An asset whole-cycle management method based on a knowledge graph

Abstract

The application discloses a kind of asset whole cycle management methods based on knowledge graph, comprising the following steps: obtaining multi-source asset data and pre-processing;Asset whole cycle time series modeling data set is carried out semantic coding;Through the subject filtering CasRel entity-relation joint extraction model, generate structured triple candidate set;According to the constraint set of pre-established execution CasRel result re-score and modify;Time information is mapped as node and is associated with the modified structured triple set;Time series multivariate set is imported into graph database and constructs asset whole cycle time series knowledge graph.The application introduces subject filtering CasRel and life cycle-time priori gating mechanism, constructs asset whole cycle time series knowledge graph, realizes the technical effect that asset relation extraction is more accurate, stage semantics is more consistent, whole cycle structure is more reliable.

Description

Technical Field

[0001] This invention relates to the field of asset management, and in particular to a knowledge graph-based method for full-lifecycle asset management. Background Technology

[0002] Currently, the asset management field generally employs traditional methods based on structured fields, rule matching, and template extraction to organize and analyze basic asset information, operational event records, and lifecycle stage information. Some advanced solutions have begun to utilize pre-trained models to encode asset-related text and construct asset knowledge bases through general entity recognition, relation extraction, and event extraction techniques. However, existing methods typically separate text encoding, entity relation extraction, lifecycle stage judgment, and time series processing into multiple independent modules, lacking a unified data semantic representation system. This makes it difficult to align cross-system, multi-source, heterogeneous asset data within the same semantic space. Furthermore, existing extraction techniques primarily focus on general entity relation patterns and are not designed for asset lifecycle characteristics, making it difficult to support consistent modeling throughout the asset lifecycle and the need for time-series semantic association.

[0003] Existing knowledge graph-based asset management methods largely rely on post-processing rules to filter, sort, and validate triples, making it difficult to guarantee the accuracy and consistency of the extraction process itself. This is especially true when there are multiple asset identifiers, fragmented lifecycle stage descriptions, and inconsistent time recording formats, which can easily lead to asset node splitting, ambiguous event relationships, lifecycle stage jumps, and temporal association errors. Furthermore, existing methods generally lack a unified modeling mechanism for the sequential constraints, mutual exclusion relationships, event-stage matching relationships, and temporal continuity of asset lifecycle stages, resulting in asset knowledge graphs that fail to fully reflect the temporal evolution logic of assets in actual operation. These shortcomings prevent existing technologies from constructing structurally rigorous, semantically consistent, and highly usable full-lifecycle temporal knowledge graphs for assets.

[0004] Therefore, how to provide a knowledge graph-based asset lifecycle management method is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a knowledge graph-based method for full-lifecycle asset management. This invention constructs a full-lifecycle temporal knowledge graph of assets by introducing subject filtering CasRel and a lifecycle-time prior gating mechanism, thereby achieving more accurate asset relationship extraction, more consistent stage semantics, and more reliable full-lifecycle structure.

[0006] According to an embodiment of the present invention, an asset lifecycle management method based on knowledge graphs includes the following steps:

[0007] Acquire multi-source asset data, including basic asset information, operation and maintenance event records, lifecycle stage identifiers, and time information; perform unified formatting, identifier merging, and time standardization processing to form a time-series modeling data set for the entire asset lifecycle.

[0008] The asset lifecycle time-series modeling dataset is input into the ERNIE encoder to perform semantic encoding on asset-related text, resulting in a set of text semantic representation sequences.

[0009] The set of text semantic representation sequences is input into the subject-filtered CasRel entity-relation joint extraction model to identify asset entity nodes, event entity nodes and lifecycle stage nodes, and extract the association relationships between asset entity nodes and event entity nodes and lifecycle stage nodes respectively, generating a candidate set of structured triples.

[0010] Based on the preset set of constraints, the CasRel results of the knowledge graph consistency constraint are re-scored on the candidate set of structured triples. Structured triples that do not meet the constraints are removed and the confidence of the remaining structured triples is adjusted to obtain the corrected set of structured triples.

[0011] Time segment alignment and aggregation are performed on the time information in the asset full-lifecycle time-series modeling dataset, mapping it to time nodes and time segment nodes, and establishing a relationship with the modified structured triple set to form a time-series multi-tuple set;

[0012] Import the time-series tuple set into a graph database, use asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes as graph nodes, and use the relationships in the time-series tuple set as graph edges to construct and output a full-lifecycle time-series knowledge graph of assets.

[0013] Optionally, the generation of the asset full-lifecycle time-series modeling dataset specifically includes:

[0014] Collect basic asset information, operation and maintenance event records, life cycle stage identifiers and time information from asset management system, operation and maintenance management system, monitoring system and file management system, classify data source types, and form original multi-source asset data set;

[0015] The fields of each data source in the original multi-source asset data set are parsed and formatted uniformly, and the fields of different systems are mapped to a preset standard field set to generate a preliminary standardized asset data set.

[0016] Based on the preliminary standardized asset data set, an asset master data table and an asset identifier mapping relationship table are constructed. The device number, asset code, tag number and geographical location identifier from different data source types are merged according to the preset mapping relationship to generate an asset unified identifier data set with the unified asset identifier as the primary key.

[0017] The operation and maintenance event records are associated and matched with the unified asset identifier data set. Using the unified asset identifier as the association key, various operation and maintenance event records are bound to the corresponding asset entities, generating an asset operation and maintenance event association data set with the unified asset identifier, event identifier, and event type as the core fields.

[0018] The lifecycle stage descriptions in the preliminary standardized asset dataset are classified and standardized, a standard vocabulary of lifecycle stages and a mapping relationship between lifecycle stages are constructed, and a unified identifier dataset of lifecycle stages is generated.

[0019] Time standardization processing is performed on the time information in the preliminary standardized asset data set, parsing the date and time strings, time interval descriptions and relative time descriptions into standard time points and standard time intervals under a unified time benchmark, and generating an asset event time series table.

