Signal system operation and maintenance auxiliary method based on knowledge graph

Abstract

The invention discloses a signal system operation and maintenance auxiliary method based on a knowledge graph, and the method comprises the following steps: S1, collecting signal system operation and maintenance data based on an EvoKG model, and converting the data into an event tetrad set; s2, constructing a multi-scale map and generating fusion embedding representation; s3, extracting context information to generate a vector sequence; s4, inputting a structure adjustment mechanism to construct a dynamic graph structure; s5, constructing a constraint relation between rule atlas definition devices; s6, inputting the dynamic atlas and the rule atlas into an EvoKG model to generate a prediction path and calculating a consistency index; s7, related information is fused to generate a to-be-predicted operation and maintenance event and path; and S8, outputting a prediction path for operation and maintenance processing. The operation and maintenance prediction accuracy and the rule consistency are improved.

Description

Technical Field

[0001] The present invention relates to the field of signal system operation and maintenance technology, and in particular to a signal system operation and maintenance auxiliary method based on knowledge graph. Background Art

[0002] Signal systems, as a core component in ensuring the safe operation of industrial control, traffic dispatching, and infrastructure, have faced a series of challenges in recent years, including large data volumes, complex state correlations, and difficulty in predicting faults, as the variety and complexity of equipment have increased. Signal systems generate a vast amount of historical operational data during operation, encompassing information such as equipment status, operation records, dispatching logic, and event timing, with strong temporal and structural dependencies. Effectively leveraging this data for efficient prediction and decision support has become a key research area in intelligent operations and maintenance.

[0003] To model entity relationships and operational events in complex systems, knowledge graph technology is widely used in equipment-level modeling and fault analysis tasks. In recent years, some graph reasoning methods, such as RGCN and CompGCN, have made progress in structural modeling. However, most of these methods remain at the static graph representation level and cannot reflect the temporal evolution and structural dynamics of events. To address dynamic graph modeling and path prediction tasks, researchers proposed the EvoKG model, which aims to improve reasoning capabilities in dynamic graphs by modeling temporal quadruplets (head entity, relationship, tail entity, time). EvoKG combines structural embedding with temporal modeling mechanisms and is representative in time-aware path generation and graph dynamic evolution modeling.

[0004] Although EvoKG has certain advantages in handling dynamic graph reasoning problems, it still has many shortcomings in actual operation and maintenance scenarios. First, EvoKG mainly relies on static structural input and simple time encoding, lacking a dynamic fusion and structural adjustment mechanism for complex contexts, making it difficult to cope with the graph structure evolution of operation and maintenance events in different time windows and context states. Second, the model design does not consider the common rule constraints in operation and maintenance logic, such as device interlocking, scheduling legitimacy, and operation dependencies. This makes its predicted path lack consistency verification with actual business rules, limiting the feasibility of deployment in actual systems. In addition, existing EvoKG implementations lack integrated modeling of multi-scale graph structures, making it difficult to uniformly handle cross-scale information at the device layer, subsystem layer, and regional layer.

[0005] Therefore, how to provide a signal system operation and maintenance assistance method based on knowledge graph is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a signal system operation and maintenance assistance method based on knowledge graph. The present invention introduces an improved EvoKG model, combines multi-scale graph modeling, context vector-driven structure adjustment mechanism and rule consistency reasoning method, constructs a dynamic graph structure suitable for complex time series operation and maintenance data, realizes path prediction and auxiliary decision-making of signal system operation and maintenance events, and has the advantages of strong structural expression ability, accurate context modeling and high consistency of reasoning result rules.

[0007] A signal system operation and maintenance assistance method based on a knowledge graph according to an embodiment of the present invention includes the following steps:

[0008] S1. Build an input structure based on the EvoKG model, collect historical operation and maintenance data of the signal system, and convert each record into an event quadruple set consisting of an entity pair, event type, and event time;

[0009] S2. Construct a multi-scale graph based on the event quadruple set, establish cross-layer connections and introduce a residual mechanism to generate a fused embedding representation;

[0010] S3. Construct an operation and maintenance event context from the event quadruple set, extract the event type, entity identifier, and time sequence information within the time window, and generate a context vector sequence;

[0011] S4. Input the context vector sequence into the structure adjustment mechanism to adjust the edge connection state and edge weight in the fused embedding representation to generate a context-driven dynamic graph structure;

[0012] S5. Construct a rule graph consisting of entity nodes and rule edges. Rule edges represent the equipment interlocking relationship, scheduling sequence relationship, and operation legality relationship.

[0013] S6. Input the dynamic graph structure and the rule graph into the EvoKG model to generate a predicted path, map the predicted path to the rule graph, calculate the path projection residual and form a consistency index;

[0014] S7, integrating the dynamic graph structure, context vector sequence and consistency index to generate the predicted operation and maintenance events and prediction paths;

[0015] S8. Output the predicted path as input data for signal system operation and maintenance processing.

