Large language model-assisted intelligent diagnosis method and system for subway equipment faults

Abstract

The present application provides a method and system for intelligent diagnosis of subway equipment faults assisted by a large language model, which relates to the field of data processing technology, including: using a large language model to extract key information, construct a multi-dimensional equipment fault knowledge graph, obtain a historical fault data set, extract key fault monitoring parameters related to the fault, obtain a sensor state data set, and determine whether the sensor state data has an anomaly; call a pre-built sensor distortion correction algorithm, generate fault monitoring correction parameters and perform parameter correction, input the multi-dimensional monitoring data into a large language model-driven diagnosis engine, and output the most likely fault type, cause analysis, recommended processing strategy, and fault identification report. The present application solves the technical problem in the prior art that the fault diagnosis has a low degree of intelligence and is difficult to achieve accurate identification and causal analysis due to the lack of the ability to fuse unstructured fault text with structured monitoring data, thereby improving the fault identification response speed and accuracy.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a method and system for intelligent diagnosis of subway equipment faults assisted by a large language model. Background Art

[0002] As a vital component of urban rail transit, subway systems feature complex equipment systems (such as signaling, power supply, train control, and environmental monitoring), face highly variable operating conditions, and place extremely high demands on safety and stability. Over long-term operation, various equipment failures, such as sensor failure, signal anomalies, and unstable power supply, are inevitable. Traditional fault diagnosis methods, which rely heavily on manual experience or rule-based diagnostic systems, struggle to efficiently process large-scale, multi-source, and heterogeneous monitoring data, and also struggle to accurately identify hidden or multi-factor coupled faults under complex operating conditions.

[0003] Existing subway equipment fault diagnosis technology primarily relies on manual experience and rule-based knowledge base systems. The diagnostic process typically involves manually comparing equipment alarm information with historical fault cases, making it difficult to implement deep semantic modeling of multi-source, heterogeneous data, limiting the accuracy and real-time nature of fault identification. In subway systems, where complex operating conditions and equipment status are closely coupled, issues such as sensor accuracy drift and missing data lead to significant anomalies in the raw monitoring data, compromising the accuracy of subsequent diagnostic reasoning. Summary of the Invention

[0004] This application provides a method and system for intelligent diagnosis of subway equipment faults assisted by a large language model, which solves the technical problems in the existing technology caused by reliance on manual rules, weak generalization ability of the diagnosis model, and inability to effectively integrate structured monitoring data and unstructured text records, resulting in low fault identification accuracy, delayed response, and difficult to interpret diagnosis results. It achieves the technical effect of integrating multi-source data for semantic modeling and improving the intelligence and accuracy of fault identification.

[0005] In view of the above problems, the first aspect of the present application provides a method for intelligent diagnosis of subway equipment faults assisted by a large language model, the method comprising: obtaining operation data of subway equipment, including sensor data, fault history records, and operation environment data of the equipment, wherein the text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; preprocessing the text of the fault history records, and using 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 operation environment, wherein the large language model is obtained by fine-tuning a special text corpus constructed for subway equipment operation and maintenance scenarios; systematically organizing the extracted key information, constructing a multidimensional equipment fault knowledge graph, defining entity nodes and semantic relationship edges of the entity nodes, and forming a structured knowledge base, wherein the entity nodes include fault type nodes, fault cause nodes, treatment measure nodes, equipment name nodes, component location nodes, and environmental condition nodes; based on the multidimensional equipment fault knowledge graph, using a large language model combined with graph neural network The system uses a network to perform deep learning and semantic modeling on equipment fault data to obtain a historical fault dataset, which includes a multidimensional feature representation, a fault causal chain, and a semantically enhanced reasoning model. The system traverses the historical fault dataset, extracts key fault monitoring parameters related to the fault, collects sensor status data corresponding to the parameters, obtains a sensor status dataset, and performs anomaly identification on the sensor status dataset to determine whether the sensor status data has data drift, data missing, signal distortion, or state anomalies. Based on the anomaly identification results, the system calls a pre-built sensor distortion correction algorithm to generate fault monitoring correction parameters and performs parameter correction. The fault monitoring correction parameters are added 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 during equipment operation and associated with the fault event. The multidimensional monitoring data in the multidimensional monitoring dataset is input into a diagnostic engine driven by a large language model. Through natural language question answering and multi-round reasoning, the system outputs the most likely fault type, cause analysis, recommended treatment strategy, and fault identification report.

[0006] The second aspect of the present application provides a subway equipment fault intelligent diagnosis system assisted by a large language model, the system comprising: an operation data acquisition module for subway equipment, the data acquisition module 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 including historical maintenance records, operation logs and fault descriptions; an equipment fault data extraction module, the equipment fault data extraction module is used to 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 equipment fault data, including equipment type, fault phenomenon, time node, processing measures, and operation environment; a multi-dimensional equipment fault knowledge graph construction module, the multi-dimensional equipment fault knowledge graph construction module 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 including fault type, cause, and measures; a historical fault data set acquisition module, the historical fault data set acquisition module is used to perform deep learning and semantic analysis on equipment fault data based on the multi-dimensional equipment fault knowledge graph using a large language model combined with a graph neural network. Modeling, obtaining a historical fault data set, the historical fault data set includes a multi-dimensional feature representation, a fault causal chain and a semantic enhancement reasoning model; a sensor state data set anomaly recognition module, the sensor state data set anomaly recognition module is used to traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect sensor state data corresponding to the parameters, obtain the sensor state data set, and perform anomaly recognition on the sensor state data set to determine whether the sensor state data has data drift, data missing, signal distortion, or state anomaly; a multi-dimensional monitoring data set formation module, the multi-dimensional monitoring data set formation module is used to Based on the abnormality identification results, a pre-built sensor distortion correction algorithm is called to generate fault monitoring correction parameters and perform parameter correction. The fault monitoring correction parameters are added 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; a fault identification module 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 through natural language question answering and multi-round reasoning, output the most likely fault type, cause analysis, recommended processing strategy and fault identification report.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: The large language model-assisted intelligent diagnosis method and system for subway equipment faults provided in this application relates to the field of data processing technology, obtains the operating data of subway equipment, uses the large language model to extract key information, 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 the fault, obtains a sensor state data set, and determines whether the sensor state data has an anomaly; based on the anomaly identification result, calls a pre-built sensor distortion correction algorithm, generates fault monitoring correction parameters and performs parameter correction, inputs the multi-dimensional monitoring data into the large language model-driven diagnosis engine, and outputs the most likely fault type, cause analysis, recommended processing strategy, and fault identification report. This application solves the technical problem in the prior art of low fault diagnosis intelligence and difficulty in achieving accurate identification and causal analysis due to the lack of fusion modeling capabilities for unstructured fault text and structured monitoring data, thereby improving the fault identification response speed and accuracy.

[0008] In summary, this application, by constructing a multi-dimensional equipment fault knowledge graph and integrating a large language model with a deep learning mechanism of a graph neural network, can accurately extract key fault information and perform semantic modeling and causal reasoning, and conduct more comprehensive fault identification, reducing missed fault detection and misjudgment rates; the diagnostic engine based on the semantically enhanced reasoning model makes the fault cause analysis more logically relevant and explanatory, avoiding misdiagnosis and missed judgment caused by sensor anomalies, while improving the robustness and accuracy of fault identification. Based on the model dynamic selection mechanism, the diagnostic bias caused by model misselection is reduced, and targeted fault identification reports and processing strategy recommendations can be quickly generated, improving diagnostic efficiency and automation level, contributing to the intelligent upgrade of subway equipment operation and maintenance management, and improving the operational safety and service reliability of urban rail transit systems.

[0009] The above description is only an overview of the technical solution of the present application. In order to more clearly understand the technical means of the present application, it can be implemented in accordance with the contents of the specification. In order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0011] Figure 1A flow chart of the large language model-assisted intelligent diagnosis method for subway equipment faults provided in an embodiment of the present application.

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

[0013] Explanation of the accompanying drawings: operation data acquisition module 10, equipment fault data extraction module 20, multi-dimensional equipment fault knowledge graph construction module 30, historical fault data set acquisition module 40, sensor status data set anomaly recognition module 50, multi-dimensional monitoring data set formation module 60, fault recognition module 70. DETAILED DESCRIPTION

[0014] This application provides a large language model-assisted intelligent diagnosis method and system for subway equipment faults, which is used to solve the technical problems in existing subway equipment operation and maintenance systems, such as the lack of semantic understanding capabilities and fault causal modeling mechanisms, which result in fault diagnosis relying on manual experience, low accuracy, and untimely response.

[0015] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only some of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this 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 are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or modules that are not clearly listed or inherent to these processes, methods, products or devices.

[0017] Example 1, as Figure 1 As shown, the present application provides a method for intelligent diagnosis of subway equipment faults assisted by a large language model, the method comprising:

[0018] P100: Acquire the operation data of subway equipment, including the equipment's sensor data, fault history records, and operating environment data. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions.

