A Method for Constructing and Semantic Completion of Industrial Knowledge Graphs for the Entire Product Lifecycle
By performing feature extraction and conflict detection on industrial product lifecycle data, a cross-modal similarity matrix is constructed, a basic entity layer and a domain relationship layer are established, and knowledge graph completion is performed using path constraints and comprehensive confidence rules. This solves the problem of correlation analysis between cross-modal knowledge elements and improves the accuracy and semantic completion rate of the industrial knowledge graph.
Patent Information
- Application Number
- CN202611588655.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-10
AI Technical Summary
Existing technologies struggle to achieve cross-modal knowledge element correlation analysis and structured reasoning completion under comprehensive confidence constraints, leading to a decline in the accuracy and semantic completion rate of industrial knowledge graph construction, which fails to meet the needs of fault diagnosis and operation and maintenance decision-making.
By preprocessing the data of the entire life cycle of industrial products, extracting and weighting cross-modal feature vectors of different knowledge elements, constructing a cross-modal similarity matrix, performing conflict detection and screening, establishing a basic entity layer and a domain relationship layer, using path constraints and comprehensive confidence rules to complete the knowledge graph, and performing version iteration management.
It achieves effective aggregation and consistency of knowledge elements of the same industrial object, improves the accuracy and reliability of the knowledge graph, ensures the reliability and timeliness of the completion results, and supports the dynamic updating and fault diagnosis of the industrial knowledge graph.
Smart Images

Figure CN123070133A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of product lifecycle management technology, and in particular to a method for constructing and semantically completing an industrial knowledge graph of the entire product lifecycle. Background Technology
[0002] In the entire lifecycle management of industrial products, knowledge graphs are widely used to integrate heterogeneous data generated during the design, manufacturing, and operation and maintenance stages. Current methods for constructing industrial knowledge graphs rely on manually defining ontology schemas and rule templates, and using natural language processing tools to extract entities, attributes, and relationships from structured or semi-structured documents. Because industrial data sources involve multiple business systems, different systems often have different naming conventions, units of measurement, and state descriptions for the same entity or attribute. Furthermore, the same term may carry different semantics at different stages, leading to frequent conflicts and ambiguities in knowledge elements during extraction. For example, the same equipment number may be assigned different model attributes, or the same process parameter may exhibit contradictory values over time. Existing technologies typically use post-hoc rule validation or manual review to address these issues. However, validation rules require enumerating all conflict patterns in advance, and manual review struggles to handle massive data volumes, resulting in insufficient coverage and timeliness of conflict detection. Furthermore, it fails to automatically distinguish the types of ambiguities and provide resolution strategies, ultimately leading to inconsistencies in factual statements within the graph.
[0003] In terms of semantic completion, existing reasoning methods largely rely on structured triples, making limited use of unstructured text, visual information, temporal evidence, and process constraints, and lacking a structured reasoning completion mechanism under comprehensive confidence constraints. The credibility, spatiotemporal conditions, stage dependencies, and conflict relationships of multi-source evidence have not been uniformly modeled, and erroneous completion results are prone to propagate along the graph. Changes in product operating conditions, equipment updates, and process adjustments cause dynamic evolution of knowledge, making cross-modal consistency and timely maintenance difficult. Graph construction and semantic completion influence each other; semantic ambiguities and missing relationships left over from the construction phase increase the completion burden, while errors introduced during the completion phase contaminate the graph structure, leading to a decrease in the accuracy of industrial knowledge graph construction and semantic completion rate, making it difficult to meet the application needs of fault diagnosis, operation and maintenance decision-making, etc.
[0004] Chinese Patent Publication No. CN122713387A discloses a method and system for constructing a geological disaster knowledge graph based on a large language model. The method includes: collecting multi-source heterogeneous data related to tunnel geological disasters; extracting geological disaster entities and their attributes and relationships after data preprocessing to construct an initial knowledge graph; building a structural constraint rule base and a disaster evolution chain rule base for geological disasters; identifying missing knowledge in the initial knowledge graph based on the rule bases and constructing a knowledge candidate space; using the knowledge candidate space as a priori constraint, generating candidate completion content for missing knowledge based on the initial knowledge graph using a large language model; verifying the structural legality, disaster evolution chain consistency, and semantic consistency of the candidate completion content; updating the initial knowledge graph based on the verified content to obtain an optimized geological disaster knowledge graph. This approach only performs completion under rule constraints and semantic verification for missing knowledge, making it difficult to achieve cross-modal knowledge element correlation analysis and structured reasoning completion under comprehensive confidence constraints, thus reducing the accuracy and semantic completion rate of industrial knowledge graph construction. Summary of the Invention
[0005] To address this, the present invention provides a method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle. This method overcomes the problem in existing technologies that make it difficult to achieve cross-modal knowledge element association analysis and structured reasoning completion under comprehensive confidence constraints, which reduces the accuracy and semantic completion rate of industrial knowledge graph construction.
