostly stored in time series form, mining knowledge and baseline rules from time-series data is of great significance for equipment health status analysis and anomaly detection. Therefore, technology based on time-series correlation has become an important direction for innovation. Its core is to use time series analysis methods to mine the patterns and regularities of equipment operation data over time, thereby achieving real-time monitoring and prediction of equipment status. The principle of this technology is based on the inherent correlation of time-series data: the vibration, temperature, and other data generated by equipment operation form a sequence over time, containing information about changes in equipment status. By analyzing the trends, periodicity, and abnormal fluctuations of these time-series data, we can gain in-depth insights into the health status of the equipment, promptly detect potential faults, and provide decision support for preventive maintenance.

[0078] Furthermore, for each fault phenomenon of the entity, this embodiment of the invention comprehensively analyzes its corresponding equipment type, operating conditions, historical fault records, and related parameter characteristics to construct an evaluation function to quantify the adaptability of different monitoring methods. The expression of the evaluation function is as follows:

[0079]

[0080] in: This indicates the suitability of monitoring method m on the target equipment, and is obtained by quantitative scoring based on the matching matrix between equipment type and monitoring method. Specifically, the data sources are equipment design documents, FMEA documents, and equipment component lists, based on which a matching matrix between equipment type and monitoring method is established. For example: rotating components (turbine bearings, thrust bearings) → suitable for adaptive observers; fluid components (pipelines, valves) → suitable for multivariate function analysis; high-risk safety components (safety valves, circuit breakers) → suitable for safety integrity analysis; finally, a matrix score (1~5 points) is used as the final score. The scoring principle should revolve around the matching degree between the characteristics of the equipment type and the core capabilities of the monitoring method. For example, combining the four principles of core function matching (40% weight), technical specification adaptation (30% weight), component structure adaptation (20% weight), and historical application adaptation (10% weight), scores should be given separately on a scale of 1 to 5 (5 points being the best), and then the scores should be converted and summed according to their weights. The final score should be determined based on the converted total score. The core is to determine the degree of adaptation between the monitoring method and the equipment type in terms of function, specifications, structure, and historical application.

[0081] This indicates the diagnostic effectiveness of monitoring method m in historical fault cases of the target equipment, calculated based on the detection rate, false alarm rate, and detection latency rate in these historical fault cases. Specifically, the data sources are historical fault data and maintenance records. The performance of different methods in similar past faults is statistically analyzed: detection rate (TPR), false alarm rate (FPR), and detection latency rate (Latency); these are then combined into a single score.

[0082]

[0083] in, , , For weighting parameters, it is generally recommended It is 0.45. It is 0.45. The value is 0.1, which can be adjusted according to different specific situations. The overall rating is mapped to a (1-5 point) scale. score.

[0084] The sensitivity of monitoring method m to key features of the fault phenomenon is represented by a sensitivity mapping table for monitoring methods to fault features. Specifically, in terms of data sources, core features corresponding to the fault phenomenon (such as parameter drift, numerical mutations, multi-parameter chain causal relationships, etc.) need to be extracted from preprocessed real-time monitoring data of the equipment as the basic input for sensitivity determination. In terms of scoring logic, a sensitivity mapping table of "monitoring method characteristics - fault features" needs to be established first, and corresponding sensitive features are matched according to the inherent technical advantages of each monitoring method. For example, adaptive observers are most sensitive to "parameter drift" features because they are good at capturing nonlinear dynamic changes; time-series correlation methods are most sensitive to "chain causal" features because they can mine the correlation patterns of data over time; and multivariate function analysis is most sensitive to "multivariate slow-changing trend" features because it can analyze the cooperative relationship between multiple variables.

[0085] Finally, the monitoring method is scored on a scale of 1-5 (5 being the highest sensitivity): If the monitoring method is highly sensitive to the key features of the current fault, and is verified by expert experience and historical fault cases, and performs stably in the identification of similar features (e.g., the accuracy rate of identification of this feature in historical faults is ≥90%), then it is scored 4-5 points; if the sensitivity is average, only meeting the basic feature identification requirements, or if there are slight deviations in historical performance, then it is scored 2-3 points; if the sensitivity is low, unable to effectively capture key features, or if there are no historical cases to support its adaptability, then it is scored 1 point.