[0019] Specifically, various sensors deployed on key subway equipment, such as temperature sensors, vibration sensors, current sensors, and displacement sensors, collect real-time data on structural status, operational status, electrical parameters, and mechanical response. This sensor data can include metadata such as timestamp, sampling frequency, data unit, and acquisition source to support subsequent data preprocessing and normalization.

[0020] The structured and unstructured fault-related information recorded throughout the life cycle of subway equipment includes historical maintenance records: covering fields such as maintenance time, maintenance personnel, maintenance parts, replacement parts, maintenance measures and effect evaluation, which are used to restore the maintenance evolution trajectory of the equipment; operation logs: automatically generated by the equipment control system or monitoring platform, recording the equipment's operating status changes, start and stop events, alarm status and dispatch instructions in each time period; fault descriptions: including text content such as fault phenomena, preliminary judgments, occurrence time, external conditions, etc. in natural language form filled in by operation and maintenance personnel or recorded by smart terminals, providing semantic modeling corpus support for large language models.

[0021] Environmental data refers to data on external interference factors related to the equipment operating environment, such as tunnel humidity, temperature, dust concentration, electromagnetic interference, earthquake impact, etc. The environmental data can be regularly collected through environmental sensors deployed on platforms, sections or around equipment to analyze the impact of external working conditions on the health status of the equipment.

[0022] P200: Preprocess the text of the fault history records, 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. The large language model is obtained by fine-tuning a dedicated text corpus built for subway equipment operation and maintenance scenarios.

[0023] Furthermore, step P200 in the embodiment of the present application further includes:

[0024] P201: Standardize and preprocess the historical fault record text collected from the subway operation and maintenance system. The preprocessing includes character cleaning, proper noun regularization, format unification, semantic error correction and segmentation labeling. P202: Input the preprocessed text into the large language model and execute a multi-layer semantic processing flow. 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 extracted key information, and label the information with low confidence or ambiguous expressions.

[0025] It should be understood that the key terms in the historical text are identified and annotated through the large language model, and proper noun entities such as equipment name, part number, maintenance tools, and personnel identity are extracted. Combined with context understanding and language structure, the semantic information such as equipment operating status, fault causality, and diagnostic conclusions implied in each record is parsed, and semantic associations between different entities are established, such as "brake → fault phenomenon → cannot be reset", "fault occurrence time → October 2024", "maintenance measures → replace sensor" and other knowledge triple structures. The extracted key information is uniformly converted into a data format that conforms to the predefined template or field, which is convenient for subsequent warehousing, training and analysis, and improves the readability, processability and semantic depth modeling capabilities of the fault corpus.

[0026] First, redundant characters (such as extra spaces, line breaks, special symbols, and non-UTF characters) in the original text are removed or replaced to improve text standardization. For grammatical errors, word confusion, or semantic ambiguity, semantic correction is performed based on the context. For example, "The brakes were unresponsive, and the current anomaly has been repaired" can be corrected to "The brakes were unresponsive, and the current anomaly has been repaired" to improve language clarity and contextual consistency. A rule-based and model-based composite segmentation algorithm is used to annotate the text based on content semantics, such as the fault phenomenon section, the handling process section, the results and suggestions section, and the relevant equipment and time nodes.

[0027] Next, the large language model uses a pre-trained language model based on the Transformer architecture, such as BERT, RoBERTa, or ChatGLM. To improve its performance in subway equipment fault diagnosis tasks, a dedicated text corpus for subway equipment operation and maintenance scenarios was first constructed. This corpus includes but is not limited to the following: historical fault records of subway equipment; operation and maintenance work orders; professional maintenance documents and manuals; maintenance reports and accident notification records; and natural language annotations in control system operation logs.

[0028] Next, fault history text collected from the subway operation and maintenance system (such as operation logs, maintenance records, and inspection reports) undergoes standardized preprocessing, including character cleaning, formatting unification, and proper noun regularization, before being input into a fine-tuned large language model. Based on the language model's contextual understanding capabilities, key information types within the text are identified. By performing association modeling on this information, logical and causal relationships between entities are identified, including: the attribution between fault phenomena and equipment types; the chronological relationship between fault occurrence time and treatment measures; and the potential causal impact between environmental factors and fault occurrence. The identification and modeling results are uniformly output as a structured set of key information, represented as key-value pairs or triples, to support subsequent graph construction and intelligent reasoning. The extracted structured key information is further used as input to construct a knowledge graph for equipment failure causality and a semantically enhanced reasoning model, improving the automation, semantic accuracy, and response efficiency of fault analysis.

[0029] Finally, the confidence assessment module of the large language model trained using multi-task learning is called upon. Combining metrics such as attention weight, contextual consistency, and predicted distribution probability, a semantic confidence score is assigned to each identified key information item. The score typically ranges from 0 to 1, with higher values ​​indicating greater confidence. Confidence assessment is performed item by item for each of the five identified key information categories (device type, fault phenomenon, time point, treatment measures, and operating environment). The confidence score and tag information are attached as auxiliary attributes to the structured information items, generating output results that can be reviewed manually or further optimized through learning. Information items marked as "fuzzy" or "low confidence" are transferred to the manual review process or fed back as training samples for fine-tuning the language model, thereby continuously optimizing model performance and text information extraction quality.

[0030] P300: Systematically organize the extracted key information, construct a multi-dimensional equipment failure 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, treatment measure nodes, equipment name nodes, component location nodes, and environmental condition nodes.

[0031] Furthermore, step P300 in the embodiment of the present application further includes:

[0032] P301: Key information extracted through the large language model is archived in a unified data structure, including fields for fault type, fault cause, treatment measures, equipment name, component location, and environmental condition parameters. P302: Based on this archived information, core entity node types in the equipment fault domain are defined, including fault type nodes, fault cause nodes, treatment measures nodes, equipment name nodes, component location nodes, and environmental condition nodes. P303: Based on semantic analysis and contextual reasoning, relationship edges between entity nodes are extracted, and each edge is annotated with a semantic type. P304: Based on these entity nodes and relationship edges, a multidimensional knowledge graph is constructed for subway equipment fault diagnosis tasks. This multidimensional knowledge graph is stored in a graph structure and can be viewed as a set of structured triples, each consisting of entity 1, relationship, and entity 2.

[0033] Optionally, a multi-dimensional equipment fault knowledge graph is generated through semantic recognition and entity relationship extraction algorithms to achieve structured expression and causal logic modeling of historical fault information, thereby improving the accuracy and interpretability of subsequent diagnostic engines in fault reasoning and strategy recommendation.

[0034] First, the key information extracted by the large language model is archived in a unified data structure to construct 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 abnormality" and "overheating alarm"; Fault cause field: used to indicate the direct or indirect cause of the fault identified by the language model, such as "cable aging" and "signal interference"; Treatment measures field: used to describe the treatment strategy or operating steps proposed by maintenance personnel or the system; Equipment name field: identifies the target equipment or system where the fault occurs, such as "shield door controller" and "train control system receiving unit"; Component location field: further identifies the specific component or module location inside the equipment, such as "door 3 of train car 1" and "left interface of the main control board";

[0035] Environmental condition parameter field: records external environmental information related to the fault, including temperature, humidity, electromagnetic interference level, etc.

[0036] Next, based on the archived structured fault information, the system defines the core entity node types in the subway equipment fault knowledge graph to build a semantic network and support subsequent reasoning and diagnosis. The core entity node types include but are not limited to the following categories: Fault Type Node: Indicates 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 causes that cause the fault type, such as "module aging", "interface cold soldering", "external strong electromagnetic interference", "component drift in high temperature environment", etc.; Handling Action Node: Indicates the disposal strategy or operation and maintenance measures taken for specific faults, such as "replace the main control board", "restart the communication link", "adjust the cable wiring", "clear dust", etc.; Device Name Node: Identifies the name of the subway system equipment that has failed, including "traction inverter", "on-board control unit", "trackside signal", "pantograph", etc.; Component Location Node (ComponentLocationNode) Node: used to further refine the internal structural location of the faulty equipment, such as "power module in the second compartment of car No. 1" or "middle port of the signal system main control cabinet." Environmental Condition Node: represents the external environmental conditions at the time of the fault, including information such as "temperature value," "humidity range," "equipment operating hours," and "on-site electromagnetic intensity level." After identifying entity nodes, the system further extracts semantic relationship edges between entity nodes based on deep semantic analysis and contextual reasoning of the archived text using a large language model, thereby constructing the edge structure within the equipment fault knowledge graph. The specific steps include: Contextual Semantic Modeling: The preprocessed fault text is fed into a fine-tuned large language model (such as BERT, K-BERT, or RoBERTa). The model uses a comprehensive understanding of sentence structure, syntactic dependencies, and contextual semantics to identify possible causal, associative, or hierarchical relationships between entity nodes. Relationship Edge Extraction: Through methods such as syntactic path tracing, semantic dependency graph analysis, and relationship extraction template matching, the system identifies significant semantic relationships between any two annotated entity nodes and generates a set of candidate edges.

