Metro catenary fault tracing method and system based on knowledge graph

Abstract

The application discloses a metro catenary fault tracing method based on a knowledge graph, and belongs to the technical field of intelligent operation and maintenance of rail transit. A catenary knowledge graph containing physical topology connection relationship and operation and maintenance semantic knowledge is constructed. Entity nodes and relationship edges in the catenary knowledge graph are mapped into embedded vectors in a low-dimensional vector space by using a RotatE algorithm. When a fault alarm signal is monitored, reverse graph traversal is performed from an alarm node as a starting point, and a candidate fault propagation path is generated. The path propagation cost is calculated in combination with physical rules, nodes in the candidate fault propagation path are converted into path vectors, the path vectors are matched with historical fault vectors in terms of similarity, and a root cause confidence score is calculated in combination with the propagation cost and the similarity to locate a fault source. The application realizes dynamic simulation of a fault propagation path and semantic correlation matching of historical experience, and significantly improves the accuracy and real-time performance of fault root cause positioning.

Description

Technical Field

[0001] This invention belongs to the field of rail transit fault diagnosis technology, specifically relating to a method and system for tracing faults in subway overhead contact lines based on knowledge graphs. Background Technology

[0002] With the rapid expansion of urban rail transit, the overhead contact system, as the core component of the traction power supply system, directly affects the safe and punctual operation of trains. The overhead contact system equipment is distributed over long distances along the track and is constantly exposed to complex and variable outdoor environments such as tunnels and elevated structures. It is highly susceptible to mechanical wear, electrical burns, foreign object intrusion, and severe weather conditions, which can lead to malfunctions. If a fault occurs in the overhead contact system and the root cause cannot be accurately located and addressed within a very short time, it can result in serious consequences such as train stoppages and track paralysis.

[0003] Currently, in the field of subway overhead contact line operation and maintenance, the mainstream fault diagnosis and tracing technologies mainly rely on Supervisory Control and Data Acquisition (SCADA) systems and simple threshold monitoring. These systems use sensors such as voltage and current transformers deployed along the contact line to collect electrical parameters in real time and transmit the data to the monitoring center. The system presets upper and lower thresholds for parameters such as voltage, current, and temperature. When the monitored data exceeds these limits, an audible and visual alarm is triggered. This technology is easy to implement, but fault location is inaccurate, and the false alarm rate is high. Furthermore, based on threshold judgment, it can logically only identify result points exceeding the threshold, but cannot find and distinguish the original fault point. This makes threshold monitoring essentially an isolated judgment based on single-point data, ignoring the complex electrical series and mechanical coupling relationships between contact line equipment. It cannot simulate the physical path of fault current or mechanical stress propagating along the line, making it difficult to provide causal tracing conclusions. To overcome this single... The limitations of threshold monitoring have led some existing technologies to introduce decision trees. These technologies involve manually compiling expert experience to build a fault diagnosis knowledge base containing a large number of identification rules. The current fault is then matched with the rules in the knowledge base to deduce possible causes. However, given the diverse types of overhead contact line equipment and complex connection relationships, manually building a rule base is extremely difficult, and rule updates are often delayed, making it hard to adapt to new types of faults. Furthermore, as the monitoring dimensions increase, the number of rules grows exponentially, drastically expanding the search space during the reasoning process and significantly increasing system response time, making it difficult to meet the real-time requirements of subway fault handling. In recent years, there have also been attempts to use purely data-driven models such as artificial neural networks for fault classification. While these models can establish a mapping relationship between input features and fault types, they lack interpretability and cannot intuitively demonstrate the fault propagation path and reasoning logic, making it difficult for maintenance personnel to be convinced and formulate accurate emergency repair strategies. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a knowledge graph-based method and system for tracing faults in subway overhead contact lines. By constructing a knowledge graph that integrates physical topology and operation and maintenance knowledge, and combining graph embedding technology and path simulation algorithms, it achieves rapid and accurate location of the root cause of the fault.

[0005] The technical solution adopted in this invention is: a method for tracing the source of subway overhead contact line faults based on knowledge graphs, comprising the following steps:

[0006] S1: Extract equipment entities, fault phenomenon entities, and handling measure entities from multi-source heterogeneous data sources, and extract the physical connection relationships and fault causal relationships between equipment to form structured triples; integrate the physical connection relationships between equipment with the preset catenary physical topology drawing data to construct a heterogeneous graph with equipment as nodes and physical connections and logical associations as edges; then perform entity alignment on the heterogeneous graph, match and merge physical topology points with operation and maintenance record entities to construct a catenary knowledge graph containing physical topology connection relationships and operation and maintenance semantic knowledge;

[0007] S2: The RotatE algorithm is used to map entity nodes and relational edges in the catenary knowledge graph into embedding vectors in a low-dimensional vector space, and each entity node and relational edge in the acquired historical fault case data is mapped into a historical fault vector.