[0086] This represents the reliability of monitoring method m under the corresponding fault phenomenon, obtained based on whether the monitoring method meets the Safety Integrity Level (SIL) requirements; its data sources include Safety Integrity Level (SIL), Risk Analysis (LOPA / FMEA), and equipment criticality labeling. The method is scored (1-5 points) based on indicators such as PFDavg and PFH under SIL requirements. The scoring logic focuses on the monitoring method's satisfaction with SIL requirements, combined with core indicators such as PFDavg (Probability of Failure Average) and PFH (Probability of Failure Per Hour): a 1-5 point scale is used (5 points being the highest reliability). If the monitoring method not only meets the SIL requirements of the current fault scenario, but also its PFDavg and PFH indicators are better than the standard thresholds (e.g., SIL 2 requirements), then the reliability is considered higher. The actual method is PFDavg= If the monitoring method meets the SIL level requirements, and the PFDavg and PFH indicators are within the standard threshold range, with no redundancy but no reliability shortcomings, then it will score 4-5 points. If the monitoring method does not meet the SIL level requirements of the current fault scenario (e.g., the fault requires SIL 2, but the method only meets SIL 1), or the PFDavg and PFH indicators exceed the standard threshold, or there is no clear SIL compatibility data, then it will score 1 point. For methods that cannot meet the SIL requirements at all and have significant reliability vulnerabilities, it will score 0 points.

[0087] The weight parameters for each indicator are 1, and the sum of the weights of each indicator is 1. The weight parameters are determined based on the fault risk level of the target equipment and industry operation and maintenance standards, and are not restricted here.

[0088] Step S4: Merge the knowledge triples and fault monitoring mapping triples into the graph database to form a device fault knowledge graph, and optimize the graph structure through entity alignment and disambiguation algorithms.

[0089] This embodiment uses the graph database platform Neo4j to perform visual modeling and graph storage of knowledge triples. To meet the requirements of structural standardization, the knowledge graph storage structure uses RDF (Resource Description Frame) to represent triples, thereby realizing the structured expression of equipment fault knowledge while ensuring the efficiency of knowledge retrieval. After the triples are stored in the graph database, a graph structure is formed with entity sets and edge sets as the core. ,in For a collection of entities, Let it be the set of edges.

[0090] To further enhance the consistency and knowledge representation capabilities of the knowledge graph, entity alignment and disambiguation algorithms are used to merge redundant nodes within the graph, effectively eliminating entity ambiguity and ensuring the accuracy of entities and relationships in the knowledge graph. This provides high-quality knowledge support for subsequent semantic retrieval, fault diagnosis, and other processes.

[0091] Step S5: Use a pre-trained semantic embedding model to encode the triples in the knowledge graph to generate normalized vectors and build a retrieval index library; receive user natural language queries and encode them into semantic vectors, retrieve relevant knowledge fragments through similarity calculation, and build a semantic subgraph that matches the query;

[0092] Specifically, in this embodiment of the invention, for each triplet Input the pre-trained semantic embedding model SentenceTransformer to generate embedding vector representations:

[0093]

[0094] in Concatenate text using triples. The d-dimensional embedding vector, after L2 normalization, is used for constructing a vector retrieval index library.

[0095]

[0096] Furthermore, such as Figure 4 As shown, this is based on question-driven semantic retrieval and subgraph expansion. Specifically, it receives user natural language questions. Embedded as semantic vectors Calculate the similarity between the vectors in the knowledge graph and the data, and retrieve the K most relevant knowledge items:

[0097]

[0098] For example, take the 5 most similar triplet segments. Using this as the core, entity backtracking operations are performed. For the retrieved entities, a first- or second-order adjacency expansion is performed according to the causal logic chain of fault phenomenon - fault cause - related components - monitoring method, constructing a semantic subgraph that matches the query. This ensures that the expanded semantic subgraph covers the cause of the fault, the adaptation monitoring method, and information about related components. In this way, a structurally complete and semantically tight knowledge context is obtained, laying a semantic and logical foundation for subsequent language model reasoning.

