Track maintenance knowledge graph maintenance optimization method

Abstract

This invention provides a method for maintaining and optimizing a track maintenance knowledge graph. First, a snapshot sequence graph of maintenance log data is constructed. Then, each edge in the snapshot sequence graph is weighted according to its exponential time-to-time weight. Based on the time-to-time weight after exponential weighting and a sliding time window, a weighted dense subgraph for each window is extracted. The edges of the weighted dense subgraphs across windows are accumulated according to their time-to-time weight and subgraph density to calculate the persistence-time-to-time cumulative weight. Then, based on the persistence-time-to-time cumulative weight, the edges of the weighted dense subgraphs are pruned according to a threshold to obtain the core subgraph. The extracted core subgraph is then written back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

Description

Technical Field

[0001] This invention belongs to the field of rail transit operation and maintenance technology, and in particular relates to a maintenance and optimization method for a knowledge graph of track maintenance. Background Technology

[0002] The operation of rail transit systems generates a large number of fault maintenance logs (such as those for escalators, power supply, signaling, and AFC systems). Establishing a knowledge graph can significantly improve the accuracy of track defect analysis. By integrating multi-source heterogeneous data and constructing a semantic association network, maintenance efficiency, accuracy, and the scientific nature of decision-making are significantly improved. With the help of the knowledge graph, multi-source data fusion and association analysis can be used to improve the accuracy of defect identification, perform dynamic risk prediction and preventive maintenance, and reduce the accident rate. Maintaining the track maintenance knowledge graph is a key link to ensure its long-term effectiveness and accuracy, involving multiple aspects such as data updates, structure optimization, quality monitoring, safety assurance, and user feedback. Maintaining the knowledge graph can ensure data quality and system performance, avoid data obsolescence and error accumulation, and match maintenance strategies and procedures for iteration, making it essential in the operation and maintenance of rail transit.

[0003] Existing patent CN119722421A discloses a knowledge graph-based intelligent operation and maintenance system for urban rail transit, comprising: a rail transit fault acquisition module, used to collect fault report information of each fault generated during the operation of the urban rail transit system in real time; a fault element extraction module, which is communicatively connected to the rail transit fault acquisition module and is used to receive fault report information and extract elements from the fault report information one by one to obtain fault element information that can comprehensively cover each fault; a fault matching module, used to match the fault element information of each fault with the knowledge graph of historical faults and output the matching result; and an operation and maintenance module, which evaluates multiple faults based on the matching result output by the fault matching module and provides maintenance plan information. This invention achieves rapid fault identification, accurate location, efficient processing, and intelligent operation and maintenance management. Summary of the Invention

[0004] Traditional rail transit operation and maintenance relies heavily on manual experience or static reports, which has the following problems: lack of timeliness, as newly emerging fault modes cannot be quickly converted into knowledge; lack of repetitive identification, as long-term recurring maintenance links fail to form knowledge fragments; lack of online maintenance, as log volumes are large and updates are frequent, making static graph methods difficult to support; and lack of closed-loop processing, as knowledge cannot directly feed back into subsequent maintenance decisions. Therefore, knowledge graphs are needed to assist in rail transit operation and maintenance. However, existing knowledge graph maintenance methods in rail transit operation and maintenance scenarios have the following drawbacks: lack of timeliness modeling, as most knowledge graph memory mechanisms use static edge weights, failing to reflect the "new facts first" characteristic; neglect of repetitiveness and stability, as existing frequent pattern mining or PageRank methods often favor single high-frequency events, making it difficult to identify long-term recurring patterns; and excessive complexity, as the maximum common subgraph (MCS) calculation is NP-hard, making real-time maintenance impossible in online scenarios.