[0016] Optionally, the multi-scale knowledge graph includes a device layer, a subsystem layer, and a regional layer, the graph nodes represent signal entities, and the graph edges represent connection relationships and scheduling relationships.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Initialize the EvoKG model structure, set the entity set, relationship set and time variable set, set the embedding dimension d and the optimization parameter set Θ;

[0019] S12. Collect historical operation and maintenance data of the signal system, including operation logs, alarm records, and control commands, and extract the start entity identifier, target entity identifier, event type, and timestamp from each record;

[0020] S13. Convert each record into an event quad<h,r,t,τ> , where h represents the starting entity identifier, r represents the event type, t represents the target entity identifier, and τ represents the event time;

[0021] S14, construct event quadruple set E = { <h i , r i , t i , τ i >|i=1,2,…,N}, where h i is the i-th starting entity identifier, r i is the i-th event type, t i is the i-th target entity identifier, τ i is the time of the ith event, and N is the total number of events.

[0022] Optionally, the S2 specifically includes:

[0023] S21. Divide the entity level according to the event quadruple set and build the device layer graph G (1) =(V (1) , E (1) ), subsystem layer graph G (2) =(V (2) , E (2) ), regional layer atlas G (3) =(V (3) , E (3) ), where V (l) Represents the entity set in the l-th layer graph, E (l) represents the edge set in the l-th layer graph, l∈{1,2,3};

[0024] S22. Establish a cross-layer connection edge set E (1→2) and E (2→3) , respectively connect the associated entity nodes in the device layer and subsystem layer, and the subsystem layer and regional layer, and build a cross-layer graph structure G = (V, E), where:

[0025] V=V (1) ∪V (2) ∪V (3) ,E=E (1) ∪E (2) ∪E(3) ∪E (1→2) ∪E (2→3) ;

[0026] S23, perform graph embedding operations on the device layer, subsystem layer, and region layer graph structures respectively to generate an embedding vector set H (1) 、H (2) 、H (3) , where the embedding vector of each node is i represents the node index, l∈{1,2,3}, and d represents the embedding dimension;

[0027] S24. Build a residual fusion model, perform weighted fusion on the three-layer embedding vectors, and generate a fused embedding representation H f :

[0028] H f =H (1) +α2·(W2H (2) +b2)+α3·(W3H (3) +b3);

[0029] Among them, α2 and α3 are fusion weight coefficients, is the linear transformation matrix, is the bias vector, H f Representation fusion embedding representation.

[0030] Optionally, the S3 specifically includes:

[0031] S31, setting the time window length and the sliding step size, filtering the event quadruple whose timestamp is within a fixed time interval before the current reference time from the event quadruple set, and generating a context event subset;

[0032] S32, extracting the starting entity identifier h from each event quadruple in the context event subset i , event type r i and target entity identifier t i , forming the event identifier combination s i =(h i ,r i ,t i ), generate a time-ordered event identifier combination sequence S = [s1, s2, ..., s n ];

[0033] S33. Construct an embedding mapping function to map each event identifier combination into a context vector:

[0034] e i =W·[E h (h i )||Er (r i )||E t (t i )]+b;

[0035] Among them, E h , E r , E t Represent the embedding functions of the starting entity, event type and target entity respectively, the output dimension is d, and the symbol || represents the splicing operation. is the linear transformation matrix, is the bias vector, is the context vector;

[0036] S34. Arrange the context vectors in ascending order according to the timestamps of the event quadruple to form a context vector sequence E = [e1, e2, ..., e n ].

[0037] Optionally, the S4 specifically includes:

[0038] S41, constructing the input of the structural adjustment mechanism, the input includes the context vector sequence E = [e1, e2, ..., e n ] and the fused embedding representation H f ;

[0039] S42. Based on the node index in the fused embedding representation, extract the starting node and target node of the connecting edge between nodes, find the context vector of the corresponding entity in the context vector sequence, calculate the context correlation between the node pairs, and generate the context adjustment matrix Where N represents the number of nodes in the graph structure, and the matrix element a ij Represents the contextual relevance between the i-th node and the j-th node;

[0040] S43. Define the initial edge weight matrix as Construct the adjusted weight matrix:

[0041] W (1) =σ(λA+(1-λ)W (0) );

[0042] Among them, λ∈[0,1] represents the adjustment coefficient, σ represents the normalization function, W (1) is the edge weight matrix after structural adjustment;

[0043] S44, according to the weight matrix W after adjustment (1) , update the edge connection status and generate a context-driven dynamic graph structure.