[0037] Finally, based on the entity nodes extracted in the previous steps and the semantic relationship edges between them, a multidimensional knowledge graph for subway equipment fault diagnosis is constructed. This knowledge graph is stored in a graph structure and organized and managed as structured triples (entity 1 - relationship - entity 2). First, all identified and annotated entity nodes are uniformly modeled, including multiple categories such as fault type, fault cause, treatment measure, equipment name, component location, and environmental conditions. Each type of node is assigned a type identifier and attribute fields (such as node unique identifier, node description, and source confidence). Simultaneously, all semantic relationship edges are uniformly coded and managed according to their type. For each pair of entity nodes with a semantic relationship, the system automatically generates a structured triple, represented as (entity 1 - relationship - entity 2). For example, from the text "Braking system failure caused by high temperature environment," a triple (brake system failure, caused by, high temperature environment) is generated and stored as an edge in the graph. All triples are imported into a graph database (such as Neo4j or GraphDB) to construct a semantic graph structure. Each node connects to all its upstream and downstream related nodes, forming multiple causal paths and processing chains, and realizing a traceable logical relationship of fault-cause-measure. To improve the expressiveness of the knowledge graph, the system supports adding additional attribute information to 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, supports retrieval through a graph query language, and encapsulates a RESTful API interface for diagnostic engine calls, supporting knowledge retrieval and path analysis in subsequent reasoning and diagnostic tasks. The construction of a multi-dimensional knowledge graph significantly enhances the structured expression and semantic association capabilities of equipment fault information, and provides a semantic basis for subsequent natural language question-answering reasoning and fault decision-making based on large language models.

[0038] P400: Based on a multi-dimensional equipment failure knowledge graph, a large language model combined with a graph neural network is used to perform deep learning and semantic modeling on equipment failure data to obtain a historical failure dataset. The historical failure dataset includes a multi-dimensional feature representation, a failure causal chain, and a semantically enhanced reasoning model.

[0039] Furthermore, step P400 in the embodiment of the present application further includes:

[0040] P401: Based on unstructured and structured data such as fault description text, maintenance records, and sensor logs, natural language processing technologies such as named entity recognition (NER) and semantic relationship extraction are used to extract information such as device type, fault phenomenon, time node, treatment measures, operating environment, and semantic relationships, and construct a knowledge graph with a multi-level semantic structure of device-phenomenon-cause-measures. P402: The knowledge graph with this multi-level semantic structure is input into a graph neural network. Through node feature aggregation and adjacency relationship modeling, a low-dimensional semantic vector representation of entity nodes and relationship edges is obtained, preserving their contextual semantics and graph structure information. P403: A large language model is used to perform deep semantic modeling and language understanding on the low-dimensional semantic vector representations of the entity nodes and relationship edges to generate context-enhanced text vectors; P404: The low-dimensional semantic vector representations of the entity nodes and relationship edges extracted by the graph neural network are integrated with the context-enhanced text vectors generated by the large language model to form a unified joint feature vector space; P405: By performing cluster analysis, fault classification and causal chain identification processing on the joint feature vector space, feature alignment and normalized modeling are performed on the operating status and fault manifestations of multiple devices to form a structured historical fault data set.

[0041] It should be understood that based on the semantic characteristics and structured requirements of subway equipment operation and maintenance data, through the joint modeling mechanism of large language models and graph neural networks, it is ensured that key information in fault data can be accurately extracted, the causal chain can be effectively modeled, and a highly reliable structured knowledge graph can be formed to support subsequent intelligent fault diagnosis and processing strategy recommendations.

[0042] First, the system acquires multi-source, heterogeneous data accumulated within the subway equipment operation and maintenance system. This data includes, but is not limited to, fault description text (e.g., operation and maintenance logs, manual repair report records), maintenance records (e.g., maintenance work orders, maintenance plans), and sensor logs (e.g., temperature, current, vibration, and other monitoring data). This dataset contains both structured and unstructured textual information. For the unstructured textual data, the system uses a text standardization module to perform preprocessing operations such as character cleaning, formatting, terminology regularization, and segment labeling. Subsequently, a fine-tuned large language model is used to perform a natural language processing pipeline, including named entity recognition (NER), semantic role labeling, and relation extraction. This automatically identifies and extracts key semantic entity information, including: equipment type (e.g., traction transformer, train signal controller); fault symptoms (e.g., voltage anomaly, communication interruption, vibration anomaly); time points (e.g., fault occurrence time, repair time); treatment measures (e.g., module replacement, equipment restart, cable reconnection), and operating environment (e.g., temperature, humidity, tunnel section, operating load). After entity extraction, the system uses a semantic relationship extraction module to identify logical relationships and causal connections between these entities. For example, contextual semantic modeling can identify relationship paths such as "a certain equipment fault occurred," "a certain fault cause led to the phenomenon," and "a certain measure was taken during maintenance to eliminate the fault." These paths are represented as structured triples: (entity 1, relationship, entity 2), such as (traction transformer, occurrence, voltage anomaly) (voltage anomaly, cause, poor heat dissipation) (poor heat dissipation, measures taken, replacement of the heat dissipation unit). Ultimately, the identified entity nodes and the semantic relationship edges between them are organized into a multi-level equipment fault knowledge graph. This knowledge graph is stored in a graph structure and supports graph visualization, semantic retrieval, and graph neural network modeling. The core semantic framework of this knowledge graph is "equipment-phenomenon-cause-measure," enabling systematic modeling and reasoning support for complex fault information.

[0043] Next, the system uses the constructed multi-level semantic structure knowledge graph for subway equipment fault diagnosis as input. A graph neural network (GNN) model is introduced for semantic embedding learning, obtaining low-dimensional semantic vector representations of entity nodes and relationship edges that are contextually and structurally aware. The "equipment-phenomenon-cause-measure" knowledge graph is loaded into the graph learning engine in a graph data format. Each entity node in the graph (e.g., "signal receiving module," "communication interruption," "EMI interference," "replace receiver") is represented as a node, and each semantic relationship (e.g., "occurrence," "cause," "take measures") is represented as a directed edge. Each node is accompanied by its initial attribute features (e.g., word embedding, entity type, context summary) as input to the GNN. Using GNN models such as GCN (Graph Convolutional Network), GAT (Graph Attention Network), or R-GCN (Relational GCN), the semantic representation of each node is updated through an iterative feature aggregation mechanism based on the node's adjacency in the graph. This process not only considers the node's own characteristics but also integrates the characteristics of its direct neighboring nodes and the types of relationships 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, typically a 128-dimensional or 256-dimensional floating-point vector. This vector not only encodes the node's own semantics but also implies its structural position, contextual causal relationships, and semantic path information within the graph. The resulting low-dimensional semantic vector is used for subsequent similarity calculations, node clustering, causal chain tracing, and graph neural network-assisted fault diagnosis tasks, providing support for subway equipment anomaly identification, fault location, and strategy recommendation.

[0044] Next, based on the low-dimensional semantic vector representation of entity nodes and relationship edges generated by the graph neural network, a large language model (LLM) was introduced to further understand the semantics and perform language modeling on this vector information to enhance the contextual expression capability and form a context-enhanced text vector for the field of subway equipment failures.

[0045] Furthermore, a large language model based on the Transformer architecture, such as BERT, RoBERTa, or GPT, is fine-tuned in a supervised manner on a large-scale corpus of subway equipment failure data (including maintenance reports, operation logs, and repair records). This enables the model to accurately identify specialized terminology, causal relationships, and contextual dependencies, improving the accuracy and robustness of language understanding. The low-dimensional semantic vectors for each entity node and relationship edge generated by the graph neural network are combined into structured triples (entity-relationship-entity) as input sequences and mapped into a token embedding space recognizable by the model through a vector-text alignment module. For example, a set of triples: ("signal module" - occurrence - "communication interruption") is converted into an embedding sequence and input into the language model. The graph structure information is also added as contextual cues, enabling the model to understand the true semantic context represented by the graph structure.

[0046] Next, a fusion mechanism is designed to construct a unified joint feature vector space based on the structured semantic information extracted by a graph neural network (GNN) and the context-enhanced semantic information extracted by a large language model (LLM). This provides highly expressive features for subsequent fault reasoning, classification, and source tracing analysis. The system first receives node vector representations from the GNN, which contain the structural semantic embeddings of entity nodes (e.g., "brake system" and "high temperature alarm") and relationship edges (e.g., "cause" and "association") in the equipment fault knowledge graph. Simultaneously, the system receives text vectors output by a fine-tuned LLM, which encode the contextual semantic information in the fault description corpus. Before entering the fusion module, both vectors are normalized using a unified dimensionality mapping layer (e.g., a fully connected layer or projection matrix) to ensure comparability within the same semantic space. A multimodal semantic fusion approach is then used to fuse the node vector representations from the GNN with the text vectors output by the LLM.