[0005] To solve the above-mentioned technical problems, the technical solution provided by this invention is: a maintenance optimization method for a track maintenance knowledge graph, comprising the following steps: S1. Construct a snapshot sequence graph of maintenance log data. Apply exponential time-increment weighting to each edge of the snapshot sequence graph. Based on the time-increment weighting and the sliding time window, extract the weighted dense subgraph of each window. S2. For the edges of the weighted dense subgraph across the window, accumulate the time-update weights and the subgraph density to calculate the persistent-time-update cumulative weights. S3. Based on the persistent-time cumulative weight, prune the edges of the weighted dense subgraph according to the threshold to obtain the core subgraph; S4. Write the extracted core subgraph back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

[0006] Specifically, when constructing a snapshot sequence diagram of maintenance log data in S1, the equipment alarm logs, maintenance processing logs, and maintenance result logs accumulated in the maintenance data according to the time sequence are used at fixed time intervals. Nodes represent equipment, fault type, maintenance action, and result status, and edges represent the causal, temporal, or co-occurrence relationships between events to construct a snapshot diagram of maintenance log data. Multiple snapshot diagrams are then combined according to the time sequence to construct a snapshot sequence diagram.

[0007] Specifically, when extracting the weighted dense subgraph in S1, the snapshot sequence graph is first initialized with weights. Then, according to the sliding time window, the edges of the snapshot sequence graph are exponentially weighted according to their timeliness. Different decay constants are set for different events. Based on the decay constants and the event occurrence time, each edge is assigned a timeliness weight that decays over time, i.e., exponential timeliness weighting, to obtain the weighted subgraph.

[0008] Specifically, in S1, after weighting and initializing the snapshot sequence graph, the newly weighted density is calculated for the weighted subgraph of each sliding time window. Then, the node with the smallest weight and its associated edges are deleted to obtain a new weighted subgraph. The density is then recalculated, and the above operation is repeated until all nodes are deleted. The weighted subgraph with the largest density is extracted as the weighted dense subgraph of the current sliding time window.

[0009] Specifically, in S2, the edges of the weighted dense subgraphs of all time windows on the same time axis are accumulated according to the time-update weight and the subgraph density, and the time-update accumulation is calculated across windows. A coefficient of variation is calculated based on the standard deviation and mean of the samples on all time window sets. The time-update accumulation weight is combined with the coefficient of variation to perform stability penalty, eliminate occasional noise, and obtain the global persistent-time-update accumulation weight.

[0010] Specifically, the coefficient of variation is obtained by dividing the standard deviation by the mean and adding a small constant close to 0.

[0011] Specifically, in S3, for the persistent-time cumulative weight of the edges of each weighted dense subgraph, the upper quartile of the distribution of the persistent-time cumulative weight of all edges is taken as the threshold; the weighted dense subgraph after retaining the edges whose persistent-time cumulative weight is greater than the threshold and deleting the other edges is taken as the core subgraph.

[0012] Specifically, when performing knowledge maintenance in S4, multiple edges may appear on the same pair of nodes in the long-term accumulated maintenance logs, representing different processing methods. Using the edge weight in the core subgraph as the anchor point, the edge with the larger weight is selected as the priority recommended repair path.

[0013] Specifically, during the retrieval optimization process in S4, when a user searches for a solution to a certain fault, there may be multiple reachable paths on the knowledge graph. If a path passes through an edge in the core subgraph, the score of that path will be improved.

[0014] Specifically, path scoring is controlled by a scoring function, which is obtained by adding the enhancement coefficient to the ordinary edge weights of the knowledge graph and multiplying them by the persistent-time cumulative weights in the core subgraph.

[0015] The beneficial effects of this invention are that it unifies timeliness, repetitiveness, and structural density into a closed loop that can be maintained online. This can be used for the close coupling of weight updates, conflict correction, and retrieval strategies in dialogue memory knowledge graphs, and has obvious engineering feasibility and innovation. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0017] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Example 1: A maintenance optimization method for a track maintenance knowledge graph, such as... Figure 1 As shown, it includes the following steps: S1. Construct a snapshot sequence graph of maintenance log data. Apply exponential time-increment weighting to each edge of the snapshot sequence graph. Based on the time-increment weighting and the sliding time window, extract the weighted dense subgraph of each window. S2. For the edges of the weighted dense subgraph across the window, accumulate the time-update weights and the subgraph density to calculate the persistent-time-update cumulative weights. S3. Based on the persistent-time cumulative weight, prune the edges of the weighted dense subgraph according to the threshold to obtain the core subgraph; S4. Write the extracted core subgraph back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

[0019] When performing knowledge maintenance, multiple edges may appear for the same pair of nodes in the long-term accumulated maintenance logs, representing different processing methods. Using the edge weight in the core subgraph as the anchor point, the edge with the larger weight is selected as the priority recommended repair path.