[0044] Optionally, the S5 specifically includes:

[0045] S51. Parse the signal system operation and maintenance rule document, extract the rule statements containing the content of equipment interlocking, scheduling sequence, and operation legality, and generate a set of rule semantic units;

[0046] S52, perform semantic relationship mapping on each rule semantic unit, extract the starting entity identifier v i , rule type identifier r k With the target entity identifier v j , forming a regular edge triple <v i , r k , v j >, where v i Indicates the rule start entity, v j represents the rule target entity, r k Indicates the rule type, including equipment interlocking, scheduling sequence and operation legality;

[0047] S53. Extract all entity identifiers from the set of regular edge triples and construct an entity node set V. r ={v1, v2, ..., v m}, where m is the number of entity nodes;

[0048] S54. Define the regular edge set f by the regular edge triple set r ,Each rule edge connects two entity nodes and corresponds to a rule type;

[0049] S55, the entity node set V r With the regular edge set f r It is organized into a directed graph structure to form a rule graph. The nodes in the graph represent signal system entities, and the edges in the graph represent rule type relationships. The direction of the edge is consistent with the semantic direction of the rule.

[0050] Optionally, the S6 specifically includes:

[0051] S61. Input the context-driven dynamic graph structure and rule graph into the EvoKG model, select entity node pairs as path start and end points, and construct the predicted path set P = {p1, p2, ..., p k}, where each path p i Represents a path sequence consisting of alternating entity nodes and relationship edges in represents the j-th entity node, represents the jth relationship edge, l i Indicates the number of relationship edges in the path;

[0052] S62. For each path p i Construct an alignment path p′ with consistent node sequences in a regular graph i,According to the relationship type comparison and matching rule edge sequence, an edge pair set is established;

[0053] S63. The path projection residual is defined as the mean squared error distance between the predicted path and the relationship edge embedding representation in the aligned path:

[0054]

[0055] in Representing relationship edges in rule graphs The embedding representation of Represents the relationship edge in the dynamic graph structure The embedding representation of is the linear transformation matrix, ‖·‖2 represents the L2 norm;

[0056] S64. Construct path complexity penalty function ψ(l i )=log(1+l i ), construct rule matching factor Where R represents the set of valid relationship types preset in the rule graph, and δ(·) is a Boolean indicator function;

[0057] S65. Construction path consistency index:

[0058] c i =exp(-βr i -γψ(l i )+λθ i );

[0059] Among them, β, γ, and λ are non-negative real weight parameters;

[0060] S66, the consistency index sequence C = {c1, c2, ..., c k}Establish a one-to-one correspondence with the predicted path set P and input the next processing step.

[0061] Optionally, the S7 specifically includes:

[0062] S71, extracting node embedding representations in the dynamic graph structure to form an embedding representation set;

[0063] S72. Extract a context vector sequence and correspond it to the node embedding representation in chronological order to form a context vector set.

[0064] S73. Extract a set of path consistency indicators and record the consistency score of each predicted path;

[0065] S74. For each predicted path, fuse the embedded representations, context vectors, and consistency scores of the start and target nodes to generate a fused representation set.

[0066] S75. Generate a set of predicted paths and a set of operation and maintenance events to be predicted based on the fused representation set.

[0067] The beneficial effects of the present invention are:

[0068] This paper builds a knowledge graph reasoning mechanism adapted to signal system operation and maintenance scenarios and introduces a structurally improved EvoKG model, enabling efficient modeling and intelligent prediction of historical operation and maintenance data. Compared to existing static graph methods, this paper employs an event quadruple representation consisting of entity pairs, event types, and time information. This accurately expresses the structural changes of operation and maintenance events over time, significantly enhancing the model's ability to model dynamic operation and maintenance behaviors.

[0069] This paper proposes a multi-scale graph construction method that integrates structural information from the device, subsystem, and region layers into a unified graph framework. Leveraging cross-scale connections and residual mechanisms, this method achieves multi-granular semantic fusion, enhancing the graph's expressive power and hierarchical adaptability. Furthermore, a structural adjustment mechanism based on context vector sequences is introduced to dynamically adjust the connection states and weights within the graph structure based on the entity, type, and sequence characteristics within the event window, enabling the dynamic graph to better reflect the true context and characteristics of the event.

[0070] During the path inference phase, this invention integrates a rule graph for the first time, constructing a set of rule edges that represent the relationships between device interlocking, scheduling sequence, and operation legitimacy. By calculating path projection residuals and extracting rule matching factors, it measures the consistency of the predicted path, effectively preventing inference results from violating business logic. This mechanism enhances the constraint and credibility of path output, providing support for explainable intelligent operation and maintenance of signal systems.

[0071] Overall, the present invention has made systematic innovations in graph structure modeling, temporal context fusion, and rule consistency reasoning. It can realize operation and maintenance path prediction tasks in highly dynamic and complex signal systems, and has the advantages of fine modeling granularity, high reasoning accuracy, and strong compliance of predicted paths. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0073] Figure 1 This is a flow chart of a signal system operation and maintenance auxiliary method based on knowledge graph proposed by the present invention;

[0074] Figure 2 This is a schematic diagram of the context vector driven structural adjustment mechanism of the knowledge graph-based signal system operation and maintenance assistance method proposed in the present invention;

[0075] Figure 3 This is a schematic diagram of the rule graph path consistency reasoning process of the knowledge graph-based signal system operation and maintenance assistance method proposed in the present invention. DETAILED DESCRIPTION

