Metro equipment fault intelligent diagnosis method and system assisted by large language model

Abstract

The invention provides an intelligent subway equipment fault diagnosis method and system assisted by a large language model, and relates to the technical field of data processing, and the method comprises the steps: extracting key information through a large language model, constructing a multi-dimensional equipment fault knowledge graph, obtaining a historical fault data set, and extracting key fault monitoring parameters related to the fault, obtaining a sensing state data set, and judging whether the sensor state data is abnormal or not; and calling a pre-constructed sensing distortion correction algorithm, generating a fault monitoring correction parameter and executing parameter correction, inputting multi-dimensional monitoring data into a diagnosis engine driven by a large language model, and outputting a most possible fault type, cause analysis and recommendation processing strategy and a fault identification report. The technical problems that in the prior art, due to the lack of fusion modeling capacity for the unstructured fault text and the structured monitoring data, the intelligent degree of fault diagnosis is low, and accurate recognition and causal analysis are difficult to achieve are solved, and the fault recognition response speed and accuracy are improved.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and particularly to an intelligent fault diagnosis method and system for subway equipment assisted by large language models. Background Art

[0002] As an important part of urban rail transit, the equipment system of the subway (such as signals, power supply, train control, environmental monitoring, etc.) is complex in composition, with changing operating conditions, and has extremely high requirements for safety and stability. During long-term operation, various faults will inevitably occur in the equipment, such as sensor failures, signal abnormalities, power supply instability, etc. Traditional fault diagnosis methods mostly rely on manual experience or rule-based diagnosis systems, which are difficult to efficiently process large-scale, multi-source heterogeneous monitoring data, and are also difficult to accurately identify hidden or multi-factor coupling faults under complex working conditions.

[0003] Existing subway equipment fault diagnosis technologies mainly rely on manual experience and rule-based knowledge base systems. The diagnosis process usually involves manually comparing equipment alarm information with historical fault cases, making it difficult to perform in-depth semantic modeling on multi-source heterogeneous data, which limits the accuracy and real-time performance of fault identification. In a subway system with complex operating conditions and dense coupling of equipment states, problems such as sensor accuracy drift and data loss result in a large amount of abnormal interference in the original monitoring data, thereby affecting the subsequent diagnostic reasoning accuracy. Summary of the Invention

[0004] This application provides an intelligent fault diagnosis method and system for subway equipment assisted by large language models, which solves the technical problems in the prior art, such as low fault identification accuracy, response lag, and difficult-to-explain diagnostic results due to reliance on manual rules, weak generalization ability of the diagnostic model, and inability to effectively integrate structured monitoring data and unstructured text records, and achieves the technical effect of performing semantic modeling by fusing multi-source data and improving the intelligence and accuracy of fault identification.

[0005] In view of the above problems, the first aspect of the present application provides an intelligent fault diagnosis method for subway equipment assisted by a large language model. The method includes: obtaining the operation data of subway equipment, including sensor data of the equipment, fault history records, and operation environment data. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; preprocessing the text of the fault history records, using a large language model to perform named entity recognition, semantic analysis, and relationship extraction, and extracting key information, including equipment type, fault phenomenon, time node, handling measures, and operation environment; systematically organizing the extracted key information, constructing a multi-dimensional equipment fault knowledge graph, defining entity nodes and semantic relationship edges of the entity nodes, and forming a structured knowledge base. The entity nodes include fault type nodes, fault cause nodes, handling measure nodes, equipment name nodes, component location nodes, and environmental condition nodes; based on the multi-dimensional equipment fault knowledge graph, using a large language model combined with a graph neural network to perform deep learning and semantic modeling on equipment fault data, obtaining a historical fault data set. The historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models; traversing the historical fault data set, extracting key fault monitoring parameters related to faults, collecting sensor status data corresponding to the parameters, obtaining a sensing status data set, and performing anomaly recognition on the sensing status data set to determine whether there are data drifts, data missing, signal distortion, and status anomalies in the sensor status data; based on the anomaly recognition result, calling a pre-constructed sensing distortion correction algorithm, generating fault monitoring correction parameters and performing parameter correction, and adding the fault monitoring correction parameters to the fault status data to form an enhanced multi-dimensional monitoring data set. The fault status data is a set of data records collected by a monitoring system and associated with fault events during the operation of the equipment; inputting the multi-dimensional monitoring data in the multi-dimensional monitoring data set into a diagnosis engine driven by a large language model, and outputting the most likely fault type, cause analysis, recommended handling strategy, and fault recognition report through natural language Q&A and multi-round reasoning.

[0006] In the second aspect of the present application, an intelligent fault diagnosis system for subway equipment assisted by a large language model is provided. The system includes: an operation data acquisition module for subway equipment, which is used to acquire the operation data of subway equipment, including sensor data, fault history records, and operation environment data of the equipment. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; an equipment fault data extraction module, which is used to preprocess the text of the fault history records, perform named entity recognition, semantic analysis, and relationship extraction using a large language model, and extract equipment fault data, including equipment type, fault phenomenon, time node, handling measures, and operation environment; a multi-dimensional equipment fault knowledge graph construction module, which is used to systematically organize the extracted equipment fault data, construct a multi-dimensional equipment fault knowledge graph, define entity nodes and semantic relationship edges of the entity nodes, and form a structured knowledge base. The entity nodes include fault types, causes, and measures; a historical fault dataset acquisition module, which is used to perform deep learning and semantic modeling on equipment fault data based on the multi-dimensional equipment fault knowledge graph, using a large language model combined with a graph neural network, to obtain a historical fault dataset, which includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models; a sensing state dataset anomaly recognition module, which is used to traverse the historical fault dataset, extract key fault monitoring parameters related to faults, collect sensor state data corresponding to the parameters, obtain a sensing state dataset, and perform anomaly recognition on the sensing state dataset to determine whether there are data drifts, data missing, signal distortion, and state anomalies in the sensor state data; a multi-dimensional monitoring dataset formation module, which is used to generate fault monitoring correction parameters and perform parameter correction by calling a pre-constructed sensing distortion correction algorithm based on the anomaly recognition result, and add the fault monitoring correction parameters to the fault state data to form an enhanced multi-dimensional monitoring dataset. The fault state data is a set of data records collected by a monitoring system and associated with fault events during the operation of the equipment; a fault recognition module, which is used to input the multi-dimensional monitoring data in the multi-dimensional monitoring dataset into a diagnostic engine driven by a large language model, and output the most likely fault type, cause analysis, recommended handling strategy, and a fault recognition report through natural language Q&A and multi-round reasoning.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: The intelligent fault diagnosis method and system for subway equipment assisted by a large language model provided in this application relate to the field of data processing technology. It acquires the operation data of subway equipment, extracts key information using a large language model, constructs a multi-dimensional equipment fault knowledge graph, obtains a historical fault data set, traverses the historical fault data set, extracts key fault monitoring parameters related to faults, obtains a sensing state data set, and determines whether there are abnormalities in the sensor state data; Based on the anomaly recognition result, it calls a pre-constructed sensing distortion correction algorithm to generate fault monitoring correction parameters and perform parameter correction, inputs the multi-dimensional monitoring data into a diagnostic engine driven by a large language model, and outputs the most likely fault type, cause analysis, recommended processing strategies, and a fault recognition report. This application solves the technical problem in the prior art that due to the lack of the ability to fuse and model unstructured fault texts and structured monitoring data, the intelligent degree of fault diagnosis is low, and it is difficult to achieve accurate identification and causal analysis, and improves the fault recognition response speed and accuracy.

[0008] In summary, by constructing a multi-dimensional equipment fault knowledge graph and integrating the deep learning mechanisms of a large language model and a graph neural network, this application can accurately extract key fault information, perform semantic modeling and causal reasoning, and conduct more comprehensive fault identification, reducing the fault omission and misjudgment rates; The diagnostic engine based on the semantic enhanced reasoning model makes the fault cause analysis more logically relevant and explanatory, avoiding misdiagnosis and missed judgment caused by sensing abnormalities, and at the same time improving the robustness and accuracy of fault recognition. Based on the model dynamic selection mechanism, it reduces the diagnostic deviation caused by model misselection, can quickly generate a highly targeted fault recognition report and processing strategy suggestions, improves the diagnostic efficiency and automation level, helps the intelligent upgrade of subway equipment operation and maintenance management, and improves the operation safety and service reliability of the urban rail transit system.

[0009] The above description is only an overview of the technical solutions of this application. In order to be able to understand the technical means of this application more clearly, it can be implemented according to the content of the specification. And in order to make the above and other purposes, features and advantages of this application more obvious and understandable, the specific implementation manners of this application are hereinafter specifically exemplified. Brief Description of the Drawings

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the following described drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0011] Figure 1Schematic flowchart of the intelligent diagnosis method for subway equipment failures assisted by a large language model provided by an embodiment of the present application.

[0012] Figure 2 Schematic structural diagram of the intelligent diagnosis system for subway equipment failures assisted by a large language model provided by an embodiment of the present application.

[0013] Explanation of reference numerals: The operation data acquisition module 10, the equipment failure data extraction module 20, the multi-dimensional equipment failure knowledge graph construction module 30, the historical failure data set acquisition module 40, the sensing state data set anomaly recognition module 50, the multi-dimensional monitoring data set formation module 60, the failure recognition module 70. Detailed implementation manners

[0014] The present application provides an intelligent diagnosis method and system for subway equipment failures assisted by a large language model, which are used to solve the technical problems in the existing subway equipment operation and maintenance system that due to the lack of semantic understanding ability and failure causality modeling mechanism, the failure diagnosis depends on manual experience, the accuracy is low, and the response is not timely.