[0020] During the retrieval optimization process, when a user searches for a solution to a certain fault, there may be multiple reachable paths on the knowledge graph. If a path passes through an edge in the core subgraph, the path's score is improved. The path score is controlled by a scoring function, which is obtained by adding the enhancement coefficient to the ordinary edge weight of the knowledge graph and multiplying it by the persistence-time cumulative weight in the core subgraph.

[0021] In rail transit operation and maintenance scenarios, a large amount of maintenance data accumulates continuously in a time-series format, for example: Equipment alarm logs (such as escalator emergency stop alarms, power supply voltage fluctuations); Maintenance and repair log (e.g., "Replace tensioner wheel" or "Switch to backup power"); Maintenance results log (e.g., "Restored to normal operation").

[0022] At fixed time intervals (e.g., 1 hour, 1 day), these logs are transformed into a "snapshot graph," where nodes represent equipment, fault type, maintenance action, and result status; edges represent causal, temporal, or co-occurrence relationships between events. As time progresses, a series of snapshot graphs are formed, constituting a snapshot sequence graph of the maintenance log data.

[0023] When extracting the weighted dense subgraph in S1, the snapshot sequence graph is first initialized with weights. Then, according to a sliding time window, the edges of the snapshot sequence graph are exponentially weighted based on their timeliness. Different decay constants are set for different events. Based on the decay constants and the event occurrence time, each edge is assigned a timeliness weight that decays over time—this is exponential timeliness weighting—resulting in the weighted subgraph. In exponential timeliness weighting, a decay constant is set to control the weight difference between "new events" and "old events." In practical applications of intelligent track maintenance, based on equipment faults (such as escalator alarms), corresponding events are required. The half-life of different events is set, and the decay constant for different events is calculated based on their half-lives. For example, if an escalator alarm needs to be handled on the same day, the half-life is set to 1. Exponential timeliness weighting ensures that recent maintenance events have a greater impact in the weighted dense subgraph calculation, while earlier events gradually weaken, meeting the requirement of intelligent track maintenance for "rapid response to the latest fault modes."

[0024] In S1, after weighting and initializing the snapshot sequence graph, the newly weighted density is calculated for the weighted subgraph of each sliding time window. Then, the node with the smallest weight and its associated edges are deleted to obtain a new weighted subgraph. The density is then recalculated, and the above operation is repeated until all nodes are deleted. The weighted subgraph with the largest density is extracted as the weighted dense subgraph of the current sliding time window.

[0025] In S2, for each edge of the weighted dense subgraph across all time windows on the same time axis, the time-update weights are accumulated along with the subgraph density, and this cross-window accumulation is used to calculate the time-update cumulative weight. A coefficient of variation is then calculated based on the standard deviation and mean of the samples across all time window sets. This time-update cumulative weight is combined with the coefficient of variation for a stability penalty to eliminate sporadic noise, resulting in the global persistent-time-update cumulative weight. The coefficient of variation is obtained by dividing the standard deviation by the mean and adding a small constant close to zero. The purpose of calculating the global persistent-time-update cumulative weight is to address the issue that some edges may suddenly appear with high weights in certain windows but then disappear, representing sporadic noise. Therefore, the coefficient of variation is introduced, and through cross-window accumulation and a stability penalty, the final global persistent-time-update cumulative weight is obtained.

[0026] In S3, the persistent-time cumulative weight of edges in each weighted dense subgraph is calculated, and the upper quartile of the distribution of persistent-time cumulative weights of all edges is taken as the threshold. Edges with persistent-time cumulative weights greater than the threshold are retained, and the other edges are deleted. The resulting weighted dense subgraph is then used as the core subgraph. The extracted core subgraph is written back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

[0027] Graph maintenance mainly involves conflict correction. In long-accumulated maintenance logs, the same pair of nodes (e.g., "Fault A → Repair Method") may have multiple edges, representing different handling methods. For example: Fault A → Repair Method 1 (Restart Controller) Fault A → Repair Method 2 (Replace Part) In this case, using the edge weights in the core subgraph as anchor points, the edges with larger weights are selected as the preferred repair paths.

