A fault diagnosis method, device, equipment, medium and product of a motor train unit

Abstract

The present application relates to the technical field of rail transit equipment fault diagnosis, and discloses a fault diagnosis method, device, equipment, medium and product of a motor train unit, a fault diagnosis method of a motor train unit comprising obtaining multi-source heterogeneous operation and maintenance data of a target motor train unit, constructing a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data; obtaining fault query information, performing semantic rewriting on the fault query information to obtain a standard multi-dimensional diagnosis abstract; based on the standard multi-dimensional diagnosis abstract, using the knowledge base with electrical topology constraints to perform multi-agent collaborative game to obtain a motor train unit fault diagnosis report. The present application greatly improves the automation level of fault diagnosis and the diagnosis accuracy in complex fault scenarios.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology for rail transit equipment, specifically to a fault diagnosis method, device, equipment, medium, and product for high-speed trains. Background Technology

[0002] With the rapid development of high-speed trains, their core traction and braking systems integrate extremely complex power electronics and microcomputer control logic. Current high-speed train fault diagnosis methods heavily rely on frontline personnel sifting through massive amounts of PDF (Portable Document Format) manuals and CAD (Computer Aided Design) drawings. This not only results in low retrieval and matching efficiency, but also in-depth logical troubleshooting heavily depends on the experience passed down from senior experts. When facing cross-system coupled faults, this can easily lead to over- or under-maintenance. Summary of the Invention

[0003] This invention provides a fault diagnosis method, device, equipment, medium, and product for high-speed trains, in order to solve the problems of low automation in current fault diagnosis methods for high-speed trains, and the tendency to over- or under-maintain when facing cross-system coupled faults.

[0004] In a first aspect, the present invention provides a fault diagnosis method for high-speed trains, the method comprising: Acquire multi-source heterogeneous operation and maintenance data of the target EMU, and construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data; Obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary; Based on standard multidimensional diagnostic summaries, a multi-agent cooperative game is conducted using a knowledge base with electrical topology constraints to obtain a fault diagnosis report for the high-speed train.

[0005] This invention provides a fault diagnosis method for high-speed trains. By constructing a knowledge base with electrical topology constraints, it correlates equipment-level faults with system-level impact paths through modeling, avoiding misjudgments caused by data gaps and noise interference in data-driven models. This significantly improves the physical interpretability and engineering credibility of the diagnostic results. Furthermore, it standardizes user queries through semantic rewriting and utilizes a large language model and multi-agent functional modules to complete collaborative game theory and interactive reasoning, generating high-speed train fault diagnosis reports that comply with safety regulations, have logical traceability, and multimodal display. This simulates the consultation and troubleshooting approach of senior human experts, greatly improving the automation level of fault diagnosis and the diagnostic accuracy in complex fault scenarios.

[0006] In one optional implementation, a knowledge base with electrical topology constraints is constructed based on multi-source heterogeneous operation and maintenance data, including: Multi-source heterogeneous operation and maintenance data is parsed and segmented to obtain standardized text fragments; these standardized text fragments are then stored in a vector database. Define the basic attribute information and relationship types corresponding to the electrical architecture characteristics of the traction and braking system of the EMU; Based on basic attribute information and relationship types, standardized text fragments are transformed into local logical topology subgraphs; By performing global knowledge integration processing on the local logical topology subgraphs, a global electrical logic topology network is obtained. A graph community discovery process is performed on the global electrical logic topology network to obtain graph functional community clusters, and functional community summary text is generated for the graph functional community clusters. A knowledge base with electrical topology constraints is constructed based on a vector database, a global electrical logic topology network, and functional community summary text.

[0007] This invention provides a fault diagnosis method for high-speed trains. By parsing and segmenting multi-source heterogeneous operation and maintenance data, it achieves standardized and searchable management of unstructured operation and maintenance data, solving the problems of data silos and chaotic formats in traditional methods. Secondly, by defining the electrical architecture characteristics of the train's traction and braking system, it transforms standardized text fragments into local logical topology subgraphs. Furthermore, by integrating global knowledge, it constructs a global electrical logical topology network, completely restoring the physical connections and signal flows of the electrical system. This overcomes the limitation of only being able to model local components, achieving global topology modeling from the component level to the system level. By constructing a knowledge graph containing electrical connection attributes, it transforms static knowledge into knowledge with physical causality. The graph logic ensures that the model can find faults by following the wires like a circuit expert. By discovering and processing graph communities, functional community clusters are obtained and functional community summaries are generated, realizing modular dimensionality reduction of complex electrical systems. This significantly reduces the computational complexity of subsequent fault diagnosis reasoning while preserving functional-level physical semantic information. Finally, a knowledge base with electrical topology constraints is constructed, which retains the efficiency of vector retrieval and introduces strong constraint verification of physical topology. This solves the pain points of poor interpretability and susceptibility to noise interference in data-driven diagnostic methods, providing a reliable reasoning basis and constraints for subsequent multi-agent collaborative diagnosis, and significantly improving the accuracy, comprehensiveness, and engineering interpretability of EMU fault diagnosis.

[0008] In one optional implementation, a global knowledge integration process is performed on the local logical topology subgraph to obtain a global electrical logic topology network, including: Entity disambiguation and global fusion are performed on the local logical topology subgraphs to obtain the global electrical topology graph. Obtain the binding relationship between components and image blocks. Based on the binding relationship between components and image blocks, bind image slice identifiers and labeled image Uniform Resource Locators to the nodes in the global electrical topology map to obtain the global electrical logic topology network.

[0009] The present invention provides a fault diagnosis method for high-speed trains, which adopts a logic-visual decoupling mapping, abandons the unstable real-time CAD parsing, and instead uses Image-ID indexing technology to call expert-level pre-annotated local slices, thereby achieving absolute determinism in visual positioning.

[0010] In one optional implementation, based on a standard multidimensional diagnostic summary, a multi-agent cooperative game is performed using a knowledge base with electrical topology constraints to obtain a train fault diagnosis report, including: Based on a knowledge base with electrical topology constraints, a retrieval expert agent is used to perform fault information retrieval and recall processing on standard multidimensional diagnostic summaries to obtain comprehensive fault context information. By using a reasoning expert intelligent agent to perform long-think-chain fault deduction processing on the comprehensive fault context information, a fault reasoning path is obtained. The compliance verification agent is used to perform operational compliance audit and logical consistency audit on the fault reasoning path to obtain the fault reasoning path that has passed compliance verification. Generate expert agents based on fault reasoning paths that have passed compliance verification; The expert intelligent agent is used to perform multimodal fault diagnosis on the target EMU and generate a fault diagnosis report for the EMU.

[0011] This invention provides a fault diagnosis method for high-speed trains. It utilizes a retrieval expert agent to retrieve and recall fault information from standard multidimensional diagnostic summaries. Employing a multi-path recall and multi-agent architecture, it eliminates the need for expensive full-scale model fine-tuning for each new train model. Only the underlying GraphRAG topology nodes and backend image assets need updating, enabling rapid migration of diagnostic capabilities and demonstrating strong vertical domain knowledge generalization and continuous iteration capabilities. Furthermore, it leverages a reasoning expert agent to perform long-chain fault deduction processing based on comprehensive fault context information, significantly improving the convergence efficiency and usefulness of fault investigation. A compliance verification agent performs operational compliance auditing and logical consistency auditing on the fault reasoning path, achieving standardization and automatic safety compliance auditing of the diagnostic process. Through the interplay of different agents—retrieval, reasoning, and compliance verification—it simulates the closed-loop process of human expert consultation. An independent compliance verification agent is introduced to conduct a secondary compliance audit of diagnostic suggestions, intercepting any operations that violate safety procedures.

[0012] In one optional implementation, based on a knowledge base with electrical topology constraints, a retrieval expert agent is used to perform fault information retrieval and recall processing on standard multidimensional diagnostic summaries to obtain comprehensive fault context information, including: The standard multidimensional diagnostic summary is converted into a high-dimensional vector, and semantic retrieval is performed on the vector database based on the high-dimensional vector to obtain historical maintenance data and unstructured expert experience text. Based on the standard multidimensional diagnostic summary, core fault-related components are extracted as seed nodes, and the functional community summary text corresponding to the seed nodes is retrieved. Using the seed node as the center, perform a multi-step neighbor traversal operation along the preset edge relationships in the global electrical logic topology network to obtain the electrical topology connection relationship; By splicing and integrating historical maintenance data, unstructured expert experience text, functional community summary text corresponding to seed nodes, and electrical topology connection relationships, comprehensive fault context information is obtained.