[0076] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0077] refer to Figure 1-3 , a signal system operation and maintenance auxiliary method based on knowledge graph, including the following steps:

[0078] S1. Build an input structure based on the EvoKG model, collect historical operation and maintenance data of the signal system, and convert each record into an event quadruple set consisting of an entity pair, event type, and event time;

[0079] S2. Construct a multi-scale graph based on the event quadruple set, establish cross-layer connections and introduce a residual mechanism to generate a fused embedding representation;

[0080] S3. Construct an operation and maintenance event context from the event quadruple set, extract the event type, entity identifier, and time sequence information within the time window, and generate a context vector sequence;

[0081] S4. Input the context vector sequence into the structure adjustment mechanism to adjust the edge connection state and edge weight in the fused embedding representation to generate a context-driven dynamic graph structure;

[0082] S5. Construct a rule graph consisting of entity nodes and rule edges. Rule edges represent the equipment interlocking relationship, scheduling sequence relationship, and operation legality relationship.

[0083] S6. Input the dynamic graph structure and the rule graph into the EvoKG model to generate a predicted path, map the predicted path to the rule graph, calculate the path projection residual and form a consistency index;

[0084] S7, integrating the dynamic graph structure, context vector sequence and consistency index to generate the predicted operation and maintenance events and prediction paths;

[0085] S8. Output the predicted path as input data for signal system operation and maintenance processing.

[0086] The present invention proposes a signal system operation and maintenance assistance method that integrates evolutionary graph reasoning and rule constraint mechanism, which can unify the modeling of structural representation, context information and consistency indicators. In the links of graph construction, context extraction, structural adjustment, path reasoning and result output, by introducing the collaborative input of EvoKG model and rule graph, the prediction results not only have temporal and structural consistency, but can also dynamically adapt to the behavioral logic in different operation and maintenance scenarios, effectively improving the recognition ability and response efficiency of complex events.

[0087] In this embodiment, the multi-scale knowledge graph includes a device layer, a subsystem layer, and a regional layer. The graph nodes represent signal entities, and the graph edges represent connection relationships and scheduling relationships.

[0088] The present invention introduces a multi-scale structure in the process of knowledge graph construction, covering the device layer, subsystem layer and regional layer, so that operation and maintenance entities of different granularities in the signal system can form a hierarchical expression. Through node semantic aggregation and cross-layer connection mechanism, the graph's modeling ability of high-dimensional relationships is enhanced, while providing a more discriminative input basis for subsequent graph reasoning, significantly improving the ability to capture complex structural dependencies.

[0089] In this embodiment, S1 specifically includes:

[0090] S11. Initialize the EvoKG model structure, set the entity set, relationship set and time variable set, set the embedding dimension d and the optimization parameter set Θ;

[0091] S12. Collect historical operation and maintenance data of the signal system, including operation logs, alarm records, and control commands, and extract the start entity identifier, target entity identifier, event type, and timestamp from each record;

[0092] S13. Convert each record into an event quad<h,r,t,τ> , where h represents the starting entity identifier, r represents the event type, t represents the target entity identifier, and τ represents the event time;

[0093] S14, construct event quadruple set E = { <h i , r i , t i , τ i >|i=1,2,…,N}, where h i is the i-th starting entity identifier, r i is the i-th event type, t i is the i-th target entity identifier, τ i is the time of the ith event, and N is the total number of events.

[0094] This paper standardizes historical signal system operation and maintenance data into uniformly structured event quads, ensuring consistent representation of event types, subject entities, object entities, and time information within the graph. Introducing normalization and embedding mapping during input structure construction helps alleviate the sparsity issues associated with heterogeneous data, providing a clear and learnable input representation for the EvoKG model, effectively improving the model's learning efficiency and generalization capabilities.

[0095] In this embodiment, S2 specifically includes:

[0096] S21. Divide the entity level according to the event quadruple set and build the device layer graph G (1) =(V (1) ,E (1) ), subsystem layer graph G (2) =(V (2) ,E (2) ), regional layer atlas G (3) =(V (3) ,E (3) ), where V (l) Represents the entity set in the l-th layer graph, E (l) represents the edge set in the l-th layer graph, l∈{1,2,3};

[0097] S22. Establish a cross-layer connection edge set E (1→2) and E (2→3) , respectively connect the associated entity nodes in the device layer and subsystem layer, and the subsystem layer and regional layer, and build a cross-layer graph structure G = (V, E), where:

[0098] V=V (1) ∪V (2) ∪V (3) ,E=E (1) ∪E (2) ∪E (3) ∪E (1→2) ∪E (2→3) ;

[0099] S23, perform graph embedding operations on the device layer, subsystem layer, and region layer graph structures respectively to generate an embedding vector set H (1) 、H (2) 、H (3) , where the embedding vector of each node is i represents the node index, l∈{1, 2, 3}, and d represents the embedding dimension;

[0100] S24. Build a residual fusion model, perform weighted fusion on the three-layer embedding vectors, and generate a fused embedding representation H f :

[0101] H f=H (1) +α2·(W2H (2) +b2)+α3·(W3H (3) +b3);

[0102] Among them, α2 and α3 are fusion weight coefficients, is the linear transformation matrix, is the bias vector, H f Representation fusion embedding representation.