[0028] Search optimization primarily involves using path scoring to prioritize paths. When a user searches for a solution to a particular fault, multiple reachable paths may exist on the knowledge graph. If a path passes through an edge in the core subgraph, its score is increased. Path scoring is controlled by a scoring function, which is calculated by adding an enhancement coefficient to the ordinary edge weights of the knowledge graph and multiplying it by the cumulative weights of persistence and relevance in the core subgraph. The core subgraph acts as a "weighted amplification" tool during the search phase, making the system more inclined to recommend maintenance paths that are stable in the long term and frequently used in the recent past.

[0029] The method provided in this embodiment integrates "novelty × repetitiveness × dense structure" within a single framework to form a unified core subgraph determination criterion. It employs exponential decay and sliding window recursion, combined with a weighted peeling method to find dense subgraphs, achieving an overall complexity that is approximately linear, meeting online real-time requirements. The coefficient of variation is introduced as a stability constraint to effectively suppress interference from occasional burst edges. The core subgraph can not only be used for retrieval but also for writing back weights to participate in conflict correction and path scoring, forming a complete memory maintenance closed loop. The core subgraph possesses clear edge weights, density, and support indicators, which can be output as evidence, improving system auditability.

[0030] Example 2: A maintenance optimization method for a track maintenance knowledge graph, comprising the following steps: S1. Construct a snapshot sequence graph of maintenance log data. Apply exponential time-increment weighting to each edge of the snapshot sequence graph. Based on the time-increment weighting and the sliding time window, extract the weighted dense subgraph of each window. S2. For the edges of the weighted dense subgraph across the window, accumulate the time-update weights and the subgraph density to calculate the persistent-time-update cumulative weights. S3. Based on the persistent-time cumulative weight, prune the edges of the weighted dense subgraph according to the threshold to obtain the core subgraph; S4. Write the extracted core subgraph back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

[0031] When performing knowledge maintenance, multiple edges may appear for the same pair of nodes in the long-term accumulated maintenance logs, representing different processing methods. Using the edge weight in the core subgraph as the anchor point, the edge with the larger weight is selected as the priority recommended repair path.

[0032] During the retrieval optimization process, when a user searches for a solution to a certain fault, there may be multiple reachable paths on the knowledge graph. If a path passes through an edge in the core subgraph, the path's score is improved. The path score is controlled by a scoring function, which is obtained by adding the enhancement coefficient to the ordinary edge weight of the knowledge graph and multiplying it by the persistence-time cumulative weight in the core subgraph.

[0033] In rail transit operation and maintenance scenarios, a large amount of maintenance data accumulates continuously in a time-series format, for example: Equipment alarm logs (such as escalator emergency stop alarms, power supply voltage fluctuations); Maintenance and repair log (e.g., "Replace tensioner wheel" or "Switch to backup power"); Maintenance results log (e.g., "Restored to normal operation").

[0034] At fixed time intervals (e.g., 1 hour, 1 day), these logs are transformed into a "snapshot graph," where nodes represent equipment, fault type, maintenance action, and result status; edges represent causal, temporal, or co-occurrence relationships between events. As time progresses, a series of snapshot graphs are formed, constituting a snapshot sequence graph of the maintenance log data.

[0035] When extracting the weighted dense subgraph in S1, the snapshot sequence graph is first initialized with weights. Then, according to a sliding time window, the edges of the snapshot sequence graph are exponentially weighted based on their timeliness. Different decay constants are set for different events. Based on the decay constants and the event occurrence time, each edge is assigned a timeliness weight that decays over time—this is exponential timeliness weighting—resulting in the weighted subgraph. In exponential timeliness weighting, a decay constant is set to control the weight difference between "new events" and "old events." In practical applications of intelligent track maintenance, based on equipment faults (such as escalator alarms), corresponding events are required. The half-life of different events is set, and the decay constant for different events is calculated based on their half-lives. For example, if an escalator alarm needs to be handled on the same day, the half-life is set to 1. Exponential timeliness weighting ensures that recent maintenance events have a greater impact in the weighted dense subgraph calculation, while earlier events gradually weaken, meeting the requirement of intelligent track maintenance for "rapid response to the latest fault modes."