[0013] This invention provides a fault diagnosis method for high-speed trains. Firstly, by vectorizing standard multidimensional diagnostic summaries and performing semantic retrieval, it achieves precise matching of fault scenarios with historical maintenance data and expert experience texts, effectively reusing fragmented maintenance knowledge and expert experience. Secondly, by extracting core fault-related components as seed nodes and retrieving functional community summaries, it quickly identifies the electrical functional module to which the fault belongs, providing a clear physical starting point for subsequent topological reasoning and significantly improving the targeting of fault location. Thirdly, through multi-step neighbor traversal operations based on the global electrical logic topology network, it fully reconstructs the connection relationships and propagation paths of faulty components in the electrical system, achieving full-link tracing from single component faults to system-level impacts. Finally, by splicing and integrating multi-source information, it constructs comprehensive fault context information including historical data, expert experience, functional semantics, and topological relationships, providing a comprehensive, structured, and physically interpretable reasoning basis for subsequent multi-agent collaborative game diagnosis, significantly improving the comprehensiveness, accuracy, and robustness of fault diagnosis.

[0014] In one optional implementation, a reasoning expert agent is used to perform long-think-chain fault deduction processing on the comprehensive fault context information to obtain the fault reasoning path, including: By using an inference expert agent to extract comprehensive fault context information, multiple potential fault causes are obtained; For each potential fault cause, a logical evolution analysis is performed to obtain the fault logic reasoning link; If the fault logic reasoning link does not converge to the core fault cause, real-time measurement data is acquired and embedded into the fault logic reasoning link. The fault logic reasoning link is dynamically corrected by using the global electrical logic topology network until it converges to the core fault cause, and the fault logic reasoning link is used as the fault reasoning path.

[0015] The present invention provides a fault diagnosis method for high-speed trains that no longer outputs broad and vague maintenance suggestions, but can simulate the troubleshooting thinking of senior experts and provide granular guidance. Through an interactive guidance mechanism, the logical path is corrected in real time based on personnel feedback, which greatly shortens the average repair time.

[0016] Secondly, the present invention provides a fault diagnosis device for high-speed trains, the device comprising: The module is used to acquire multi-source heterogeneous operation and maintenance data of the target EMU and build a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data. The semantic rewriting module is used to obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary. The diagnostic module is used to obtain a fault diagnosis report for the EMU by using a knowledge base with electrical topology constraints to conduct multi-agent cooperative game based on standard multidimensional diagnostic summaries.

[0017] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, wherein the memory and the processor are communicatively connected to each other, the memory stores computer instructions, and the processor executes the computer instructions to perform the fault diagnosis method for high-speed trains described in the first aspect or any corresponding embodiment thereof.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the fault diagnosis method for a high-speed train according to the first aspect or any corresponding embodiment described above.

[0019] Fifthly, the present invention provides a computer program product, including computer instructions, which are used to cause a computer to execute the fault diagnosis method for a high-speed train according to the first aspect or any corresponding embodiment described above. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of an application scenario according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of a first method for diagnosing faults in a high-speed train according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the second process of a fault diagnosis method for a high-speed train according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the third process of a fault diagnosis method for a high-speed train according to an embodiment of the present invention. Figure 5 This is a structural block diagram of a fault diagnosis device for a high-speed train according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0024] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0025] As the core components of a high-speed train, the traction system provides power output, and the braking system ensures driving safety. These two systems integrate multiple technologies such as power electronics, microcomputer control, and pneumatic control, and their internal logic is extremely complex.

[0026] In current operation and maintenance practices, fault diagnosis and troubleshooting of high-speed trains mainly rely on front-line maintenance personnel. The onboard monitoring system uses TCMS (Train Control and Management System) / HMI (Human Machine Interaction). When the onboard monitoring system (TCMS / HMI) reports a fault code, maintenance personnel need to consult paper or electronic versions of the "Technical Maintenance Manual" (PDF format), which can easily be thousands of pages long, and compare it with the "Electrical Principle CAD Drawings." This approach faces the following prominent problems: Information retrieval and matching efficiency is extremely low: the causes of failures are often scattered in texts, tables and multiple electrical circuit diagrams in different chapters. Manual cross-document retrieval is not only time-consuming and labor-intensive, but also prone to missing key criteria under fatigue.

[0027] Expert experience is difficult to pass on effectively: In-depth logical troubleshooting of traction and braking systems often relies on the intuition and experience of senior technical experts. Newly hired or junior technicians often cannot quickly establish the cause-and-effect chain when faced with non-standard, cross-system coupled faults (such as traction power blocking triggered by grid voltage fluctuations), leading to over- or under-repair.

[0028] To address the aforementioned issues, Retrieval-Augmented Generation (RAG) technology was introduced to assist in maintenance. However, in the demanding field of rail transit, this approach has significant limitations: Lack of understanding of topological logic (semantic silo problem): Vector-based RAGs match fragments based on semantic similarity. However, the essence of electrical faults is the failure of current or signal conduction along the topological path. The relevant RAG can only recall discrete component descriptions and cannot understand the electrical logic topology between the pantograph, traction transformer, and traction converter, resulting in the model output suggestions lacking the support of physical laws.

[0029] The "illusion" risk of large language models: When generating maintenance instructions, large language models (LLMs) lack strong rule constraints in vertical domains, making it easy to fabricate seemingly professional but actually erroneous measurement points or thresholds. In the railway industry, where safety is paramount, this "illusion of fact" may lead to serious personal or equipment safety accidents.

[0030] The digital bottleneck of CAD circuit diagrams: When processing high-density vector CAD circuit diagrams, relevant computer vision recognition technologies are prone to misreading text (such as mistaking terminal number X10 for X70) or failing to track line number trajectories. The solution of real-time parsing of drawings and generation of visual annotations can hardly achieve 100% accuracy in engineering practice, and there is a huge risk of misleading.

[0031] The lack of a scientific industrial evaluation system: related large language model evaluations focus more on language fluency and lack quantitative evaluation standards for the completeness of diagnostic suggestions, whether the vehicle can be repaired, and the accuracy of the suggestions in maintenance scenarios, which makes it difficult for the system performance to pass industrial audits.

[0032] To address the industry pain points of high electrical logic coupling, unstable CAD circuit diagram recognition, and the susceptibility of large language models to logical illusions in the vertical field of rail transit, this invention provides an intelligent diagnostic solution for high-speed trains that combines advanced large language model generation capabilities with rigorous electrical topology logic and industrial safety red line constraints. This enables a leap from experience-driven to logic-driven intelligent maintenance of high-speed trains. The invention provides a fault diagnosis method for high-speed trains, utilizing GraphRAG (Graph-based Retrieval-Augmented Generation) and multi-agent collaboration for fault diagnosis. It employs a novel diagnostic paradigm that combines topology constraint logic, agent-simulated experts, decoupled mapping vision, and three-dimensional index evaluation. This deeply couples the general reasoning capabilities of large language models (LLM) with industry knowledge possessing physical topology constraints, resolving the challenges of logical illusions and diagram interpretation in the complex operating conditions of high-speed train traction and braking systems.

[0033] As an optional application scenario of this invention, such as Figure 1 As shown, the intelligent system for EMU condition diagnosis adopts a distributed or integrated diagnostic system architecture, including: a distributed knowledge storage module 101, a large language model inference engine 102, an image back-end storage server 103, a multi-agent collaborative processing module 104, and an interactive terminal 105.

[0034] The distributed knowledge storage module 101 includes a vector database 1011 (storing document embedding vectors) and a graph database 1012 (storing GraphRAG indexes with topological attributes). The distributed knowledge storage module 101 is connected to the multi-agent collaborative processing module 104 through a data bus, providing static factual evidence with physical law constraints for diagnosis, and is the underlying data source for suppressing "illusions".