[0103] This paper introduces a residual mechanism and cross-scale projection mapping in the multi-scale graph modeling process. By learning the graph structure representations at different levels and then performing residual fusion, local semantic features and global structural relationships can be preserved simultaneously. The introduction of residual paths effectively avoids the problem of information degradation in the process of deep graph feature fusion, and provides more stable embedding input for the subsequent structural adjustment and path reasoning of the EvoKG model.

[0104] In this embodiment, S3 specifically includes:

[0105] S31, setting the time window length and the sliding step size, filtering the event quadruple whose timestamp is within a fixed time interval before the current reference time from the event quadruple set, and generating a context event subset;

[0106] S32, extracting the starting entity identifier h from each event quadruple in the context event subset i , event type r i and target entity identifier t i , forming the event identifier combination s i =(h i , r i , t i ), generate a time-ordered event identifier combination sequence S = [s1, s2, ..., s n ];

[0107] S33. Construct an embedding mapping function to map each event identifier combination into a context vector:

[0108] e i =W·[E h (h i )||E r (r i )||E t (t i )]+b;

[0109] Among them, E h , E r , E tRepresent the embedding functions of the starting entity, event type and target entity respectively, the output dimension is d, and the symbol || represents the splicing operation. is the linear transformation matrix, is the bias vector, is the context vector;

[0110] S34. Arrange the context vectors in ascending order according to the timestamps of the event quadruple to form a context vector sequence E = [e1, e2, ..., e n ].

[0111] This paper introduces a time window filtering mechanism when constructing context vector sequences, aggregating historical event fragments based on event type, entity identifier, and occurrence sequence. A context embedding mapping function transforms raw event features into a unified time series vector representation, providing a temporal logic basis for dynamic graph structure adjustment and path screening. This effectively enhances the model's ability to model event evolution and improves the temporal consistency of predicted paths.

[0112] In this embodiment, the S4 specifically includes:

[0113] S41, constructing the input of the structural adjustment mechanism, the input includes the context vector sequence E = [e1, e2, ..., e n ] and the fused embedding representation H f ;

[0114] S42. Based on the node index in the fused embedding representation, extract the starting node and target node of the connecting edge between nodes, find the context vector of the corresponding entity in the context vector sequence, calculate the context correlation between the node pairs, and generate the context adjustment matrix Where N represents the number of nodes in the graph structure, and the matrix element a ij Represents the contextual relevance between the i-th node and the j-th node;

[0115] S43. Define the initial edge weight matrix as Construct the adjusted weight matrix:

[0116] W (1) =σ(λA+(1-λ)W (0) );

[0117] Among them, λ∈[0,1] represents the adjustment coefficient, σ represents the normalization function, W (1) is the edge weight matrix after structural adjustment;

[0118] S44, according to the weight matrix W after adjustment (1) , update the edge connection status and generate a context-driven dynamic graph structure.

[0119] The proposed structural adjustment mechanism uses a sequence of context vectors as input and dynamically adjusts the edge connection states and edge weights in the fused embedding representation. By introducing edge state masks and edge weight reconstruction modules, the graph structure can adjust its connection structure in real time based on context changes, improving the adaptability of the graph representation to the current task objectives and supporting the generation of dynamic graph structures.

[0120] In this embodiment, the S5 specifically includes:

[0121] S51. Parse the signal system operation and maintenance rule document, extract the rule statements containing the content of equipment interlocking, scheduling sequence, and operation legality, and generate a set of rule semantic units;

[0122] S52, perform semantic relationship mapping on each rule semantic unit, extract the starting entity identifier v i , rule type identifier r k With the target entity identifier v j , forming a regular edge triple <v i , r k ,v j >, where v i Indicates the rule start entity, v j represents the rule target entity, r k Indicates the rule type, including equipment interlocking, scheduling sequence and operation legality;

[0123] S53. Extract all entity identifiers from the set of regular edge triples and construct an entity node set V. r ={v1,v2,…,v m}, where m is the number of entity nodes;

[0124] S54. Define the regular edge set f by the regular edge triple set r ,Each rule edge connects two entity nodes and corresponds to a rule type;

[0125] S55, the entity node set V r With the regular edge set f r It is organized into a directed graph structure to form a rule graph. The nodes in the graph represent signal system entities, and the edges in the graph represent rule type relationships. The direction of the edge is consistent with the semantic direction of the rule.

[0126] During the construction of the rule graph, this method encodes device interlocking, scheduling sequence, and operation legality as rule-edge triplets, enabling existing knowledge in the system to participate in the reasoning process in a structured manner. The rule graph, embedded as a symbolic constraint structure within the EvoKG model, effectively ensures that the predicted path does not violate known operation and maintenance rules, improving the interpretability and reliability of the reasoning results while also enhancing the logical consistency of the graph reasoning.