[0006] To achieve the above objectives, this invention provides a method for constructing and semantically completing an industrial knowledge graph throughout the product lifecycle, comprising the following steps: S1. Collect the operating status of industrial products to obtain full life cycle data of industrial products, and preprocess the full life cycle data of industrial products to obtain preprocessed data. S2. Based on the preprocessed data, determine the list of available feature extractors, extract features from each knowledge element in the preprocessed data, linearly transform the initial feature vectors of each modality to a unified target dimension, weight the feature vectors of the target dimension, and perform time-series encoding on the weighted feature vectors according to the generation time of each knowledge element to determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object. Then, perform weighted aggregation on the time-series enhanced feature vectors according to the cross-modal similarity matrix and the preset feature weight rules to obtain the similarity analysis results. S3. Based on the similarity analysis results, determine the set of candidate knowledge elements belonging to the same industrial object, perform conflict detection on the set of candidate knowledge elements, obtain the conflict detection results, and filter the set of candidate knowledge elements based on the conflict detection results to obtain the target set of knowledge elements. S4. Construct a basic entity layer based on the target knowledge element set, and construct a domain relationship layer according to the structural composition relationship and temporal sequence relationship between entities in the basic entity layer. Establish multidimensional semantic relationships between entities based on the basic entity layer and the domain relationship layer to obtain the initial knowledge graph. S5. Perform structural constraint reasoning on entities with missing attributes and entities with missing relationships using the initial knowledge graph. Determine the set of path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph. Predict the candidate relationship types for entity pairs with missing relationships. Determine the comprehensive confidence level based on the cross-modal matching score, graph structure consistency score, industrial rule compliance, historical case similarity, and data source credibility of each candidate completion result in the completion candidate set. Validate the completion candidate set based on the comprehensive confidence level and preset multiple constraint rules to obtain the candidate completion results that pass the validation and the candidate knowledge that fails the validation and is pending review. S6. Based on the candidate completion results that have passed the verification, perform semantic completion on the initial knowledge graph to obtain the semantically completed knowledge graph, and mark the candidate knowledge that has not passed the verification as pending manual review. S7. Iterate the semantically completed knowledge graph according to the preset version identifier and effective time interval to obtain the iterated entities and relations, and mark the confidence level of the correlation between the iterated entities and relations to obtain the updated knowledge graph.
[0007] The technical principle of this application is as follows: After preprocessing the full lifecycle data of industrial products, heterogeneous feature extraction channels are designed for different knowledge elements such as the structural composition, functional attributes, operating status, and time sequence processes of the same industrial object. Cross-modal information such as text descriptions, numerical parameters, and state sequences are mapped to a unified vector space. Based on this, the comprehensive similarity between different knowledge elements is calculated according to preset feature weight rules to determine candidate knowledge elements belonging to the same industrial object. Conflict detection is performed on the candidate set, and elements with irreconcilable conflicts are eliminated, resulting in a consistent set of target knowledge elements. A basic entity layer is constructed based on the target knowledge element set, and a domain relationship layer is established according to the inherent structural dependencies and production sequence relationships between entities, thus forming a knowledge graph with initial semantic associations. For entities or entity pairs in the knowledge graph that lack attributes or relationships, inference is performed using path constraints, type constraints, and cardinality constraints in the graph structure. Candidate results are then filtered using comprehensive confidence and preset multiple constraint rules. The verified candidate completion results are used for knowledge graph completion, while those that fail are transferred to manual review. Meanwhile, each completion or editing operation is assigned a version identifier and a valid time interval is set, enabling the knowledge graph to be managed in a version-by-version manner in the time dimension. After each iteration, the confidence of entities and relationships is recalculated and marked, realizing the traceable dynamic update of the knowledge graph.