[0015] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0016] It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not have to be used to describe a specific order or sequence. It should be understood that such data used can be interchanged under appropriate circumstances so that the embodiments of the present application described here can be implemented in an order other than those illustrated or described here. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or are inherent to these processes, methods, products or devices.

[0017] Embodiment 1, as Figure 1 shown, the present application provides an intelligent diagnosis method for subway equipment failures assisted by a large language model, and the method includes: P100: Obtain the operation data of subway equipment, including sensor data of the equipment, failure history records, and operation environment data, and the text of the failure history records includes historical maintenance records, operation logs, and failure descriptions.

[0018] Specifically, first, various sensors installed on key subway equipment, such as temperature sensors, vibration sensors, current sensors, displacement sensors, etc., are used to collect data on the structural state, operating state, electrical parameters, mechanical responses, etc. in real time. The sensor data can include meta-information such as timestamps, sampling frequencies, data units, and collection sources to support subsequent data preprocessing and normalization operations.

[0019] Structured and unstructured information related to faults recorded during the entire life cycle of subway equipment, specifically including historical maintenance records: covering field information such as maintenance time, maintenance personnel, maintenance parts, replaced components, maintenance measures, and effect evaluation, used to restore the evolution track of equipment maintenance; operation logs: automatically generated by the equipment control system or monitoring platform, recording changes in the operating state, start-stop events, alarm status, and dispatching instructions of the equipment during each time period; fault descriptions: including text content such as fault phenomena, preliminary judgments, occurrence times, and external conditions in natural language filled in by operation and maintenance personnel or recorded by intelligent terminals, providing semantic modeling corpus support for large language models.

[0020] Environmental data and data on external interference factors related to the equipment operating environment, such as tunnel humidity, temperature, dust concentration, electromagnetic interference, seismic shocks, etc. The environmental data can be regularly collected through environmental sensors installed on platforms, sections, or around equipment, and is used to analyze the impact of external working conditions on the equipment health status.

[0021] P200: Preprocess the text of the fault history record, and use a large language model to perform named entity recognition, semantic analysis, and relationship extraction to extract key information, including equipment type, fault phenomenon, time node, treatment measures, and operating environment.

[0022] Furthermore, step P200 of the embodiment of the present application further includes: P201: Perform standardized preprocessing on the historical fault record text collected from the subway operation and maintenance system. The preprocessing includes character cleaning, regularization of proper nouns, format unification, semantic error correction, and segment annotation; P202: Input the preprocessed text into a fine-tuned large language model to perform a multi-layer semantic processing process; P203: Through semantic recognition and relationship modeling, extract and output a structured set of key information. The key information includes equipment type, fault phenomenon, time node, treatment measures, and operating environment; P204: Assign a semantic confidence score to each type of key information extracted, and tag information with low confidence or ambiguous expressions.

[0023] It should be understood that by using a large language model to identify and label key terms in historical texts, proprietary noun entities such as equipment names, component numbers, maintenance tools, and personnel identities are extracted. Combining context understanding and language structure, semantic information such as equipment operating status, fault causality, and diagnostic conclusions implicit in each record is parsed, and semantic associations between different entities are established, such as knowledge triple structures like "brake → fault phenomenon → unable to reset", "fault occurrence time → October 2024", "maintenance measure → replace sensor", etc. The extracted key information is uniformly converted into a data format that conforms to a predefined template or field, facilitating subsequent storage, training, and analysis, and enhancing the readability, processability, and semantic deep modeling ability of fault corpora.

[0024] First, redundant characters (such as extra spaces, line breaks, special symbols, non-UTF characters) in the original text are removed or replaced to improve text standardization. For grammar errors, word confusion, or semantic ambiguity, semantic error correction is performed in combination with the context. For example, "The brake has no response, and the current anomaly has been repaired" is corrected to "The brake is unresponsive, and the current anomaly has been repaired after that" to enhance the clarity of language logic and context consistency. A composite segmentation algorithm based on rules and models is used to annotate the text according to content semantics, such as: fault phenomenon paragraphs, processing process paragraphs, results and suggestions paragraphs, relevant equipment and time nodes, etc.

[0025] Next, the large language model uses a pre-trained language model based on the Transformer architecture, such as BERT, RoBERTa, or ChatGLM-like models. To improve its performance in subway equipment fault diagnosis tasks, a dedicated text corpus for subway equipment operation and maintenance scenarios is first constructed. The corpus includes, but is not limited to, the following: historical fault record texts of subway equipment; content of operation and maintenance work orders; professional maintenance documents and manuals; maintenance reports and accident notification records; natural language annotation segments in control system operation logs.

[0026] Next, the fault history texts collected from the subway operation and maintenance system (such as operation logs, maintenance records, inspection reports, etc.) are preprocessed through standardization processes such as character cleaning, format unification, and proper noun regularization, and then input into the fine-tuned large language model. Based on the context understanding ability of the language model, the types of key information in the text are identified. By performing association modeling on the above information, the logical and causal relationships between entities are identified, including: the attribution relationship between the fault phenomenon and the equipment type; the chronological relationship between the fault occurrence time and the handling measures; the potential causal impact between environmental factors and the occurrence of faults. The results of identification and modeling are uniformly output as a structured set of key information, represented in the form of key-value pairs or triples, to support subsequent graph construction and intelligent reasoning. The extracted structured key information will be used as input to further construct an equipment fault causal knowledge graph and a semantic-enhanced reasoning model, improving the automation, semantic accuracy, and response efficiency of fault analysis.

[0027] Finally, call the confidence evaluation module of the large language model trained based on multi-task learning, and combine indicators such as attention weights, context consistency, and prediction distribution probability to perform semantic confidence scoring on each identified key information item. The scoring range is usually 0 to 1, and the higher the value, the higher the confidence. Perform confidence evaluation item by item on the five identified types of key information (equipment type, fault phenomenon, time node, handling measures, operating environment). Attach the confidence score and the marked information as auxiliary attributes to the structured information item to form an output result for manual review or further learning and optimization. Information items marked as "ambiguous" or "low confidence" will be transferred to the manual review process or used as training samples to be fed back for fine-tuning the language model, thereby continuously optimizing the model performance and the quality of text information extraction.

[0028] P300: Systematically organize the extracted key information, construct a multi-dimensional equipment fault knowledge graph, define entity nodes and the semantic relationship edges of the entity nodes to form a structured knowledge base. The entity nodes include fault type nodes, fault cause nodes, handling measure nodes, equipment name nodes, component location nodes, and environmental condition nodes.

[0029] Furthermore, step P300 of the embodiment of the present application further includes: P301: Archive the key information extracted by the large language model in a unified data structure, including a fault type field, a fault cause field, a handling measure field, a device name field, a component location field, and an environmental condition parameter field; P302: Define the core entity node types in the device fault domain according to the archived information, and the node types include a fault type node, a fault cause node, a handling measure node, a device name node, a component location node, and an environmental condition node; P303: Based on the semantic analysis and context reasoning results, extract the relationship edges between entity nodes and annotate the semantic types of each edge; P304: Build a multi-dimensional knowledge graph for the subway device fault diagnosis task based on the entity nodes and relationship edges. The multi-dimensional knowledge graph is stored in a graph structure and can be regarded as a set of structured triples, where the triple is entity 1 - relationship - entity 2.

[0030] Optionally, generate a multi-dimensional device fault knowledge graph through a semantic recognition and entity relationship extraction algorithm to achieve a structured expression of historical fault information and a causal logic modeling, improving the accuracy and interpretability of subsequent diagnostic engines in fault reasoning and strategy recommendation.

[0031] First, archive the key information extracted by the large language model in a unified data structure to build a standardized structured data form. The data structure includes, but is not limited to, the following fields: Fault type field: used to record the specific type of fault event, such as "communication anomaly", "overheat alarm", etc.; Fault cause field: used to represent the direct or indirect cause of the fault identified by the language model, such as "cable aging", "signal interference"; Handling measure field: used to describe the handling strategies or operation steps proposed by maintenance personnel or the system; Device name field: identifies the target device or system where the fault occurs, such as "platform screen door controller", "train control system receiving unit"; Component location field: further identifies the specific component or module location inside the device, such as "the 3rd door of car 1 of the train", "the left interface of the main control board"; Environmental condition parameter field: records the external environmental information related to the time of fault occurrence, including temperature, humidity, electromagnetic interference level, etc.

[0032] Next, based on the archived structured fault information, the system defines the core entity node types in the subway equipment fault knowledge graph for constructing a semantic network and supporting subsequent reasoning and diagnosis. The core entity node types include, but are not limited to, the following categories: Fault Type Node: Represents the specific category of abnormal or fault events that occur during the operation of the equipment, such as "signal interruption", "power failure", "communication delay", "brake failure", etc.; Fault Cause Node: Describes the technical or environmental incentives that cause the occurrence of the fault type, such as "module aging", "interface soldering joint looseness", "external strong electromagnetic interference", "component drift under high temperature environment", etc.; Handling Action Node: Represents the disposal strategies or operation and maintenance means taken for specific faults, such as "replace the main control board", "restart the communication link", "adjust the cable wiring", "clean the accumulated dust", etc.; Device Node: Identifies the name of the subway system equipment where the fault occurs, including "traction inverter", "on-vehicle control unit", "trackside signal machine", "pantograph", etc.; Component Location Node: Used to further refine the internal structural location of the faulty equipment, such as "power supply module in the second car of train No. 1", "mid-section port of the main control cabinet of the signal system", etc.; Environment Node: Characterizes the external environmental state when the fault occurs, including information such as "temperature value", "humidity range", "equipment operation period", "on-site electromagnetic intensity level", etc. After completing the entity node recognition, based on the deep semantic analysis and context reasoning results of the archived text by the large language model, further extract the semantic relationship edges between entity nodes to construct the edge structure in the equipment fault knowledge graph. The specific steps include: Context Semantic Modeling: Input the preprocessed fault text into a fine-tuned large language model (such as BERT, K-BERT, or RoBERTa, etc.), and the model comprehensively understands the sentence structure, syntactic dependencies, and context semantics to identify possible causal, associative, or hierarchical relationships between entity nodes. Relationship Edge Extraction: Through methods such as syntactic path tracking, semantic dependency graph analysis, and relationship extraction template matching, the system identifies the significant semantic relationships existing between any two labeled entity nodes and generates a candidate edge set.