[0035] The large language model inference engine 102 is deployed on the DeepSeek-V3 / R1 (large language model) parameter matrix of the GPU (Graphics Processing Unit) cluster. It is called by the multi-agent collaborative processing module 104 through the API (Application Programming Interface) to provide low-level natural language understanding and long thought chain reasoning capabilities, and to perform deep causal logic analysis.

[0036] The image back-end storage server 103 includes an independent file storage array that stores pre-annotated partial slices of CAD circuit diagrams and actual installation images of components. The image back-end storage server 103 is connected to the generating expert intelligent agent through the Image-ID index interface to achieve decoupling of logical conclusions and visual assets, thereby avoiding the high risk of misjudgment caused by virtual real-time analysis of drawings. The multi-agent collaborative processing module 104 (the system processing core) consists of multiple functionally decoupled agent runtime environments, including retrieval experts, reasoning experts, compliance verification experts, and generation experts. The multi-agent collaborative processing module 104 runs on the processor of a high-performance computing server. It interfaces with the large language model inference engine 102 at the upper layer and with the storage module at the lower layer. It is responsible for task routing, intent rewriting, logical deduction, security auditing, and multimodal data synthesis. It is the central hub that simulates the diagnostic thinking of human experts.

[0037] The interactive terminal 105 includes a computer and an interactive application running on it. It synchronizes data with the back-end server via a wireless network (5G / Wi-Fi), collects real-time measurement values ​​fed back by operators, and displays the final integrated graphic and text diagnostic report.

[0038] Among them, the heterogeneous knowledge base consisting of vector database 1011, graph database 1012, large language model inference engine 102 and image backend storage server 103 serves as the knowledge storage engine; the multi-agent (intelligent agent) collaborative scheduling center in the multi-agent collaborative processing module 104 is responsible for coordinating tasks such as retrieval enhancement, logical reasoning, and compliance filtering, and managing multi-round interaction states; the interactive terminal 105 adopts a multimodal human-computer interaction interface, supports maintenance personnel to input natural language, and displays the diagnostic report in the form of a combination of logical conclusions, labeled circuit diagrams and physical location maps.

[0039] According to an embodiment of the present invention, a fault diagnosis method for high-speed trains is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0040] This embodiment provides a fault diagnosis method for high-speed trains, which can be used in the aforementioned intelligent system for high-speed train condition diagnosis. Figure 2 This is a flowchart of a fault diagnosis method for high-speed trains according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: Obtain multi-source heterogeneous operation and maintenance data of the target EMU, and construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data.

[0041] Specifically, during the offline phase of the intelligent system for EMU condition diagnosis, a knowledge base with electrical topology constraints is constructed. Among them, the logic of constructing multi-source data logic is to protect how to transform EMU maintenance manuals, fault history and CAD circuit diagrams into a GraphRAG knowledge base with electrical topology attributes, and to achieve decoupling and binding of knowledge nodes and expert-level labeled visual assets through Image-ID.

[0042] Furthermore, the Retrieval Enhanced Generation Flow (RAGFlow) engine is used to uniformly access and differentiate the multi-source heterogeneous operation and maintenance data of the EMU. Multi-source heterogeneous data refers to materials with different sources, different formats, and close logical connections, including: PDF technical manuals of traction and braking systems (unstructured text), Excel fault code tables (semi-structured data), historical maintenance records and expert experience records (structured data), and CAD circuit schematics (graphical data).

[0043] Step S202: Obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary.

[0044] Specifically, relying on the large language model, the system performs a diagnostic element integrity assessment on the initial fault query information (conversational fault description / alarm code) input by maintenance personnel. If the large language model determines that the current input information lacks key diagnostic constraints, the system will proactively initiate multiple rounds of interactive follow-up questions to the maintenance personnel through the interactive terminal to obtain complete fault context information (such as vehicle model, batch number, software version, mileage, fault condition, etc.).

[0045] Furthermore, after acquiring and integrating all supplementary information, the system filters out environmental noise words and redundant interjections in multi-turn dialogues using a large language model. It then merges and standardizes the ambiguous business language, detailed vehicle history, and fault condition parameters to generate a standard multidimensional diagnostic summary (example format: "CR400BF EMU, BCU software version V2.1, 300,000 km in operation, under normal braking conditions, after the BCU issues a relief command, the brake cylinder pressure sensor feedback value does not drop below the threshold within the specified time"). This diagnostic summary is then transmitted to the message bus of its built-in multi-agent collaborative scheduling center (i.e., multi-agent collaborative processing module) and the expert agent is activated for retrieval.

[0046] Step S203: Based on the standard multidimensional diagnostic summary, a multi-agent cooperative game is conducted using a knowledge base with electrical topology constraints to obtain a fault diagnosis report for the EMU.

[0047] Specifically, the multi-agent collaborative game adopts a task flow mechanism among four types of expert agents: retrieval, reasoning, compliance verification, and generation.

[0048] Furthermore, by acquiring fault query information input by maintenance personnel, and relying on the built-in large language model and multi-agent collaborative processing module, collaborative game and interactive reasoning are completed to generate a high-speed train fault diagnosis report that complies with safety procedures, has logical traceability, and multimodal display, simulating the consultation and troubleshooting thinking of senior human experts.

[0049] This embodiment provides a fault diagnosis method for high-speed trains. By constructing a knowledge base with electrical topology constraints, it associates equipment-level faults with system-level impact paths in a modeling manner. This avoids misjudgments caused by data gaps and noise interference in data-driven models, significantly improving the physical interpretability and engineering credibility of the diagnostic results. Furthermore, it standardizes user queries through semantic rewriting and utilizes a large language model and multi-agent functional modules to complete collaborative game theory and interactive reasoning. This generates a high-speed train fault diagnosis report that complies with safety regulations, has logical traceability, and multimodal display. It simulates the consultation and troubleshooting approach of senior human experts, greatly improving the automation level of fault diagnosis and the diagnostic accuracy in complex fault scenarios.

[0050] This embodiment provides a fault diagnosis method for high-speed trains, which can be used in the aforementioned intelligent system for high-speed train condition diagnosis. Figure 3 This is a flowchart of a fault diagnosis method for high-speed trains according to an embodiment of the present invention, such as... Figure 3 As shown, the process includes the following steps: Step S301: Obtain multi-source heterogeneous operation and maintenance data of the target EMU, and construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data.

[0051] Specifically, step S301 includes: Step S3011: Parse and segment the multi-source heterogeneous operation and maintenance data to obtain standardized text fragments; wherein, the standardized text fragments are stored in a vector database.

[0052] Specifically, the parsing strategies for multi-source heterogeneous operation and maintenance data in different formats are as follows: The Traction and Braking System PDF Technical Manual utilizes a deep parsing algorithm based on visual layout recognition to remove headers, footers, and irrelevant watermarks, accurately identifying heading levels, paragraphs, tables, and image positions. It employs a semantic boundary segmentation strategy, treating complete equipment descriptions / operating procedures as a single text block. Warning / attention-related strong safety constraints are extracted and converted into independent text blocks, tagged with high-security-priority metadata. All text slices automatically inherit the chapter title as a contextual prefix, preventing loss of context.

[0053] Excel Fault Code Table: Combine single-row and column fault codes into a complete statement according to a preset template. Example format: "When the fault code is XXXX, the alarm name is traction converter fault, and the triggering logic condition is that the intermediate DC link voltage is lower than XXX volts."

[0054] Historical maintenance records and expert experience records: Extract discrete fields from work orders in the database (occurrence time, assigned section, vehicle type, fault symptoms, located parts, handling measures, expert cause analysis, etc.) and concatenate them to rewrite them into coherent expert maintenance log-style text fragments; convert the text into high-dimensional vectors (Embedding) and store them in a vector database.

[0055] CAD circuit schematic diagram: Extract the text number of the components (such as relay K1, contactor KM2, etc.) in the drawing as the index key and bind it to the storage path of the corresponding local image block; experts preprocess the connection relationship of the components and convert it into descriptive text.

[0056] Furthermore, the text or high-dimensional vectors obtained from parsing the multi-source heterogeneous operation and maintenance data of different formats are stored in a vector database.