[0047] Finally, based on the joint feature vector space generated by the fusion of a graph neural network and a large language model, semantic analysis and modeling of the operating status, fault phenomena, and environmental context of multiple devices were performed. A structured historical fault dataset tailored to the diagnostic task was constructed using a variety of techniques, including clustering, classification, and causal chain identification. Unsupervised learning algorithms (such as K-Means, DBSCAN, and spectral clustering) were used to cluster fault-related samples across all vector samples in the joint feature vector space, identifying fault groups with similar characteristic patterns. Through the implementation of these steps, the deep fusion and semantic modeling of multi-source fault information for subway equipment were completed, resulting in a structured historical fault dataset consisting of clustering results, classification labels, and causal chains. This provides a high-quality data foundation and semantic support for subsequent intelligent diagnostic reasoning and strategy generation.

[0048] Furthermore, step P405 of the embodiment of the present application further includes:

[0049] P405-1: The historical fault dataset includes multi-dimensional feature representation, fault causal chain, and semantically enhanced reasoning model; P405-2: The multi-dimensional feature representation is a joint representation vector constructed based on dimensions such as equipment type, key operating parameters, and working environment characteristics, which is used to reflect the fault manifestation characteristics under different equipment states; P405-3: The fault causal chain is an equipment failure chain that is collaboratively mined through a graph neural network and a large language model to obtain the semantic causal relationship of initial cause-propagation path-final failure; P405-4: The semantically enhanced reasoning model is a K-BERT architecture that adopts a large language model and a knowledge graph to obtain a multi-dimensional semantic feature representation with fault causal identification capabilities, semantic ambiguity resolution capabilities, and cross-path reasoning capabilities.

[0050] Specifically, by performing cluster analysis, fault classification and causal chain identification processing on the joint feature vector space, feature alignment and normalized modeling of the operating status and fault manifestations of multiple devices are performed to form a structured historical fault data set, which can further improve the fault knowledge transfer and generalization capabilities across devices and scenarios. By constructing a unified fault feature representation framework and standardized modeling mechanism, the adaptability and interpretability of historical fault knowledge under different equipment types and working conditions can be ensured.

[0051] First, a structured data set is constructed by deeply processing multi-source heterogeneous data generated during subway equipment operation, including sensor data, maintenance records, and fault text. Using a joint modeling mechanism of graph neural networks and large language models, this data from various sources (such as sensor data such as temperature, current, and vibration, fault description text, and maintenance logs) is vectorized and uniformly represented in a feature space. Each equipment sample is mapped into a high-dimensional feature vector containing operating parameters, environmental conditions, historical maintenance status, and textual semantics to comprehensively reflect the equipment status. Through cluster analysis, semantic relationship modeling, and graph structure reasoning, fault triggering conditions, evolution paths, and causal relationships are mined from historical data, constructing a multi-node causal path from "equipment status → fault symptoms → fault type → handling outcome." This causal chain, as the core relational structure in the knowledge graph, facilitates fault location, tracing, and prediction. This inference model combines the learning capabilities of graph neural networks (such as GCN and GAT) with pre-trained large language models (such as K-BERT or T5) for node semantics and structural dependencies. Through context-aware mechanisms and knowledge graph infusion strategies, it improves the expressiveness and inference accuracy of fault knowledge. This model supports semantic completion, anomaly interpretation, and the generation of action suggestions based on historical knowledge. Through structured modeling, it achieves deep representation and intelligent reasoning of historical fault knowledge, providing foundational support for subsequent real-time diagnosis and policy recommendations.

[0052] Next, a joint representation vector is constructed based on multiple dimensions, including equipment type, key operating parameters, and operating environment characteristics, to reflect the operating status and potential fault manifestations of subway equipment under different operating conditions. Information such as equipment structural attributes, functional classification, and part number is extracted as part of the equipment category embedding vector to distinguish the structural differences in feature representation between different types of subway equipment (such as traction systems, braking systems, and air conditioning systems). Core indicators reflecting the operational health of the equipment, such as voltage, current, temperature, vibration frequency, and acceleration, are selected from sensor data and normalized and aggregated with time series features to form an operational status vector subspace. Environmental contextual information, such as operating time period, car load, external temperature and humidity, and tunnel location, is combined to construct an environmental semantic vector to characterize the potential impact of external conditions on the equipment's operating status and fault manifestations. This information from different dimensions is integrated into a unified joint representation vector through feature concatenation or attention fusion mechanisms. This represents the equipment's state at a specific moment in operation and is used for subsequent tasks such as cluster analysis, classification recognition, and causal modeling.

[0053] Next, based on subway equipment operation data and historical fault text, a structured causal sequence was constructed by combining a graph neural network (GNN) with a large language model (LLM) for collaborative modeling and reasoning. Based on semantic entity nodes extracted from the knowledge graph, such as equipment operating status, anomalies, operational behaviors, and environmental disturbances, the GNN modeled and aggregated features of potential propagation paths between nodes, initially constructing a possible causal propagation graph. This graph structure information was integrated with the LLM and fed into a fine-tuned Transformer model. Leveraging its capabilities in natural language understanding and semantic relationship modeling, the model classified and annotated the causal semantics between nodes, identifying a ternary causal chain structure: "cause-intermediate state-resultant fault." Ultimately, a fault causal chain with the backbone of "initial cause → fault propagation path → final fault" was generated. Each chain consists of multiple semantically related nodes, annotated with the causal relationship type (e.g., trigger, influence, aggravate, inhibit), its occurrence order, and confidence score. The generated fault causal chain is systematically organized into structured sequence data, which facilitates the call and reuse of subsequent fault prediction, anomaly tracing and recommendation strategy generation tasks.

[0054] Furthermore, to address the challenges of ambiguous representations, complex causal chains, and multi-path intersections in subway equipment operation data, an improved K-BERT (Knowledge-enhanced BERT) model architecture is employed as the core of the semantically enhanced reasoning module, enabling deep fusion of language understanding and structural knowledge. The input of large amounts of fault description text, maintenance records, and operation logs is first encoded into token sequences required by the pre-trained BERT language model. Simultaneously, based on the constructed multi-dimensional equipment fault knowledge graph, key entities involved (such as "brake failure," "high temperature environment," and "motor failure") are identified and matched. The relevant knowledge subgraphs are then soft-inserted into the representation of the original sentence 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 through a self-attention mechanism, thereby preventing noise propagation and achieving joint learning of language context and structural knowledge. Through a fusion representation learning mechanism, the model outputs a multi-dimensional semantic feature representation vector. Multidimensional semantic feature representation vectors can identify implicit causal triggering structures in sentences (e.g., "due to...", "due to...") and establish semantic relationships between corresponding nodes. They leverage context and graph constraints to resolve the specific meaning of, for example, "power outage" in different devices. Furthermore, they can perform multi-hop semantic reasoning on graph structures for nonlinear, cross-segment causal relationships, revealing complex propagation paths. The resulting semantic vector not only retains the contextual information of the text but also incorporates the entity and relationship semantics of the structured knowledge graph. This vector drives subsequent modules such as anomaly identification, cause location, and policy recommendation. Furthermore, it can be integrated into multi-level reasoning processes for more accurate fault diagnosis decision support.

[0055] P500: Traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect sensor status data corresponding to the parameters, obtain a sensor status data set, and perform anomaly identification on the sensor status data set to determine whether the sensor status data has data drift, data missing, signal distortion, or status abnormality.

[0056] Furthermore, step P500 in the embodiment of the present application further includes:

[0057] P501: Traverse the equipment fault data set, analyze the real-time collected equipment operation data one by one, and extract key fault monitoring parameters associated with the fault event. The key fault monitoring parameters include but are not limited to vibration amplitude, current fluctuation, temperature rise gradient, speed deviation, voltage anomaly, leakage volume and other indicator parameters; P502: Extract key fault monitoring parameters, locate the sensor channel corresponding to the key fault monitoring parameters, collect the current status data of the sensor, and form a sensor status data set corresponding to the monitoring parameters one by one. The sensor status data includes sensor value, timestamp, signal quality, communication status, and sensor operation status; P503: Perform abnormality identification processing on the sensor status data, and use a multi-dimensional detection mechanism to determine whether there is an abnormality. The abnormality includes data drift, data missing, signal distortion, and status abnormality; P504: Mark the detected abnormal state and output the corresponding abnormality category, impact range and confidence level.

[0058] Optionally, through collaborative modeling of graph neural networks and large language models, a joint feature vector representation that integrates contextual semantics and structured knowledge is generated. Through multidimensional clustering analysis and semantic similarity calculation, historical fault cases and causal chains that are highly correlated with the target device status are screened out, thereby achieving accurate matching and traceable analysis of the current abnormal status, providing data support and logical basis for subsequent fault prediction and strategy recommendation.