[0036] In S1, after weighting and initializing the snapshot sequence graph, the newly weighted density is calculated for the weighted subgraph of each sliding time window. Then, the node with the smallest weight and its associated edges are deleted to obtain a new weighted subgraph. The density is then recalculated, and the above operation is repeated until all nodes are deleted. The weighted subgraph with the largest density is extracted as the weighted dense subgraph of the current sliding time window.

[0037] In S2, for each edge of the weighted dense subgraph across all time windows on the same time axis, the time-update weights are accumulated along with the subgraph density, and this accumulation is performed across windows to calculate the time-update cumulative weight. A coefficient of variation is then calculated based on the standard deviation and mean of the samples across all time window sets. This time-update cumulative weight is combined with the coefficient of variation for a stability penalty to eliminate occasional noise, resulting in the global persistent-time-update cumulative weight. The coefficient of variation is obtained by dividing the standard deviation by the mean and adding a small constant close to zero. The purpose of calculating the global persistent-time-update cumulative weight is to address the issue that some edges may suddenly appear with high weights in certain windows but then disappear, representing occasional noise. The coefficient of variation is introduced for this purpose, and through cross-window accumulation and a stability penalty, the final global persistent-time-update cumulative weight is obtained. During implementation, the specific sliding window length and number of windows are directly related to the associated fault log time range; in practical applications, maintenance personnel can flexibly set these based on maintenance cycles, equipment characteristics, and log granularity.

[0038] In S3, the persistent-time cumulative weight of edges in each weighted dense subgraph is calculated, and the upper quartile of the distribution of persistent-time cumulative weights of all edges is taken as the threshold. Edges with persistent-time cumulative weights greater than the threshold are retained, and the other edges are deleted. The resulting weighted dense subgraph is then used as the core subgraph. The extracted core subgraph is written back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

[0039] Graph maintenance mainly involves conflict correction. In long-accumulated maintenance logs, the same pair of nodes (e.g., "Fault A → Repair Method") may have multiple edges, representing different handling methods. For example: Fault A → Repair Method 1 (Restart Controller) Fault A → Repair Method 2 (Replace Part) In this case, using the edge weights in the core subgraph as anchor points, the edges with larger weights are selected as the preferred repair paths.

[0040] Search optimization primarily involves using path scoring to prioritize paths. When a user searches for a solution to a particular fault, multiple reachable paths may exist on the knowledge graph. If a path passes through an edge in the core subgraph, its score is increased. Path scoring is controlled by a scoring function, which is calculated by adding an enhancement coefficient to the ordinary edge weights of the knowledge graph and multiplying it by the cumulative weights of persistence and relevance in the core subgraph. The core subgraph acts as a "weighted amplification" tool during the search phase, making the system more inclined to recommend maintenance paths that are stable in the long term and frequently used in the recent past.