[0099] The vector retrieval mechanism provided in this invention can quickly narrow down the range of candidate knowledge, preventing the model from getting lost in a large amount of redundant information; it can make up for the problem of limited context length and incomplete knowledge in large models; and it makes knowledge updates no longer dependent on model retraining, but only on updating the semantic vector library.

[0100] Step S6: Transform the semantic subgraph into a structured context, combine it with the user query to input prompt words into the large language model, generate recommendation results containing fault causes, impact analysis, warning levels and maintenance suggestions, and update the knowledge graph or semantic index based on user feedback.

[0101] Specifically, after subgraph retrieval is completed, all relevant triples are sorted and concatenated according to semantic logic or structural path to construct paragraph-style knowledge description text, which is then input into large language models such as DeepSeek as system prompt words. Specifically, the subgraph... The triples in the text are sorted according to causal paths to construct natural language paragraphs. Combined with user questions, these become input prompts for the large language model:

[0102]

[0103] Calling the large language model Perform semantic reasoning and recommendation generation:

[0104]

[0105] The output includes potential anomaly causes, impact analysis, priority assessment, and maintenance suggestions. It retains users' historical questions and answers, enabling continuous modeling of multi-turn dialogue context. The final large model's input includes system prompts, historical dialogue content, and the current question; these three elements work together within the language model's context window, ensuring logical coherence, knowledge accuracy, and natural language in the recommendation results.

[0106] Specifically, after matching the user input with the entities in the knowledge graph, the entity with the highest similarity is selected as the central entity. Starting from this central entity, a two-hop subgraph is extracted from the knowledge graph, containing the entity and its directly related first- and second-order adjacent entities and their relationships. This subgraph comprehensively covers contextual information highly relevant to the current problem, including possible causes of failure, historical failure modes, hazard levels, and associated components. Subsequently, this two-hop subgraph is input as structured text into the DeepSeek large language model. The large language model possesses powerful knowledge understanding and reasoning capabilities, enabling it to fully learn the semantic relationships and causal chains contained in the knowledge graph, comprehensively analyze the user input and the background information of the knowledge graph, and generate diagnostic results.

[0107] The user-input natural language question is "Why is the temperature of the turbine thrust bearing high?" (e.g., Figure 5 (as shown) and "Monitoring technology for high turbine thrust bearing temperature" (such as...) Figure 6 Taking the example shown, the input text is first semantically encoded using the SentenceTransformer model, and then vector matching is performed in the knowledge graph to identify the core entity most similar to it, "high temperature of turbine thrust bearing". Starting with this core entity, a two-hop subgraph is constructed from the knowledge graph, extracting related entities and relationships directly or indirectly connected to it, including potential causes of failure, associated components, historical cases, possible operating conditions, and their corresponding hazard levels. This two-hop subgraph effectively captures the semantic environment and knowledge context surrounding the failure phenomenon. Next, this subgraph is input into the DeepSeek large language model in a structured format. Based on understanding the graph structure and semantic content, the large model, combined with the user's questions, performs multi-hop causal reasoning and semantic integration to generate a fault diagnosis result. The final output is:

[0108] Fault cause identification mainly includes the following aspects:

[0109] I. Lubrication system problems (most common)

[0110] 1. Low lubricating oil pressure

[0111] Phenomenon: Insufficient oil supply to the thrust bearing prevents the formation of a complete oil film, resulting in direct friction and heat generation between the bearing pad and the thrust disc.

[0112] Possible causes: main oil pump failure, low lubricating oil pump outlet pressure, pipeline leakage, pressure regulating valve malfunction, filter blockage, etc.

[0113] 2. The oil quality is substandard.

[0114] Phenomenon: Emulsification, water content, impurities, and oxidation of lubricating oil lead to a decline in its lubrication, cooling, and anti-wear properties.