[0008] S3: When a fault alarm signal is detected, starting from the alarm node, a reverse graph traversal is performed based on the physical connection relationship in the contact network knowledge graph. During the graph traversal, the path cost of fault propagation is calculated based on Kirchhoff's laws. Propagation paths with path costs exceeding a preset threshold are eliminated, and at least one candidate fault propagation path is generated.

[0009] S4: Map the entity nodes and relation edges in the candidate fault propagation path to path vectors using the RotatE algorithm, and calculate the similarity between the path vectors and historical fault vectors in the same low-dimensional vector space.

[0010] S5: Based on the propagation cost of the candidate fault propagation path and the similarity obtained in S4, a weighted fusion model is used to calculate the confidence score of each entity node in the fault propagation path as a potential root cause node, and the root cause of the fault is determined by sorting according to the confidence score.

[0011] Preferably, in step S1, matching and merging physical topology points with maintenance record entities includes using an entity alignment algorithm to calculate the similarity between the names of physical topology points and maintenance entities. When the similarity exceeds a preset threshold, the two are merged into the same entity node. Through step S1, data silos can be broken down, and multi-source heterogeneous data scattered in SCADA systems, maintenance logs, and drawings can be unified and integrated at the semantic level, providing a structured data foundation for subsequent inference calculations.

[0012] Preferably, in step S2, the RotatE algorithm defines the relationship as a rotation operation from the head entity to the tail entity. It trains the entity vector and relationship vector by minimizing the difference between the distance metric of the true triplet and the distance metric of the negative sampled triplet. The formula for calculating the distance metric is:

[0013] ,

[0014] in, For the head entity vector, For the tail entity vector, For relation vectors, The element product is represented by the RotatE algorithm, which can make full use of the rotation characteristics of complex space to accurately model the complex symmetric, antisymmetric and reciprocal relationships between contact network equipment, and solve the defect of traditional translation model in being unable to handle reciprocal relationships (such as power supply and power receiving).

[0015] Preferably, in step S3, the path cost is calculated based on the path length constraint, where the path length is the connection relationship between devices at levels 3 to 7. By introducing physical rules to constrain the path cost calculation, the propagation and attenuation characteristics of faults in the physical topology can be effectively simulated, eliminating physically unreachable or extremely low-probability paths, thus significantly reducing the subsequent search space.

[0016] Preferably, in step S4, obtaining the path vector includes obtaining the entity embedding vector corresponding to each entity node in the candidate fault propagation path, setting a weight coefficient according to the position of each entity node in the propagation path, and calculating the path vector using a weighted average method, wherein the weight coefficient is based on the principle that the closer the node is to the alarm node, the higher the weight; this can transform the discrete graph path into a computable numerical vector, making it possible to perform efficient similarity matching using vector operations.

[0017] Preferably, in step S5, the weighted fusion model is:

[0018] ,

[0019] in, As a potential root cause node, For normalization function, For path cost, For similarity, and The weighted fusion model takes into account both the hard constraints of physical propagation mechanisms and the soft constraints of historical operation and maintenance experience, thereby improving the accuracy and robustness of root cause localization.

[0020] Preferably, the knowledge graph-based subway catenary fault tracing method further includes an incremental learning step. The incremental learning involves: when the similarity of all candidate fault propagation paths is lower than a preset matching threshold, determining that the current fault is an unknown type of fault; storing the path vector and feature data of the current fault in a database to be confirmed, and generating a manual investigation prompt; receiving manually confirmed fault root cause labels, using the confirmed data as new positive samples to fill back into the catenary knowledge graph, and triggering incremental update training of the embedded vectors; through incremental learning, it is endowed with the ability to self-evolve, and can continuously enrich the knowledge base and adapt to new fault modes as the operation time progresses.

[0021] A knowledge graph-based fault tracing system for subway overhead contact lines includes:

[0022] The knowledge graph construction module is used to build a knowledge graph of the overhead contact system that includes physical topology connections and operational semantics.