[0059] First, the constructed historical equipment failure dataset is traversed to extract key monitoring dimensions and their evolutionary characteristics associated with known failure events. The real-time collected equipment operation data is parsed and processed one by one, and each record is matched and analyzed with the trigger conditions in the historical failure causal chain to identify possible abnormal evolution paths. This process is based on a unified data structure and semantic labels to ensure the consistency of real-time data and historical knowledge. Key fault monitoring parameters highly correlated with typical failure events are extracted, including but not limited to: vibration amplitude: used to identify problems such as abnormal wear, imbalance, or looseness of mechanical components; current fluctuation: reflecting load fluctuations in the drive system, motor overload, or control loop anomalies; temperature rise gradient: used to monitor changes in the equipment's operating thermal state and identify poor heat dissipation or abnormal heating; speed deviation: used to determine load anomalies, mechanical jamming, or instability of rotating equipment; voltage anomalies: used to identify electrical problems such as unstable power input and electrical control unit failure; and air leakage: used to assess the sealing status of pneumatic systems and the degree of pipeline aging. Parameters are collected synchronously from multiple sensor channels. After preprocessing steps such as data synchronization, noise removal, and time alignment, they are uniformly mapped into the equipment status analysis model, providing the core input for subsequent fault trend prediction, causal chain location, and intelligent diagnosis. This approach enables the transition from historical data-driven fault-sensitive parameter modeling to the effective extraction of potential fault signals from real-time data, enhancing the ability to identify early signs of faults under complex operating conditions.

[0060] Next, based on the predefined fault diagnosis parameter mapping relationship library, the key fault monitoring parameters extracted from the historical fault data and real-time data analysis are processed. Each key monitoring parameter is associated with one or more specific sensor channels, which are uniquely identified in the subway equipment operation and maintenance system. The sensor channel corresponding to each key fault monitoring parameter is located by the equipment number and the monitoring parameter field index. For example: vibration amplitude Accelerometer channel; current fluctuation Current transformer channel; temperature rise gradient Thermistor channel; Leakage Airflow pressure / flow sensor channels, etc. After completing the channel positioning, the system dispatches the data acquisition module to perform status query and real-time data acquisition on the sensors. During the acquisition process, the system will record and generate a set of sensor status data sets that correspond one to one with the key monitoring parameters. Each record in the data set includes the following: Sensor value: the physical quantity measurement value output by the current sensor, which is used to characterize the current state of the monitoring indicator; Timestamp: the acquisition time, which is used to support time series analysis and fault trend backtracking; Signal quality: indicates whether the data is interfered with by noise, whether the signal amplitude is abnormal, and other quality indicators; Communication status: reflects whether the data transmission process is complete, such as whether there is communication packet loss, delay, etc.; Sensor operation status: records whether the sensor itself is operating normally, including calibration status, power status, working mode, etc.

[0061] Furthermore, an anomaly detection process is executed on the resulting sensor status dataset to ensure the accuracy and reliability of subway equipment operation data. Anomaly detection, based on a multi-source fusion data quality control mechanism, employs a multi-dimensional detection mechanism to verify the status data of each sensor channel individually, determining whether data drift, missing data, signal distortion, and state anomalies are present. Data drift detection analyzes the short-term or long-term mean / variance trends of sensor values ​​by comparing them against reference thresholds over a sliding window. If the current value consistently deviates from the historical stability range and is not explained by environmental factors, it is considered a data drift anomaly. For example, a sensor outputting a continuously increasing vibration amplitude when the equipment is unloaded.

[0062] Next, for sensor state anomalies identified by the multi-dimensional detection mechanism, the system executes the abnormal state annotation module's processing flow, outputting structured annotations for each abnormal data item to form standardized abnormal information items. Based on the anomaly detection results and preset classification rules, each abnormal data item is labeled with its category, which mainly includes but is not limited to: data drift (Drift), data missing (Missing), signal distortion (Distortion), and state anomaly (FaultyState). Each type of anomaly corresponds to a specific detection dimension and indicator, such as drift corresponding to numerical trend variation, and distortion corresponding to frequency spectrum anomaly.

[0063] Finally, each anomaly type corresponds to specific detection dimensions and indicators, such as drift for numerical trend variations and distortion for frequency spectrum anomalies. By combining the structural hierarchy of the sensor's component, the correlation between monitored parameters, and the equipment topology, the system automatically assesses the operational objects and scope of the anomaly's potential impact. For example, if a vibration sensor anomaly is located at the traction motor bearing, the impact scope is at the "traction system subcomponent level." If multiple adjacent parameters are abnormal, the system can label it as "equipment system-level impact," with the affected component being the traction motor bearing and the impact scope being at the component / system level. Based on the confidence output of the anomaly detection algorithm (such as an SVM probability output, a neural network softmax result, or an empirical threshold), the system calculates the anomaly confidence level for each piece of anomaly data and categorizes it into levels: high (≥0.85), medium (0.6-0.85), and low (<0.6). The original confidence score (e.g., 0.923) is retained to support subsequent confidence-based decision-making. Ultimately, the system stores the anomaly status annotation information uniformly as structured record items. Each record includes at least the following fields: anomaly timestamp, sensor channel ID, anomaly category, affected components / scope, anomaly confidence level and value, and a brief description of the anomaly. For example, the anomaly data item labeled "continuously rising deviation exceeds the 5% threshold" will serve as a key input for subsequent equipment fault diagnosis, early warning triggering, and root cause analysis.

[0064] P600: Based on the anomaly identification results, call the pre-built sensor distortion correction algorithm to generate fault monitoring correction parameters and perform parameter correction. The fault monitoring correction parameters are added 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 during equipment operation and associated with the fault event.

[0065] Furthermore, step P600 in the embodiment of the present application further includes:

[0066] P601: Based on the anomaly type marked in the anomaly identification result, the sensor distortion correction algorithm is input. The sensor distortion correction algorithm is constructed based on a multi-model integration mechanism driven by the anomaly type. For different types of abnormal states, the corresponding correction sub-algorithm is called to obtain fault monitoring correction parameters. P602: Combined with the reconstructed residual of the abnormal feature vector, the historical error distribution curve, the real-time working condition context information and the prediction confidence assessment results of the correction sub-algorithm, dynamic model selection and correction parameter output are performed.

[0067] Specifically, for the abnormal sensor status data output by the abnormality recognition module, the system automatically calls the pre-built sensor distortion correction algorithm to realize dynamic correction processing of fault monitoring parameters, thereby improving the accuracy and availability of subsequent data analysis.

[0068] First, based on the anomaly labeling output by the anomaly identification module, the anomaly type field in each sensor status data item is parsed to identify specific anomaly categories, such as data drift, signal distortion, missing data, or abnormal state. Based on these identified anomaly types, the system executes model matching logic within the built-in correction model scheduler. This scheduler, based on the "anomaly type → correction sub-model" mapping rule, activates the following sub-algorithm modules as needed: data drift: invokes the trend regression fitting and residual compensation model; signal distortion: invokes the denoising and restoration model based on wavelet multiscale analysis; missing data: invokes the time series interpolation and prediction reconstruction model; and abnormal state: invokes the state envelope extraction and correction fitting model. The called correction sub-algorithm processes the original monitoring data of the corresponding sensor channel and outputs fault monitoring correction parameters. The correction parameters include but are not limited to: numerical offset correction value (such as: +0.85), filter repair factor (such as: wavelet threshold, order parameter), missing completion value and confidence score, abnormal point repair identifier and replacement 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 to support subsequent data enhancement, modeling training and visual analysis.

[0069] Next, to achieve higher accuracy and greater adaptability in sensor anomaly correction, the system further incorporates a dynamic model selection mechanism within its multi-model integration process. This mechanism comprehensively considers multiple evaluation metrics to dynamically determine and accurately select the optimal correction algorithm, ultimately outputting correction parameters that best suit the current fault scenario. The input sensor anomaly feature vector is first reconstructed using each candidate correction algorithm. The reconstructed residual between the original anomaly vector and the input is calculated to measure the degree of recovery after correction. Smaller reconstructed residuals indicate a stronger fit and restoration capability for the model. Historical records similar to the current device, sensor channel, and anomaly type are retrieved from the historical fault database. A corresponding error distribution reference curve is constructed and compared with the error distribution of the current model's processing results. A matching score is calculated to aid in determining model suitability. The system then obtains the current device's real-time operating context (including operating mode, load level, ambient temperature, vibration frequency, etc.) and compares it with the operating conditions supported by the correction algorithm. Only models with operating condition compatibility exceeding a set threshold are retained as candidate models. For each candidate sub-algorithm's correction results, its internal confidence prediction module is invoked to calculate a prediction confidence score based on factors such as residual distribution, feature coverage, and model training weight stability. This score is used to comprehensively assess the reliability and acceptability of the correction results. The aforementioned dimensional indicators are weighted and integrated to form a comprehensive score. A ranking and screening process is then performed, and the sub-algorithm with the highest comprehensive score is dynamically selected as the execution model for this correction task. The corresponding fault monitoring correction parameters are then output, including the correction deviation value, time period weight adjustment factor, anomaly mask annotation, and signal restoration curve parameters.