[0115] Specific effects: Abnormal viscosity: If the viscosity is too high, the fluidity is poor and the cooling effect is not good; if the viscosity is too low, it is difficult to form an oil film with sufficient strength.

[0116] Excessive moisture content can cause oil emulsification, disrupt the stability of the oil film, and potentially lead to component corrosion.

[0117] Excessive impurities: wear down the gold layer (Babbitt alloy) on the bearing surface and scratch the surface of the thrust disc.

[0118] 3. High lubricating oil temperature

[0119] Phenomenon: When the oil inlet temperature is too high, the oil viscosity decreases, reducing its heat carrying capacity and causing the overall bearing operating temperature to rise.

[0120] Possible causes: decreased heat exchange efficiency of the oil cooler, insufficient cooling water volume, or high water temperature.

[0121] II. Problems with the Mechanical Body

[0122] 1. Defects in the thrust bearing itself

[0123] Large deviation in tile thickness: uneven stress on each tile, with some tiles operating under overload.

[0124] Insufficient contact area: Poor contact between the tile and the thrust plate leads to excessive local pressure, resulting in a rapid increase in temperature.

[0125] Babbitt alloy detachment, wear, or scratches on the bearing pads: This damages the oil film and causes frictional heat.

[0126] 2. Thrust plate problem

[0127] Thrust disc drift exceeding limits: This is a key reason explicitly stated in the knowledge base. Excessive runout of the thrust disc end face generates huge axial pulsating forces during operation, continuously impacting the thrust bearings, damaging the oil film, and causing the bearing temperature to rise and fluctuate.

[0128] The thrust disc has poor surface finish: it has scratches, rust, or wear, which disrupts the formation of the oil film.

[0129] III. Operation and Maintenance Issues

[0130] 1. Excessive axial thrust

[0131] Phenomenon: The load acting on the thrust bearing exceeds the design value.

[0132] Possible cause: Scaling in the flow passage: especially in high and medium pressure cylinders, which will increase the pressure difference before and after the blades and increase the axial thrust.

[0133] Drastic load changes: such as sudden load shedding or rapid load increase.

[0134] Abnormal steam parameters: such as main steam temperature and pressure being too low.

[0135] Diaphragm seal wear: Increased interstage leakage alters pressure distribution and increases axial thrust.

[0136] 2. Improper installation and maintenance

[0137] Improper adjustment of thrust bearing clearance (too small or too large).

[0138] The new tile was not properly scraped after replacement, resulting in poor contact.

[0139] The lubricating oil lines or nozzles are clogged or incorrectly installed.

[0140] Summary and Diagnostic Approach: When high thrust bearing temperature is observed, troubleshooting should proceed from the outside in and from easy to difficult areas.

[0141] First, check the lubrication system: confirm that the oil pressure, oil temperature, and oil quality are normal. This is the most common and easiest cause to resolve.

[0142] Next, analyze the operating conditions: check whether the unit load and axial displacement values ​​are abnormal, and determine whether the axial thrust is too large.

[0143] Finally, consider the mechanical body: if there are no problems with the above, then it is likely that there is a defect in the thrust bearing, thrust disc itself, or its installation, which usually requires shutdown and maintenance to resolve completely.

[0144] The knowledge base emphasizes three core causes: substandard oil quality, excessive thrust disc drift, and defects in the thrust bearing itself. These should be given high priority in actual fault analysis. Recommended monitoring technologies mainly include:

[0145] Core technology: Multivariate function analytical monitoring technology. This technology does not simply monitor whether a single parameter such as temperature exceeds the limit, but rather establishes a mathematical model containing multiple related variables and analyzes the dynamic relationship between these variables to detect anomalies at an early stage.

[0146] 1. Monitoring Principle:

[0147] This technology uses thrust bearing temperature as a dependent variable and multiple related operating parameters as independent variables to establish a functional relationship (e.g., Tbearing = f(lubricating oil pressure, lubricating oil temperature, unit load, axial displacement, ...)). The system establishes a baseline model by learning the relationships between these parameters under normal conditions in real time. When the relationships between operating parameters deviate from the baseline model, the system can issue an early warning even if the bearing temperature itself has not yet exceeded the limit.