[0023] The graph embedding processing module is used to map entity nodes and relation edges in the catenary knowledge graph into embedding vectors in a low-dimensional vector space, and to map each entity node and relation edge in the acquired historical fault case data into a historical fault vector.

[0024] The path simulation module is used to perform a reverse graph traversal based on the physical connection relationship in the contact network knowledge graph, starting from the alarm node when a fault alarm signal is detected, and generate at least one candidate fault propagation path.

[0025] The case matching module is used to convert entity nodes in candidate fault propagation paths into path vectors and calculate the similarity between path vectors and historical fault vectors in a low-dimensional vector space. The root cause localization module is used to calculate the confidence score of potential root cause nodes based on the propagation cost and similarity of candidate fault propagation paths, and determine the root cause of the fault by sorting them according to the confidence score.

[0026] Preferably, the knowledge graph-based subway catenary fault tracing system adopts an end-edge-cloud collaborative architecture. At the perception layer, sensor groups are deployed to collect real-time status data of the catenary. At the network and edge layer, edge computing gateways are deployed to preprocess and cache the real-time status data locally. At the platform and application layer, a cloud server cluster is deployed to execute the construction of the catenary knowledge graph, graph embedding training, and root cause localization algorithms. The end-edge-cloud collaborative architecture ensures the real-time performance and reliability of the system. The preprocessing at the edge layer reduces the computing pressure on the cloud, while the large-scale computing power of the cloud supports complex graph training.

[0027] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of a knowledge graph-based method for tracing faults in subway overhead contact lines.

[0028] The beneficial effects of this invention are as follows:

[0029] (1) The hybrid graph construction technology of "physical topology + operation and maintenance knowledge" is adopted. By integrating multi-source heterogeneous data through a unified ontology model, the problem of data semantic fragmentation in the existing technology is solved. It can link information scattered in SCADA system, maintenance log and drawings, so that fault diagnosis is no longer limited to single-dimensional signal analysis, but has cross-domain knowledge reasoning ability. Compared with single data source analysis, the completeness of contextual information for fault location is significantly improved.

[0030] (2) The RotatE knowledge graph embedding algorithm was adopted. Through complex space rotation modeling, the accurate representation of complex relationships was realized. Since the RotatE algorithm naturally supports the modeling of symmetric and inverse relationships, the system can accurately distinguish the direction of fault propagation. The accuracy in processing multi-relationship paths is significantly improved. The complex graph matching problem is transformed into an efficient vector operation problem, and the computational complexity is greatly reduced.

[0031] (3) By calculating the path cost constrained by physical rules, the search space is reduced to a local subgraph, and then the vector retrieval technology is used for confirmation, so that the fault tracing response time is shortened from minutes to seconds;

[0032] (4) An incremental learning mechanism for “unknown faults” was designed. Through the closed loop of manual feedback, the system’s self-evolution was realized, which enabled the system’s knowledge base to be continuously enriched as the operation time progressed, and the fault diagnosis accuracy to be improved year by year with the operating years, effectively solving the problem of the lag in knowledge updates of expert systems.

[0033] This invention achieves dynamic simulation of fault propagation paths and semantic association matching of historical experience, significantly improving the accuracy and real-time performance of fault root cause localization. Detailed Implementation

[0034] The present invention will now be described in further detail with reference to specific embodiments.

[0035] Example

[0036] The knowledge graph-based subway catenary fault tracing system provided in this embodiment adopts an end-edge-cloud collaborative layered architecture. In terms of physical entities, it mainly includes a perception layer deployed on field equipment, a network and edge layer deployed on the trackside or substation, and a platform and application layer deployed in the control center or data center.

[0037] The sensing layer is mainly composed of various sensors, including voltage transformers, current transformers, temperature sensors, anemometers, and video monitoring devices. These sensors are distributed along the contact network to collect electrical parameters (such as voltage, current, and impedance), physical environmental parameters (such as temperature, wind speed, and foreign object intrusion), and equipment appearance status in real time. After collection, the sensors communicate with the network and edge layer through industrial Ethernet or wireless sensor networks to upload the original analog or digital signals.

[0038] The network and edge layer mainly consists of an edge computing gateway, which integrates a preprocessing unit and a local caching unit. The preprocessing unit is used to clean, filter, and convert the protocol of the raw data uploaded by the perception layer. The local caching unit is used to temporarily store data when the network is interrupted to ensure data integrity. The network and edge layer communicates with the platform and application layer to upload the preprocessed feature data to the cloud.