[0057] Step S3012: Define the basic attribute information and association types corresponding to the electrical architecture characteristics of the EMU traction and braking system.

[0058] Specifically, based on the electrical architecture characteristics of the EMU traction and braking system, standardized definition operations for ontological attributes and relationships are performed to set unified attribute rules and relationship paradigms for subsequent electrical topology GraphRAG modeling. Among them, basic attribute information includes component types (such as core components, sensors, relays, logic gates, etc.) and component functions (such as pantograph raising / lowering power supply function, relay contact on / off control function, etc.). Relationship types include four core relationships that can be extracted in subsequent graph construction: power supply connection, signal control, data communication, and fault causality. At the same time, unified attributes are also defined for fault-related objects, including fault phenomena, alarm code type classification, trigger condition attributes, etc.

[0059] Furthermore, by defining the ontological attributes of the EMU electrical domain as described above, hard attribute and relation constraint standards are set for the subsequent large language model to perform structured information extraction and generate entity-relation-entity triples. This ensures that all information extracted from the text conforms to the actual characteristics of the EMU electrical and logical architecture, avoids the generation of meaningless and generalized open domain entities in the large model, and provides a unified definition basis for the subsequent node creation and edge relationship establishment of the GraphRAG graph, ensuring the standardization and consistency of graph construction.

[0060] Step S3013: Based on the basic attribute information and the type of association, the standardized text fragments are transformed into local logical topology subgraphs.

[0061] Specifically, structured information extraction based on a large language model is performed, transforming natural language into entity-relation-entity triples. The specific steps are as follows: standardized text fragments are input into the large language model one by one, supplemented with targeted prompts, and the model decomposes them into nodes and edges; at the same time, the model is forced to generate only four specific edge relations according to the industrial semantic ontology framework, namely: power supply connection edge: describing the current energy flow from the pantograph to the transformer; signal control edge: extracting the hard-wired action logic of relay contacts in the text; data communication edge: extracting the multifunction vehicle bus (MVB) interaction instructions between microcomputer control units (such as traction control unit and brake control unit); fault causal edge: establishing the causal connection between sensor abnormal conditions and the final triggered fault code.

[0062] Furthermore, based on the nodes, edges, and specific edge relationships obtained above, a local logical topology subgraph is constructed, which realizes the transformation of independent text slices into local logical topology subgraphs. The local logical topology subgraph is used as the electrical topology GraphRAG, laying the foundation for the construction of the global graph.

[0063] Step S3014: Perform global knowledge integration processing on the local logic topology subgraph to obtain the global electrical logic topology network.

[0064] In some optional implementations, step S3014 above includes: Step a1: Perform entity disambiguation and global fusion processing on the local logical topology subgraph to obtain the global electrical topology map.

[0065] Specifically, due to the different sources of standardized text fragments, the extracted local logical topology subgraphs have problems of isolation and naming redundancy (such as "traction converter", "TCU" and "converter device" all referring to the same device). Therefore, entity disambiguation and knowledge fusion processing are performed on the local logical topology subgraphs. The specific steps include: comparing entity semantics across text blocks, merging nodes that refer to the same physical device, and using core control modules such as TCU (Telematics Control Unit) and BCU (Brake Control Unit) as hubs to assemble discrete triples into a global electrical topology map covering the entire vehicle system.

[0066] Step a2: Obtain the binding relationship between components and image blocks. Based on the binding relationship between components and image blocks, bind image slice identifiers and labeled image Uniform Resource Locators to the nodes in the global electrical topology map to obtain the global electrical logic topology network.

[0067] Specifically, to solve the problem of real-time accurate identification of complex CAD circuit diagrams, a logic-visual decoupling mapping is performed synchronously. The specific steps are as follows: the binding relationship between components and image blocks obtained in step S3011 is integrated into the global electrical topology map. Each node is bound with a corresponding image slice ID (Image-ID) and a Uniform Resource Locator (URL) for the labeled image. This achieves index-based parsing, enabling online diagnosis to instantly call the visualized labeled slice through the index ID, thereby obtaining the global electrical logic topology network and storing it in the graph database.

[0068] Because real-time AI analysis of vector CAD drawings suffers from fatal flaws such as low recognition rates and easy confusion of line numbers, the aforementioned optional implementation adopts a logic-visual decoupling mapping strategy that uses indexes instead of parsing. Through Image-ID binding in the preprocessing stage, it ensures that after diagnosing a fault, the system consistently pushes high-precision circuit slices and physical installation diagrams verified by business experts to frontline personnel. This approach avoids the error risks of algorithmic parsing, achieves 100% determinism in visual guidance, and significantly reduces the cognitive load on maintenance personnel during on-site operations.

[0069] Step S3015: Perform graph community discovery processing on the global electrical logic topology network to obtain graph functional community clusters, and generate functional community summary text for the graph functional community clusters.

[0070] Specifically, the Leiden graph clustering algorithm (a graph clustering algorithm for community detection) is used to perform graph community detection processing on the global electrical logic topology network. That is, based on the edge connection density in the global electrical logic topology network, physically / logically closely connected node groups are divided into different graph functional community clusters (such as the pantograph lifting logic community). Then, a large language model is called to traverse all nodes and edges in each community and generate a high-level summary text for each graph functional community cluster. This enables the global electrical logic topology network to have both micro-physical connection relationships and macro-semantic capabilities, improving the efficiency and global relevance of subsequent retrieval.

[0071] Furthermore, the specific steps for graph community discovery processing of the global electrical logic topology network include: fully loading all entity nodes (electrical components, control units, sensors, relays, etc.) and four types of relationship edges (power supply connection, signal control, data communication, fault causation) of the global electrical logic topology network, and constructing a complete graph structure calculation matrix; quantifying and statistically analyzing the number of edge connections, association frequency, and topological coupling degree between each node, using local edge connection density as the core partitioning criterion to measure the strength of physical wiring associations and logical control associations between different electrical nodes; and using the Leiden algorithm's local movement iteration mechanism to perform graph community discovery on each of the scattered nodes. Belonging Traversal: Closely connected and highly coupled neighboring nodes are initially aggregated into several small initial node clusters, completing the first round of coarse-grained community aggregation; Based on the principle of maximizing modularity, continuous iterative optimization is carried out: unreasonable cross-domain nodes are split, highly coupled small clusters are merged, and isolated nodes with weak logical connections and cross-subsystems are removed to ensure that the connections within the same community are tight and the boundaries between communities are clear; According to the dual dimensions of physical installation partitioning and electrical functional logic, independent and semantically complete functional community clusters are finally divided, such as pantograph lifting logic community, traction converter control community, brake pressure closed-loop control community and other professional subsystem communities.

[0072] Furthermore, after completing the functional community cluster division, the large language model is automatically invoked to perform a full-domain information traversal reading for each graph functional community cluster. The model fully reads all the topological elements contained within a single community, including all device entity nodes, four types of topological connection relationships, component functional attributes, interaction logic, and fault association rules. Based on this, the large language model performs semantic condensation and integration of fragmented topological details and scattered device interaction relationships within the community, eliminates redundant fine-grained information at the underlying level, and provides a unified summary around the core role, operating logic, control flow, and association scope of the community. Finally, a high-level semantic summary text is generated for each graph functional community cluster.

[0073] Step S3016: Construct a knowledge base with electrical topology constraints based on the vector database, the global electrical logic topology network, and the functional community summary text.

[0074] Specifically, a heterogeneous knowledge base is formed by a vector database, a graph database, a large language model inference engine, and an image backend storage server.

[0075] Step S302: Obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary. For details, please refer to [link to relevant documentation]. Figure 2 Step S202 of the illustrated embodiment will not be described again here.

[0076] Step S303: Based on the standard multidimensional diagnostic summary, multi-agent cooperative game is performed using a knowledge base with electrical topology constraints to obtain the EMU fault diagnosis report. For details, please refer to [link to relevant documentation]. Figure 2 Step S203 of the illustrated embodiment will not be described again here.