[0127] In this embodiment, S6 specifically includes:

[0128] S61. Input the context-driven dynamic graph structure and rule graph into the EvoKG model, select entity node pairs as path start and end points, and construct the predicted path set P = {p1, p2, ..., p k}, where each path p i Represents a path sequence consisting of alternating entity nodes and relationship edges in represents the j-th entity node, represents the jth relationship edge, l i Indicates the number of relationship edges in the path;

[0129] S62. For each path p i Construct an alignment path p′ with consistent node sequences in a regular graph i ,According to the relationship type comparison and matching rule edge sequence, an edge pair set is established;

[0130] S63. The path projection residual is defined as the mean squared error distance between the predicted path and the relationship edge embedding representation in the aligned path:

[0131]

[0132] in Representing relationship edges in rule graphs The embedding representation of Represents the relationship edge in the dynamic graph structure The embedding representation of is the linear transformation matrix, ‖·‖2 represents the L2 norm;

[0133] S64. Construct path complexity penalty function ψ(l i )=log(1+l i ), construct rule matching factor Where R represents the set of valid relationship types preset in the rule graph, and δ(·) is a Boolean indicator function;

[0134] S65. Construction path consistency index:

[0135] c i =exp(-βr i -γψ(l i )+λθ i );

[0136] Among them, β, γ, and λ are non-negative real weight parameters;

[0137] S66, the consistency index sequence C={c1,c2,…,c k}Establish a one-to-one correspondence with the predicted path set P and input the next processing step.

[0138] This paper introduces a path complexity penalty function and a rule matching factor into the path inference process, and calculates a consistency index based on the set of valid relationships in the rule graph. The path projection residual reflects the difference between the current predicted path and the rule path, the complexity penalty term controls path length redundancy, and the rule matching factor screens out path candidates that do not meet the constraints. Together, they achieve a well-structured and logically feasible predicted path generation.

[0139] In this embodiment, the S7 specifically includes:

[0140] S71, extracting node embedding representations in the dynamic graph structure to form an embedding representation set;

[0141] S72. Extract a context vector sequence and correspond it to the node embedding representation in chronological order to form a context vector set.

[0142] S73. Extract a set of path consistency indicators and record the consistency score of each predicted path;

[0143] S74. For each predicted path, fuse the embedded representations, context vectors, and consistency scores of the start and target nodes to generate a fused representation set.

[0144] S75. Generate a set of predicted paths and a set of operation and maintenance events to be predicted based on the fused representation set.

[0145] The present invention jointly models the dynamic graph structure, context vector sequence and consistency index, and finally outputs the predicted operation and maintenance events and their corresponding paths through feature fusion and inference path scoring mechanism. This fusion method improves the degree of consistency between the predicted output and the actual operation and maintenance situation, enhances the model's response ability to key events, and provides a supporting basis for the intelligent prediction scheduling and maintenance operations of the signal system.

[0146] Example 1:

[0147] To verify the feasibility of this invention, we selected three months of historical operation and maintenance data from a power dispatch center within its coverage area. The data sources included equipment monitoring records, operation logs, fault alarms, and dispatch instruction texts, totaling over two million entries, covering over 400 signaling system entities and over 160 equipment subsystems. Traditional operation and maintenance methods rely on manual analysis of system logs and event records, with personnel conducting event classification, path identification, and response recommendations based on their experience. This approach suffers from slow response, incomplete identification, and low utilization of relevant information.

[0148] Based on the knowledge graph construction mechanism, this paper extracts the structural and temporal relationships between event entities and establishes a dynamic graph representation that integrates structural expression with semantic context. On this basis, it uses an improved EvoKG model to model and predict the evolutionary paths of signal events, generate structural consistency evaluation indicators, and combine rule knowledge to generate operation and maintenance recommendations.

[0149] During the experiment, a comparison was conducted between the traditional manual assistance method and the proposed method. The evaluation dimensions covered analysis efficiency, path accuracy, warning effectiveness, and consistency indicator distribution. The specific comparison is shown in Table 1:

[0150] Table 1 Comparative analysis of the effects of intelligent operation and maintenance of signal systems

[0151]

[0152] In terms of analytical efficiency, traditional methods take an average of over nine minutes per dispatcher to process a complex event. However, the proposed method, which incorporates a dynamic graph reasoning mechanism and combines semantic contextual representation, rapidly constructs event paths and assesses their plausibility, reducing the average processing time for a single event to 2.1 minutes. For an entire batch of sample data, traditional manual methods require nearly 15 hours, while the automated method takes just over three hours, significantly alleviating the burden of manual processing.