[0008] Compared with existing technologies, the beneficial effects of this application are as follows: By performing hierarchical cross-modal feature extraction and similarity analysis on different knowledge elements belonging to the same industrial object, it is possible to effectively aggregate knowledge elements of the same object scattered in multi-source heterogeneous data, avoiding the problem of knowledge fragmentation caused by data dispersion. By determining the list of available feature extractors in the feature processing stage, the initial feature vectors of each modality are linearly transformed to a unified target dimension and then weighted, and combined with the time of knowledge element generation for temporal encoding, a cross-modal similarity matrix of different knowledge elements of the same industrial object is constructed. The similarity analysis results are obtained by weighting and aggregating the temporal enhanced feature vectors according to the preset feature weight rules, which can effectively realize the temporal alignment and cross-modal association matching of multimodal industrial knowledge. By performing conflict detection and screening on the candidate knowledge element set, it is possible to automatically identify and eliminate contradictory or redundant information generated during the knowledge fusion process, ensuring the consistency and accuracy of the target knowledge element set. By constructing a domain relationship layer based on structural composition and temporal sequence relationships and establishing multi-dimensional semantic associations, the initial knowledge graph can not only contain the entities themselves but also the multi-level semantic connections between entities, enhancing the structured expression capability of the knowledge graph. By performing structural constraint reasoning on missing attribute entities and missing relation entity pairs using an initial knowledge graph, and determining the path feature vector set by combining known relationship paths between entities, candidate relationship types are predicted for missing relation entity pairs. A comprehensive confidence score is calculated by integrating cross-modal matching scores, graph structure consistency scores, industrial rule compliance, historical case similarity, and data source credibility. The candidate set is then validated and completed according to preset multi-constraint rules. The validated candidate completion results and candidate knowledge awaiting review are output. This approach enables automatic semantic completion of the knowledge graph while ensuring the reliability of the completion results. Marking candidate knowledge that fails validation as awaiting manual review balances automation efficiency with the accuracy of manual review. By iterating through versions based on preset version identifiers and valid time intervals, and marking the confidence level of the association between entities and relationships after iteration, the knowledge graph can dynamically evolve with the product lifecycle. This ensures the timeliness and traceability of the association credibility of entities and relationships in each version, thereby improving the accuracy and semantic completion rate of industrial knowledge graph construction.
[0009] Furthermore, S1 includes the following steps: S11. Collect raw operating status signals from each monitoring node of the industrial product according to the preset acquisition frequency and preset acquisition channel, and perform analog-to-digital conversion on the raw operating status signals to obtain digital operating status signals. S12. Based on the time-domain and frequency-domain characteristics of each signal component in the digital operation status signal, noise component identification is performed on the digital operation status signal to obtain noise-marked signal components. Then, the digital operation status signal is filtered according to the noise-marked signal components to obtain the filtered operation status signal. S13. Divide the filtered running status signal into time windows according to the preset time window length and preset sliding step size to obtain window signal segments corresponding to each time window, and perform missing value detection on the window signal segments to obtain missing value detection results. Then, fill in the missing value segments with data according to the missing value detection results to obtain filled signal segments. S14. Perform outlier detection on the filled signal segment, obtain the outlier detection results, and perform outlier correction on the filled signal segment based on the outlier detection results to obtain preprocessed data.
[0010] In this solution, the original signal is acquired and converted from analog to digital by a preset acquisition frequency and a preset acquisition channel. Noise components are identified and filtered out, then the signal is segmented according to a time window and missing values are filled in. Finally, outliers are detected and corrected to obtain complete and accurate preprocessed data, thereby improving the quality of the operating status signal.
[0011] Furthermore, S2 includes the following steps: S21. Determine the feature extractor corresponding to each knowledge element based on the modality type of each knowledge element in the preprocessed data, obtain a candidate list of feature extractors, and filter the candidate list of feature extractors according to the available status identifier of each feature extractor in the candidate list to obtain a list of available feature extractors. S22. Based on the output dimensions of each feature extractor in the list of available feature extractors, perform feature extraction on each knowledge element in the preprocessed data to obtain the initial feature vector of each modality, and perform linear transformation on the initial feature vector of each modality to obtain the target dimension feature vector. S23. The target dimension feature vector is weighted according to the source identifier of each knowledge element in the preprocessed data to obtain a weighted feature vector. The weighted feature vector is then time-encoded according to the generation time of each knowledge element in the preprocessed data to obtain a time-enhanced feature vector. S24. Based on the time-series enhanced feature vector, determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object, and perform weighted aggregation of the time-series enhanced feature vector according to the cross-modal similarity matrix and the preset feature weight rules to obtain the similarity analysis results.
[0012] In this scheme, by selecting available feature extractors and unifying feature dimensions, source weighting and temporal encoding are performed on knowledge elements, a cross-modal similarity matrix is constructed and weighted aggregation is performed, thereby improving the accuracy of multimodal feature fusion of industrial objects and the reliability of similarity analysis results.
[0013] Furthermore, S3 includes the following steps: S31. Based on the similarity analysis results, perform entity type identification on each candidate knowledge element in the candidate knowledge element set belonging to the same industrial object, obtain the entity type label of each candidate knowledge element, and extract the entity attribute value of each candidate knowledge element based on the entity type label of each candidate knowledge element to obtain the entity attribute value set. S32. Classify the entity attribute values of each candidate knowledge element based on the semantic type corresponding to each entity attribute value in the entity attribute value set to obtain classified entity attribute values. Extract relation paths from the entity relation types in the candidate knowledge element set based on the classified entity attribute values to obtain a relation path set. S33. Based on the source identifier of each candidate knowledge element in the candidate knowledge element set, trace the source of the entity attribute value set and the relation path set to obtain the source-marked entity attribute value set and the source-marked relation path set. S34. Conflict detection is performed on the entity attribute values and entity relationship types in the candidate knowledge element set by using the data type of each entity attribute value in the source tag entity attribute value set and the semantic direction of each relationship path in the source tag relationship path set, and the conflict detection results are obtained. S35. Based on the conflict detection results, filter the entity attribute values and entity relationship types in the candidate knowledge element set to obtain the target knowledge element set.