[0077] This embodiment provides a fault diagnosis method for high-speed trains. By parsing and segmenting multi-source heterogeneous operation and maintenance data, it achieves standardized and searchable management of unstructured operation and maintenance data, solving the problems of traditional data silos and chaotic formats. Secondly, by defining the electrical architecture characteristics of the high-speed train's traction and braking system, standardized text fragments are transformed into local logical topology subgraphs. A global electrical logical topology network is constructed through global knowledge integration, fully restoring the physical connection relationships and signal flow of the electrical system. This overcomes the limitation of only being able to model local components, achieving global topology modeling from the component level to the system level. By constructing a knowledge graph containing electrical connection attributes, static knowledge is transformed into knowledge with physical causality. The graph logic ensures that the model can find faults by following the wires, just like a circuit expert. By discovering and processing graph communities, functional community clusters are obtained and functional community summaries are generated, realizing modular dimensionality reduction of complex electrical systems. This significantly reduces the computational complexity of subsequent fault diagnosis reasoning while preserving functional-level physical semantic information. Finally, a knowledge base with electrical topology constraints is constructed, which retains the efficiency of vector retrieval and introduces strong constraint verification of physical topology. This solves the pain points of poor interpretability and susceptibility to noise interference in data-driven diagnostic methods, providing a reliable reasoning basis and constraints for subsequent multi-agent collaborative diagnosis, and significantly improving the accuracy, comprehensiveness, and engineering interpretability of EMU fault diagnosis.

[0078] This embodiment provides a fault diagnosis method for high-speed trains, which can be used in the aforementioned intelligent system for high-speed train condition diagnosis. Figure 4 This is a flowchart of a fault diagnosis method for high-speed trains according to an embodiment of the present invention, such as... Figure 4 As shown, the process includes the following steps: Step S401: Obtain multi-source heterogeneous operation and maintenance data of the target EMU, and construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data. For details, please refer to [link to relevant documentation]. Figure 3 Step S301 of the illustrated embodiment will not be described again here.

[0079] Step S402: Obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary.

[0080] Specifically, after maintenance personnel input a fault description or upload an alarm code through the interactive terminal, the fault query information is captured, and then the built-in large language model and multi-agent collaborative diagnosis process is initiated.

[0081] Step S403: Based on the standard multidimensional diagnostic summary, a multi-agent cooperative game is conducted using a knowledge base with electrical topology constraints to obtain a fault diagnosis report for the EMU.

[0082] Specifically, step S403 includes: Step S4031: Based on the knowledge base with electrical topology constraints, the retrieval expert agent is used to perform fault information retrieval and recall processing on the standard multidimensional diagnostic summary to obtain comprehensive fault context information.

[0083] Specifically, after receiving the diagnostic summary through the retrieval expert agent, the retrieval expert agent initiates a dual-path retrieval strategy. Relying on its own underlying vector database and graph database, the retrieval expert agent performs in-depth fault information mining, retrieval, and recall operations, providing comprehensive fault context information such as macro-level community summaries, micro-level connectivity relationships, and historical similar cases for subsequent fault reasoning.

[0084] The retrieval agent performs graph-based topology path detection and multi-strategy recall; the inference agent utilizes the long chain of thought (CoT) capability of the large AI model, combining real-time on-site measurements for interactive troubleshooting and deduction; the compliance verification agent performs security audits and corrections on inference instructions based on a pre-set SOP (Standard Operating Procedure) rule base, i.e., it is responsible for verifying the consistency of maintenance procedures and intercepting conclusions that violate electrical and physical laws; and the generation agent performs standardized report synthesis and dynamically embeds visual guidance evidence, synthesizing and displaying textual investigation plans with retrieved Image-ID associated images.

[0085] In some optional implementations, step S4031 above includes: Step b1: Convert the standard multidimensional diagnostic summary into a high-dimensional vector, and perform semantic retrieval on the vector database based on the high-dimensional vector to obtain historical maintenance data and unstructured expert experience text.

[0086] Specifically, by retrieving expert intelligent agents, the fault phenomena in the standard multidimensional diagnostic summary are transformed into high-dimensional vectors. The cosine similarity between this vector and the data in the vector database is calculated. Based on the similarity ranking, the Top-K most similar historical maintenance data and unstructured expert experience texts are recalled to provide historical statistical probability reference support for fault reasoning.

[0087] Step b2: Extract core fault-related components as seed nodes based on standard multidimensional diagnostic summaries, and retrieve the functional community summary text corresponding to the seed nodes.

[0088] Specifically, by retrieving expert intelligent agents, core fault-related components are extracted from standard multidimensional diagnostic summaries as seed nodes. The functional community logical summary (i.e., high-level semantic summary text) to which the seed node belongs is retrieved from the graph database. This provides the macroscopic operating logic of the corresponding subsystem for fault reasoning. In other words, relying on the functional community system divided by Leiden clustering in the offline stage, the electrical functional community to which the seed node belongs is quickly matched, and the high-level logical summary text pre-generated by the community is directly retrieved. This allows for the rapid acquisition of the overall architecture, working principle, and collaborative control logic of the corresponding traction and braking subsystems, as well as other macroscopic operating rules, thus building a top-level understanding of fault analysis.

[0089] Step b3: Using the seed node as the center, perform a multi-step neighbor traversal operation along the preset edge relationships in the global electrical logic topology network to obtain the electrical topology connection relationship.

[0090] Specifically, with the seed node as the center, a multi-step neighbor traversal operation is performed along four preset edge relationships: power supply, signal, communication, and fault causation. This retrieves and obtains the microscopic electrical topology connection relationships related to the core component. That is, with the selected seed node as the central origin, the system strictly follows the preset four types of topology edge relationships: power supply connection, signal control, data communication, and fault causation. Multi-hop neighbor traversal retrieval is carried out in the global electrical logic topology network. By extending the node association range layer by layer, the system traverses the upstream and downstream related components, control loops, signal links, and fault propagation paths level by level. This comprehensively explores and obtains the microscopic and refined electrical topology connection relationships of the upstream and downstream of the core fault component, such as hardware wiring, command interaction, signal transmission, and fault linkage, thus filling in the details of the underlying physical and logical coupling between devices.

[0091] Step b4 involves concatenating and integrating historical maintenance data, unstructured expert experience text, functional community summary text corresponding to seed nodes, and electrical topology connection relationships to obtain comprehensive fault context information.

[0092] Specifically, the dual-path retrieval results completed by the retrieval expert agent are spliced ​​and integrated to form comprehensive fault context information, and this information is pushed to the built-in reasoning expert agent.

[0093] In the above optional implementations, by vectorizing the standard multidimensional diagnostic summary and performing semantic retrieval, accurate matching of fault scenarios with historical maintenance data and expert experience texts is achieved, effectively reusing fragmented maintenance knowledge and expert experience, and solving the problem that fault diagnosis in the prior art relies on fixed rules and is difficult to adapt to complex scenarios. Secondly, by extracting core fault-related components as seed nodes and retrieving functional community summaries, the electrical functional module to which the fault belongs is quickly identified, providing a clear physical starting point for subsequent topology reasoning and significantly improving the targeting of fault location. Thirdly, through multi-step neighbor traversal operations based on the global electrical logic topology network, the connection relationship and propagation path of the faulty component in the electrical system are fully restored, realizing full-link tracing from single component faults to system-level impacts, solving the pain point of difficulty in analyzing the impact of cross-component faults in the prior art. Finally, by splicing and integrating multi-source information, a comprehensive fault context information including historical data, expert experience, functional semantics, and topological relationships is constructed, providing a comprehensive, structured, and physically interpretable reasoning basis for subsequent multi-agent collaborative game diagnosis, significantly improving the comprehensiveness, accuracy, and robustness of fault diagnosis.

[0094] Step S4032: Use the reasoning expert agent to perform long thought chain fault inference processing on the comprehensive fault context information to obtain the fault reasoning path.

[0095] Specifically, after receiving comprehensive fault context information, the reasoning expert agent utilizes the long thought chain capability of the large language model to initiate a hypothesis-verification cyclical game fault deduction operation. Its core function is to simulate the troubleshooting thinking of a senior expert and achieve dynamic convergence of the fault reasoning path through interactive guidance.

[0096] In some optional implementations, step S4032 above includes: Step c1 involves using an inference expert agent to extract comprehensive fault context information and obtain multiple potential fault causes.