[0020] Using the unified asset identifier as the primary key, the unified asset identifier data set, the asset operation and maintenance event associated data set, the lifecycle stage unified identifier data set, and the asset event time series table are associated, merged, and sorted in chronological order to construct a full lifecycle time series modeling data set for assets.

[0021] Optionally, the generation of the text semantic representation sequence set specifically includes:

[0022] From the asset lifecycle time-series modeling dataset, extract asset-related texts based on the unified asset identifier, including asset basic information fields, operation and maintenance event record fields, and lifecycle stage identifier fields. Fill in empty spaces and trim redundant fields to construct an initial set of asset-related texts indexed by the unified asset identifier.

[0023] Sentence segmentation and word segmentation are performed on the initial set of asset-related texts. Long words not in the preset vocabulary are segmented into sub-words. Numbers, special symbols and time descriptions are normalized and replaced to obtain a set of asset-related text segmentation sequences composed of word sequence and sub-word sequence.

[0024] Based on the source type and the asset entity to which each asset-related text belongs in the asset-related text segmentation sequence set, a segment label is assigned to each asset-related text and a position index corresponding to the word position is generated. The life cycle stage identifier is transcribed into the life cycle stage category code. The input feature sequence of the ERNIE encoder is constructed with the unified asset identifier, segment label, position index and life cycle stage category code as constraints.

[0025] The input feature sequence of the ERNIE encoder is fed into the multi-layer coding network of the ERNIE encoder. Attention calculation and feedforward transformation are performed sequentially according to a fixed number of network layers and network structure to obtain a set of word-level semantic vector sequences corresponding to each asset-related text.

[0026] Based on the set of word-level semantic vector sequences, for each asset-related text, the word-level semantic vector corresponding to the aggregation marker at a preset position is selected as the sentence-level semantic vector, and the sentence-level semantic vectors of different source types under the same unified asset identifier are serialized and arranged in chronological order to generate a set of sentence-level semantic vector sequences indexed by the unified asset identifier;

[0027] The set of word-level semantic vector sequences and the set of sentence-level semantic vector sequences are aligned and merged according to unified asset identifiers, time information and lifecycle stage identifiers. Through index binding and structured storage, a set of text semantic representation sequences covering the semantic features of basic asset information, operation and maintenance event records and lifecycle stage identifiers is constructed.

[0028] Optionally, the generation of the candidate set of structured triples specifically includes:

[0029] Based on the asset full-lifecycle time-series modeling dataset and the text semantic representation sequence set, an asset entity candidate set is constructed. The asset name is matched with the corresponding original text of each text semantic representation sequence to form a subject candidate position set. This set is then indexed and bound to the text semantic representation sequence set to generate a text semantic representation sequence set with subject candidate tags.

[0030] Based on the CasRel entity-relation joint extraction model for subject filtering, a subject candidate filtering mask is constructed on the basis of the set of text semantic representation sequences with subject candidate tags. The word positions that are not in the set of subject candidate positions are masked. The subject candidate filtering mask is associated with the corresponding text semantic representation sequence to form a set of text semantic inputs with subject filtering constraints.

[0031] The text semantic input set with subject filtering constraints is fed into the subject entity recognition subnetwork. Based on the subject candidate filtering mask, the subject start position prediction and subject end position prediction are performed on each text semantic representation sequence to obtain the subject entity interval set, which is mapped to form the asset entity node candidate set.

[0032] The candidate set of asset entity nodes and the text semantic input set with subject filtering constraints are fed into the relation type gating prediction subnetwork. The life cycle stage prior gating unit and the time prior gating unit are introduced. The relation type prediction results are gating and filtering based on the unified identifier of the life cycle stage and the life cycle stage category. The time prior features are constructed based on the time information of the asset entity nodes. The preset relation type set is judged one by one to obtain the event relation type subset and the life cycle stage relation type subset corresponding to the asset entity node candidates. These are then bound to the corresponding asset entity node candidates to form an asset entity node candidate set with relation type labels.

[0033] The candidate set of asset entity nodes with relation type labels and the text semantic input set with subject filtering constraints are fed into the object entity extraction subnetwork. For the event association relationship type subset, the event object start position prediction and event object end position prediction are performed in the text semantic representation sequence to extract event entity node candidates. For the life cycle stage association relationship type subset, the life cycle stage object start position prediction and life cycle stage object end position prediction are performed in the text semantic representation sequence to extract life cycle stage node candidates.

[0034] The candidate asset entity nodes, candidate event entity nodes, and candidate lifecycle stage nodes are combined according to the relationship type subset. The structured triplet candidates between asset entity nodes and event entity nodes and lifecycle stage nodes are generated according to the corresponding event relationship type and lifecycle stage relationship type. The three triplet candidates are then summarized to form a set of structured triplet candidates.

[0035] Optionally, the generation of the modified structured triple set specifically includes:

[0036] Based on the asset lifecycle time-series modeling dataset, a constraint set is constructed, which is divided into a lifecycle stage sequential constraint subset, a lifecycle stage mutual exclusion constraint subset, an operation and maintenance event and lifecycle stage matching constraint subset, and a time continuity constraint subset.

[0037] Read the basic confidence scores of the structured triplet candidates, combine them with the source location, asset entity node, event entity node, lifecycle stage node, time prior features and time information to generate a constraint test index list, and form an update confidence score calculation rule based on the weighted summation rule of the basic confidence scores and various consistency scores.

[0038] According to the constraint test index list, the life cycle stage sequence constraint, life cycle stage mutual exclusion constraint, operation and maintenance event and life cycle stage matching constraint and time continuity constraint are retrieved one by one. The corresponding consistency score is calculated respectively. Under the modulation of the time prior feature, each consistency score is substituted into the update confidence calculation rule to obtain the update confidence of the structured triplet candidate.

[0039] The updated confidence level is compared with the preset confidence threshold. Candidate structured triples whose updated confidence level does not reach the threshold and violates any constraint are eliminated. The updated confidence level is used to replace the base confidence level, forming a set of consistent triples with updated confidence level labels.