[0039] The platform and application layer are deployed on a cloud server cluster, including a memory, a processor, and multiple functional modules running on the processor. These functional modules include a knowledge graph construction module, a graph embedding processing module, a path simulation module, a case matching module, and a root cause localization module. The memory stores the overhead contact line knowledge graph, a historical fault case library, and a vector database. The processor is used to execute computer programs, coordinate the work of each module, and implement the fault tracing method.

[0040] The knowledge graph construction module has its input end connected to a multi-source heterogeneous data source and its output end connected to a memory. It is used to construct a catenary knowledge graph that includes physical topology connections and operational semantic knowledge. In this embodiment, the knowledge graph construction module includes a data submodule and a fusion submodule.

[0041] The data submodule is used to extract equipment entities, fault phenomenon entities, and handling measure entities from the multi-source heterogeneous data sources received, and to extract the physical connection relationships and fault causal relationships between equipment to form structured triples. In this embodiment, the extracted multi-source data includes structured data, real-time telemetry data (voltage, current RMS values, etc.) from the SCADA system and equipment ledger relationship tables; semi-structured data, from maintenance logs; and unstructured data, from equipment manuals and fault case texts.

[0042] The fusion submodule is used to fuse the feature information extracted by the data submodule, that is, to fuse the physical connection relationships between devices with the preset catenary physical topology drawing data, and construct a heterogeneous graph with devices as nodes and physical connections and logical associations as edges. Then, entity alignment is performed on the heterogeneous graph to match and merge physical topology points with maintenance record entities, and construct a catenary knowledge graph containing physical topology connection relationships and maintenance semantic knowledge. In this embodiment, specifically, a physical topology subgraph is constructed based on the connection relationships exported from the catenary line drawings or GIS system. In the graph, nodes represent physical devices (such as contact wires, catenary cables, droppers, disconnectors, etc.), and edges represent physical connection relationships. The attributes of the edges include parameters such as connection impedance and physical distance. At the same time, an maintenance knowledge subgraph is constructed. In the graph, nodes include fault phenomena, fault causes, and handling measures, and edges represent logical associations. Finally, using an entity alignment algorithm, the similarity between the names of physical nodes and maintenance entities is calculated. When the similarity exceeds a threshold, the two are merged into the same entity node, thereby realizing the unified mounting of physical attributes and semantic knowledge and forming a complete catenary knowledge graph.

[0043] The input end of the graph embedding processing module is connected to the memory for reading the overhead contact network knowledge graph, and the output end is connected to the case matching module.

[0044] The graph embedding processing module uses the RotatE algorithm to map entity nodes and relation edges in the catenary knowledge graph into embedding vectors in a low-dimensional vector space; specifically, for triples in the catenary knowledge graph... The RotatE algorithm will determine the relationship Model as a de novo entity End-to-end entity Rotation operation; define the embedding vector as a complex vector, satisfying: in, The Hadamard product represents the element-wise multiplication in the complex space, which is equivalent to rotating the head entity vector by a certain angle in the complex plane to obtain the tail entity vector.

[0045] By defining a distance metric function, the validity of the triples is ensured. The formula for the distance metric function is as follows:

[0046] ,

[0047] in, For the head entity vector, For the tail entity vector, For relation vectors, Represents the element-wise product;

[0048] It employs a self-adversarial negative sampling technique, training entity and relation vectors—the embedding vectors—by minimizing the difference between the distance metrics of the true triples and the negatively sampled triples. Specifically, it optimizes the vector representations of entities and relations by minimizing the loss function.

[0049] ,

[0050] in, This is the interval hyperparameter, with a preferred value of 6.0; These are generated negative samples; after training, each entity and relation obtains a... Dimensional vector representation, in this embodiment, the embedded dimension The optimal dimension is 200. When the dimension is less than 50, the vector space cannot accommodate complex semantic information, which leads to the failure of similarity calculation. When the dimension is greater than 300, the training time increases exponentially and overfitting occurs. When the dimension is selected as 200, the best balance is achieved between expressive power and computational efficiency.

[0051] The graph embedding processing module maps each entity node and relation edge in the acquired historical fault case data into a historical fault vector and stores it in memory.

[0052] The input of the path simulation module is connected to the edge computing gateway and the memory, and the output is connected to the case matching module.