[0033] Finally, based on the entity nodes extracted in the foregoing steps and the semantic relationship edges between them, a multi-dimensional knowledge graph for subway equipment fault diagnosis tasks is constructed. The knowledge graph is stored in a graph structure and organized and managed in the form of structured triples (entity 1 - relationship - entity 2). First, all the identified and labeled entity nodes are uniformly modeled, including multiple types of nodes such as fault types, fault causes, handling measures, equipment names, component locations, environmental conditions, etc., and type identifiers and attribute fields (such as node unique identifiers, node descriptions, source confidence levels, etc.) are set for each type of node; at the same time, all semantic relationship edges are uniformly encoded and managed according to their types. For each pair of entity nodes with a semantic relationship, the system automatically generates a structured triple, expressed as (entity 1 - relationship - entity 2). For example, from the text "The brake system failure is induced by a high-temperature environment", the triple (brake system failure, caused_by, high-temperature environment) can be generated and stored as an edge in the graph. All triples are imported into a graph database (such as Neo4j, GraphDB, etc.) to construct a semantic graph structure. Each node connects all its upstream and downstream related nodes to form multiple causal paths and handling chains, realizing the traceable logical relationship of fault - cause - measure. To enhance the expression ability of the knowledge graph, the system supports adding additional attribute information to the triples, such as time tags (indicating the time period when the fault event occurred), confidence scores, data source identifiers (such as operation logs, maintenance records, etc.), applicable equipment models, etc., to form a knowledge graph structure with multi-dimensional semantics such as time, space, logic, and environment. Finally, the constructed multi-dimensional knowledge graph is stored in a structured graph database, supporting retrieval through the graph query language, and at the same time, RESTful API interfaces are encapsulated for the diagnostic engine to call, supporting knowledge retrieval and path analysis in subsequent inference and diagnosis tasks. The construction of the multi-dimensional knowledge graph significantly enhances the structured expression and semantic association ability of equipment fault information, providing a semantic basis for subsequent natural language question answering inference and fault decision-making based on large language models.

[0034] P400: Based on the multi-dimensional equipment fault knowledge graph, using a large language model combined with a graph neural network, deep learning and semantic modeling are performed on equipment fault data to obtain a historical fault data set, and the historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models.

[0035] Furthermore, step P400 of the embodiment of the present application further includes: P401: Based on unstructured and structured data such as fault description texts, maintenance records, and sensor logs, natural language processing technologies such as named entity recognition (NER) and semantic relation extraction are used to extract information and semantic relations such as equipment type, fault phenomenon, time node, handling measures, operating environment, etc. from them, and a knowledge graph containing a multi-level semantic structure of equipment - phenomenon - cause - measure is constructed; P402: The knowledge graph with the multi-level semantic structure is input into a graph neural network, and through node feature aggregation and adjacency relation modeling, low-dimensional semantic vector representations of entity nodes and relation edges are obtained, retaining their context semantics and graph structure information; P403: A large language model is used to perform in-depth semantic modeling and language understanding on the low-dimensional semantic vector representations of the entity nodes and relation edges to generate context-enhanced text vectors, and the large language model is a Transformer structure model fine-tuned for fault corpora; P404: The low-dimensional semantic vector representations of entity nodes and relation edges extracted by the graph neural network are fused with the context-enhanced text vectors extracted by the large language model to form a unified joint feature vector space; P405: Through clustering analysis, fault classification, and causal chain identification processing on the joint feature vector space, feature alignment and normalization modeling are performed on the operating states and fault manifestations of multiple devices to form a structured historical fault data set.

[0036] It should be understood that according to the semantic characteristics and structured requirements of subway equipment operation and maintenance data, through the joint modeling mechanism of a large language model and a graph neural network, key information in fault data can be accurately extracted, causal chains can be effectively modeled, and a highly reliable structured knowledge graph can be formed to support subsequent intelligent fault diagnosis and processing strategy recommendation.

[0037] First, obtain the multi-source heterogeneous data accumulated in the subway equipment operation and maintenance system. The data includes, but is not limited to, fault description texts (such as operation and maintenance logs, manual repair reports), repair records (such as repair work orders, maintenance plans), and sensor logs (such as monitoring data of temperature, current, vibration, etc.). This dataset contains both structured information and unstructured text information. For unstructured text data, the system performs preprocessing operations such as character cleaning, format unification, term regularization, and segmentation annotation through a text standardization processing module. Subsequently, use the fine-tuned large language model to execute the natural language processing process, including operations such as named entity recognition (NER), semantic role annotation, and relation extraction, to automatically identify and extract key semantic entity information, specifically including: equipment type (such as: traction transformer, train signal controller); fault phenomenon (such as: voltage anomaly, communication interruption, vibration anomaly); time node (such as: fault occurrence time, repair time); treatment measures (such as: replace module, restart equipment, reconnect cable), operating environment (such as: temperature, humidity, tunnel section, operation load). After entity extraction is completed, the system identifies the logical relationships and causal connections between various entities through a semantic relation extraction module. For example, identify relationship paths such as "a certain fault phenomenon occurred in the equipment", "a certain fault cause led to this phenomenon", "a certain measure was taken during repair to eliminate this fault" through context semantic modeling, and represent them in the form of structured triples: (entity 1, relation, entity 2), such as: (traction transformer, appears, voltage anomaly)(voltage anomaly, cause, poor heat dissipation)(poor heat dissipation, take measure, replace heat dissipation unit) Finally, organize the identified various entity nodes and the semantic relation edges between them into a multi-level structured equipment fault knowledge graph. The knowledge graph is stored in a graph structure and supports graph visualization, semantic retrieval, and graph neural network modeling. The core semantic skeleton of this knowledge graph is "equipment - phenomenon - cause - measure", which can realize systematic modeling and reasoning support for complex fault information.

[0038] Next, based on the multi-level semantic structure knowledge graph constructed for the subway equipment fault diagnosis task, the system uses it as the input graph structure and further introduces a Graph Neural Network (GNN) model for semantic embedding learning to obtain low-dimensional semantic vector representations of entity nodes and relationship edges with context information and structure perception capabilities. The "equipment - phenomenon - cause - measure" knowledge graph is loaded into the graph learning engine in graph data format. Each entity node in the graph (such as "signal reception module", "communication interruption", "EMI interference", "replace receiver") is represented as a node, and each semantic relationship (such as "appears", "cause", "take measures") is represented as a directed edge. Each node is attached with its initial attribute features (such as word embedding, entity type, context summary) as the input of the graph neural network. Graph neural network models such as GCN (Graph Convolutional Network), GAT (Graph Attention Network), or R-GCN (Relational GCN) are used to update the semantic representation of each node through an iterative feature aggregation mechanism according to the adjacency relationship of the node in the graph. This process not only considers its own features but also synthesizes the features and relationship types of its direct neighbor nodes to model the semantic propagation path of fault information in the graph. After multiple rounds of aggregation, each entity node and relationship edge in the graph is assigned a low-dimensional semantic embedding vector, usually a 128-dimensional or 256-dimensional floating-point vector. This vector not only encodes the semantics of the node itself but also implicitly contains its structural position, context causal relationship, and the semantic path information in the graph. The finally generated low-dimensional semantic vectors are used for subsequent similarity calculation, node clustering, causal chain tracing, and graph neural network-assisted fault diagnosis tasks, providing support for subway equipment anomaly recognition, fault location, and strategy recommendation.

[0039] Next, based on the low-dimensional semantic vector representations of entity nodes and relationship edges generated by the graph neural network, a Large Language Model (LLM) is introduced to further semantically understand and linguistically model this vector information to enhance the context expression ability and form context-enhanced text vectors for the subway equipment fault domain.

[0040] Furthermore, a large language model based on the Transformer architecture, such as BERT, RoBERTa, or GPT-like models, is adopted and fine-tuned with supervision on a large-scale subway equipment failure-related corpus (including maintenance reports, operation logs, repair records, etc.) to enable it to have the capabilities of fault semantic modeling and context-aware reasoning. Through fine-tuning, the model can accurately identify professional terms, causal relationships, and context dependencies, improving the accuracy and robustness of language understanding. The low-dimensional semantic vectors of each entity node and relationship edge generated by the graph neural network are combined into an input sequence according to the structured triple (entity-relationship-entity) and mapped into the Token Embedding space recognizable by the model through the vector-text alignment module. For example, for a set of triples: ("signal module" — occurred — "communication interruption"), it is transformed into an embedded sequence form and input into the language model, and the graph structure information is attached as a context prompt to enable the model to understand the real semantic background represented by the graph structure.