[0148] 2. Main parameters to be monitored:

[0149] Direct parameter: Thrust pad temperature (multiple measuring points).

[0150] Lubrication system parameters: lubricating oil pressure, lubricating oil inlet temperature, and oil quality indicators (online viscosity and moisture monitoring).

[0151] Load parameters: unit power, main steam flow.

[0152] Mechanical condition parameters: rotor axial displacement, thrust bearing return oil temperature, vibration value.

[0153] Process parameters: main steam pressure, temperature, etc. (used to assess changes in axial thrust).

[0154] 3. Technological advantages:

[0155] Early warning: It can detect potential faults, such as poor lubrication and abnormal load changes, by observing subtle changes in the relationship between parameters before the temperature rises significantly.

[0156] Reduce false alarms: It can distinguish between temperature rise caused by normal load changes and true fault-related temperature rise, thus improving alarm accuracy.

[0157] Fault diagnosis: By analyzing which parameter relationships have deviated, we can provide direction for determining the cause of the fault (such as whether it is a hydraulic problem or a mechanical problem).

[0158] Supplement: In addition to the advanced technologies mentioned above, traditional monitoring methods remain fundamental and necessary:

[0159] Temperature monitoring: Thermocouples or resistance temperature detectors (PT100) are embedded inside the thrust bearing to collect real-time temperature data and trigger alarms for exceeding limits.

[0160] Trend analysis: Monitor the rate and trend of temperature change. Even if the absolute value does not exceed the standard, a continued rapid increase in temperature should be taken seriously.

[0161] Regular sampling and testing: Offline laboratory analysis of lubricating oil quality, checking viscosity, moisture content, impurity content, etc.

[0162] In summary, modern steam turbine thrust bearing temperature monitoring typically employs an intelligent early warning system based on multivariate function analytical monitoring technology, supplemented by traditional direct temperature measurement and oil quality analysis, forming a comprehensive and reliable monitoring and protection system that effectively ensures the safe operation of the unit.

[0163] This invention also updates the knowledge graph or semantic index based on user feedback, including: recording the user's adoption tag for the recommendation result; if adopted, increasing the weight of the corresponding fault phenomenon-monitoring method mapping relationship by a preset ratio; if not adopted, adding or correcting triples in the knowledge graph based on the fault handling information supplemented by the user; and simultaneously, when the user feedback data reaches a preset amount, fine-tuning the encoding parameters of the pre-trained semantic embedding model based on the user feedback data, and updating the vector data in the retrieval index.

[0164] Suppose a user queries "How to diagnose high thrust bearing temperature in steam turbines," the system recommends "using the adaptive observer method for monitoring" based on a knowledge graph and generates a recommendation result. If the user marks the adoption: the original mapping relationship "high thrust bearing temperature - adaptive monitoring method - adaptive observer method" has a weight of 0.8 (out of 1). After being increased by a preset proportion (e.g., 20%), the new weight is updated to 0.96, strengthening the priority of this mapping in subsequent recommendations;

[0165] If the user's annotation is not adopted and they add "the multivariate function analysis method is more effective in practice because the fault is accompanied by slow changes in multiple parameters": based on this feedback, a new triplet "high thrust bearing temperature - suitable monitoring method - multivariate function analysis method (fit score 0.9)" is added, and the original "high thrust bearing temperature - suitable monitoring method - adaptive observer method" fit score is corrected from 0.8 to 0.6.

[0166] The preset data volume is set at 100 user feedback entries. Once 100 entries are collected: Text pairs of "fault phenomenon description - monitoring method adaptation" (e.g., "high thrust bearing temperature - multivariate function analytical method") are extracted from the feedback data and used as fine-tuning data. This data is then used to fine-tune the encoding parameters of the SentenceTransformer model, enabling the model to more accurately capture the semantic relationship between multi-parameter slowly varying faults and the multivariate function analytical method. The triplet related to "high thrust bearing temperature" in the knowledge graph is re-encoded to generate a new normalized vector, replacing the old vector in the retrieval index. This ensures that when users subsequently query similar faults, they can prioritize retrieving knowledge related to the multivariate function analytical method. This eliminates the need for frequent manual updates to the knowledge graph and model, achieving self-iteration through user feedback. This allows the fault diagnosis method to quickly adapt to new equipment fault modes, extending the practical application cycle of the technical solution and reducing the manual maintenance costs for operation and maintenance personnel.