[0053] When a fault alarm signal is detected, the path simulation module is triggered. The core task of this module is to generate a set of candidate fault propagation paths. Due to the complex connection relationship between the contact network equipment, directly traversing the entire map would require a huge amount of calculation. Therefore, physical rules are used for constraint, namely, the path length is used for constraint. In this embodiment, the path length is set to 5. The setting is based on the typical length of the contact network power supply arm and the number of cascaded equipment. Fault propagation exceeding level 5 is extremely rare and will be cut off by the protection device.

[0054] Therefore, when a node issues an alarm, a reverse graph traversal is performed starting from the alarm node, based on the physical connection relationships in the contact network knowledge graph. During the graph traversal, the path cost of fault propagation is calculated based on Kirchhoff's laws, and the path cost is calculated based on a preset 5-level path length constraint. In this embodiment, the path cost function considers electrical distance and attenuation effects, specifically:

[0055] ,

[0056] in, For nodes and The connection impedance between them (obtained from the properties of the edges in the contact network knowledge graph). The attenuation factor is used; the above function simulates the loss of fault energy during propagation. Paths farther from the alarm point and with higher impedance are less likely to become actual fault paths. Propagation paths with path costs exceeding a preset threshold are eliminated, generating at least one candidate fault propagation path. In this embodiment, 10 candidate fault propagation paths are preferably generated. Simultaneously...

[0057] The input of the case matching module is connected to the path simulation module and the memory, and the output is connected to the root cause localization module.

[0058] The entity nodes and relation edges in each candidate fault propagation path are mapped to path vectors using the RotatE algorithm. In this embodiment, a weight coefficient is set according to the position of each entity node in the propagation path, and a weighted average method is used for calculation, as follows:

[0059] ,

[0060] in, For path vectors, It is a node Embedded vector, These are weighting coefficients.

[0061] The weighting follows the principle of "nearer is larger, farther is smaller", meaning that the closer the node is to the alarm point, the higher the weight, because the closer the node is to the alarm point, the more direct the impact on the current fault state.

[0062] Calculate the cosine similarity between the path vector of each candidate fault propagation path and the historical fault vector, and set a similarity threshold. If the calculated similarity is greater than or equal to the similarity threshold, the current candidate fault propagation path is determined to be highly similar to the historical fault case. If it is less than the similarity threshold, the corresponding candidate fault propagation path is removed.

[0063] The root cause localization module is the decision output of the entire system, and its input is connected to the case matching module. The root cause localization module combines the propagation cost of the integrated path simulation module with the semantic similarity of the case matching module, and uses a weighted fusion model to calculate the confidence score of each entity node in the candidate fault propagation path as a potential root cause node, as follows:

[0064] ,

[0065] in, As a potential root cause node, For normalization function, For path cost, For similarity, and These are preset weighting coefficients;

[0066] The normalization function normalizes the propagation cost to Functions over an interval; and These are the weighting coefficients, and + =1;

[0067] Finally, the calculated confidence scores are sorted in descending order, and the top three results are output. At the same time, a source tracing report can be generated, which shows the location of the fault source, the fault type, the confidence level, a visual topology diagram of the propagation path, and recommended handling measures. All of this data can be extracted from the overhead contact network knowledge graph.

[0068] The knowledge graph-based subway catenary fault tracing system also includes a reinforcement learning mechanism. When the case matching module calculates that the similarity of all candidate fault propagation paths is less than the similarity threshold, it determines that the current fault is an unknown type of fault, and triggers the anomaly handling process. The path vector and feature data of the current fault are stored in the database to be confirmed, and a prompt message is generated by the manual investigation prompt module. After on-site confirmation by maintenance personnel, the actual fault root cause label is input into the system. The system uses the confirmed data as a new positive sample, backfills it into the catenary knowledge graph and historical fault case database, and triggers the incremental update training of the graph embedding processing module.