[0041] Next, for the structured semantic information extracted by the Graph Neural Network (GNN) and the context-enhanced semantic information extracted by the Large Language Model (LLM), a fusion mechanism is designed to construct a unified joint feature vector space to provide feature support with high semantic expression ability for subsequent tasks such as fault reasoning, classification, and traceability analysis. First, the node vector representation from the graph neural network is received, which contains the structural semantic embeddings of entity nodes (such as "brake system", "high-temperature alarm") and relationship edges (such as "causes", "associated") in the equipment failure knowledge graph. At the same time, the system receives the text vector output by the fine-tuned large language model, which encodes the context semantic information in the fault description corpus. Before entering the fusion module, the two types of vectors are respectively standardized through a unified dimension mapping layer (such as a fully connected layer or a projection matrix) to ensure their comparability within the same semantic space dimension, and a multimodal semantic fusion method is used to fuse the node vector representation of the graph neural network and the text vector output by the large language model.

[0042] Finally, based on the joint feature vector space generated by the fusion of graph neural networks and large language models, semantic analysis and modeling are performed on the operating states, fault phenomena, and environmental contexts of multiple devices. Using various technical means such as clustering, classification, and causal chain identification, a structured historical fault dataset for diagnostic tasks is constructed. For all vector samples in the joint feature vector space, unsupervised learning algorithms (such as K-Means, DBSCAN, spectral clustering, etc.) are used to cluster the fault-related samples and divide them into fault groups with similar feature patterns. Through the implementation of the above steps, the deep fusion and semantic modeling of multi-source fault information of subway equipment are completed, and a structured historical fault dataset containing clustering results, classification labels, and causal chains is established, providing a high-quality data foundation and semantic support for subsequent intelligent diagnostic reasoning and strategy generation.

[0043] Further, step P405 of the embodiment of the present application further includes: P405-1: The historical fault dataset includes multi-dimensional feature representations, fault causal chains, and semantic enhanced reasoning models; P405-2: The multi-dimensional feature representation is a joint representation vector constructed based on dimensions such as device type, key operating parameters, and working condition environment characteristics, and is used to reflect the fault manifestation characteristics under different device states; P405-3: The fault causal chain is the device fault chain mined through the collaboration of graph neural networks and large language models, obtaining the semantic causal relationship of initial cause - propagation path - final fault; P405-4: The semantic enhanced reasoning model is the K-BERT architecture that fuses pre-trained language models and knowledge graphs, obtaining multi-dimensional semantic feature representations with the capabilities of fault causal identification, semantic ambiguity resolution, and cross-path reasoning.

[0044] Specifically, through clustering analysis, fault classification, and causal chain identification processing of the joint feature vector space, feature alignment and normalization modeling are performed on the operating states and fault manifestations of multiple devices, forming a structured historical fault dataset, which can further improve the fault knowledge transfer and generalization capabilities across devices and scenarios. By constructing a unified fault feature representation framework and a standardized modeling mechanism, the adaptability and interpretability of historical fault knowledge under different device types and working conditions are ensured.

[0045] First, a structured data set is constructed by deeply processing multi-source heterogeneous data such as sensor data, operation and maintenance records, and fault texts generated during the operation of subway equipment. Based on the joint modeling mechanism of graph neural networks and large language models, vectorization processing is performed on data from different sources (such as sensor data like temperature, current, vibration, etc., fault description texts, maintenance logs, etc.), and they are uniformly represented in the feature space. Each equipment sample is mapped to a high-dimensional feature vector containing operating parameters, environmental conditions, historical maintenance status, text semantics, etc., to comprehensively reflect the equipment state. Through clustering analysis, semantic relationship modeling, and graph structure reasoning, fault triggering conditions, evolution paths, and their causal relationships are mined from historical data, and a multi-node causal path of "equipment state → fault symptom → fault type → treatment result" is constructed. This causal chain, as the core relationship structure in the knowledge graph, helps with fault location, tracing, and prediction. The reasoning model integrates the learning abilities of graph neural networks (such as GCN, GAT) and pre-trained large language models (such as K-BERT or T5) for node semantics and structural dependencies. Through the context-aware mechanism and knowledge graph injection strategy, the expression ability and reasoning accuracy of fault knowledge are improved. This model supports semantic completion, anomaly explanation, and generation of treatment suggestions based on historical knowledge. Through structured modeling, deep representation and intelligent reasoning of historical fault knowledge are achieved, providing basic support for subsequent real-time diagnosis and strategy recommendation.

[0046] Next, a joint representation vector constructed based on multiple dimensions such as the type of equipment, key operating parameters, and characteristics of the operating condition environment is used to reflect the operating state and potential fault manifestations of subway equipment under different operating conditions. Information such as the structural attributes, function classification, and component numbers of the equipment is extracted and used as part of the equipment category embedding vector to distinguish the structural differences in feature representations of different types of subway equipment (such as traction systems, braking systems, air conditioning systems, etc.). Core indicators reflecting the operating health status of the equipment, such as voltage, current, temperature, vibration frequency, acceleration, etc., are selected from the sensor data, and after normalization processing and time series feature aggregation, an operating trend vector subspace is formed. Combining environmental context information such as the operating time period, carriage load, external temperature and humidity, and tunnel location, an environmental semantic vector is constructed to depict the potential impact of external conditions on the operating state and fault manifestations of the equipment. Information from different dimensions is integrated into a unified joint representation vector through feature concatenation or attention fusion mechanisms, serving as the state representation of the equipment at a certain operating moment for subsequent tasks such as clustering analysis, classification recognition, and causal modeling.

[0047] Next, based on the subway equipment operation data and historical fault texts, a structured causal sequence is constructed by combining the collaborative modeling and reasoning mechanisms of the graph neural network (GNN) and the large language model (LLM). Based on the semantic entity nodes such as equipment operation status, abnormal phenomena, operation behaviors, and environmental disturbances extracted from the knowledge graph, GNN is used to model the potential propagation paths between nodes and aggregate features, initially constructing a possible causal propagation graph. The graph structure information is fused with the large language model and input into the fine-tuned Transformer model. Utilizing its capabilities in natural language understanding and semantic relationship modeling, the causal semantics between nodes are classified and labeled to identify the ternary causal chain structure of "inducing factor - intermediate state - resulting fault". Finally, a fault causal chain with the main structure of "initial inducing factor → fault propagation path → final fault" is generated. Each chain consists of multiple semantically related nodes, annotated with the type of causal relationship (such as: triggering, influencing, intensifying, suppressing), its occurrence order, and confidence score. The generated fault causal chains are systematically organized into structured sequence data for subsequent calls and reuse in tasks such as fault prediction, anomaly tracing, and recommendation strategy generation.

[0048] Furthermore, to address the difficulties in subway equipment operation data, such as ambiguous expressions, complex causal chains, and multi-path intersections, an improved K-BERT (Knowledge-enhanced BERT) model architecture is adopted as the core of the semantic enhancement reasoning module to achieve in-depth fusion reasoning of language understanding and structural knowledge. For a large number of input fault description texts, maintenance records, and operation logs, they are first encoded into the Token sequences required by the pre-trained language model BERT. At the same time, based on the constructed multi-dimensional equipment fault knowledge graph, key entities involved therein (such as "brake failure", "high-temperature environment", "motor failure", etc.) are identified and matched, and the relevant knowledge sub-graphs are fused into the representation of the original sentence in a soft insertion manner to construct an enhanced input. The "visibility matrix" mechanism introduced in K-BERT is used to control the knowledge injection process, enabling limited information interaction between knowledge nodes and text nodes in the self-attention mechanism, thereby avoiding noise propagation and realizing joint learning of language context and structural knowledge. The model outputs a multi-dimensional semantic feature representation vector through a fusion representation learning mechanism. The multi-dimensional semantic feature representation vector can identify the implicit causal trigger structure in the sentence (such as "caused by...", "appeared due to...") and establish the semantic relationship between the corresponding nodes; use context and graph constraint resolution to clarify the specific meaning of "power off" in different devices; perform multi-hop semantic reasoning on the graph structure for non-linear and cross-segment causal relationships, thereby revealing complex propagation paths. The finally generated semantic vector not only retains the context information of the text but also integrates the entity and relationship semantics in the structured knowledge graph, which is used to drive subsequent modules such as anomaly recognition, cause location, and strategy recommendation, and can be integrated into a multi-level reasoning process to achieve more accurate fault diagnosis decision support.

[0049] P500: Traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect the sensor status data corresponding to the parameters to obtain a sensing status data set, and perform anomaly recognition on the sensing status data set to determine whether there are data drifts, data missing, signal distortions, and status anomalies in the sensor status data.

[0050] Furthermore, step P500 of the embodiment of the present application further includes: P501: Traverse the device fault dataset, parse the device operation data collected in real time item by item, and extract the key fault monitoring parameters associated with the fault events. The key fault monitoring parameters include, but are not limited to, index parameters such as vibration amplitude, current fluctuation, temperature rise gradient, rotational speed deviation, voltage anomaly, air leakage volume, etc.; P502: Extract the key fault monitoring parameters, locate the sensor channels corresponding to the key fault monitoring parameters, collect the current state data of the sensors, and form a sensing state dataset corresponding one-to-one with the monitoring parameters. The sensing state data includes sensing values, timestamps, signal quality, communication status, and sensor operating status; P503: Perform anomaly recognition processing on the sensing state data, and use a multi-dimensional detection mechanism to determine whether there are anomalies. The anomalies include data drift, data missing, signal distortion, and status anomaly; P504: Label the detected abnormal states and output the corresponding abnormal categories, influence ranges, and confidence levels.