[0040] The set of consistent triples with updated confidence labels is grouped and sorted according to the unified asset identifier, event entity node and lifecycle stage node. Adjacent triples of the same asset entity node are rearranged according to the lifecycle stage order constraint. Triples with stage jumps or rollbacks are truncated to generate a modified set of structured triples.

[0041] Optionally, the generation of the time-series tuple set specifically includes:

[0042] Extract time information corresponding to standard time points and standard time intervals from the asset full-lifecycle time series modeling data set, collect and organize the time information according to the unified asset identifier, and construct the asset time original sequence set indexed by the unified asset identifier;

[0043] Based on the preset time granularity configuration and time benchmark, time segmentation is performed on each time series in the original asset time sequence set. Standard time points are mapped to discrete time nodes, and the time interval covering continuous standard time points is divided into several time segments according to the time granularity, generating a set of time nodes and a set of time segmentation results.

[0044] Based on the time segment division result set, the start time, end time, unified asset identifier, and covered time node index of each time segment are structured and represented to construct a time segment node set with time segment as the basic unit, and the time segment node set is indexed and bound to the time node set;

[0045] The modified set of structured triples is mapped to the corresponding time nodes and time segment nodes according to the unified asset identifier and time information. Within each time segment node, the associated asset entity nodes, event entity nodes and life cycle stage nodes are summarized. Based on the summary results, a state description and state attributes are generated for the time segment node, and a set of time segment state nodes is constructed.

[0046] Using asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment status nodes as elements, an initial set of time-series tuples is generated based on the association and index relationships in the modified set of structured triples.

[0047] The initial set of time-series tuples is sorted and deduplicated according to the unified asset identifier and time node order. Multivariate association records that do not meet the time continuity requirements due to missing time information and time conflicts are removed, and a time-series tuple set is constructed.

[0048] Optionally, the generation of the asset lifecycle time-series knowledge graph specifically includes:

[0049] Establish a graph schema for a full-lifecycle time-series knowledge graph in a graph database. Define the node types and attribute fields for asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment status nodes. Also define the association relationship types and relationship attribute fields between asset entity nodes and the above nodes to form the graph schema configuration result.

[0050] Based on the time-series tuple set, deduplication and aggregation are performed on asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes. Corresponding graph database node objects are created for each type of node and node attributes are written to form a node writing set.

[0051] Based on the association types in the time-series multi-tuple set, the asset entity nodes are mapped to event entity nodes, lifecycle stage nodes, and time nodes to time segment status nodes. Corresponding relationship edges are created for each multi-tuple association record in the graph database. The updated confidence in the corrected structured triple set is written into the relationship attribute, and the relationship edges are written into the set.

[0052] Consistency checks are performed on the node writing set and the relation edge writing set. The connectivity of nodes and the temporal continuity of relation edges are checked according to the unified asset identifier and time order. Isolated nodes and relation edges that do not meet the time continuity requirements are removed, and duplicate nodes are merged to form a graph dataset which is then registered in the graph database as an asset full-lifecycle time-series knowledge graph.

[0053] The beneficial effects of this invention are:

[0054] This invention constructs a time-series modeling dataset for the entire asset lifecycle, introduces an ERNIE encoder to obtain a set of textual semantic representation sequences covering basic asset information, operation and maintenance event records, and lifecycle stage identifiers. Based on this, a subject-filtering CasRel entity-relation joint extraction model is adopted, combining lifecycle stage prior gating units and time prior gating units. In the entity-relation extraction stage, candidate relations are screened and constrained using unified lifecycle stage identifiers and time prior features, effectively compressing the search space, reducing the probability of invalid triples and contradictory triples, and improving the accuracy and confidence of extracting the associations between asset entity nodes, event entity nodes, and lifecycle stage nodes, providing a high-quality structured foundation for subsequent knowledge graph time-series modeling.

[0055] Building upon this foundation, this invention further utilizes a constraint set comprised of lifecycle phase sequence constraints, lifecycle phase mutual exclusion constraints, operation and maintenance event and lifecycle phase matching constraints, and temporal continuity constraints. Combined with temporal prior features, it performs CasRel result rescoring and sequence rearrangement on structured triple candidates using knowledge graph consistency constraints. Through time segment alignment and aggregation, it generates time nodes, time segment nodes, and time segment state nodes, constructing a temporal tuple set containing asset entity nodes, event entity nodes, lifecycle phase nodes, time nodes, and time segment state nodes. Ultimately, this forms a full-lifecycle temporal knowledge graph in the graph database, significantly improving the temporal continuity and lifecycle consistency of asset state evolution descriptions. This provides structured, queryable, and reasonable temporal knowledge support for refined full-lifecycle asset management, operation and maintenance decisions, and risk tracing. Attached Figure Description

[0056] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0057] Figure 1 This is a flowchart of a knowledge graph-based asset lifecycle management method proposed in this invention;

[0058] Figure 2 This is a schematic diagram illustrating the generation of a structured triplet candidate set for an asset lifecycle management method based on knowledge graphs proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figures 1-2A knowledge graph-based method for full-lifecycle asset management includes the following steps:

[0061] Acquire multi-source asset data, including basic asset information, operation and maintenance event records, lifecycle stage identifiers, and time information; perform unified formatting, identifier merging, and time standardization processing to form a time-series modeling data set for the entire asset lifecycle.

[0062] The asset lifecycle time-series modeling dataset is input into the ERNIE encoder to perform semantic encoding on asset-related text, resulting in a set of text semantic representation sequences.

[0063] The set of text semantic representation sequences is input into the subject-filtered CasRel entity-relation joint extraction model to identify asset entity nodes, event entity nodes and lifecycle stage nodes, and extract the association relationships between asset entity nodes and event entity nodes and lifecycle stage nodes respectively, generating a candidate set of structured triples.

[0064] Based on the preset set of constraints, the CasRel results of the knowledge graph consistency constraint are re-scored on the candidate set of structured triples. Structured triples that do not meet the constraints are removed and the confidence of the remaining structured triples is adjusted to obtain the corrected set of structured triples.