[0097] Specifically, the reasoning expert agent, based on the electrical topology relationships and fault causal links contained in the comprehensive fault context information, combined with the control logic and equipment coupling rules of the EMU's traction and braking systems, diverges and disassembles multiple fault propagation paths at the system level, component level, and part level, focusing on the current fault phenomenon, alarm type, and operating conditions. It comprehensively lists and outputs multiple reasonable potential fault causes, avoiding simplistic and one-sided assumptions, and covering all high-probability fault sources.

[0098] Step c2 involves performing logical evolution analysis on each potential fault cause to obtain the fault logic reasoning link.

[0099] Specifically, the reasoning expert agent proposes multiple potential fault causes based on the fault causal chain, and relies on a large model to simulate the expert thinking mode to conduct logical evolution analysis on each cause. That is, for each hypothetical fault cause (i.e., potential fault cause), the entire evolution path of component abnormality generation—signal command abnormality—loop logic failure—final fault phenomenon triggering is completely deduced. Combined with electrical topology constraints, equipment working mechanism and control interaction rules, the occurrence conditions, mechanism of action and phenomenon matching degree of each cause are demonstrated one by one, and the systematic deduction and feasibility assessment of all fault hypotheses are completed.

[0100] Step c3: If the fault logic reasoning link does not converge to the core fault cause, then obtain real-time measurement data and embed the real-time measurement data into the fault logic reasoning link.

[0101] Specifically, if the reasoning expert agent determines that the existing topological facts are insufficient to uniquely determine the cause of the fault, the computer generates interactive operation instructions through the agent and displays them to the maintenance personnel through the front-end interface, prompting the maintenance personnel to perform the specified on-site measurement operation (such as measuring the voltage of pin 5 of plug X102).

[0102] Furthermore, after completing the multi-causal logical evolution comparison, the reasoning expert agent horizontally compares multiple inference links: if the evolution logic of multiple fault hypotheses all conforms to the existing topological relationships, control rules and fault representations, and the current graph facts, case data and text information lack key differentiating conditions that can achieve exclusive screening, and it is impossible to screen out a unique root cause from multiple reasonable hypotheses, then it is determined that there are blind spots in the existing information and the fault cause cannot be accurately located. At this time, the reasoning expert agent automatically generates targeted and implementable on-site detection instructions, accurately specifying the measurement points, detection objects and monitoring parameters. The instructions are intuitively displayed to maintenance personnel through the interactive terminal, guiding them to carry out targeted and precise on-site measurement operations, and supplementing the physical measurement data required for reasoning.

[0103] Step c4: Use the global electrical logic topology network to dynamically correct the fault logic reasoning link until it converges to the core fault cause, and use the fault logic reasoning link as the fault reasoning path.

[0104] Specifically, after receiving real-time on-site measurement data from maintenance personnel, the data is immediately injected into the fault logic reasoning chain of the reasoning expert agent, and linked with the GraphRAG knowledge base (i.e., graph database) for analysis, dynamically correcting the fault reasoning path until it converges to the core fault cause.

[0105] Furthermore, after collecting real measurement data such as voltage, pressure, on / off signals, and module status from maintenance personnel on-site, the objective measured parameters are used as rigid constraints and embedded in the long-chain fault logic reasoning link of the reasoning expert agent in real time. This achieves a deep integration of the actual on-site working conditions and the theoretical reasoning process. Then, the injected measured data is linked and verified with the electrical topology GraphRAG knowledge base built in the offline stage. Matching analysis is carried out against standard electrical parameter thresholds, loop on / off logic, signal transmission specifications, and normal operating characteristics of components to verify whether the measured results meet the normal operating boundary conditions of the equipment.

[0106] Furthermore, based on the verification results of measured objective data, erroneous fault assumptions and invalid deduction branches that do not conform to actual working conditions are automatically eliminated, the original reasoning direction and investigation priority are corrected and optimized, the scope of suspected faults is continuously compressed, and through the closed-loop iteration of assumption deduction, on-site supplementary testing, data verification and path correction, the reasoning boundary is continuously converged, non-related fault factors are gradually eliminated, and finally the unique and real core fault cause is locked and determined, thus completing the accurate location of the root cause of the fault.

[0107] In the above optional implementation, instead of outputting broad and vague maintenance suggestions (such as "check the relay"), it can simulate the troubleshooting process of a senior expert and provide granular guidance that is precise to "terminal blocks, wire numbers, pins" and specific "voltage / resistance measurement thresholds". Through an interactive guidance mechanism, the logic path is corrected in real time based on personnel feedback. Experimental data shows that the system achieves a decision convergence gain of 0.91 in assisted diagnosis. Compared with the manual manual mode, the average number of steps to locate the root cause is reduced by about 40% to 60%, which greatly shortens the average repair time and significantly improves the convergence efficiency and usefulness of fault diagnosis.

[0108] Step S4033: Use the compliance verification agent to perform operational compliance auditing and logical consistency auditing on the fault reasoning path to obtain the fault reasoning path that has passed compliance verification.

[0109] Specifically, by verifying the compliance of the intelligent agent, mandatory safety and authenticity audits are performed on each step of the causal chain in the fault reasoning process output by the reasoning expert intelligent agent. The core function is to intercept fault reasoning conclusions that violate safety procedures and electrical and physical laws, thereby eliminating the risk of logical illusion in large model reasoning from the process perspective. The specific audit operations are divided into operational compliance audit and logical consistency audit.

[0110] Furthermore, the specific steps of the operational compliance audit are as follows: The compliance verification agent compares the fault reasoning suggestions given by the reasoning expert agent with the preset EMU safety operation procedures and examines whether they include prerequisite safety actions such as "lowering the pantograph, disconnecting power, and checking for power". If they are found to be missing, the current fault reasoning conclusion is forcibly rejected. After the reasoning conclusion is rejected, the compliance verification agent continuously monitors the feedback from the reasoning expert agent and urges it to re-optimize the fault reasoning suggestions. If the reasoning suggestions fully include all necessary prerequisite safety operations and the order of operations conforms to the EMU safety operation procedures, then the operation of this step is deemed compliant and proceeds to the next level of logical consistency audit.

[0111] Furthermore, the specific steps of the logical consistency audit process are as follows: The compliance verification agent fully extracts the entire fault reasoning process chain output by the reasoning expert agent, including all details such as fault cause assumptions, electrical logic evolution analysis, voltage / current direction derivation, and fault causal transmission relationship. It focuses on dissecting key contents such as electrical parameters (voltage, current values, etc.), loop connection relationships, and signal interaction logic involved in the reasoning process, transforming the abstract reasoning text into a quantifiable and verifiable electrical logic model. Through the compliance verification agent, it calls relevant algorithms based on basic electrical laws (such as Kirchhoff's current law) to verify whether the voltage / current direction deduced by the reasoning expert agent violates physical common sense. If a logical jump / error is detected in the reasoning process, the computer generates feedback prompts containing "logical contradictions" through the agent, driving the reasoning expert agent to review and optimize the fault reasoning chain.

[0112] Furthermore, upon receiving feedback prompts, the reasoning expert agent initiates a review of the reasoning chain. For the logical contradictions indicated, it uses fundamental electrical laws and the GraphRAG knowledge base to correct erroneous voltage / current flow derivations, supplement missing logical derivation steps, and adjust fault propagation relationships. Simultaneously, the compliance verification agent performs a secondary verification of the corrected reasoning chain. This process is repeated until the reasoning process is free of logical errors and logical jumps, and fully conforms to the laws of electrical physics.

[0113] Furthermore, the compliance verification agent continuously executes the aforementioned adversarial verification operations until the fault reasoning conclusion output by the reasoning expert agent meets the dual requirements of security compliance and logical consistency.

[0114] The aforementioned compliance verification agent can automatically check whether the generated maintenance steps cover the safety red lines of EMU operation such as "lowering the pantograph, disconnecting the power, and grounding". If the steps generated by the reasoning expert agent lack necessary safety protection measures, the compliance verification agent will automatically intercept and force the completion. This mechanism upgrades the safety review from post-event manual verification to pre-event automatic verification, ensuring that the generated report has a completeness score of over 0.95, thus protecting personal and equipment safety.