[0051] Optionally, through collaborative modeling of a graph neural network and a large language model, a joint feature vector representation that fuses context semantics and structured knowledge is generated. Through multi-dimensional clustering analysis and semantic similarity calculation, historical fault cases and causal chains highly associated with the target device state are screened out, so as to achieve precise matching and root cause analysis of the current abnormal state, providing data support and logical basis for subsequent fault prediction and strategy recommendation.

[0052] First, traverse the constructed historical device fault dataset, and extract the key monitoring dimensions and their evolution characteristics associated with known fault events. Parse the device operation data collected in real time item by item, and perform matching analysis on each record with the triggering conditions in the historical fault causal chain to identify possible abnormal evolution paths. This process is based on a unified data structure and semantic tags to ensure the consistency of real-time data and historical knowledge. Extract key fault monitoring parameters highly related to typical fault events, including but not limited to: Vibration amplitude: used to identify problems such as abnormal wear, imbalance, or looseness of mechanical components; Current fluctuation: reflects load fluctuations in the drive system, motor overload, or abnormal control circuits; Temperature rise gradient: used to monitor the thermal state changes of the device and identify poor heat dissipation or abnormal heating phenomena; Rotational speed deviation: used to judge problems such as load anomalies, mechanical jams, or instability of rotating equipment; Voltage anomaly: identify electrical problems such as unstable power input and faults in the electrical control unit; Air leakage volume: used to evaluate the sealing state of the pneumatic system and the aging degree of the pipeline. The parameters are synchronously collected by multiple sensor channels, and after preprocessing steps such as data synchronization, denoising, and time alignment, they are uniformly mapped into the device state analysis model, providing core inputs for subsequent fault trend prediction, causal chain positioning, and intelligent diagnosis. In this way, it realizes the modeling of fault-sensitive parameters driven by historical data and the effective extraction of potential fault signals in real-time data, enhancing the ability to identify early fault signs under complex operating conditions.

[0053] Next, based on a predefined library of fault diagnosis parameter mapping relationships, key fault monitoring parameters extracted from the analysis of historical fault data and real-time data are processed. Each key monitoring parameter is associated with one or more specific sensor channels, and these channels have unique identifiers in the operation and maintenance system of subway equipment. Through the equipment number and the monitoring parameter field index, the sensor channels corresponding to each key fault monitoring parameter are located. For example: vibration amplitude Acceleration sensor channel; current fluctuation Current transformer channel; temperature rise gradient Thermistor channel; air leakage volume Airflow pressure / flow sensor channel, etc. After the channel positioning is completed, the system schedules the data acquisition module to query the status of the sensors and collect real-time data. During the acquisition process, the system records and generates a set of sensing status data sets corresponding one-to-one to the key monitoring parameters. Each record in this data set includes the following content: Sensing value: The physical quantity measurement value output by the current sensor, used to characterize the current status of the monitoring index; Timestamp: The acquisition time, used to support time series analysis and fault trend backtracking; Signal quality: Indicates quality indicators such as whether the data is affected by noise interference and whether the signal amplitude is abnormal; Communication status: Reflects whether the data transmission process is complete, such as whether there are communication packet losses, delays, etc.; Sensor operating status: Records whether the sensor itself is operating normally, including calibration status, power status, working mode, etc.

[0054] Furthermore, an abnormal identification processing process is executed for the formed sensing status data set to ensure the accuracy and reliability of the operation data of subway equipment. Abnormal identification is based on a multi-source fusion data quality control mechanism, and a multi-dimensional detection mechanism is used to verify each piece of status data of each sensing channel one by one to determine whether there are data drift (Data Drift) detection, missing data (Missing Data) identification, signal distortion (Signal Distortion) detection, and state anomaly (StateAnomaly) identification. Data drift (Data Drift) detection compares the sliding window with a reference threshold to analyze the mean / variance change trend of the sensing value in the short term or long term. If the current value continuously deviates from the historical stable range and cannot be explained by environmental change factors, it is determined as a data drift anomaly. For example, a certain sensor outputs a continuously increasing vibration amplitude under the no-load state of the equipment.

[0055] Next, for the sensing state anomalies identified by the multi-dimensional detection mechanism, the system executes the processing flow of the anomaly state annotation module, performs structured annotation on each abnormal data item, and outputs standardized anomaly information items. Based on the anomaly detection results and preset classification rules, each abnormal data is labeled with its corresponding category, mainly including but not limited to: data drift (Drift), data missing (Missing), signal distortion (Distortion), and state anomaly (FaultyState). Each type of anomaly corresponds to specific detection dimensions and metrics. For example, drift corresponds to numerical trend variation, and distortion corresponds to abnormal frequency spectrum, etc.

[0056] Finally, each type of anomaly corresponds to specific detection dimensions and metrics. For example, drift corresponds to numerical trend variation, and distortion corresponds to abnormal frequency spectrum, etc. Combining the device component structure level where the sensor is located, the correlation degree of monitoring parameters, and the device topology information, automatically evaluate the operating objects and scope that may be affected by the anomaly. For example, if the vibration sensor anomaly is located at the traction motor bearing position, the affected scope is "traction system sub-component level"; if multiple adjacent parameters are abnormal, the system can be labeled as "device system-level impact", the affected component: traction motor bearing; the affected scope level: component level / system level. Based on the confidence output value of the anomaly detection algorithm (such as SVM probability output, neural network softmax result, or empirical threshold judgment), calculate the anomaly confidence corresponding to each abnormal data, and perform level division, usually including: high (≥0.85), medium (0.6 - 0.85), low (<0.6), and at the same time retain the original confidence score (such as 0.923) to support subsequent confidence-aware decision-making processing. Finally, the system stores the annotation information of the anomaly state as a structured record item. Each record includes at least the following fields: anomaly timestamp, sensor channel ID, anomaly category, affected component / scope, anomaly confidence level and value, and anomaly brief description, such as "continuous upward deviation exceeds 5% threshold". The labeled abnormal data items will be used as the key inputs for subsequent device fault diagnosis, early warning trigger, and root cause analysis.

[0057] P600: Based on the anomaly recognition result, call the pre-constructed sensing distortion correction algorithm to generate fault monitoring correction parameters and perform parameter correction, and add the fault monitoring correction parameters to the fault state data to form an enhanced multi-dimensional monitoring data set. The fault state data is a set of data records collected by the monitoring system and associated with fault events during the operation of the device.

[0058] Furthermore, step P600 of the embodiment of the present application further includes: P601: According to the abnormal types marked in the abnormal recognition result, input the sensing distortion correction algorithm. The sensing distortion correction algorithm is constructed based on a multi-model integration mechanism driven by abnormal types. For different types of abnormal states, the corresponding correction sub-algorithms are called respectively to obtain the fault monitoring correction parameters; P602: Combine the reconstruction residual of the abnormal feature vector, the historical error distribution curve, the real-time working condition context information, and the prediction confidence evaluation result of the correction sub-algorithm to perform model dynamic selection and correction parameter output.

[0059] Specifically, for the sensing state abnormal data output by the abnormal recognition module, the system automatically calls the pre-constructed sensing distortion correction algorithm to implement dynamic correction processing of the fault monitoring parameters, improving the accuracy and availability of subsequent data analysis.

[0060] First, according to the abnormal annotation result output by the abnormal recognition module, parse the abnormal type field in each sensing state data to identify specific abnormal categories such as data drift, signal distortion, data missing, or state abnormality existing in the current data. Based on the above-identified abnormal types, the system executes model matching logic in the built-in correction model scheduler. The scheduler activates the following sub-algorithm modules as needed according to the "abnormal type → correction sub-model" mapping rule: Data drift: Call the trend regression fitting and residual compensation model; Signal distortion: Call the denoising and restoration model based on wavelet multi-scale analysis; Data missing: Call the time series interpolation and prediction reconstruction model; State abnormality: Call the state envelope extraction and correction fitting model. The called correction sub-algorithm processes the original monitoring data of the corresponding sensing channel and outputs the fault monitoring correction parameters. The correction parameters include but are not limited to: numerical offset correction value (such as: +0.85), filtering repair factor (such as: wavelet threshold, order parameter), missing completion value and confidence score, abnormal point repair flag and substitution value. The generated correction parameters are output in a structured format and bound to the corresponding fields of the original monitoring data to form a traceable and interpretable correction result record, supporting subsequent data enhancement, modeling training, and visualization analysis.

[0061] Next, to achieve more accurate and adaptable correction processing for sensing abnormal states, during the execution of the multi-model integration mechanism, the system further introduces a dynamic model selection mechanism. This mechanism comprehensively considers evaluation indicators from multiple dimensions to dynamically judge and accurately call the optimal correction sub-algorithm, thereby outputting correction parameters most suitable for the current fault scenario. For the input sensing abnormal feature vector, first, perform reconstruction processing through each candidate correction sub-algorithm respectively, and calculate the reconstruction residual value between it and the original abnormal vector, which is used to measure the recovery degree after correction. The smaller the reconstruction residual, the stronger the fitting and restoration ability of the model for the current abnormal type. Retrieve historical records similar to the current device, sensing channel, and abnormal type from the historical fault database, construct a corresponding error distribution reference curve, and compare it with the error distribution of the current model processing result to calculate the matching degree score, which is used as an auxiliary judgment basis for model adaptability. Obtain the real-time operating condition context information of the current device (including operating mode, load level, environmental temperature, vibration frequency, etc.), and perform matching analysis with the operating conditions supported by the correction sub-algorithm. Only retain the models with a compatibility degree higher than the set threshold as the candidate model set. For the correction results of each candidate sub-algorithm, call its internal confidence prediction module, and calculate the predicted confidence score based on elements such as the residual distribution form, feature coverage rate, and model training weight stability, which is used to comprehensively evaluate the reliability and acceptability of the correction results. Weight and fuse the above-mentioned indicators from each dimension to form a comprehensive score result, perform sorting and screening, and dynamically select the sub-algorithm with the highest comprehensive score as the execution model for this correction task, and output the corresponding fault monitoring correction parameters, including: correction deviation value, time period weight adjustment factor, abnormal point mask annotation, signal restoration curve parameters, etc.