[0065] Time segment alignment and aggregation are performed on the time information in the asset full-lifecycle time-series modeling dataset, mapping it to time nodes and time segment nodes, and establishing a relationship with the modified structured triple set to form a time-series multi-tuple set;

[0066] Import the time-series tuple set into a graph database, use asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes as graph nodes, and use the relationships in the time-series tuple set as graph edges to construct and output a full-lifecycle time-series knowledge graph of assets.

[0067] In this embodiment, the generation of the asset full-lifecycle time-series modeling dataset specifically includes:

[0068] Collect basic asset information, operation and maintenance event records, life cycle stage identifiers and time information from asset management system, operation and maintenance management system, monitoring system and file management system, classify data source types, and form original multi-source asset data set;

[0069] The fields of each data source in the original multi-source asset data set are parsed and formatted uniformly, and the fields of different systems are mapped to a preset standard field set to generate a preliminary standardized asset data set.

[0070] Based on the preliminary standardized asset data set, an asset master data table and an asset identifier mapping relationship table are constructed. The device number, asset code, tag number and geographical location identifier from different data source types are merged according to the preset mapping relationship to generate an asset unified identifier data set with the unified asset identifier as the primary key.

[0071] The operation and maintenance event records are associated and matched with the unified asset identifier data set. Using the unified asset identifier as the association key, various operation and maintenance event records are bound to the corresponding asset entities, generating an asset operation and maintenance event association data set with the unified asset identifier, event identifier, and event type as the core fields.

[0072] The lifecycle stage descriptions in the preliminary standardized asset dataset are classified and standardized, a standard vocabulary of lifecycle stages and a mapping relationship between lifecycle stages are constructed, and a unified identifier dataset of lifecycle stages is generated.

[0073] Time standardization processing is performed on the time information in the preliminary standardized asset data set, parsing the date and time strings, time interval descriptions and relative time descriptions into standard time points and standard time intervals under a unified time benchmark, and generating an asset event time series table.

[0074] Using the unified asset identifier as the primary key, the unified asset identifier data set, the asset operation and maintenance event associated data set, the lifecycle stage unified identifier data set, and the asset event time series table are associated, merged, and sorted in chronological order to construct a full lifecycle time series modeling data set for assets.

[0075] In this embodiment, the generation of the text semantic representation sequence set specifically includes:

[0076] From the asset lifecycle time-series modeling dataset, extract asset-related texts based on the unified asset identifier, including asset basic information fields, operation and maintenance event record fields, and lifecycle stage identifier fields. Fill in empty spaces and trim redundant fields to construct an initial set of asset-related texts indexed by the unified asset identifier.

[0077] Sentence segmentation and word segmentation are performed on the initial set of asset-related texts. Long words not in the preset vocabulary are segmented into sub-words. Numbers, special symbols and time descriptions are normalized and replaced to obtain a set of asset-related text segmentation sequences composed of word sequence and sub-word sequence.

[0078] Based on the source type and the asset entity to which each asset-related text belongs in the asset-related text segmentation sequence set, a segment label is assigned to each asset-related text and a position index corresponding to the word position is generated. The life cycle stage identifier is transcribed into the life cycle stage category code. The input feature sequence of the ERNIE encoder is constructed with the unified asset identifier, segment label, position index and life cycle stage category code as constraints.

[0079] The input feature sequence of the ERNIE encoder is fed into the multi-layer coding network of the ERNIE encoder. Attention calculation and feedforward transformation are performed sequentially according to a fixed number of network layers and network structure to obtain a set of word-level semantic vector sequences corresponding to each asset-related text.

[0080] Based on the set of word-level semantic vector sequences, for each asset-related text, the word-level semantic vector corresponding to the aggregation marker at a preset position is selected as the sentence-level semantic vector, and the sentence-level semantic vectors of different source types under the same unified asset identifier are serialized and arranged in chronological order to generate a set of sentence-level semantic vector sequences indexed by the unified asset identifier;

[0081] The set of word-level semantic vector sequences and the set of sentence-level semantic vector sequences are aligned and merged according to unified asset identifiers, time information and lifecycle stage identifiers. Through index binding and structured storage, a set of text semantic representation sequences covering the semantic features of basic asset information, operation and maintenance event records and lifecycle stage identifiers is constructed.

[0082] In this embodiment, the generation of the candidate set of structured triples specifically includes:

[0083] Based on the asset full-lifecycle time-series modeling dataset and the text semantic representation sequence set, an asset entity candidate set is constructed. The asset name is matched with the corresponding original text of each text semantic representation sequence to form a subject candidate position set. This set is then indexed and bound to the text semantic representation sequence set to generate a text semantic representation sequence set with subject candidate tags.

[0084] Based on the CasRel entity-relation joint extraction model for subject filtering, a subject candidate filtering mask is constructed on the basis of the set of text semantic representation sequences with subject candidate tags. The word positions that are not in the set of subject candidate positions are masked. The subject candidate filtering mask is associated with the corresponding text semantic representation sequence to form a set of text semantic inputs with subject filtering constraints.

[0085] The text semantic input set with subject filtering constraints is fed into the subject entity recognition subnetwork. Based on the subject candidate filtering mask, the subject start position prediction and subject end position prediction are performed on each text semantic representation sequence to obtain the subject entity interval set, which is mapped to form the asset entity node candidate set.

[0086] The candidate set of asset entity nodes and the text semantic input set with subject filtering constraints are fed into the relation type gating prediction subnetwork. The life cycle stage prior gating unit and the time prior gating unit are introduced. The relation type prediction results are gating and filtering based on the unified identifier of the life cycle stage and the life cycle stage category. The time prior features are constructed based on the time information of the asset entity nodes. The preset relation type set is judged one by one to obtain the event relation type subset and the life cycle stage relation type subset corresponding to the asset entity node candidates. These are then bound to the corresponding asset entity node candidates to form an asset entity node candidate set with relation type labels.