[0070] Furthermore, the sensor distortion correction algorithm described in step P601 of the embodiment of the present application further includes:

[0071] P601-1: Acquire historical monitoring data and sensor raw sampling values ​​of multiple subway equipment under normal operation and fault conditions; P601-2: Label the abnormal conditions appearing in the historical monitoring data to obtain an abnormal label data set, which includes identification data of data drift, data loss, signal distortion, and state abnormality; P601-3: Preset the abnormal label data set as L, L= ,in, Normal, For data drift, For missing data, For signal distortion, The status is abnormal; P601-4: Get the original sensor value at time t ;

[0072] Calculate the historical mean , the formula is:

[0073] = × ;

[0074] in, is the historical average, is the number of historical samples;

[0075] When the exception label is When , the calculation formula of fault monitoring correction parameter is:

[0076] = - × ;

[0077] in, Correction parameters for fault monitoring, is the original sensor value at time t, is the drift correction coefficient, ,initialization =0.6 , is the historical mean;

[0078] P601-5: When the exception label is When , the calculation formula of fault monitoring correction parameter is:

[0079] = × + × ;

[0080] in, Correction parameters for fault monitoring, is the interpolation balance coefficient, ,initialization =0.5 , for The original sensor value at the moment, for Raw sensor value at the moment;

[0081] P601-6: When the exception label is When , the calculation formula of fault monitoring correction parameter is:

[0082] = × - × Δ ;

[0083] Among them, Δ is the local deviation, Δ = - × , is the signal suppression coefficient, ,initialization =0.3 ;

[0084] P601-7: When the exception label is When , the calculation formula of fault monitoring correction parameter is:

[0085] = × × ;

[0086] in, Correction parameters for fault monitoring, is the smoothing parameter, ,initialization =0.7;

[0087] P601-8: When the exception label is When , the calculation formula of fault monitoring correction parameter is:

[0088] = ;

[0089] in, Correction parameters for fault monitoring, is the original sensor value at time t;

[0090] Output fault monitoring correction parameters ;

[0091] P601-9: If , ,... , >= 3, it is judged as abnormal and marked as a continuous abnormal segment;

[0092] Set the sliding window to W, W= ;

[0093] The local polynomial regression method is used to reconstruct the data of the continuous abnormal segments, and the value sequence generated by the sliding window as a whole is used to replace the original continuous abnormal segments.

[0094] First, a multi-model integration mechanism driven by anomaly type is employed, enabling adaptive selection and precise correction capabilities. This includes steps such as anomaly identification, sub-algorithm selection, correction parameter generation, and correction execution. The system receives the annotation results output by the anomaly identification component and identifies the anomaly type corresponding to the current sensor state data, including but not limited to drift, missing data, signal distortion, and malfunction. For each anomaly type, multiple specialized 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; for missing data, time series interpolation (such as spline interpolation, Kalman filtering) and Transformer prediction model reconstruction; for signal distortion, wavelet transform reconstruction and bandpass filtering recovery algorithms; and for malfunction, AutoEncoder reconstruction networks or clustering models (such as DBSCAN) are used to eliminate and fill in anomalies. Based on the currently identified anomaly type, the corresponding candidate sub-algorithm set is invoked. Using multiple evaluation indicators such as reconstruction residual, historical matching degree, and confidence score, the output results of the sub-algorithms are weighted and scored to select the optimal model or perform multi-model weighted fusion.

[0095] Next, the correction process includes inputting anomaly annotation results, identifying candidate sub-algorithms, performing reconstruction prediction, evaluating correction quality, and outputting a set of correction parameters. The input anomaly annotation results include anomaly type, time period, anomaly level, and impact range. Determining candidate sub-algorithms involves screening available models from a sub-algorithm library based on anomaly type. Reconstruction prediction involves reconstructing, fitting, or infilling the sensor data for the anomaly segment using the candidate models to generate preliminary correction values. Correction quality evaluation involves calculating the deviation between the reconstructed residual and the original data, comparing it to a historical error distribution curve, evaluating the correction effect, and determining the credibility of the correction using a comprehensive confidence scoring mechanism. The output correction parameter set includes offset correction factors, missing interpolation values, spectrum correction curves, and outlier masks. Data updating and storage involves replacing the original anomaly data with the correction results and appending identification fields to form an enhanced sensor state dataset. By deploying this sensor distortion correction algorithm, real-time repair and enhancement of equipment fault monitoring data can be achieved, providing a high-quality, low-noise input foundation for subsequent fault identification, semantic modeling, and intelligent diagnosis.

[0096] Furthermore, abnormal situations appearing in the historical monitoring data are labeled to obtain an abnormal label data set, wherein the abnormal label data set includes identification data of data drift, data missing, signal distortion, and state abnormality.

[0097] If , ,... , >= 3, it is judged as abnormal and marked as a continuous abnormal segment;

[0098] Set the sliding window to W, W= ;

[0099] The local polynomial regression method is used to reconstruct the data of continuous abnormal segments, and the original abnormal segments are replaced by the value sequence generated by the sliding window as a whole.

[0100] P700: Inputs multidimensional monitoring data from a multidimensional monitoring dataset into a diagnostic engine driven by a large language model. Through natural language question answering and multi-round reasoning, it outputs the most likely fault type, cause analysis, recommended processing strategy, and fault identification report.

[0101] Furthermore, step P700 in the embodiment of the present application further includes:

[0102] P701: Input the multidimensional monitoring data set that integrates sensor status information, fault monitoring parameters and sensor correction parameters into the diagnosis engine driven by the large language model; P702: Based on the large language model, the diagnosis engine performs semantic analysis and intent understanding on the multidimensional monitoring data through a natural language question-answering mechanism, and extracts characteristic patterns, contextual information and potential semantic logic associated with the fault; P703: Utilizing the context retention capability and long text comprehension capability of the large language model, combined with the multidimensional equipment fault knowledge graph, fault causal chain and multidimensional monitoring data set, perform multiple rounds of causal reasoning, fault identification question-answering process and fault causal chain deduction, and output structured fault diagnosis results. The structured fault diagnosis results include the most likely fault type of the target equipment, the cause analysis of the corresponding fault, executable recommended processing strategies, and a fault identification report in natural language form for operation and maintenance personnel.

[0103] Optionally, semantic understanding and contextual reasoning modeling of multi-dimensional monitoring data ensures the accuracy and interpretability of fault diagnosis results, meeting the high-reliability operation and maintenance requirements of subway equipment, thereby achieving precise fault identification and efficient handling. When input data exhibits abnormal drift or contextual ambiguity, semantic error correction and confidence calibration are performed to ensure the stability and credibility of diagnostic results even in abnormal scenarios.

[0104] First, a multidimensional monitoring dataset, integrating sensor status information, fault monitoring parameters, and sensor correction parameters, is input into an intelligent diagnostic engine driven by a large language model. This multidimensional monitoring dataset includes, but is not limited to, real-time sensor data, timestamps, communication status, signal quality indicators, monitoring parameters (such as vibration amplitude, current fluctuation, and temperature gradient), and correction parameters generated based on anomaly identification results. The intelligent diagnostic engine is built on a pre-trained large language model fine-tuned with subway fault data and possesses natural language understanding and multimodal information fusion capabilities. After receiving this multidimensional data, the diagnostic engine first performs semantic parsing and context modeling, using the large language model to structure the input information and semantically align key data fields. Subsequently, the diagnostic engine combines a knowledge graph of historical equipment faults with contextual information, performing natural language question answering, multi-round reasoning, and causal chain deduction to output the most likely fault type, cause analysis, impact scope, and recommended treatment strategy for the current equipment state. It also automatically generates a fault identification report, which includes the diagnostic conclusion, reference basis, confidence score, and recommended actions.

[0105] This process enables intelligent interpretation of monitoring data and semantically enhanced decision-making, effectively improving the accuracy, interpretability and response speed of equipment fault identification.

[0106] Next, the diagnostic engine is built on a pre-trained large language model, enabling semantic understanding and question-answering reasoning for multimodal input data. After receiving multi-dimensional monitoring data that integrates sensor status information, fault monitoring parameters, and correction parameters, the diagnostic engine uses a built-in natural language question-answering mechanism to perform item-by-item semantic parsing and intent recognition on this data. Specifically, the natural language question-answering mechanism first models the semantic structure of the input data, extracting keywords, hyponyms, and related semantics from the text or structured description. Subsequently, it integrates annotated fault instances from historical data to understand the fault-indicating significance of key parameters (such as abnormal vibration, current fluctuation, and temperature gradient) in a specific context. Through this processing, the diagnostic engine automatically identifies characteristic patterns, operational context, and underlying semantic logic chains related to the current fault (e.g., "unstable operating voltage + abnormal communication status → signal interference → equipment failure"), providing a semantic foundation for subsequent fault type determination and recommended handling strategies.