[0062] Further, the sensing distortion correction algorithm described in step P601 of the embodiment of the present application further includes: P601-1: Obtain the historical monitoring data and original sensor sampling values of multiple subway devices under normal operation and fault conditions; P601-2: Perform tagging processing on the abnormal situations in the historical monitoring data to obtain an abnormal label data set, and the abnormal label data set includes identification data of data drift, data missing, signal distortion, and state abnormality; P601-3: Preset the abnormal label data set as L, L = , where is normal, is data drift, is data missing, is signal distortion, is state abnormality; P601-4: Obtain the original sensor value at time t; Calculate the historical mean as , and the formula is: = × ; Among them, is the historical mean value, is the historical sample quantity; When the anomaly label is , the calculation formula for the fault monitoring correction parameter is: = - × ; Among them, is the fault monitoring correction parameter, is the original sensor value at time t, is the drift correction coefficient, , initialize =0.6, is the historical mean value; P601-5: When the anomaly label is , the calculation formula for the fault monitoring correction parameter is: = × + × ; Among them, is the fault monitoring correction parameter, is the interpolation balance coefficient, , initialize =0.5, is the original sensor value at the moment, is the original sensor value at the moment; P601-6: When the anomaly label is , the calculation formula for the fault monitoring correction parameter is: = × - × Δ ; Among them, Δ is the local deviation, Δ = - × , is the signal suppression coefficient, , initialize =0.3; P601-7: When the abnormal label is , the calculation formula for the fault monitoring correction parameter is: = × × ; Wherein, is the fault monitoring correction parameter, is the smoothing parameter, , = 0.7; P601-8: When the abnormal label is , the calculation formula for the fault monitoring correction parameter is: = ; Wherein, is the fault monitoring correction parameter, is the original sensor value at time t; Output the fault monitoring correction parameter ; P601-9: Judge if , ,... , >= 3, it is determined as abnormal, then mark it as a continuous abnormal section; Set the sliding window as W, W = ; Use the local polynomial regression method to reconstruct the data of the continuous abnormal section, and replace the original abnormal section with the value sequence generated by the whole sliding window.

[0063] First, an exception type-driven multi-model integration mechanism is adopted, which has the capabilities of adaptive selection and precise correction. It mainly includes steps such as anomaly recognition, sub-algorithm selection, correction parameter generation, and correction execution. Receive the annotation results output by the anomaly recognition component, and identify the anomaly type corresponding to the current sensing state data, including but not limited to: data drift, data missing, signal distortion, and status anomaly. For each anomaly type, multiple dedicated correction sub-algorithms are pre-designed and deployed, including but not limited to: for data drift, trend line regression (such as sliding linear regression) and exponential smoothing reconstruction algorithms are used; for data missing, time series interpolation (such as spline interpolation, Kalman filter) and Transformer prediction models are used for reconstruction; for signal distortion, wavelet transform reconstruction and band-pass filter restoration algorithms are used; for status anomaly, an AutoEncoder reconstruction network or a clustering model (such as DBSCAN) is used to eliminate and fill anomalies. According to the currently identified anomaly type, call the corresponding candidate sub-algorithm set. Use multiple evaluation indicators such as reconstruction residuals, historical matching degrees, and confidence scores to perform weighted scoring on the output results of the sub-algorithms, and select the optimal model or perform multi-model weighted fusion.

[0064] Next, the correction process includes inputting anomaly annotation results, determining candidate sub-algorithms, performing reconstruction prediction, evaluating correction quality, and outputting a set of correction parameters. Among them, inputting anomaly annotation results: includes anomaly type, time period, anomaly level, and influence range; determining candidate sub-algorithms is to screen available models in the sub-algorithm library according to the anomaly type; performing reconstruction prediction is to use the candidate model to reconstruct, fit, or fill the sensing data of the anomaly segment to generate preliminary correction values; evaluating correction quality includes calculating the deviation between the reconstruction residuals and the original data, comparing the historical error distribution curve and evaluating the correction effect, and comprehensively using the confidence scoring mechanism to judge the correction credibility; outputting a set of correction parameters includes offset correction factors, missing interpolation values, spectral correction curves, anomaly point masks, etc.; data update and storage is to replace the original anomaly data with the correction results and append identification fields to form an enhanced sensing state data set. By deploying this sensing distortion correction algorithm, real-time repair and enhancement processing of equipment fault monitoring data can be realized, providing a high-quality and low-noise input basis for subsequent fault identification, semantic modeling, and intelligent diagnosis.

[0065] Furthermore, the abnormal situations that occur in the historical monitoring data are labeled to obtain an abnormal label data set, and the abnormal label data set includes identification data of data drift, data missing, signal distortion, and status anomaly.

[0066] Judge if , ,... , >= 3, it is determined as an anomaly and marked as a continuous anomaly section; Set the sliding window as W, W = ; Use the local polynomial regression method to reconstruct the data of the continuous anomaly section, and replace the original anomaly section with the value sequence generated by the sliding window as a whole.

[0067] P700: Input the multi-dimensional monitoring data in the multi-dimensional monitoring dataset into the diagnostic engine driven by the large language model. Through natural language Q&A and multi-round reasoning, output the most likely fault type, cause analysis, recommended handling strategies, and fault identification reports.

[0068] Furthermore, step P700 of the embodiment of the present application further includes: P701: Input the multi-dimensional monitoring dataset that combines sensor status information, fault monitoring parameters, and sensing correction parameters into the diagnostic engine driven by the large language model; P702: The diagnostic engine, based on the pre-trained large language model, performs semantic parsing and intention understanding on the multi-dimensional monitoring data through the natural language Q&A mechanism, and extracts the feature patterns, context information, and potential semantic logics associated with faults; P703: Utilize the context retention ability and long text understanding ability of the large language model, combine the multi-dimensional device fault knowledge graph, fault causal chain, and multi-dimensional monitoring dataset, and execute a multi-round causal reasoning and fault identification Q&A process to form a dynamic reasoning path; P704: Based on the results of the multi-round reasoning, output a structured fault diagnosis result, and the structured fault diagnosis result includes the most likely fault type of the target device, the cause analysis of the corresponding fault, the executable recommended handling strategies, and a fault identification report in the form of natural language for operation and maintenance personnel.

[0069] Optionally, through semantic understanding and context reasoning modeling of multi-dimensional monitoring data, ensure that the accuracy and interpretability of the fault diagnosis results meet the high-reliability operation and maintenance requirements of subway equipment, so as to achieve accurate identification and efficient disposal of faults. When there are abnormal drifts or context ambiguities in the input data, perform semantic error correction and confidence calibration processing to ensure the stability and credibility of the diagnosis results in abnormal scenarios.

[0070] First, input the multi-dimensional monitoring data set that integrates sensor status information, fault monitoring parameters, and sensing calibration parameters into the intelligent diagnosis engine driven by a large language model. The multi-dimensional monitoring data set includes, but is not limited to: real-time sensor acquisition values, timestamps, communication status, signal quality indicators, monitoring parameter items (such as vibration amplitude, current fluctuation, temperature rise gradient, etc.), and calibration parameters generated based on anomaly recognition results. The intelligent diagnosis engine is constructed based on a large language model that is pre-trained and fine-tuned with subway fault corpora, and has the capabilities of natural language understanding and multi-modal information fusion. After receiving the above multi-dimensional data, the diagnosis engine first performs semantic parsing and context modeling, structurally interprets the input information through the large language model, and performs semantic alignment on the key data fields. Subsequently, the diagnosis engine combines the device historical fault knowledge graph and context environment information, performs natural language question answering, multi-round reasoning, and causal chain deduction, outputs the most likely fault type, cause analysis, impact scope, and recommended handling strategies under the current device state, and automatically generates a fault recognition report, which includes diagnosis conclusions, reference bases, confidence scores, and recommended measures.

[0071] This process realizes the intelligent interpretation and semantic-enhanced decision-making of monitoring data, effectively improving the accuracy, interpretability, and response speed of device fault recognition.

[0072] Next, the diagnosis engine is constructed based on a pre-trained large language model and has the capabilities of semantic understanding and question-answering reasoning for multi-modal input data. After receiving the multi-dimensional monitoring data that integrates sensor status information, fault monitoring parameters, and calibration parameters, the diagnosis engine calls the built-in natural language question-answering mechanism to perform item-by-item semantic parsing and intent recognition on the data. Specifically, the natural language question-answering mechanism first performs semantic structure modeling on the input data, extracts keywords, hyponyms, and associated semantics in the text or structured description; subsequently, combined with the fault instances annotated in the historical corpus, it realizes the understanding of the fault indication meaning of key parameters (such as abnormal vibration, current fluctuation, temperature rise gradient, etc.) in a specific context. Through the above processing, the diagnosis engine can automatically identify the feature patterns, operating background context information, and potential semantic logic chains related to the current fault (such as "unstable operating voltage + abnormal communication status → signal interference → device failure"), providing a semantic foundation for subsequent fault type determination and handling strategy recommendation.