[0087] The candidate set of asset entity nodes with relation type labels and the text semantic input set with subject filtering constraints are fed into the object entity extraction subnetwork. For the event association relationship type subset, the event object start position prediction and event object end position prediction are performed in the text semantic representation sequence to extract event entity node candidates. For the life cycle stage association relationship type subset, the life cycle stage object start position prediction and life cycle stage object end position prediction are performed in the text semantic representation sequence to extract life cycle stage node candidates.

[0088] The candidate asset entity nodes, candidate event entity nodes, and candidate lifecycle stage nodes are combined according to the relationship type subset. The structured triplet candidates between asset entity nodes and event entity nodes and lifecycle stage nodes are generated according to the corresponding event relationship type and lifecycle stage relationship type. The three triplet candidates are then summarized to form a set of structured triplet candidates.

[0089] In this embodiment, the generation of the modified structured triplet set specifically includes:

[0090] Based on the asset lifecycle time-series modeling dataset, a constraint set is constructed, which is divided into a lifecycle stage sequential constraint subset, a lifecycle stage mutual exclusion constraint subset, an operation and maintenance event and lifecycle stage matching constraint subset, and a time continuity constraint subset.

[0091] Among them, the lifecycle phase sequence constraint subset is used to limit the sequential relationship between lifecycle phase identifiers, the lifecycle phase mutual exclusion constraint subset is used to limit the lifecycle phase nodes of the same asset entity node in the same time segment cannot be in a mutually exclusive state at the same time, the operation and maintenance event and lifecycle phase matching constraint subset is used to limit the matching relationship between different operation and maintenance event records and corresponding lifecycle phase identifiers, and the time continuity constraint subset is used to limit the continuity and irreversibility of the lifecycle phase nodes of the same asset entity node on the time axis.

[0092] Read the basic confidence scores of the structured triplet candidates, combine them with the source location, asset entity node, event entity node, lifecycle stage node, time prior features and time information to generate a constraint test index list, and form an update confidence score calculation rule based on the weighted summation rule of the basic confidence scores and various consistency scores.

[0093] According to the constraint test index list, the life cycle stage sequence constraint, life cycle stage mutual exclusion constraint, operation and maintenance event and life cycle stage matching constraint and time continuity constraint are retrieved one by one. The corresponding consistency score is calculated respectively. Under the modulation of the time prior feature, each consistency score is substituted into the update confidence calculation rule to obtain the update confidence of the structured triplet candidate.

[0094] The updated confidence level is compared with the preset confidence threshold. Candidate structured triples whose updated confidence level does not reach the threshold and violates any constraint are eliminated. The updated confidence level is used to replace the base confidence level, forming a set of consistent triples with updated confidence level labels.

[0095] The set of consistent triples with updated confidence labels is grouped and sorted according to the unified asset identifier, event entity node and lifecycle stage node. Adjacent triples of the same asset entity node are rearranged according to the lifecycle stage order constraint. Triples with stage jumps or rollbacks are truncated to generate a modified set of structured triples.

[0096] In this embodiment, the generation of the time-series tuple set specifically includes:

[0097] Extract time information corresponding to standard time points and standard time intervals from the asset full-lifecycle time series modeling data set, collect and organize the time information according to the unified asset identifier, and construct the asset time original sequence set indexed by the unified asset identifier;

[0098] Based on the preset time granularity configuration and time benchmark, time segmentation is performed on each time series in the original asset time sequence set. Standard time points are mapped to discrete time nodes, and the time interval covering continuous standard time points is divided into several time segments according to the time granularity, generating a set of time nodes and a set of time segmentation results.

[0099] Based on the time segment division result set, the start time, end time, unified asset identifier, and covered time node index of each time segment are structured and represented to construct a time segment node set with time segment as the basic unit, and the time segment node set is indexed and bound to the time node set;

[0100] The modified set of structured triples is mapped to the corresponding time nodes and time segment nodes according to the unified asset identifier and time information. Within each time segment node, the associated asset entity nodes, event entity nodes and life cycle stage nodes are summarized. Based on the summary results, a state description and state attributes are generated for the time segment node, and a set of time segment state nodes is constructed.

[0101] Using asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment status nodes as elements, an initial set of time-series tuples is generated based on the association and index relationships in the modified set of structured triples.

[0102] The initial set of the time series tuple includes a unified asset identifier, event entity node, lifecycle stage node, time node identifier, and time segment status node identifier.

[0103] The initial set of time-series tuples is sorted and deduplicated according to the unified asset identifier and time node order. Multivariate association records that do not meet the time continuity requirements due to missing time information and time conflicts are removed, and a time-series tuple set is constructed.

[0104] In this embodiment, the generation of the asset lifecycle time-series knowledge graph specifically includes:

[0105] Establish a graph schema for a full-lifecycle time-series knowledge graph in a graph database. Define the node types and attribute fields for asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment status nodes. Also define the association relationship types and relationship attribute fields between asset entity nodes and the above nodes to form the graph schema configuration result.

[0106] Based on the time-series tuple set, deduplication and aggregation are performed on asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes. Corresponding graph database node objects are created for each type of node and node attributes are written to form a node writing set.

[0107] Based on the association types in the time-series multi-tuple set, the asset entity nodes are mapped to event entity nodes, lifecycle stage nodes, and time nodes to time segment status nodes. Corresponding relationship edges are created for each multi-tuple association record in the graph database. The updated confidence in the corrected structured triple set is written into the relationship attribute, and the relationship edges are written into the set.

[0108] Consistency checks are performed on the node writing set and the relation edge writing set. The connectivity of nodes and the temporal continuity of relation edges are checked according to the unified asset identifier and time order. Isolated nodes and relation edges that do not meet the time continuity requirements are removed, and duplicate nodes are merged to form a graph dataset which is then registered in the graph database as an asset full-lifecycle time-series knowledge graph.