[0115] Step S4034: Generate an expert intelligent agent based on the fault reasoning path that has passed compliance verification.

[0116] Specifically, after the fault reasoning conclusion of the reasoning expert agent passes the compliance verification, the generation of the expert agent is initiated. This agent then performs the multimodal fault diagnosis report synthesis operation. Its core function is to transform the fault logic reasoning chain into a fault diagnosis solution that conforms to industrial standards, is highly readable, and is visual, making it easier for maintenance personnel to perform operations on-site.

[0117] Step S4035: Use an expert intelligent agent to perform multimodal fault diagnosis on the target EMU and generate a fault diagnosis report for the EMU.

[0118] Specifically, by generating an expert intelligent agent, the Image-IDs corresponding to the components involved in the fault reasoning path are extracted. Based on the Image-ID, the pre-stored expert-level annotated circuit diagrams are accurately retrieved from the image backend storage server. Based on the expert intelligent agent, the conclusive text descriptions, complete long thought chain fault reasoning trajectories (explaining why this judgment is made) and highlighted circuit slice diagrams (indicating where to repair) are integrated to finally generate a structured multimodal fault diagnosis scheme, namely, the EMU fault diagnosis report.

[0119] This embodiment provides a fault diagnosis method for high-speed trains, which introduces GraphRAG technology with electrical topology constraints and an independent compliance verification agent. GraphRAG uses physical laws of the train's traction and braking system to rigidly constrain the retrieval process, requiring inference experts to follow electrical topology logic when generating conclusions. Combined with secondary auditing by the verification agent, the system successfully controls the illusion rate of generative large models in industrial scenarios from over 10% to less than 0.5%, ensuring the inherent security of the diagnostic results and avoiding the shortcomings of related retrieval enhancement generation schemes, which can only perform semantic similarity matching and cannot understand the circuit flow.

[0120] The following specific embodiments will be used to verify a fault diagnosis method for high-speed trains.

[0121] Example 1: To address the fault diagnosis problem of the traction and braking system of high-speed trains, various fault diagnosis solutions are provided and compared in depth with the embodiments of this invention through experimental analysis. The specific details and limitations are analyzed as follows: Domain-wide large model based on supervised fine-tuning (SFT): This approach attempts to conduct supervised deep fine-tuning of a general large language model by collecting massive amounts of EMU maintenance cases, regulations, and expert experience, so as to internalize the professional knowledge of the railway vertical domain.

[0122] Limitations Analysis: "Knowledge Black Box" and Lack of Interpretability: The diagnostic logic of the fine-tuned model is hidden in the parameter weights, and cannot provide a clear topological evidence chain like GraphRAG in this embodiment. This does not meet the strict auditing requirements of railway operation and maintenance for attribution and traceability of diagnostic conclusions; Uncontrollability of the Illusion Problem: Although fine-tuning can improve the professional language, it cannot fundamentally eliminate the model's arbitrary fabrication when the logic is complex. When high-voltage equipment such as traction converters are involved, this uncontrollable illusion will bring fatal safety risks; Extremely High Knowledge Update Costs: When the EMU undergoes hardware modification or maintenance procedures are revised, the fine-tuned model needs to re-execute the expensive training process, while the RAG architecture used in this embodiment only needs to update the underlying graph nodes, which has greater engineering flexibility.

[0123] The solution is based on a hard-coded rule engine and a traditional knowledge graph: This solution uses graph databases such as Neo4j to store component relationships and performs fault logic retrieval through preset "If-Then" expert rules.

[0124] Limitations Analysis: Extremely poor ability to handle non-standard faults: The rule engine can only handle predefined fault scenarios. Faced with complex, multi-factor-laden non-standard operating conditions encountered in the maintenance of the Fuxing bullet train, hard-coded rules often fail to cover all branches, leading to retrieval failures. Insufficient user-friendliness: This solution cannot understand the colloquial fault descriptions of frontline personnel and lacks the natural language processing and intent rewriting capabilities provided by the large language model in this embodiment, resulting in a significant technical hurdle for human-computer interaction. Lack of logical convergence capability: The correlation graph search can only provide paths and cannot dynamically correct logical chains based on on-site measurements like the reasoning expert agent in this embodiment.

[0125] A solution based on real-time analysis of CAD circuit diagrams using computer vision: This solution attempts to use OCR (Optical Character Recognition) and large-scale visual models to scan, recognize, and annotate vector or bitmap circuit diagrams retrieved on-site in real time.

[0126] Limitations Analysis: Recognition accuracy cannot reach industrial red lines: The circuit diagrams of high-speed trains are extremely detailed, and wire numbers and terminal numbers (such as X101 and X107) are easily misread under low light or high compression ratios. Experiments have shown that the accuracy of computer vision technology fluctuates greatly when processing such high-density vector graphics; Disconnection from logical understanding: Recognizing lines and text does not mean understanding the electrical logic. Real-time parsing solutions often cannot deeply align the symbols in the image with the semantic knowledge in the maintenance manual.

[0127] The Image-ID visual asset decoupling mapping scheme used in this embodiment perfectly avoids the randomness errors of real-time parsing through pre-annotation + index calling, ensuring 100% determinism of visual guidance.

[0128] The solution based on Standard Vector RAG: uses semantic vector similarity matching to match user queries with text blocks in the repair manual.

[0129] Limitations Analysis: Semantic silo effect: Standard RAG can only find similar descriptive fragments, but cannot see the physical connections between electrical components. For example, it can find the description of the traction motor, but may miss the status of the current transformer upstream of its power supply; Loss of causal inference capability: Electrical fault diagnosis is essentially a causal flow. Standard RAG lacks topological constraints and cannot generate diagnostic reports with a tight causal chain.

[0130] The GraphRAG (Graph Augmented Retrieval) introduced in this embodiment fills this gap by providing a global perspective through graph community summaries, which is something that standard vector retrieval cannot achieve.

[0131] In summary, the integrated solution of GraphRAG topological constraints, multi-agent collaboration, and visual asset decoupling mapping proposed in this embodiment is the best path after balancing industrial safety, engineering implementation difficulty, and logical accuracy. It not only leverages the intuitive reasoning of large language models but also adds physical locks by utilizing graphs and verification agents, making it the most practical and valuable technical solution in the field of high-speed train fault diagnosis.

[0132] This embodiment also provides a fault diagnosis device for high-speed trains, which is used to implement the above embodiments and preferred embodiments; details already described will not be repeated. As used below, the term "module" can refer to a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0133] This embodiment provides a fault diagnosis device for high-speed trains, such as... Figure 5 As shown, it includes: Module 501 is used to acquire multi-source heterogeneous operation and maintenance data of the target EMU and build a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data. Semantic rewriting module 502 is used to obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary. The diagnostic module 503 is used to obtain a fault diagnosis report for the EMU by using a knowledge base with electrical topology constraints to conduct multi-agent cooperative game based on standard multidimensional diagnostic summaries.

[0134] In some alternative implementations, the construction module 501 includes: The parsing unit is used to parse and divide multi-source heterogeneous operation and maintenance data into blocks to obtain standardized text fragments; the standardized text fragments are then stored in a vector database. The definition unit is used to define the basic attribute information and association types corresponding to the electrical architecture characteristics of the traction and braking system of the EMU; The transformation unit is used to transform standardized text fragments into local logical topology subgraphs based on basic attribute information and relationship types. The integration unit is used to perform global knowledge integration processing on the local logical topology subgraph to obtain the global electrical logic topology network. The generation unit is used to perform graph community discovery processing on the global electrical logic topology network to obtain graph functional community clusters and generate functional community summary text for the graph functional community clusters. The building unit is used to construct a knowledge base with electrical topology constraints based on a vector database, a global electrical logic topology network, and functional community summary text.

[0135] In some alternative implementations, the integration unit includes: The fusion sub-unit is used to perform entity disambiguation and global fusion processing on the local logical topology subgraph to obtain the global electrical topology graph. The binding sub-unit is used to obtain the binding relationship between components and image blocks. Based on the binding relationship between components and image blocks, image slice identifiers and labeled image Uniform Resource Locators are bound to the nodes in the global electrical topology map to obtain the global electrical logic topology network.