[0073] Then, use a large language model (LLM) that is pre-trained and fine-tuned for device fault corpora, and combine its context retention ability and long text understanding ability to perform in-depth semantic modeling and reasoning processing on the multi-source fused device fault information. In the processing flow, the multi-dimensional device fault knowledge graph, the historical fault causal chain constructed, and the enhanced multi-dimensional monitoring data set are jointly used as inputs and input into the fault recognition module driven by the large language model.

[0074] Finally, based on the output results of the multi-round causal reasoning process driven by the large language model, a structured fault diagnosis result for the device intelligent diagnosis task is constructed. Based on the fused multi-dimensional monitoring data, the device fault knowledge graph, and the historical causal chain, the diagnostic engine synthesizes the semantic aggregation results in the reasoning path to determine the most likely fault state type of the current target device and outputs a clear fault classification label (such as: drive motor overheating, sensor malfunction, control module communication anomaly, etc.). Combining the semantic relationship of entity nodes extracted by the graph neural network, the dynamic semantic chain constructed by the large language model, and the context information, the content of the fault cause analysis is output, clearly pointing out the initial cause, the intermediate propagation path, and the key impact nodes, forming a complete causal chain description of "fault cause - propagation process - manifestation result". According to the identified fault type and cause path, historical processing cases and expert rules are matched in the knowledge graph, and the executable processing strategies in the current fault state are output, covering emergency response measures, component replacement suggestions, operating parameter adjustment plans, etc. For operation and maintenance technicians, a structured semantic-enhanced natural language fault identification report is generated. The report content includes device name, identification time, fault type, cause path, monitoring index summary, processing suggestions, etc., with the characteristics of strong readability and clear context logic, supporting system archiving and manual review. The structured fault diagnosis result can be called by the automated decision-making module and also displayed to on-site operation and maintenance personnel through the human-machine interaction interface, realizing the closed-loop support of intelligent auxiliary diagnosis and disposal.

[0075] In summary, the embodiments of the present application at least have the following technical effects: The present application determines the key monitoring parameters and semantic feature patterns related to faults by fusing multi-source device operation data, fault text descriptions, and sensor status information, and constructs a multi-dimensional knowledge graph and a structured historical fault data set for subway device intelligent diagnosis.

[0076] Through the collaborative modeling mechanism of the graph neural network and the large language model, the identification of the fault causal chain and multi-round semantic reasoning are further realized, and finally a structured fault diagnosis result with context understanding ability and executable processing strategies are output. Combining the sensing anomaly correction mechanism, the semantic enhanced reasoning of the knowledge graph, and the fault feature vector aggregation method, it is ensured that the diagnostic engine has high robustness and interpretability for fault types under complex working conditions, thereby significantly improving the accuracy and intelligent level of subway device fault identification. It achieves accurate, fast, and interpretable device fault diagnosis, realizes a comprehensive automated process from multi-dimensional monitoring data to fault diagnosis, and effectively improves the accuracy, reliability, and processing efficiency of subway device fault identification.

[0077] Embodiment 2 is based on the same inventive concept as the subway device fault intelligent diagnosis method assisted by the large language model in the foregoing embodiment, as Figure 2As shown in the figure, the present application provides an intelligent diagnosis system for subway equipment failures assisted by a large language model. The system and method embodiments in the embodiments of the present application are based on the same inventive concept. Among them, the system includes: An operation data acquisition module 10, which is used to acquire the operation data of subway equipment, including sensor data, fault history records, and operation environment data of the equipment. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; A device fault data extraction module 20, which is used to preprocess the text of the fault history records, perform named entity recognition, semantic analysis, and relationship extraction using a large language model, and extract device fault data, including device type, fault phenomenon, time node, treatment measures, and operation environment; A multi-dimensional device fault knowledge graph construction module 30, which is used to systematically organize the extracted device fault data, construct a multi-dimensional device fault knowledge graph, define entity nodes and semantic relationship edges of the entity nodes, and form a structured knowledge base. The entity nodes include fault types, causes, and measures; A historical fault data set acquisition module 40, which is used to perform deep learning and semantic modeling on device fault data based on the multi-dimensional device fault knowledge graph, using a large language model combined with a graph neural network, to obtain a historical fault data set. The historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models; A sensing state data set anomaly recognition module 50, which is used to traverse the historical fault data set, extract key fault monitoring parameters related to faults, collect sensor state data corresponding to the parameters, obtain a sensing state data set, and perform anomaly recognition on the sensing state data set to determine whether there are data drifts, data missing, signal distortion, and state anomalies in the sensor state data; A multi-dimensional monitoring data set formation module 60, which is used to generate fault monitoring correction parameters and perform parameter correction by calling a pre-constructed sensing distortion correction algorithm based on the anomaly recognition result, and add the fault monitoring correction parameters to the fault state data to form an enhanced multi-dimensional monitoring data set. The fault state data is a set of data records collected by the monitoring system and associated with fault events during the operation of the equipment; A fault identification module 70, which is used to input the multi-dimensional monitoring data in the multi-dimensional monitoring data set into a diagnosis engine driven by a large language model, and output the most likely fault type, cause analysis, recommended treatment strategy, and a fault identification report through natural language Q&A and multi-round reasoning.

[0078] It should be noted that the above order of the embodiments of the present application is only for description and does not represent the superiority or inferiority of the embodiments. In addition, the above description of specific embodiments of this specification has been made. Moreover, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0079] The above are only the preferred embodiments of the present application and are not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included within the protection scope of the present application.

[0080] This specification and the drawings are only exemplary descriptions of the present application and are considered to have covered any and all modifications, variations, combinations, or equivalents within the scope of the present application. Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the present application and its equivalent technologies, the present application is intended to include these changes and modifications.

Claims

1. An intelligent diagnosis method for subway equipment faults assisted by large language models, characterized in that, The method includes: P100. Obtain the operation data of subway equipment, including sensor data of the equipment, fault history records, and operation environment data. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; P200. Preprocess the text of the fault history records, use a large language model to perform named entity recognition, semantic analysis, and relationship extraction, and extract key information. The key information includes equipment type, fault phenomenon, time node, handling measures, and operation environment; P300. Systematically organize the extracted key information, construct a multi-dimensional equipment fault knowledge graph, define entity nodes and semantic relationship edges of the entity nodes, and form a structured knowledge base. The entity nodes include fault type nodes, fault cause nodes, handling measures nodes, equipment name nodes, component location nodes, and environmental condition nodes; P400. Based on the multi-dimensional equipment fault knowledge graph, use a large language model combined with a graph neural network to perform deep learning and semantic modeling on equipment fault data, and obtain a historical fault data set. The historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models; P500. Traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect the sensor status data corresponding to the parameters, obtain a sensing status data set, and perform anomaly recognition on the sensing status data set to determine whether there are data drifts, data missing, signal distortion, and status anomalies in the sensor status data; P600. Based on the anomaly recognition result, call a pre-constructed sensing distortion correction algorithm, generate fault monitoring correction parameters and perform parameter correction, and add the fault monitoring correction parameters to the fault status data to form an enhanced multi-dimensional monitoring data set. The fault status data is a set of data records collected by the monitoring system and associated with fault events during the operation of the equipment; P700. Input the multi-dimensional monitoring data in the multi-dimensional monitoring data set into a diagnostic engine driven by a large language model, and output the most likely fault type, cause analysis, recommended handling strategy, and fault identification report through natural language question answering and multi-round reasoning.

2. The intelligent fault diagnosis method for subway equipment assisted by large language models according to claim 1, wherein, Preprocess the text of the fault history records, use a large language model to perform named entity recognition, semantic analysis, and relationship extraction, and extract key information. The key information includes equipment type, fault phenomenon, time node, handling measures, and operation environment, including: Perform standardized preprocessing on the text of historical fault records collected from the subway operation and maintenance system. The preprocessing includes character cleaning, regularization of proper nouns, format unification, semantic error correction, and segment annotation; Input the preprocessed text into a fine-tuned large language model and perform a multi-layer semantic processing process; Through semantic recognition and relationship modeling, extract and output a structured set of key information. The key information includes equipment type, fault phenomenon, time node, handling measures, and operation environment; Assign semantic confidence scores to each type of key information extracted, and label the information with low confidence or ambiguous expressions.

3. The intelligent fault diagnosis method for subway equipment assisted by a large language model according to claim 1, characterized in that, Systematically organize the extracted key information, construct a multi-dimensional equipment fault knowledge graph, define entity nodes and semantic relationship edges of the entity nodes, and form a structured knowledge base. The entity nodes include fault type nodes, fault cause nodes, handling measure nodes, equipment name nodes, component location nodes, and environmental condition nodes, including: Archive the key information extracted by the large language model in a unified data structure, including a fault type field, a fault cause field, a handling measure field, an equipment name field, a component location field, and an environmental condition parameter field; According to the archived information, define the core entity node types in the field of equipment faults. The node types include fault type nodes, fault cause nodes, handling measure nodes, equipment name nodes, component location nodes, and environmental condition nodes; Based on the semantic analysis and context reasoning results, extract the relationship edges between entity nodes and label the semantic types of each edge; Based on the above entity nodes and relationship edges, construct a multi-dimensional knowledge graph for subway equipment fault diagnosis tasks. The multi-dimensional knowledge graph is stored in a graph structure and can be regarded as a set of structured triples, where the triple is entity 1 - relationship - entity 2.