[0109] Example 1:

[0110] To verify the feasibility of this invention in practice, it was applied to a large-scale energy equipment management center in a coastal area. This center is responsible for the long-term status management of wind turbine units, substation equipment, and grid-connected facilities. The center has a large number of devices, complex sources of maintenance records, and the same equipment spans multiple business departments at different stages of its lifecycle. This results in fragmented historical records, non-standardized event content, and chaotic time stamps, making it impossible to form a coherent asset lifecycle chain. Previously, when analyzing the health status of primary equipment or tracing the root cause of failures, problems such as contradictory event sequences for the same equipment, disordered lifecycle stage stamps, and inconsistent record times often occurred, seriously affecting the accuracy of maintenance scheduling. Therefore, the knowledge graph-based asset lifecycle management method proposed in this invention was applied to the above scenario, aiming to solve core challenges such as difficulty in aligning cross-source data, difficulty in extracting asset relationships, difficulty in unifying lifecycle stages, and difficulty in connecting time series.

[0111] In this scenario, asset basic information, operation and maintenance event descriptions, and stage identifiers are first collected from the monitoring center, operation and maintenance management platform, and archives. These are then uniformly organized and input into the ERNIE encoder to generate a semantically consistent text representation sequence. Subsequently, a subject-filtered CasRel entity-relation joint extraction model is used to identify asset entities, event types, and lifecycle stages from a large amount of natural language operation and maintenance records. Furthermore, prior gating of lifecycle stages and time is used to filter out unreasonable relation combinations, making the extracted structured triples semantically closer to the actual state of the equipment. For example, in a certain operation and maintenance cycle, contradictory descriptions often appear where a device suddenly jumps back to the "installation stage" after the "commissioning stage." This invention can use prior gating to filter out such contradictory relationships during extraction, reducing the generation of invalid relationships at the source. Further, combining lifecycle sequence and time continuity constraints, the re-scoring mechanism of this invention performs consistency verification on the extraction results, ensuring that the description of the equipment's state evolution remains logically coherent.

[0112] After data extraction and constraint verification, time segment alignment and aggregation are used to achieve a unified representation of data at different time granularities, ensuring that each time node and time segment node accurately maps to the actual state of the equipment at that time. The resulting asset lifecycle time-series knowledge graph is deployed and runs in the management center for daily inspection support analysis, fault tracing, pre-prediction preparation for status prediction, and cross-departmental collaboration. In actual use, managers have found a significant improvement in data query efficiency. Previously, determining equipment stage changes required manually reviewing numerous maintenance records; now, the knowledge graph clearly shows the complete chain of time sequence, event nodes, and stage nodes. Simultaneously, inconsistencies in cross-departmental understanding of the same equipment's history have been greatly reduced, the consistency of lifecycle stages has been significantly improved, and time conflicts between multi-source records are automatically identified and eliminated during the re-scoring stage.

[0113] From the perspective of continuous operation in the implementation scenario, the management center has successfully reduced the amount of manual review and improved the efficiency of equipment status verification through the asset lifecycle time-series knowledge graph constructed by this invention during the past few months of operation and maintenance. It has also enabled the rapid identification of key event nodes in multiple fault investigations. Thanks to the collaborative mechanism of time-prior gating, lifecycle-prior gating, and consistency constraint rescoring, this invention ensures that asset data possesses semantic consistency, logical coherence, and temporal reliability across the entire lifecycle, making the asset status evolution chain clearer and the decision-making basis more reliable, thus providing significant value support for actual operation and maintenance management.

[0114] Table 1. Performance Comparison of Knowledge Graph-Based Asset Lifecycle Management Method and Traditional Methods

[0115]

[0116] In the overall comparison shown in Table 1, this invention demonstrates significant improvements in key element extraction tasks such as asset entity recognition, lifecycle stage recognition, and operation and maintenance event recognition. For example, the accuracy of asset entity recognition increased from approximately 82% to over 94%. The core reason for this is the introduction of a subject filtering mechanism, which pre-locates the positions of potential asset names in the text and filters out words and characters unrelated to the target asset, thus avoiding a large amount of irrelevant noise interference from the input stage. At the same time, the addition of a subject filtering mask makes the range of subject entity prediction more focused, thereby reducing the probability of misidentification.

[0117] In terms of lifecycle phase identification and operational event identification, this invention achieves an improvement of over 15 percentage points compared to traditional methods, primarily due to the introduction of prior gating for lifecycle phases. The model incorporates phase semantic constraints during the relation type determination stage, suppressing unreasonable relation combinations at the extraction stage and significantly reducing the number of erroneous triples that need to be removed later. Furthermore, this invention combines phase subset guidance in the object entity extraction stage, narrowing the search space and achieving more precise localization.

[0118] The advantages of the relationship consistency dimension are even more apparent. In traditional methods, the lifecycle order conflict rate is close to 15%, while this invention reduces it to about 2%. This significant improvement comes from two key mechanisms: first, time prior gating provides a time constraint reference in the relationship type prediction stage, making the prediction results more consistent with the natural evolution logic of the device lifecycle; second, in the re-scoring stage, the confidence of conflicting relationships is readjusted based on the consistency score, so that the description differences of cross-source texts do not lead to confusion of structured knowledge.

[0119] Regarding the quality of the time dimension, this invention improves the accuracy of time node mapping from approximately 80% to over 95% through time segment alignment and unified granularity mapping strategies, while also significantly enhancing the continuity and completeness of time segments. This stems from the fact that this invention retains prior temporal features in the early extraction stage and utilizes temporal consistency for modulation in triple rescoring, making the time chains in the final knowledge graph smoother and more coherent.

[0120] Regarding the quality of the knowledge graph structure, this invention significantly reduces isolated nodes and erroneous relation edges, increasing the coverage of the complete lifecycle chain from 67% to 89%. This means that the historical trajectory of the same asset across stages and events can be continuously tracked, greatly improving the readability and usability of asset evolution analysis and operation and maintenance decisions. The fundamental reason for this is that this invention implements a two-stage quality control mechanism of "extraction-end prior constraints + graph construction-end consistency correction," ensuring consistency throughout the entire process from the source to the database.