[0136] In some alternative implementations, the diagnostic module 503 includes: The retrieval unit is used to perform fault information retrieval and recall processing on standard multidimensional diagnostic summaries based on a knowledge base with electrical topology constraints, and to obtain comprehensive fault context information by using a retrieval expert intelligent agent. The reasoning unit is used to perform long-think-chain fault reasoning processing on the comprehensive fault context information using a reasoning expert intelligent agent to obtain the fault reasoning path. The audit unit is used to perform operational compliance auditing and logical consistency auditing on the fault reasoning path using the compliance verification agent, so as to obtain the fault reasoning path that has passed compliance verification. The generation unit is used to generate expert agents based on the fault reasoning path that has passed compliance verification; The diagnostic unit is used to perform multimodal fault diagnosis on the target EMU using an expert intelligent agent and generate a fault diagnosis report for the EMU.

[0137] In some optional implementations, the retrieval unit includes: The semantic retrieval subunit is used to convert standard multidimensional diagnostic summaries into high-dimensional vectors, and based on the high-dimensional vectors, perform semantic retrieval on the vector database to obtain historical maintenance data and unstructured expert experience text. The sub-unit is retrieved to extract core fault-related components as seed nodes based on standard multidimensional diagnostic summaries, and to retrieve the functional community summary text corresponding to the seed nodes. Traverse sub-cells, which are used to perform multi-step neighbor traversal operations with the seed node as the center and along the preset edge relationships in the global electrical logic topology network to obtain the electrical topology connection relationships; The splicing subunit is used to splice and integrate historical maintenance data, unstructured expert experience text, functional community summary text corresponding to seed nodes, and electrical topology connection relationships to obtain comprehensive fault context information.

[0138] In some optional implementations, the deduction unit includes: The extraction subunit is used to extract comprehensive fault context information using a reasoning expert agent to obtain multiple potential fault causes; The logical evolution analysis subunit is used to perform logical evolution analysis on each potential fault cause to obtain the fault logic reasoning link; The embedded subunit is used to acquire real-time measurement data and embed the real-time measurement data into the fault logic reasoning link if the fault logic reasoning link does not converge to the core fault cause. The correction sub-unit is used to dynamically correct the fault logic reasoning link using the global electrical logic topology network until it converges to the core fault cause, and uses the fault logic reasoning link as the fault reasoning path.

[0139] The fault diagnosis device for high-speed trains provided in this embodiment of the invention can execute the fault diagnosis method for high-speed trains provided in any embodiment of the invention, and has the corresponding functional modules and beneficial effects for executing the method. Further functional descriptions of the various modules and units are the same as those in the corresponding embodiments described above, and will not be repeated here.

[0140] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0141] The following is a detailed reference. Figure 6This diagram illustrates a suitable structural design for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 601, which can perform various appropriate actions and processes based on a program stored in read-only memory (ROM) 602 or a program loaded from memory 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of the electronic device. The processor 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0142] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0143] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 609, or installed from a memory 608, or installed from a ROM 602. When the computer program is executed by the processor 601, it performs the functions defined in a fault diagnosis method for a high-speed train according to embodiments of the present invention.

[0144] Figure 6 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0145] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, it implements the fault diagnosis method for high-speed trains shown in the above embodiments.

[0146] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0147] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A fault diagnosis method for high-speed trains, characterized in that, The method includes: Acquire multi-source heterogeneous operation and maintenance data of the target EMU, and construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data; Obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary; Based on the aforementioned standard multidimensional diagnostic summary, a multi-agent cooperative game is conducted using the knowledge base with electrical topology constraints to obtain a fault diagnosis report for the high-speed train.

2. The method according to claim 1, characterized in that, The construction of a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data includes: The multi-source heterogeneous operation and maintenance data is parsed and segmented to obtain standardized text fragments; wherein, the standardized text fragments are stored in a vector database; Define the basic attribute information and relationship types corresponding to the electrical architecture characteristics of the traction and braking system of the EMU; Based on the basic attribute information and the association type, the standardized text fragments are transformed into local logical topology subgraphs; The local logic topology subgraph is subjected to global knowledge integration processing to obtain a global electrical logic topology network; The global electrical logic topology network is subjected to graph community discovery processing to obtain graph functional community clusters, and functional community summary text is generated for the graph functional community clusters. The knowledge base with electrical topology constraints is constructed based on the vector database, the global electrical logic topology network, and the functional community summary text.

3. The method according to claim 2, characterized in that, The step of performing global knowledge integration processing on the local logical topology subgraph to obtain a global electrical logic topology network includes: The local logical topology subgraph is subjected to entity disambiguation and global fusion processing to obtain a global electrical topology graph; Obtain the binding relationship between components and image blocks, and based on the binding relationship between components and image blocks, bind image slice identifiers and labeled image Uniform Resource Locators to the nodes in the global electrical topology map to obtain the global electrical logic topology network.

4. The method according to claim 2, characterized in that, The train fault diagnosis report is obtained by using the knowledge base with electrical topology constraints to conduct multi-agent cooperative game based on the standard multidimensional diagnostic summary, including: Based on the knowledge base with electrical topology constraints, a retrieval expert agent is used to perform fault information retrieval and recall processing on the standard multidimensional diagnostic summary to obtain comprehensive fault context information. The reasoning expert intelligent agent is used to perform long thought chain fault inference processing on the comprehensive fault context information to obtain the fault reasoning path; The compliance verification agent is used to perform operational compliance auditing and logical consistency auditing on the fault reasoning path to obtain a fault reasoning path that has passed compliance verification. An expert agent is generated based on the fault reasoning path that has passed the compliance verification. The expert intelligent agent is used to perform multimodal fault diagnosis on the target EMU and generate a fault diagnosis report for the EMU.

5. The method according to claim 4, characterized in that, Based on the knowledge base with electrical topology constraints, the retrieval expert agent performs fault information retrieval and recall processing on the standard multidimensional diagnostic summary to obtain comprehensive fault context information, including: The standard multidimensional diagnostic summary is converted into a high-dimensional vector, and based on the high-dimensional vector, semantic retrieval is performed on the vector database to obtain historical maintenance data and unstructured expert experience text. Based on the standard multidimensional diagnostic summary, core fault-related components are extracted as seed nodes, and the functional community summary text corresponding to the seed nodes is retrieved. Using the seed node as the center, perform a multi-step neighbor traversal operation along the preset edge relationship in the global electrical logic topology network to obtain the electrical topology connection relationship; The historical maintenance data, the unstructured expert experience text, the functional community summary text corresponding to the seed node, and the electrical topology connection relationship are spliced ​​and integrated to obtain the comprehensive fault context information.

6. The method according to claim 4, characterized in that, The process of using a reasoning expert agent to perform long-think-chain fault deduction on the comprehensive fault context information to obtain the fault reasoning path includes: By using a reasoning expert intelligent agent to extract the comprehensive fault context information, multiple potential fault causes are obtained; For each of the potential fault causes, a logical evolution analysis is performed to obtain the fault logic reasoning link; If the fault logic reasoning link does not converge to the core fault cause, then real-time measurement data is acquired and embedded into the fault logic reasoning link. The fault logic reasoning link is dynamically corrected using the global electrical logic topology network until it converges to the core fault cause, and the fault logic reasoning link is used as the fault reasoning path.

7. A fault diagnosis device for high-speed trains, characterized in that, The device includes: The module is used to acquire multi-source heterogeneous operation and maintenance data of the target EMU, and to construct a knowledge base with electrical topology constraints based on the multi-source heterogeneous operation and maintenance data; The semantic rewriting module is used to obtain fault query information, perform semantic rewriting on the fault query information, and obtain a standard multidimensional diagnostic summary. The diagnostic module is used to perform multi-agent cooperative game based on the standard multidimensional diagnostic summary and the knowledge base with electrical topology constraints to obtain a fault diagnosis report for the EMU.

8. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the fault diagnosis method for the EMU as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the fault diagnosis method for the EMU as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions, which are used to cause a computer to execute the fault diagnosis method for the EMU as described in any one of claims 1 to 6.

Images