[0153] From an accuracy perspective, traditional manual analysis is limited by the information processing dimension and often suffers from path omissions or misconnections. This method integrates event entities, operation times, and scheduling rules into the graph modeling process, resulting in more reasonable and comprehensive predicted paths. The method accurately identified 90 paths overall, significantly exceeding the 68 identified by manual analysis. There were only two instances of redundant paths or misidentifications, a reduction of over 80% compared to manual methods.

[0154] In terms of assisted operations and maintenance, the device targets recommended based on the predicted paths are more targeted, with a hit rate of 92.4%, significantly better than the 65.8% of traditional methods. For points with unreasonable configurations or scheduling logic vulnerabilities, this method can screen them through rule matching and path residual indicators, identifying 19 valid anomalies, a significant improvement over the manual method, which only identified 7.

[0155] In addition, through consistency index evaluation, only 38.1% of the manually constructed paths meet the scheduling rule conditions, while 71.6% of the paths output by this method remain within the rule consistency threshold, showing that it has stronger intrinsic control capabilities in automatic path generation and rule alignment.

[0156] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A signal system operation and maintenance auxiliary method based on knowledge graph, characterized in that: The steps include: S1. Build an input structure based on the EvoKG model, collect historical operation and maintenance data of the signal system, and convert each record into an event quadruple set consisting of an entity pair, event type, and event time; S2. Construct a multi-scale graph based on the event quadruple set, establish cross-layer connections and introduce a residual mechanism to generate a fused embedding representation; S3. Construct an operation and maintenance event context from the event quadruple set, extract the event type, entity identifier, and time sequence information within the time window, and generate a context vector sequence; S4. Input the context vector sequence into the structure adjustment mechanism to adjust the edge connection state and edge weight in the fused embedding representation to generate a context-driven dynamic graph structure; S5. Construct a rule graph consisting of entity nodes and rule edges. Rule edges represent the equipment interlocking relationship, scheduling sequence relationship, and operation legality relationship. S6. Input the dynamic graph structure and the rule graph into the EvoKG model to generate a predicted path, map the predicted path to the rule graph, calculate the path projection residual and form a consistency index; S7, integrating the dynamic graph structure, context vector sequence and consistency index to generate the predicted operation and maintenance events and prediction paths; S8. Output the predicted path as input data for signal system operation and maintenance processing.

2. A signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1, characterized in that: The multi-scale knowledge graph includes a device layer, a subsystem layer, and a regional layer. The graph nodes represent signal entities, and the graph edges represent connection relationships and scheduling relationships.

3. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: Said S1 specifically includes: S11. Initialize the EvoKG model structure, set the entity set, relationship set and time variable set, set the embedding dimension d and the optimization parameter set Θ; S12. Collect historical operation and maintenance data of the signal system, including operation logs, alarm records, and control commands, and extract the start entity identifier, target entity identifier, event type, and timestamp from each record; S13. Convert each record into an event quad<h,r,t,τ> , where h represents the starting entity identifier, r represents the event type, t represents the target entity identifier, and τ represents the event time; S14, construct event quadruple set E = { <h i , r i ,t i , τ i >|i=1,2,…,N}, where h i is the i-th starting entity identifier, r i is the i-th event type, t i is the i-th target entity identifier, τ i is the time of the ith event, and N is the total number of events.

4. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S2 specifically includes: S21. Divide the entity level according to the event quadruple set and build the device layer graph G (1) =(V (1) , E (1) ), subsystem layer graph G (2) =(V (2) , E (2) ), regional layer atlas G (3) =(V (3) , E (3) ), where V (l) Represents the entity set in the l-th layer graph, E (l) represents the edge set in the l-th layer graph, l∈{1,2,3}; S22. Establish a cross-layer connection edge set E (1→2) and E (2→3) , respectively connect the associated entity nodes in the device layer and subsystem layer, and in the subsystem layer and regional layer, and construct a cross-layer graph structure G = (V, E), where: V=V (1) ∪V (2) ∪V (3) ,E=E (1) ∪E (2) ∪E (3) ∪E (1→2) ∪E (2→3) ; S23, perform graph embedding operations on the device layer, subsystem layer, and region layer graph structures respectively to generate an embedding vector set H (1) 、H (2) 、H (3) , where the embedding vector of each node is i represents the node index, l∈{1, 2, 3}, and d represents the embedding dimension; S24. Build a residual fusion model, perform weighted fusion on the three-layer embedding vectors, and generate a fused embedding representation H f : <h2 style=";text-align:left;direction:ltr">H<h2 style=";text-align:left;direction:ltr"> f <h2 style=";text-align:left;direction:ltr"> =H<h2 style=";text-align:left;direction:ltr"> (1) <h2 style=";text-align:left;direction:ltr"> +α2·(W2H<h2 style=";text-align:left;direction:ltr"> (2) <h2 style=";text-align:left;direction:ltr"> +b2)+α3·(W3H<h2 style=";text-align:left;direction:ltr"> (3) <h2 style=";text-align:left;direction:ltr"> +b3); Among them, α2 and α3 are fusion weight coefficients, is the linear transformation matrix, is the bias vector, H f Representation fusion embedding representation.

5. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S3 specifically includes: S31, setting the time window length and the sliding step size, filtering the event quadruple whose timestamp is within a fixed time interval before the current reference time from the event quadruple set, and generating a context event subset; S32, extracting the starting entity identifier h from each event quadruple in the context event subset i , event type r i and target entity identifier t i , forming the event identifier combination s i =(h i , r i , t i ), generate a time-ordered event identifier combination sequence S = [s1, s2, ..., s n ]; S33. Construct an embedding mapping function to map each event identifier combination into a context vector: e i =W·[E h (h i )||E r (r i )||E t (t i )]+b; Among them, E h ,E r , E t Represent the embedding functions of the starting entity, event type and target entity respectively, the output dimension is d, and the symbol || represents the splicing operation. is the linear transformation matrix, is the bias vector, is the context vector; S34. Arrange the context vectors in ascending order according to the timestamps of the event quadruple to form a context vector sequence E = [e1, e2, ..., e n ].

6. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S4 specifically includes: S41, constructing the input of the structural adjustment mechanism, the input includes the context vector sequence E = [e1, e2, ..., e n ] and the fused embedding representation H f ; S42. Based on the node index in the fused embedding representation, extract the starting node and target node of the connecting edge between nodes, find the context vector of the corresponding entity in the context vector sequence, calculate the context correlation between the node pairs, and generate the context adjustment matrix Where N represents the number of nodes in the graph structure, and the matrix element a ij Represents the contextual relevance between the i-th node and the j-th node; S43. Define the initial edge weight matrix as Construct the adjusted weight matrix: W (1) =σ(λA+(1-λ)W (0) ); Among them, λ∈[0,1] represents the adjustment coefficient, σ represents the normalization function, W (1) is the edge weight matrix after structural adjustment; S44, according to the weight matrix W after adjustment (1) , update the edge connection status and generate a context-driven dynamic graph structure.

7. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S5 specifically includes: S51. Parse the signal system operation and maintenance rule document, extract the rule statements containing the content of equipment interlocking, scheduling sequence, and operation legality, and generate a set of rule semantic units; S52, perform semantic relationship mapping on each rule semantic unit, extract the starting entity identifier v i , rule type identifier r k With the target entity identifier v j , forming a regular edge triple <v i ,r k ,v j >, where v i Indicates the rule start entity, v j represents the rule target entity, r k Indicates the rule type, including equipment interlocking, scheduling sequence and operation legality; S53. Extract all entity identifiers from the set of regular edge triples and construct an entity node set V. r ={v1,v2,…,v m }, where m is the number of entity nodes; S54. Define the regular edge set f by the regular edge triple set r ,Each rule edge connects two entity nodes and corresponds to a rule type; S55, the entity node set V r With the regular edge set f r It is organized into a directed graph structure to form a rule graph. The nodes in the graph represent signal system entities, and the edges in the graph represent rule type relationships. The direction of the edge is consistent with the semantic direction of the rule.

8. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S6 specifically includes: S61. Input the context-driven dynamic graph structure and rule graph into the EvoKG model, select entity node pairs as path start and end points, and construct the predicted path set P = {p1, p2, ..., p k }, where each path p i Represents a path sequence consisting of alternating entity nodes and relationship edges in represents the j-th entity node, represents the jth relationship edge, l i Indicates the number of relationship edges in the path; S62. For each path p i Construct an alignment path p′ with consistent node sequences in a regular graph i ,According to the relationship type comparison and matching rule edge sequence, an edge pair set is established; S63. The path projection residual is defined as the mean squared error distance between the predicted path and the relationship edge embedding representation in the aligned path: in Representing relationship edges in rule graphs The embedding representation of Represents the relationship edge in the dynamic graph structure The embedding representation of is the linear transformation matrix, ‖·‖2 represents the L2 norm; S64. Construct path complexity penalty function ψ(l i )=log(1+l i ), construct rule matching factor Where R represents the set of valid relationship types preset in the rule graph, and δ(·) is a Boolean indicator function; S65. Construction path consistency index: c i =exp(-βr i -γψ(l i )+λθ i ); Among them, β, γ, and λ are non-negative real weight parameters; S66, the consistency index sequence C = {c1, c2, ..., c k }Establish a one-to-one correspondence with the predicted path set P and input the next processing step.

9. The signal system operation and maintenance auxiliary method based on knowledge graph according to claim 1 is characterized in that: The S7 specifically includes: S71, extracting node embedding representations in the dynamic graph structure to form an embedding representation set; S72. Extract a context vector sequence and correspond it to the node embedding representation in chronological order to form a context vector set. S73. Extract a set of path consistency indicators and record the consistency score of each predicted path; S74. For each predicted path, fuse the embedded representations, context vectors, and consistency scores of the start and target nodes to generate a fused representation set. S75. Generate a set of predicted paths and a set of operation and maintenance events to be predicted based on the fused representation set.

Images