[0069] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for tracing faults in subway overhead contact lines based on knowledge graphs, characterized in that: Includes the following steps, S1: Extract equipment entities, fault phenomenon entities, and handling measure entities from multi-source heterogeneous data sources, and extract the physical connection relationships and fault causal relationships between equipment to form structured triples; integrate the physical connection relationships between equipment with the preset catenary physical topology drawing data to construct a heterogeneous graph with equipment as nodes and physical connections and logical associations as edges; then perform entity alignment on the heterogeneous graph, match and merge physical topology points with operation and maintenance record entities to construct a catenary knowledge graph containing physical topology connection relationships and operation and maintenance semantic knowledge; S2: The RotatE algorithm is used to map entity nodes and relational edges in the catenary knowledge graph into embedding vectors in a low-dimensional vector space, and each entity node and relational edge in the acquired historical fault case data is mapped into a historical fault vector. S3: When a fault alarm signal is detected, starting from the alarm node, a reverse graph traversal is performed based on the physical connection relationship in the contact network knowledge graph. During the graph traversal, the path cost of fault propagation is calculated based on Kirchhoff's laws. Propagation paths with path costs exceeding a preset threshold are eliminated, and at least one candidate fault propagation path is generated. S4: Map the entity nodes and relation edges in the candidate fault propagation path to path vectors using the RotatE algorithm, and calculate the similarity between the path vectors and historical fault vectors in the same low-dimensional vector space. S5: Based on the propagation cost of the candidate fault propagation path and the similarity obtained in S4, a weighted fusion model is used to calculate the confidence score of each entity node in the fault propagation path as a potential root cause node, and the root cause of the fault is determined by sorting according to the confidence score.

2. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: In step S1, matching and merging physical topology points with operation and maintenance record entities includes using an entity alignment algorithm to calculate the similarity between the names of physical topology points and operation and maintenance entities. When the similarity exceeds a preset threshold, the two are merged into the same entity node.

3. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: In step S2, the RotatE algorithm defines the relation as a rotation operation from the head entity to the tail entity. It trains the entity vector and relation vector by minimizing the difference between the distance metric of the true triplet and the distance metric of the negatively sampled triplet. The formula for calculating the distance metric is: ， in, For the head entity vector, For the tail entity vector, For relation vectors, Represents the element product.

4. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: In step S3, the path cost is calculated based on the path length constraint, where the path length is the connection relationship between devices at levels 3 to 7.

5. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: In step S4, obtaining the path vector includes obtaining the entity embedding vector corresponding to each entity node in the candidate fault propagation path, setting a weight coefficient according to the position of each entity node in the propagation path, and calculating the path vector using a weighted average method, wherein the weight coefficient is based on the principle that the closer the node is to the alarm node, the higher the weight.

6. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: In step S5, the weighted fusion model is as follows: ， in, As a potential root cause node, For normalization function, For path cost, For similarity, and These are the preset weighting coefficients.

7. The method for tracing subway overhead contact line faults based on knowledge graphs according to claim 1, characterized in that: The knowledge graph-based subway catenary fault tracing method further includes an incremental learning step. The incremental learning is as follows: when the similarity of all candidate fault propagation paths is lower than a preset matching threshold, the current fault is determined to be an unknown type of fault; the path vector and feature data of the current fault are stored in the database to be confirmed, and a manual investigation prompt is generated; the fault root cause label confirmed by manual investigation is received, and the confirmed data is used as a new positive sample to fill back into the catenary knowledge graph, and the incremental update training of the embedded vector is triggered.

8. A knowledge graph-based fault tracing system for subway overhead contact lines, characterized in that: include: The knowledge graph construction module is used to build a knowledge graph of the overhead contact system that includes physical topology connections and operational semantics. The graph embedding processing module is used to map entity nodes and relation edges in the overhead contact network knowledge graph into embedding vectors in a low-dimensional vector space, namely historical fault vectors. The path simulation module is used to perform a reverse graph traversal based on the physical connection relationship in the contact network knowledge graph, starting from the alarm node when a fault alarm signal is detected, and generate at least one candidate fault propagation path. The case matching module is used to convert entity nodes in candidate fault propagation paths into path vectors and calculate the similarity between path vectors and historical fault vectors in a low-dimensional vector space. The root cause localization module is used to calculate the confidence score of potential root cause nodes based on the propagation cost and similarity of candidate fault propagation paths, and to determine the root cause of the fault by sorting them according to the confidence scores.

9. The knowledge graph-based subway overhead contact line fault tracing system according to claim 8, characterized in that: The knowledge graph-based subway catenary fault tracing system adopts an end-edge-cloud collaborative architecture. At the perception layer, a sensor group is deployed to collect real-time status data of the catenary. At the network and edge layer, an edge computing gateway is deployed to preprocess and cache the real-time status data locally. At the platform and application layer, a cloud server cluster is deployed to execute the catenary knowledge graph construction, graph embedding training, and root cause localization algorithm.

10. A computer-readable storage medium having a computer program stored thereon, the computer program being executed by a processor to implement the steps of the knowledge graph-based subway overhead contact line fault tracing method as described in any one of claims 1-7.

Images