[0121] In summary, this invention achieves systematic improvements in four key dimensions: entity recognition, relation extraction, temporal processing, and knowledge graph construction. Its core advantages come from the synergistic effect of subject filtering, lifecycle prior gating, temporal prior gating, and consistency re-scoring mechanism, ultimately constructing a logically continuous, temporally reliable, and structurally robust asset lifecycle temporal knowledge graph.

[0122] 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 equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge graph-based method for full-lifecycle asset management, characterized in that, Includes the following steps: Acquire multi-source asset data, including basic asset information, operation and maintenance event records, lifecycle stage identifiers and time information, and perform unified formatting, identifier merging and time standardization processing to form a full lifecycle time-series modeling data set for assets. The asset lifecycle time-series modeling dataset is input into the ERNIE encoder to perform semantic encoding on asset-related text, resulting in a set of text semantic representation sequences. The set of text semantic representation sequences is input into the subject-filtered CasRel entity-relation joint extraction model to identify asset entity nodes, event entity nodes and lifecycle stage nodes, and extract the association relationships between asset entity nodes and event entity nodes and lifecycle stage nodes respectively, generating a candidate set of structured triples. Based on the preset set of constraints, the candidate set of structured triples is re-scored, structured triples that do not meet the constraints are eliminated, and the confidence of the remaining structured triples is adjusted to obtain the corrected set of structured triples. Time segment alignment and aggregation are performed on the time information in the asset full-lifecycle time-series modeling dataset, mapping it to time nodes and time segment nodes, and establishing a relationship with the modified structured triple set to form a time-series multi-tuple set; Import the time-series tuple set into the graph database, use asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes as graph nodes, and use the relationships in the time-series tuple set as graph edges to construct and output a full lifecycle time-series knowledge graph of assets. The time segment status node is obtained by aggregating the asset entity nodes, event entity nodes, and lifecycle stage nodes associated with the time segment node to generate a status description and status attributes for the time segment node.

2. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the asset full-lifecycle time-series modeling dataset specifically includes: Collect basic asset information, operation and maintenance event records, life cycle stage identifiers and time information from asset management system, operation and maintenance management system, monitoring system and file management system, classify data source types, and form original multi-source asset data set; The fields of each data source in the original multi-source asset data set are parsed and formatted uniformly, and the fields of different systems are mapped to a preset standard field set to generate a preliminary standardized asset data set. Based on the preliminary standardized asset data set, an asset master data table and an asset identifier mapping relationship table are constructed. The device number, asset code, tag number and geographical location identifier from different data source types are merged according to the preset mapping relationship to generate an asset unified identifier data set with the unified asset identifier as the primary key. The operation and maintenance event records are associated and matched with the unified asset identifier data set. Using the unified asset identifier as the association key, various operation and maintenance event records are bound to the corresponding asset entities, generating an asset operation and maintenance event association data set with the unified asset identifier, event identifier, and event type as the core fields. The lifecycle stage descriptions in the preliminary standardized asset dataset are classified and standardized, a standard vocabulary of lifecycle stages and a mapping relationship between lifecycle stages are constructed, and a unified identifier dataset of lifecycle stages is generated. Time standardization processing is performed on the time information in the preliminary standardized asset data set, parsing the date and time strings, time interval descriptions and relative time descriptions into standard time points and standard time intervals under a unified time benchmark, and generating an asset event time series table. Using the unified asset identifier as the primary key, the unified asset identifier data set, the asset operation and maintenance event associated data set, the lifecycle stage unified identifier data set, and the asset event time series table are associated, merged, and sorted in chronological order to construct a full lifecycle time series modeling data set for assets.

3. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the text semantic representation sequence set specifically includes: From the asset lifecycle time-series modeling dataset, extract asset-related texts based on the unified asset identifier, including asset basic information fields, operation and maintenance event record fields, and lifecycle stage identifier fields. Fill in empty spaces and trim redundant fields to construct an initial set of asset-related texts indexed by the unified asset identifier. Sentence segmentation and word segmentation are performed on the initial set of asset-related texts. Long words not in the preset vocabulary are segmented into sub-words. Numbers, special symbols and time descriptions are normalized and replaced to obtain a set of asset-related text segmentation sequences composed of word sequence and sub-word sequence. Based on the source type and the asset entity to which each asset-related text belongs in the asset-related text segmentation sequence set, a segment label is assigned to each asset-related text and a position index corresponding to the word position is generated. The life cycle stage identifier is transcribed into the life cycle stage category code. The input feature sequence of the ERNIE encoder is constructed with the unified asset identifier, segment label, position index and life cycle stage category code as constraints. The input feature sequence of the ERNIE encoder is fed into the multi-layer coding network of the ERNIE encoder. Attention calculation and feedforward transformation are performed sequentially according to a fixed number of network layers and network structure to obtain a set of word-level semantic vector sequences corresponding to each asset-related text. Based on the set of word-level semantic vector sequences, for each asset-related text, the word-level semantic vector corresponding to the aggregation marker at a preset position is selected as the sentence-level semantic vector, and the sentence-level semantic vectors of different source types under the same unified asset identifier are serialized and arranged in chronological order to generate a set of sentence-level semantic vector sequences indexed by the unified asset identifier; The set of word-level semantic vector sequences and the set of sentence-level semantic vector sequences are aligned and merged according to unified asset identifiers, time information and lifecycle stage identifiers. Through index binding and structured storage, a set of text semantic representation sequences covering the semantic features of basic asset information, operation and maintenance event records and lifecycle stage identifiers is constructed.

4. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the candidate set of structured triples specifically includes: Based on the asset full-lifecycle time-series modeling data set and the text semantic representation sequence set, an asset entity candidate set is constructed. The asset name is matched with the corresponding original text of each text semantic representation sequence to form a subject candidate position set, which is then indexed and bound to the text semantic representation sequence set to generate a text semantic representation sequence set with subject candidate tags. Based on the CasRel entity-relation joint extraction model for subject filtering, a subject candidate filtering mask is constructed on the basis of the set of text semantic representation sequences with subject candidate tags. The word positions that are not in the set of subject candidate positions are masked. The subject candidate filtering mask is associated with the corresponding text semantic representation sequence to form a set of text semantic inputs with subject filtering constraints. The text semantic input set with subject filtering constraints is fed into the subject entity recognition subnetwork. Based on the subject candidate filtering mask, the subject start position prediction and subject end position prediction are performed on each text semantic representation sequence to obtain the subject entity interval set, which is mapped to form the asset entity node candidate set. The candidate set of asset entity nodes and the text semantic input set with subject filtering constraints are fed into the relation type gating prediction subnetwork. The life cycle stage prior gating unit and the time prior gating unit are introduced. The relation type prediction results are gating and filtering based on the unified identifier of the life cycle stage and the life cycle stage category. The time prior features are constructed based on the time information of the asset entity nodes. The preset relation type set is judged one by one to obtain the event relation type subset and the life cycle stage relation type subset corresponding to the asset entity node candidates. These are then bound to the corresponding asset entity node candidates to form an asset entity node candidate set with relation type labels. The candidate set of asset entity nodes with relation type labels and the text semantic input set with subject filtering constraints are fed into the object entity extraction subnetwork. For the event relation type subset, the event object start position prediction and event object end position prediction are performed in the text semantic representation sequence to extract event entity node candidates. For the life cycle stage relation type subset, the life cycle stage object start position prediction and life cycle stage object end position prediction are performed in the text semantic representation sequence to extract life cycle stage node candidates. The candidate asset entity nodes, candidate event entity nodes, and candidate lifecycle stage nodes are combined according to the relationship type subset. The structured triplet candidates between asset entity nodes and event entity nodes and lifecycle stage nodes are generated according to the corresponding event relationship type and lifecycle stage relationship type. The three triplet candidates are then summarized to form a set of structured triplet candidates.

5. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the modified structured triplet set specifically includes: Based on the asset lifecycle time-series modeling dataset, a constraint set is constructed, which is divided into a lifecycle stage sequential constraint subset, a lifecycle stage mutual exclusion constraint subset, an operation and maintenance event and lifecycle stage matching constraint subset, and a time continuity constraint subset. Read the basic confidence scores of the structured triplet candidates, combine them with the source location, asset entity node, event entity node, lifecycle stage node, time prior features and time information to generate a constraint test index list, and form the update confidence score calculation rule based on the weighted sum of the basic confidence scores and each consistency score. The constraint check index list is searched item by item for lifecycle phase sequence constraints, lifecycle phase mutual exclusion constraints, operation and maintenance event and lifecycle phase matching constraints and time continuity constraints. The corresponding consistency scores are calculated respectively, and each consistency score is substituted into the update confidence calculation rule to obtain the update confidence of the structured triplet candidate. The updated confidence level is compared with the preset confidence threshold. Candidate structured triples whose updated confidence level does not reach the threshold and violates any constraint are eliminated. The updated confidence level is used to replace the base confidence level, forming a set of consistent triples with updated confidence level labels. The set of consistent triples with updated confidence labels is grouped and sorted according to the unified asset identifier, event entity node and lifecycle stage node. Adjacent triples of the same asset entity node are rearranged according to the lifecycle stage order constraint. Triples with stage jumps or rollbacks are truncated to generate a modified set of structured triples.

6. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the time-series tuple set specifically includes: Extract time information corresponding to standard time points and standard time intervals from the asset full-lifecycle time series modeling data set, collect and organize the time information according to the unified asset identifier, and construct the asset time original sequence set indexed by the unified asset identifier; Based on the preset time granularity configuration and time benchmark, time segmentation is performed on each time series in the original asset time sequence set. Standard time points are mapped to discrete time nodes, and the time interval covering continuous standard time points is divided into several time segments according to the time granularity, generating a set of time nodes and a set of time segmentation results. Based on the time segment division result set, the start time, end time, unified asset identifier, and covered time node index of each time segment are structured and represented to construct a time segment node set with time segment as the basic unit, and the time segment node set is indexed and bound to the time node set; The modified set of structured triples is mapped to the corresponding time nodes and time segment nodes according to the unified asset identifier and time information. Within each time segment node, the associated asset entity nodes, event entity nodes and life cycle stage nodes are summarized. Based on the summary results, a state description and state attributes are generated for the time segment node, and a set of time segment state nodes is constructed. Using asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment status nodes as elements, an initial set of time-series tuples is generated based on the association and index relationships in the modified set of structured triples. The initial set of time-series tuples is sorted and deduplicated according to the unified asset identifier and time node order. Multivariate association records that do not meet the time continuity requirements due to missing time information and time conflicts are removed, and a time-series tuple set is constructed.

7. The asset lifecycle management method based on knowledge graphs according to claim 1, characterized in that, The generation of the asset lifecycle time-series knowledge graph specifically includes: Establish a graph schema for a full-lifecycle time-series knowledge graph in a graph database. Define the node types and attribute fields for asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes, and time segment state nodes. Also define the association relationship types and relationship attribute fields between asset entity nodes and event entity nodes, lifecycle stage nodes, time nodes, and time segment state nodes to form the graph schema configuration result. Based on the time-series tuple set, deduplication and aggregation are performed on asset entity nodes, event entity nodes, lifecycle stage nodes, time nodes and time segment status nodes. Corresponding graph database node objects are created for each type of node and node attributes are written to form a node writing set. Based on the association types in the time-series multi-source group set, the asset entity nodes are mapped to event entity nodes, lifecycle stage nodes, and time nodes to time segment status nodes. Corresponding relationship edges are created in the graph database for each multi-source association record in the time-series multi-source group set. The updated confidence in the corrected structured triple set is written into the relationship attribute, forming relationship edges that are written into the set. Consistency checks are performed on the node writing set and the relation edge writing set. The connectivity of nodes and the temporal continuity of relation edges are checked according to the unified asset identifier and time order. Isolated nodes and relation edges that do not meet the time continuity requirements are removed, and duplicate nodes are merged to form a graph dataset which is then registered in the graph database as an asset full-lifecycle time-series knowledge graph.

Images