[0041] The specific implementation process of this embodiment is as follows: for a 3-day track maintenance log (switching the window by 1 hour each day). Day 1 09:00–10:00 * Emergency stop alarm → Maintenance personnel arrive. * Maintenance personnel arrive → Replace emergency stop switch * Replace the emergency stop switch → Resume operation * Emergency stop alarm → Remote restart * Remote restart → Restart controller → Resume operation Day 2 09:00–10:00 * Emergency stop alarm → Maintenance personnel arrive. * Maintenance personnel arrive → Replace emergency stop switch * Replace the emergency stop switch → Resume operation Day 3 09:00–10:00 * Emergency stop alarm → Maintenance personnel arrive. * Maintenance personnel arrive → Replace emergency stop switch * Replace the emergency stop switch → Resume operation * Emergency stop alarm → System self-test → Reset button → Resume operation The above maintenance log graph is used for single-window edge weight calculation, i.e., weighting is applied exponentially. The half-life H=1 is set, so the edge weight of the nearest window is 1; it decays by half (0.5) the day before, and by 0.25 the day before that; each time an edge appears, the corresponding weight is added. Then, the results are accumulated across windows, and the accumulated weights are calculated. The results are as follows: Side e Day 1 Day 2 Day 3 Accumulated weight W*(e) Emergency stop alarm → Maintenance personnel arrive at the scene 1 0.5 0.25 1.75 Maintenance personnel arrived → Replaced emergency stop switch 1 0.5 0.25 1.75 Replace emergency stop switch → Resume operation 1 0.5 0.25 1.75 Emergency stop alarm → Remote restart 1 0 0 1 Remote restart → Restart controller 1 0 0 1 Restart controller → Resume operation 1 0 0 1 Emergency stop alarm → System self-check 0 0 0.25 0.25 System self-test → Reset button 0 0 0.25 0.25 Reset button → Restore operation 0 0 0.25 0.25 Calculate the mean μ and standard deviation σ of each edge in different windows to obtain the coefficient of variation CV: Side e Weighted sequence (Day1, Day2, Day3) μ σ CV Adjusted weights Emergency stop alarm → Maintenance personnel arrive at the scene [1,0.5,0.25] 0.58 0.31 0.53 1.14 Repairman arrives → Replaces switch [1,0.5,0.25] 0.58 0.31 0.53 1.14 Replace the switch → Restore operation [1,0.5,0.25] 0.58 0.31 0.53 1.14 Emergency stop alarm → Remote restart [1,0,0] 0.33 0.47 1.41 0.41 Remote restart → Restart controller [1,0,0] 0.33 0.47 1.41 0.41 Restart controller → Resume operation [1,0,0] 0.33 0.47 1.41 0.41 Emergency stop alarm → System self-check [0,0,0.25] 0.08 0.12 1.5 0.1 System self-test → Reset button [0,0,0.25] 0.08 0.12 1.5 0.1 Reset button → Restore operation [0,0,0.25] 0.08 0.12 1.5 0.1 As can be seen: The weight of the core link (emergency stop alarm → maintenance personnel arrive → switch replacement → operation resumes) remains stable and remains high (≈1.14) even after penalties. The "remote restart" factor only appeared once, with large fluctuations, and its weight dropped to 0.41. The edge for "System Self-Test / Reset Button" is even rarer, with a weight of only 0.10.

[0042] Based on the above results, threshold pruning is performed. Assuming the pruning threshold = upper quartile ≈ 0.5, the following edges are retained: 1.14 (three core links); and 0.41 (remote restart related edges) and 0.10 (system self-check related edges) are removed, resulting in the final core subgraph: Emergency stop alarm → Maintenance personnel arrive → Replace emergency stop switch → Operation resumes In long-accumulated maintenance logs, the same fault may correspond to multiple repair methods. For example: Fault: Escalator emergency stop alarm Repair method 1: Restart the controller Repair method 2: Replace the emergency stop switch Based on the edge weights of the core subgraph, the edge with the higher weight is selected as the recommended path, i.e., "replace the emergency stop switch" is selected as the recommended path.

[0043] When a user queries a solution for "emergency stop alarm", the knowledge graph may have multiple reachable paths: Emergency stop alarm → Restart controller → Resume operation Emergency stop alarm → Replace emergency stop switch → Restore operation If an edge in the path is within the core subgraph, the score is improved, i.e., the score for "emergency stop alarm → replace emergency stop switch → resume operation" is improved.

[0044] The method provided in this embodiment integrates "novelty × repetitiveness × dense structure" within a single framework to form a unified core subgraph determination criterion. It employs exponential decay and sliding window recursion, combined with a weighted peeling method to find dense subgraphs, achieving an overall complexity that is approximately linear, meeting online real-time requirements. The coefficient of variation is introduced as a stability constraint to effectively suppress interference from occasional burst edges. The core subgraph can not only be used for retrieval but also for writing back weights to participate in conflict correction and path scoring, forming a complete memory maintenance closed loop. The core subgraph possesses clear edge weights, density, and support indicators, which can be output as evidence, improving system auditability.