[0107] Then, leveraging a large language model (LLM) pre-trained and fine-tuned for equipment failure data, combined with its context-preserving and long-text understanding capabilities, deep semantic modeling and reasoning are performed on the multi-source fused equipment failure information. The processing flow incorporates a multi-dimensional equipment failure knowledge graph, a historically constructed fault causal chain, and an enhanced multi-dimensional monitoring dataset as input to the fault identification module driven by the LLM.

[0108] Finally, based on the output of a multi-round causal reasoning process driven by a large language model, a structured fault diagnosis result is constructed for intelligent equipment diagnosis tasks. Based on the fused multi-dimensional monitoring data, the equipment fault knowledge graph, and the historical causal chain, the diagnostic engine integrates the semantic aggregation results from the reasoning path to determine the most likely fault type of the target device and output a clear fault classification label (e.g., drive motor overheating, sensor failure, control module communication anomaly, etc.). Combining the semantic relationships between entity nodes extracted by the graph neural network, the dynamic semantic chain constructed by the large language model, and contextual information, the engine outputs a fault cause analysis, clearly identifying the initial cause, intermediate propagation path, and key influencing nodes, forming a complete causal chain description from "fault cause-propagation process-performance outcome." Based on the identified fault type and causal path, the engine matches historical handling cases with expert rules in the knowledge graph to output an executable handling strategy for the current fault state, including emergency response measures, component replacement recommendations, and operating parameter adjustment plans. A structured, semantically enhanced, natural language fault identification report is generated for maintenance technicians. The report includes the device name, identification time, fault type, cause path, monitoring indicator summary, and treatment recommendations. The report is highly readable and has clear contextual logic, supporting system archiving and manual review. Structured fault diagnosis results can be used by automated decision-making modules and presented to on-site maintenance personnel through a human-computer interaction interface, achieving a closed-loop support for intelligently assisted diagnosis and resolution.

[0109] In summary, the embodiments of the present application have at least the following technical effects:

[0110] This application determines the key monitoring parameters and semantic feature patterns related to faults by integrating multi-source equipment operation data, fault text descriptions and sensor status information, and constructs a multi-dimensional knowledge graph and structured historical fault dataset for intelligent diagnosis of subway equipment.

[0111] Through the collaborative modeling mechanism of graph neural networks and large language models, the fault causal chain is identified and multi-round semantic reasoning is achieved, ultimately outputting structured fault diagnosis results and executable processing strategies with contextual understanding capabilities. Combining a sensor anomaly correction mechanism, knowledge graph semantic enhancement reasoning, and fault feature vector aggregation methods, the diagnostic engine ensures high robustness and interpretability for fault types under complex working conditions, significantly improving the accuracy and intelligence level of subway equipment fault identification. This achieves accurate, fast, and interpretable equipment fault diagnosis, realizing a fully automated process from multi-dimensional monitoring data to fault diagnosis, effectively improving the accuracy, reliability, and processing efficiency of subway equipment fault identification.

[0112] Example 2, based on the same inventive concept as the large language model-assisted intelligent diagnosis method for subway equipment faults in the previous embodiment, Figure 2 As shown, the present application provides a large language model-assisted intelligent diagnosis system for subway equipment faults. The system and method embodiments in the present application are based on the same inventive concept. The system includes:

[0113] Operation data acquisition module 10, which is used to acquire 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;

[0114] An equipment failure data extraction module 20 is used to pre-process the text of the fault history records, perform named entity recognition, semantic analysis, and relationship extraction using a large language model, and extract equipment failure data, including equipment type, failure phenomenon, time node, treatment measures, and operating environment;

[0115] A multi-dimensional equipment failure knowledge graph construction module 30 is used to systematically organize the extracted equipment failure data, construct a multi-dimensional equipment failure knowledge graph, define entity nodes and semantic relationship edges between the entity nodes, and form a structured knowledge base. The entity nodes include fault type, cause, and measure.

[0116] A historical fault dataset acquisition module 40 is configured to perform deep learning and semantic modeling on equipment fault data based on a multi-dimensional equipment fault knowledge graph using a large language model combined with a graph neural network to obtain a historical fault dataset. The historical fault dataset includes a multi-dimensional feature representation, a fault causal chain, and a semantically enhanced reasoning model.

[0117] A sensor state data set anomaly identification module 50 is used to traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect sensor state data corresponding to the parameters, obtain the sensor state data set, and perform anomaly identification on the sensor state data set to determine whether the sensor state data has data drift, data missing, signal distortion, or abnormal state;

[0118] a multi-dimensional monitoring data set forming module 60, configured to invoke a pre-built sensor distortion correction algorithm based on the anomaly identification results, generate fault monitoring correction parameters, 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 collection of data records collected by the monitoring system during equipment operation and associated with the fault event.

[0119] The fault identification module 70 is used to input the multidimensional monitoring data in the multidimensional monitoring data set into the diagnosis engine driven by the large language model, and output the most likely fault type, cause analysis, recommended processing strategy and fault identification report through natural language question answering and multi-round reasoning.

[0120] It should be noted that the order in which the embodiments of the present application are presented is for illustrative purposes only and does not necessarily represent the superiority or inferiority of the embodiments. Furthermore, the foregoing descriptions of specific embodiments of this specification are provided. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order or sequential sequence shown to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0121] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

[0122] This specification and drawings are merely illustrative of the present application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Obviously, those skilled in the art may make various modifications and variations to this application without departing from the scope of this application. Thus, this application is intended to include such modifications and variations as fall within the scope of this application and its equivalents.

Claims

1. A large language model-assisted intelligent diagnosis method for subway equipment faults, characterized in that: The method comprises: P100. Obtaining subway equipment operating data, including equipment sensor data, fault history records, and operating environment data. The text of the fault history records includes historical maintenance records, operating logs, and fault descriptions. P200. Preprocess the text of the fault history records, using a large language model to perform named entity recognition, semantic analysis, and relationship extraction to extract key information, including equipment type, fault phenomenon, time point, treatment measures, and operating environment. The large language model is fine-tuned using a specialized text corpus built for subway equipment operation and maintenance scenarios. P300: Systematize the extracted key information to construct a multi-dimensional equipment fault knowledge graph, define entity nodes and semantic relationship edges between the entity nodes, and form a structured knowledge base. The entity nodes include fault type nodes, fault cause nodes, treatment measure nodes, equipment name nodes, component location nodes, and environmental condition nodes. P400, based on a multi-dimensional equipment failure knowledge graph, uses a large language model combined with a graph neural network to perform deep learning and semantic modeling on equipment failure data to obtain a historical failure dataset. The historical failure dataset includes a multi-dimensional feature representation, a failure causal chain, and a semantically enhanced reasoning model. P500: traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect sensor status data corresponding to the parameters, obtain a sensor status data set, and perform anomaly identification on the sensor status data set to determine whether the sensor status data has data drift, data missing, signal distortion, or status abnormality; P600: Based on the anomaly identification results, invoke a pre-built sensor and distortion correction algorithm to generate fault monitoring correction parameters and perform parameter correction. The fault monitoring correction parameters are added to the fault status data to form an enhanced multi-dimensional monitoring data set. The fault status data is a collection of data records collected by the monitoring system during equipment operation and associated with the fault event. P700 inputs the multidimensional monitoring data from the multidimensional monitoring data set into the diagnostic engine driven by the large language model. Through natural language question answering and multiple rounds of reasoning, it outputs the most likely fault type, cause analysis, recommended processing strategy and fault identification report.

2. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: 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 device type, fault phenomenon, time node, treatment measures, and operating environment, including: Performing standardized preprocessing on the text of historical fault records collected from the subway operation and maintenance system, including character cleaning, proper noun regularization, format unification, semantic error correction, and segmentation annotation; The pre-processed text is fed into a large language model to perform a multi-layer semantic processing flow. Through semantic recognition and relationship modeling, a structured set of key information is extracted and output, including equipment type, fault phenomenon, time node, treatment measures, and operating environment; Each type of extracted key information is assigned a semantic confidence score, and information with low confidence or ambiguous expressions is labeled.

3. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: The extracted key information is systematically organized to 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, treatment measure nodes, equipment name nodes, component location nodes, and environmental condition nodes, including: The key information extracted through the large language model is archived in a unified data structure, including the fault type field, fault cause field, treatment measure field, equipment name field, component location field, and environmental condition parameter field; Based on the archived information, core entity node types in the field of equipment failure are defined, including a failure type node, a failure cause node, a treatment measure node, a device name node, a component location node, and an environmental condition node; Based on the results of semantic analysis and contextual reasoning, the relationship edges between entity nodes are extracted, and the semantic type of each edge is annotated; Based on the above entity nodes and relationship edges, a multidimensional knowledge graph for subway equipment fault diagnosis tasks is constructed. The multidimensional knowledge graph is stored in a graph structure and can be regarded as a set of structured triples, where the triples are entity 1-relationship-entity 2.

4. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: Based on the multi-dimensional equipment fault knowledge graph, the large language model is combined with the graph neural network to perform deep learning and semantic modeling on the equipment fault data to obtain a historical fault dataset. The historical fault dataset includes multi-dimensional feature representation, fault causal chain and semantic enhancement reasoning model, including: Based on unstructured and structured data from fault description text, maintenance records, and sensor logs, we use natural language processing technologies such as named entity recognition (NER) and semantic relationship extraction to extract device type, fault phenomenon, time node, treatment measures, operating environment information, and semantic relationships, and construct a knowledge graph with a multi-level semantic structure of device-phenomenon-cause-measure. Inputting the multi-level semantic structured knowledge graph into a graph neural network, and obtaining low-dimensional semantic vector representations of entity nodes and relationship edges through node feature aggregation and adjacency relationship modeling, thereby retaining their contextual semantics and graph structure information; Using a large language model to perform deep semantic modeling and language understanding on the low-dimensional semantic vector representations of the entity nodes and relationship edges, and generate context-enhanced text vectors; The low-dimensional semantic vector representation of entity nodes and relationship edges extracted by the graph neural network is fused with the context-enhanced text vector generated by the large language model to form a unified joint feature vector space; By performing cluster analysis, fault classification and causal chain identification processing on the joint feature vector space, feature alignment and normalized modeling are performed on the operating status and fault manifestations of multiple devices to form a structured historical fault data set.

5. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 4, characterized in that: The historical fault data set includes: The historical fault dataset includes multi-dimensional feature representation, fault causal chain, and semantic enhancement reasoning model; The multi-dimensional feature representation is a joint representation vector constructed based on the device type, key operating parameters, and working environment feature dimensions, which is used to reflect the fault manifestation characteristics under different device states; The fault causal chain is a device fault chain mined collaboratively by a graph neural network and a large language model, obtaining the semantic causal relationship of initial cause, propagation path, and final fault. The semantically enhanced reasoning model adopts the K-BERT architecture that integrates a large language model with a knowledge graph to obtain a multi-dimensional semantic feature representation with the ability to identify fault causality, resolve semantic ambiguity, and perform cross-path reasoning.

6. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: The traversing of the equipment fault data, extracting fault monitoring parameters, collecting sensor status information corresponding to the parameters, obtaining sensor status data, and performing abnormality identification to determine whether the status data has data drift, data missing, signal distortion, or status abnormality, includes: Traversing the equipment fault data set, parsing the real-time collected equipment operation data one by one, and extracting key fault monitoring parameters associated with the fault event, such as but not limited to indicator parameters of vibration amplitude, current fluctuation, temperature rise gradient, speed deviation, voltage anomaly, and air leakage; Extract 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 sensor state data set corresponding to the monitoring parameters. The sensor state data includes sensor value, timestamp, signal quality, communication status, and sensor operation status. Performing abnormality identification processing on the sensor status data, using a multi-dimensional detection mechanism to determine whether there are abnormalities, the abnormalities include data drift, data loss, signal distortion, and abnormal status; The detected abnormal state is marked and the corresponding abnormality category, impact range and confidence level are output.

7. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: Based on the anomaly identification results, a pre-built sensor distortion correction algorithm is called to generate fault monitoring correction parameters and perform parameter correction. The fault monitoring correction parameters are added 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 during equipment operation and associated with the fault event, including: According to the anomaly type marked in the anomaly recognition result, the sensor distortion correction algorithm is input. The sensor distortion correction algorithm is constructed based on a multi-model integration mechanism driven by the anomaly type. For different types of abnormal states, the corresponding syndrome algorithm is called to obtain fault monitoring correction parameters; The model is dynamically selected and correction parameters are output based on the reconstructed residual of the abnormal feature vector, the historical error distribution curve, the real-time working condition context information and the prediction confidence evaluation results of the correction sub-algorithm.

8. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 7, characterized in that: The sensor distortion correction algorithm includes: Obtain historical monitoring data and sensor raw sampling values ​​of multiple subway equipment under normal operation and fault conditions; Labeling abnormal situations appearing in the historical monitoring data to obtain an abnormal label data set, wherein the abnormal label data set includes identification data of data drift, data missing, signal distortion, and state abnormality; The default abnormal label data set is L, L= ,in, Normal, For data drift, For missing data, For signal distortion, Abnormal status; Get the raw sensor value at time t ; Calculate the historical mean , the formula is: = × ； in, is the historical average, is the number of historical samples; When the exception label is When , the calculation formula of fault monitoring correction parameter is: = - × ； in, Correction parameters for fault monitoring, is the original sensor value at time t, is the drift correction coefficient, ,initialization =0.6, is the historical mean; When the exception label is When , the calculation formula of fault monitoring correction parameter is: = × + × ； in, Correction parameters for fault monitoring, is the interpolation balance coefficient, ,initialization =0.5, for The original sensor value at the moment, for Raw sensor value at the moment; When the exception label is When , the calculation formula of fault monitoring correction parameter is: = × - × D ; Among them, Δ is the local deviation, Δ = - × , is the signal suppression coefficient, ,initialization =0.3 ; When the exception label is When , the calculation formula of fault monitoring correction parameter is: = × × ； in, Correction parameters for fault monitoring, is the smoothing parameter, ,initialization =0.7 ; When the exception label is When , the calculation formula of fault monitoring correction parameter is: = ； in, Correction parameters for fault monitoring, is the original sensor value at time t; Output fault monitoring correction parameters ; If , ,... , >= 3, it is judged as abnormal and marked as a continuous abnormal segment; Set the sliding window to W, W= ; The local polynomial regression method is used to reconstruct the data of the continuous abnormal segments, and the value sequence generated by the sliding window as a whole is used to replace the original continuous abnormal segments.

9. The large language model-assisted intelligent diagnosis method for subway equipment faults according to claim 1, characterized in that: The multi-dimensional monitoring data in the multi-dimensional monitoring data set is input into the diagnosis engine driven by the large language model. Through natural language question answering and multiple rounds of reasoning, the most likely fault type, cause analysis, recommended treatment strategy and fault identification report are output, including: Input the multi-dimensional monitoring data set integrating sensor status information, fault monitoring parameters and sensor correction parameters into the diagnosis engine driven by the large language model; The diagnostic engine, based on a large language model, performs semantic analysis and intent understanding on the multi-dimensional monitoring data through a natural language question-answering mechanism, extracting characteristic patterns, contextual information, and latent semantic logic associated with the fault; By utilizing the context retention and long text comprehension capabilities of the large language model, combined with the multi-dimensional equipment fault knowledge graph, fault causal chain and multi-dimensional monitoring data set, multiple rounds of causal reasoning, fault identification question-and-answer process and fault causal chain deduction are performed to output structured fault diagnosis results. The structured fault diagnosis results include the most likely fault type of the target equipment, the cause analysis of the corresponding fault, executable recommended processing strategies, and a fault identification report in natural language form for operation and maintenance personnel.

10. A large language model-assisted intelligent diagnosis system for subway equipment faults, characterized by: The system comprises: A subway equipment operation data acquisition module, which is used to acquire subway equipment operation data, including equipment sensor data, fault history records, and operating environment data. The text of the fault history records includes historical maintenance records, operation logs, and fault descriptions; An equipment failure data extraction module is used to pre-process the text of the fault history records, perform named entity recognition, semantic analysis, and relationship extraction using a large language model, and extract equipment failure data, including equipment type, failure phenomenon, time node, treatment measures, and operating environment; A multi-dimensional equipment failure knowledge graph construction module is used to systematically organize the extracted equipment failure data, construct a multi-dimensional equipment failure 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, cause, and measure; A historical fault dataset acquisition module, which is used to perform deep learning and semantic modeling on equipment fault data based on a multi-dimensional equipment fault knowledge graph using a large language model combined with a graph neural network to obtain a historical fault dataset. The historical fault dataset includes a multi-dimensional feature representation, a fault causal chain, and a semantically enhanced reasoning model; A sensor state data set anomaly identification module is used to traverse the historical fault data set, extract key fault monitoring parameters related to the fault, collect sensor state data corresponding to the parameters, obtain the sensor state data set, and perform anomaly identification on the sensor state data set to determine whether the sensor state data has data drift, data missing, signal distortion, or state anomaly; a multidimensional monitoring data set formation module, the multidimensional monitoring data set formation module being configured to invoke a pre-built sensor distortion correction algorithm based on the anomaly identification results, generate fault monitoring correction parameters, perform parameter correction, and add the fault monitoring correction parameters to the fault status data to form an enhanced multidimensional monitoring data set. The fault status data is a collection of data records collected by the monitoring system during equipment operation and associated with the fault event; A fault identification module is used to input multidimensional monitoring data from a multidimensional monitoring data set into a diagnostic engine driven by a large language model, and output the most likely fault type, cause analysis, recommended processing strategy, and fault identification report through natural language question answering and multiple rounds of reasoning.

Images