4. The intelligent diagnosis method for subway equipment faults assisted by a large language model according to claim 1, wherein, Based on the multi-dimensional equipment fault knowledge graph, use the large language model combined with the graph neural network to perform deep learning and semantic modeling on equipment fault data to obtain a historical fault data set. The historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models, including: Based on unstructured and structured data such as fault description texts, maintenance records, and sensor logs, use natural language processing technologies such as named entity recognition NER and semantic relationship extraction to extract information and semantic relationships such as equipment types, fault phenomena, time nodes, handling measures, and operating environments, and construct a knowledge graph containing a multi-level semantic structure of equipment - phenomenon - cause - measure; Input the knowledge graph with the multi-level semantic structure into the graph neural network, and obtain the low-dimensional semantic vector representations of entity nodes and relationship edges through node feature aggregation and adjacency relationship modeling, retaining their context semantics and graph structure information; Use the large language model to perform deep semantic modeling and language understanding on the low-dimensional semantic vector representations of entity nodes and relationship edges to generate context-enhanced text vectors. The large language model is a Transformer structure model fine-tuned for fault corpora; Fuse the low-dimensional semantic vector representations of entity nodes and relationship edges extracted by the graph neural network with the context-enhanced text vectors extracted by the large language model to form a unified joint feature vector space; Through clustering analysis, fault classification, and causal chain identification processing of the joint feature vector space, perform feature alignment and normalization modeling on the operating states and fault manifestations of multiple devices to form a structured historical fault data set.

5. The intelligent diagnosis method for subway equipment faults assisted by a large language model according to claim 4, characterized in that The historical fault data set includes: The historical fault data set includes multi-dimensional feature representations, fault causal chains, and semantic enhanced inference models; The multi-dimensional feature representation is a joint representation vector constructed based on dimensions such as device type, key operating parameters, and working condition environment characteristics, and is used to reflect the fault manifestation characteristics under different device states; The fault causal chain is a device fault chain jointly mined by a graph neural network and a large language model, obtaining the semantic causal relationship of initial cause - propagation path - final fault; The semantic enhanced reasoning model is the K-BERT architecture that integrates a pre-trained language model and a knowledge graph, obtaining a multi-dimensional semantic feature representation with the capabilities of fault causal recognition, semantic ambiguity resolution, and cross-path reasoning; 6. The intelligent fault diagnosis method for subway equipment assisted by a large language model according to claim 1, characterized in that, Traverse the device fault data, extract fault monitoring parameters, collect the sensor status information corresponding to the parameters, obtain the sensing status data, and perform anomaly identification to determine whether there are data drift, data missing, signal distortion, and status anomalies in the status data, including: Traverse the device fault data set, parse the device operation data collected in real time item by item, and extract the key fault monitoring parameters associated with the fault event. The key fault monitoring parameters include but are not limited to index parameters such as vibration amplitude, current fluctuation, temperature rise gradient, rotational speed deviation, voltage anomaly, and air leakage volume; Extract the key fault monitoring parameters, locate the sensor channels corresponding to the key fault monitoring parameters, collect the current status data of the sensors, and form a sensing status data set corresponding one-to-one with the monitoring parameters. The sensing status data includes sensing values, timestamps, signal quality, communication status, and sensor operating status; Perform anomaly identification processing on the sensing status data, and use a multi-dimensional detection mechanism to determine whether there are anomalies. The anomalies include data drift, data missing, signal distortion, and status anomalies; Label the detected abnormal status and output the corresponding abnormal category, influence range, and confidence level; 7. The intelligent diagnosis method for subway equipment faults assisted by a large language model according to claim 1, wherein, Based on the anomaly identification result, call the pre-constructed sensing distortion correction algorithm, generate the fault monitoring correction parameters and execute parameter correction, and add the fault monitoring correction parameters to the fault status data to form an enhanced multi-dimensional monitoring data set. The fault status data is a set of data records collected by the monitoring system and associated with the fault event during the operation of the device, including: According to the abnormal type labeled in the anomaly identification result, input the sensing distortion correction algorithm. The sensing distortion correction algorithm is constructed based on a multi-model integration mechanism driven by the abnormal type. For different types of abnormal statuses, call the corresponding correction sub-algorithms respectively to obtain the fault monitoring correction parameters; Combine the reconstruction residual of the abnormal feature vector, the historical error distribution curve, the real-time working condition context information, and the prediction confidence evaluation result of the correction sub-algorithm to perform model dynamic selection and correction parameter output; 8. The intelligent fault diagnosis method for subway equipment assisted by large language model according to claim 7, characterized in that, The sensing distortion correction algorithm includes: Obtain the historical monitoring data and sensor original sampling values of multiple subway devices under normal operation and fault conditions; Perform labeling processing on the abnormal situations in the historical monitoring data to obtain an abnormal label data set. The abnormal label data set includes identification data of data drift, data missing, signal distortion, and status anomalies; The preset abnormal exception label data set is L, L = , where is normal, is data drift, is data missing, is signal distortion, is status abnormal; Obtain the original sensor value at time t ; Calculate the historical mean as , and the formula is: = × ； Among them, is the historical average value, is the historical sample quantity; When the abnormal label is , the calculation formula for the fault monitoring and correction parameter is: = - × ； Among them, is the fault monitoring and correction parameter, is the original sensor value at time t, is the drift correction coefficient, , initialize = 0.6, is the historical mean; When the abnormal label is , the calculation formula for the fault monitoring and correction parameter is: = × + × ； Among them, is the fault monitoring and correction parameter, is the interpolation balance coefficient, , initialize = 0.5, is the original sensor value at the moment, is the original sensor value at the moment; When the abnormal label is , the calculation formula for the fault monitoring and correction parameter is: = × - × Δ ; Among them, Δ is the local deviation, Δ = - × , is the signal suppression coefficient, During initialization, = 0.3; When the abnormal label is , the calculation formula for the fault monitoring and correction parameter is: = × × ； Among them, is the fault monitoring and correction parameter, is the smoothing parameter, , = 0.7; When the abnormal label is , the calculation formula for the fault monitoring and correction parameter is: = ； Among them, is the fault monitoring and correction parameter, is the original sensor value at time t; Output fault monitoring and calibration parameters ; Determine if , ,... , >= 3, it is determined as abnormal, and is marked as a continuous abnormal section; Set the sliding window as W, W = ; Reconstruct the data of the continuous abnormal section using the local polynomial regression method, and replace the original abnormal section with the value sequence generated by the sliding window as a whole.

9. The intelligent fault diagnosis method for subway equipment assisted by a large language model according to claim 1, characterized in that Input the multi-dimensional monitoring data in the multi-dimensional monitoring dataset into the large language model-driven diagnostic engine, and through natural language Q&A and multi-round reasoning, output the most likely fault type, cause analysis, recommended handling strategies, and fault identification reports, including: Input the multi-dimensional monitoring dataset that integrates sensor status information, fault monitoring parameters, and sensing correction parameters into the large language model-driven diagnostic engine; Based on the pre-trained large language model, the diagnostic engine performs semantic parsing and intention understanding on the multi-dimensional monitoring data through the natural language Q&A mechanism, and extracts the feature patterns, context information, and potential semantic logics related to faults; Utilize the context preservation ability and long text understanding ability of the large language model, combine with the multi-dimensional equipment fault knowledge graph, fault causal chain, and multi-dimensional monitoring dataset, and execute a multi-round causal reasoning and fault identification Q&A process to form a dynamic reasoning path; Based on the results of the multi-round reasoning, output a structured fault diagnosis result, and the structured fault diagnosis result includes the most likely fault type of the target device, cause analysis of the corresponding fault, executable recommended handling strategies, and a fault identification report in natural language form for operation and maintenance personnel.

10. An intelligent fault diagnosis system for subway equipment assisted by large language models, characterized in that, The system includes: An operating data acquisition module for subway equipment, which is used to acquire the operating data of subway equipment, including sensor data, fault history records, and operating environment data of the equipment. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; An equipment fault data extraction module, which is used to preprocess the text of the fault history records, use the large language model to perform named entity recognition, semantic analysis, and relationship extraction, and extract equipment fault data, including equipment type, fault phenomenon, time node, handling measures, and operating environment; A multi-dimensional equipment fault knowledge graph construction module, which is used to systematically organize the extracted equipment fault data, construct a multi-dimensional equipment fault knowledge graph, define entity nodes and semantic relationship edges of the entity nodes, and form a structured knowledge base. The entity nodes include fault types, causes, and measures; A historical fault dataset acquisition module, which is used to perform deep learning and semantic modeling on equipment fault data based on the multi-dimensional equipment fault knowledge graph, using the large language model combined with the graph neural network, to obtain a historical fault dataset, and the historical fault dataset includes multi-dimensional feature representations, fault causal chains, and semantic enhanced reasoning models; A sensing status dataset anomaly identification module, which is used to traverse the historical fault dataset, extract key fault monitoring parameters related to faults, collect the sensor status data corresponding to the parameters, obtain the sensing status dataset, and perform anomaly identification on the sensing status dataset to determine whether there are data drifts, data missing, signal distortions, and status anomalies in the sensor status data; Multidimensional Monitoring Dataset Formation Module, which is used to generate fault monitoring correction parameters and perform parameter correction by calling a pre-built sensing distortion correction algorithm based on the anomaly recognition results, and add the fault monitoring correction parameters to the fault status data to form an enhanced multidimensional monitoring dataset. The fault status data is a set of data records collected by the monitoring system and associated with fault events during the operation of the device; Fault Identification Module, which is used to input the multidimensional monitoring data in the multidimensional monitoring dataset into a large language model-driven diagnostic engine, and output the most likely fault type, cause analysis, recommended handling strategies and a fault identification report through natural language Q&A and multi-round reasoning.

Images