Claims

1. A maintenance optimization method for a track maintenance knowledge graph, characterized in that, Includes the following steps: S1. Construct a snapshot sequence graph of maintenance log data, apply exponential time-increment weighting to each edge of the snapshot sequence graph, and extract the weighted dense subgraph of each window based on the time-increment weighting and the sliding time window. S2. For the edges of the weighted dense subgraph across the window, accumulate the time-update weights and the subgraph density to calculate the persistent-time-update cumulative weights. S3. Based on the persistent-time cumulative weight, prune the edges of the weighted dense subgraph according to the threshold to obtain the core subgraph; S4. Write the extracted core subgraph back to the maintenance memory knowledge graph for knowledge maintenance and retrieval optimization.

2. The maintenance optimization method for track maintenance knowledge graph according to claim 1, characterized in that, When constructing a snapshot sequence diagram of maintenance log data in S1, the equipment alarm logs, maintenance processing logs, and maintenance result logs accumulated in the maintenance data according to the time sequence are used to construct a snapshot diagram of maintenance log data at fixed time intervals. Nodes represent equipment, fault type, maintenance action, and result status, and edges represent the causal, temporal, or co-occurrence relationships between events. Multiple snapshot diagrams are then combined according to the time sequence to construct a snapshot sequence diagram.

3. The maintenance optimization method for track maintenance knowledge graph according to claim 1 or 2, characterized in that, When extracting the weighted dense subgraph in S1, the snapshot sequence graph is first initialized with weights. Then, the edges of the snapshot sequence graph are exponentially weighted according to their timeliness according to the sliding time window. Different decay constants are set for different events. Based on the decay constants and the event occurrence time, each edge is assigned a timeliness weight that decays over time, i.e., exponential timeliness weighting, to obtain the weighted subgraph.

4. The maintenance optimization method for the track maintenance knowledge graph according to claim 3, characterized in that, In S1, after weighting and initializing the snapshot sequence graph, the newly weighted density is calculated for the weighted subgraph of each sliding time window. Then, the node with the smallest weight and its associated edges are deleted to obtain a new weighted subgraph. The density is then recalculated, and the above operation is repeated until all nodes are deleted. The weighted subgraph with the largest density is extracted as the weighted dense subgraph of the current sliding time window.

5. The maintenance optimization method for track maintenance knowledge graph according to claim 1, characterized in that, In S2, for each edge of the weighted dense subgraph of all time windows on the same time axis, the time-update weight and the subgraph density are accumulated, and the time-update cumulative weight is calculated by accumulating across windows. A coefficient of variation is calculated based on the standard deviation and mean of each edge in different windows. The time-update cumulative weight is combined with the coefficient of variation to perform stability penalty, eliminate occasional noise, and obtain the global persistent-time-update cumulative weight.

6. The maintenance optimization method for track maintenance knowledge graph according to claim 5, characterized in that, The coefficient of variation is obtained by dividing the standard deviation by the mean and adding a small constant that approaches zero.

7. The maintenance optimization method for track maintenance knowledge graph according to claim 1, characterized in that, In S3, for each weighted dense subgraph, the upper quartile of the distribution of the persistent-time cumulative weights of all edges is taken as the threshold; the weighted dense subgraph after retaining the edges whose persistent-time cumulative weights are greater than the threshold and deleting the other edges is taken as the core subgraph.

8. The maintenance optimization method for track maintenance knowledge graph according to claim 1, characterized in that, When performing knowledge maintenance in S4, multiple edges may appear for the same pair of nodes in the long-term accumulated maintenance logs, representing different processing methods. Using the edge weight in the core subgraph as the anchor point, the edge with the larger weight is selected as the priority recommended repair path.

9. The maintenance optimization method for track maintenance knowledge graph according to claim 1, characterized in that, During the retrieval optimization process in S4, when a user searches for a solution to a certain fault, there may be multiple reachable paths on the knowledge graph. If a path passes through an edge in the core subgraph, the score of that path will be improved.

10. The maintenance optimization method for the track maintenance knowledge graph according to claim 9, characterized in that, Path scoring is controlled by a scoring function, which is obtained by multiplying the ordinary edge weights of the knowledge graph by the enhancement coefficients and the persistent-time cumulative weights in the core subgraph.

Images