[0167] This embodiment also provides a device fault diagnosis system based on knowledge graphs and large language models. This system is used to implement the above embodiments and preferred embodiments, and details already described will not be repeated. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the system described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0168] This embodiment provides a device fault diagnosis system based on knowledge graphs and large language models, such as... Figure 7 As shown, it includes:

[0169] The target equipment data acquisition and processing module is used to acquire multiple types of data sources from the target equipment, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals and equipment design data, and to perform preprocessing and feature extraction to obtain a structured dataset.

[0170] The knowledge triple construction module is used to extract knowledge triples in the form of head entity-relation-tail entity from the structured dataset based on the natural language processing model, and to construct a device health status ontology based on the knowledge triples.

[0171] The fault monitoring mapping triplet construction module is used to comprehensively analyze the corresponding equipment type, operating conditions, historical fault records and related parameter characteristics for each fault phenomenon of an entity in the knowledge triplet. It constructs an evaluation function to quantify the adaptability of different monitoring methods to the fault phenomenon, forms a "fault phenomenon - monitoring method" mapping relationship, and generates a fault monitoring mapping triplet in the form of fault phenomenon - suitable monitoring method - adaptability score.

[0172] The equipment fault knowledge graph construction module is used to merge and store knowledge triples and fault monitoring mapping triples into a graph database to form an equipment fault knowledge graph, and optimize the graph structure through entity alignment and disambiguation algorithms.

[0173] The indexing module is used to encode triples in the knowledge graph using a pre-trained semantic embedding model to generate normalized vectors and build a retrieval index library; it receives user natural language queries and encodes them into semantic vectors, retrieves relevant knowledge fragments through similarity calculation, and builds a semantic subgraph that matches the query;

[0174] The recommendation result output and update optimization module is used to transform the semantic subgraph into a structured context, combine it with the user query to input prompt words into the large language model, generate recommendation results containing fault causes, impact analysis, warning levels and maintenance suggestions, and update the knowledge graph or semantic index library based on user feedback.

[0175] Further functional descriptions of the various modules and units described above are the same as those in the corresponding embodiments described above, and will not be repeated here. The device fault diagnosis system based on knowledge graphs and large language models in this embodiment is presented in the form of functional units. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0176] This invention also provides a computer device having the above-described features. Figure 7 The system shown is a device fault diagnosis system based on knowledge graphs and large language models.

[0177] Please see Figure 8 , Figure 8 This is a schematic diagram of the structure of a computer device provided in an optional embodiment of the present invention, such as... Figure 8As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 8 Take a processor 10 as an example.

[0178] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GDA), or any combination thereof.

[0179] The memory 20 stores instructions executable by at least one processor 10 to cause at least one processor 10 to perform the method shown in the above embodiments.

[0180] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0181] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include combinations of the above types of memory. The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0182] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0183] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0184] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for diagnosing equipment faults based on knowledge graphs and large language models, characterized in that, include: Acquire multiple types of data sources from the target equipment, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals, and equipment design documents, and perform preprocessing and feature extraction to obtain a structured dataset; Based on a natural language processing model, knowledge triples in the form of head entity-relation-tail entity are extracted from the structured dataset, and a device health status ontology is constructed based on the knowledge triples. For each fault phenomenon of an entity in the knowledge triple, the corresponding equipment type, operating conditions, historical fault records and related parameter characteristics are comprehensively analyzed. An evaluation function is constructed to quantify the adaptability of different monitoring methods to it, forming a fault phenomenon-monitoring method mapping relationship, and generating a fault monitoring mapping triple in the form of fault phenomenon-adapted monitoring method-adaptability score. The expression for the evaluation function is: in: This indicates the degree of suitability of monitoring method m on the target equipment. It is obtained by quantitative scoring based on the matching matrix between equipment type and monitoring method. The data sources are equipment design data, FMEA documents, and equipment parts list. Based on this, a matching matrix between equipment type and monitoring method is established. This represents the diagnostic effectiveness of monitoring method m in historical fault cases of the target equipment, calculated comprehensively based on the detection rate, false alarm rate, and detection delay rate of historical fault cases; This indicates the sensitivity of monitoring method m to the key features of the fault phenomenon. It is obtained from a sensitivity mapping table of monitoring methods to fault features. The data source is the pre-processed real-time monitoring data of the equipment, from which the core features corresponding to the fault phenomenon are extracted. The scoring logic establishes a sensitivity mapping table of "monitoring method characteristics - fault features". Based on the inherent technical advantages of each monitoring method, corresponding sensitive features are matched. The monitoring method is scored according to the scoring system. If the monitoring method is highly sensitive to the key features of the current fault and is verified by expert experience and historical fault cases, it is scored 4-5 points; if the sensitivity is average and can only meet the basic feature recognition requirements, or if there are slight deviations in historical performance, it is scored 2-3 points; if the sensitivity is low and cannot effectively capture key features, or if there are no historical cases to support its adaptability, it is scored 1 point. This indicates the reliability of monitoring method m under the corresponding fault phenomenon, and is obtained based on whether the monitoring method meets the safety integrity level requirements. The weight parameters for each indicator are defined, and the sum of the weights of each indicator is 1. The weight parameters are determined based on the fault risk level of the target equipment and industry operation and maintenance standards. The monitoring method includes at least one of security integrity analysis, multivariate function analysis, adaptive observer, and time series correlation methods. Knowledge triples and fault monitoring mapping triples are merged and stored in a graph database to form a device fault knowledge graph, and the graph structure is optimized by entity alignment and disambiguation algorithms. A pre-trained semantic embedding model is used to encode triples in the knowledge graph to generate normalized vectors and build a retrieval index library; user natural language queries are received and encoded into semantic vectors, and relevant knowledge fragments are retrieved through similarity calculation to build a semantic subgraph that matches the query; The semantic subgraph is transformed into a structured context and combined with the user query to form prompt words input into the large language model. This generates recommendation results that include fault causes, impact analysis, warning levels, and maintenance suggestions. At the same time, the knowledge graph or semantic index is updated based on user feedback. The step of updating the knowledge graph or semantic index based on user feedback includes: Record the user's adoption tag for the recommendation result. If adopted, the weight of the corresponding fault phenomenon-monitoring method mapping relationship is increased by a preset ratio. If not adopted, the triples in the knowledge graph are added or corrected based on the fault handling information supplemented by the user. At the same time, when the user feedback data reaches the preset data volume, the encoding parameters of the pre-trained semantic embedding model are fine-tuned based on the user feedback data, and the vector data in the retrieval index is updated.

2. The method according to claim 1, characterized in that, The pre-trained semantic embedding model is the SentenceTransformer model, which processes triples in the knowledge graph. The expression for encoding and generating the embedding vector is: in, Concatenate text using triples. It is a d-dimensional embedding vector; The embedded vectors are normalized, and a retrieval index is constructed based on the normalized vectors.

3. The method according to claim 1, characterized in that, The retrieved entities are expanded using a first- or second-order adjacency extension based on the causal logic chain of fault phenomenon - fault cause - associated components - monitoring method, ensuring that the expanded semantic subgraph covers fault cause, adapted monitoring method and related component information.

4. A device fault diagnosis system based on knowledge graphs and large language models, characterized in that, include: The target equipment data acquisition and processing module is used to acquire multiple types of data sources from the target equipment, including real-time monitoring data, historical fault data, FMEA documents, maintenance manuals and equipment design data, and to perform preprocessing and feature extraction to obtain a structured dataset. The knowledge triple construction module is used to extract knowledge triples in the form of head entity-relation-tail entity from the structured dataset based on the natural language processing model, and to construct a device health status ontology based on the knowledge triples. The fault monitoring mapping triplet construction module is used to comprehensively analyze the corresponding equipment type, operating conditions, historical fault records and related parameter characteristics for each fault phenomenon of an entity in the knowledge triplet. It constructs an evaluation function to quantify the adaptability of different monitoring methods to it, forms a "fault phenomenon - monitoring method" mapping relationship, and generates a fault monitoring mapping triplet in the form of fault phenomenon - suitable monitoring method - adaptability score. The expression for the evaluation function is: in: This indicates the degree of suitability of monitoring method m on the target equipment. It is obtained by quantitative scoring based on the matching matrix between equipment type and monitoring method. The data sources are equipment design data, FMEA documents, and equipment parts list. Based on this, a matching matrix between equipment type and monitoring method is established. This represents the diagnostic effectiveness of monitoring method m in historical fault cases of the target equipment, calculated comprehensively based on the detection rate, false alarm rate, and detection delay rate of historical fault cases; This indicates the sensitivity of monitoring method m to the key features of the fault phenomenon. It is obtained from a sensitivity mapping table of monitoring methods to fault features. The data source is the pre-processed real-time monitoring data of the equipment, from which the core features corresponding to the fault phenomenon are extracted. The scoring logic establishes a sensitivity mapping table of "monitoring method characteristics - fault features". Based on the inherent technical advantages of each monitoring method, corresponding sensitive features are matched. The monitoring method is scored according to the scoring system. If the monitoring method is highly sensitive to the key features of the current fault and is verified by expert experience and historical fault cases, it is scored 4-5 points; if the sensitivity is average and can only meet the basic feature recognition requirements, or if there are slight deviations in historical performance, it is scored 2-3 points; if the sensitivity is low and cannot effectively capture key features, or if there are no historical cases to support its adaptability, it is scored 1 point. This indicates the reliability of monitoring method m under the corresponding fault phenomenon, and is obtained based on whether the monitoring method meets the safety integrity level requirements. The weight parameters for each indicator are defined, and the sum of the weights of each indicator is 1. The weight parameters are determined based on the fault risk level of the target equipment and industry operation and maintenance standards. The monitoring method includes at least one of security integrity analysis, multivariate function analysis, adaptive observer, and time series correlation methods. The equipment fault knowledge graph construction module is used to merge and store knowledge triples and fault monitoring mapping triples into a graph database to form an equipment fault knowledge graph, and optimize the graph structure through entity alignment and disambiguation algorithms. The indexing module is used to encode triples in the knowledge graph using a pre-trained semantic embedding model to generate normalized vectors and build a retrieval index library; it receives user natural language queries and encodes them into semantic vectors, retrieves relevant knowledge fragments through similarity calculation, and builds a semantic subgraph that matches the query; The recommendation result output and update optimization module is used to transform the semantic subgraph into a structured context, combine it with the user query to input prompt words into the large language model, generate recommendation results containing fault causes, impact analysis, warning levels and maintenance suggestions, and update the knowledge graph or semantic index library based on user feedback. The step of updating the knowledge graph or semantic index based on user feedback includes: Record the user's adoption tag for the recommendation result. If adopted, the weight of the corresponding fault phenomenon-monitoring method mapping relationship is increased by a preset ratio. If not adopted, the triples in the knowledge graph are added or corrected based on the fault handling information supplemented by the user. At the same time, when the user feedback data reaches the preset data volume, the encoding parameters of the pre-trained semantic embedding model are fine-tuned based on the user feedback data, and the vector data in the retrieval index is updated.

5. A computer device, characterized in that, include: The device fault diagnosis method based on knowledge graph and large language model as described in any one of claims 1 to 4 is provided, wherein the memory and the processor are interconnected, the memory stores computer instructions, and the processor executes the computer instructions.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the device fault diagnosis method based on knowledge graph and large language model as described in any one of claims 1 to 4.

7. A computer program product, characterized in that, It includes computer instructions for causing a computer to execute any one of the device fault diagnosis methods based on knowledge graphs and large language models as described in claims 1 to 4.

Images