[0014] In this solution, the entity types of candidate knowledge elements are identified and attribute values are extracted. After classifying the attribute values, the relationship paths are extracted, the source marker attributes and relationship paths are traced, data type and semantic direction conflicts are detected, and the target knowledge element set is selected to improve the accuracy of knowledge elements.
[0015] Furthermore, S4 includes the following steps: S41. Establish an entity index table based on the unique identifier of each entity in the target knowledge element set, and generate an initial entity embedding vector based on the entity type and entity attribute value of each entity in the entity index table to obtain a set of initial entity embedding vectors. S42. Cluster the entities in the target knowledge element set according to the semantic distance corresponding to the initial embedding vectors of each entity in the initial embedding vector set, obtain the entity clustering results, determine the structural composition relationship between each entity according to the entity clustering results, obtain the structural composition relationship set, determine the temporal sequence relationship between each entity according to the generation time of each knowledge element in the target knowledge element set, obtain the temporal sequence relationship set. S43. Construct a domain relation layer based on the set of structural composition relations and the set of temporal sequence relations, and establish multidimensional semantic relations between entities according to the set of initial entity embedding vectors and the relation types in the domain relation layer to obtain the initial knowledge graph.
[0016] In this scheme, an entity index table is established and an initial embedding vector is generated. The structural composition relationship is analyzed by clustering. A domain relationship layer is constructed by combining the temporal sequence relationship. Multidimensional semantic relationships are established to form an initial knowledge graph, which enhances the structural and temporal nature of knowledge organization and improves the semantic expression ability and overall association accuracy of the initial knowledge graph.
[0017] Furthermore, S5 includes the following steps: S51. Determine the attribute statistical features of known entities based on the attribute distribution of known entities in the initial knowledge graph, and generate attribute candidate values for entities with missing attributes in the initial knowledge graph based on the attribute statistical feature vectors to obtain an attribute completion candidate set. Determine the path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph to obtain a path feature vector set. S52. Based on the path feature vector set, predict the candidate types of entity pairs with missing relations in the initial knowledge graph to obtain the candidate set of relation completion. Determine the candidate set of completion based on the candidate set of attribute completion and the candidate set of relation completion. S53. Determine the overall confidence level of each candidate completion result based on its cross-modal matching score, graph structure consistency score, industry rule compliance, historical case similarity, and data source credibility. The mathematical expression for the overall confidence level of each candidate completion result is as follows: In the formula, Indicates the first The overall confidence level of the candidate completion results The weighting coefficients represent the cross-modal matching scores. Indicates the first Cross-modal matching scores of candidate completion results The weighting coefficients represent the graph structure consistency score. Indicates the first The graph structure consistency score of each candidate completion result. Weighting coefficients representing the degree of compliance with industrial regulations. Indicates the first The degree of compliance with industry rules for each candidate completion result. The weighting coefficients representing the similarity of historical cases. Indicates the first Historical case similarity of each candidate completion result Weighting coefficients representing the credibility of data sources. Indicates the first The credibility of the data sources for each candidate completion result. ; S54. The candidate completion set is initially screened based on the comprehensive confidence level and the preset confidence threshold to obtain the initial candidate set. The candidate completion results in the initial candidate set are then constrained and verified according to the preset multiple constraint rules. Candidate completion results whose comprehensive confidence level is not less than the preset confidence threshold and which pass all constraint verifications are marked as verified candidate completion results. Candidate completion results whose comprehensive confidence level is less than the preset confidence threshold or which fail any constraint verification are marked as unverified candidate knowledge to be reviewed.
[0018] In this scheme, candidate sets for attribute completion and candidate sets for relationship completion are generated by statistical features of entity attributes and path features of entity pairs in the knowledge graph, respectively. The comprehensive confidence of the candidate completion results is obtained by integrating multi-dimensional evaluation dimensions. The candidate completion results are then screened and verified in a hierarchical manner by combining confidence thresholds and preset multiple constraint rules, thereby improving the rationality and reliability of the knowledge graph completion results and reducing the probability of outputting incorrect completion content.
[0019] Furthermore, S6 includes the following steps: S61. Based on the entity identifier and relation identifier in the verified candidate completion results, locate the nodes to be updated and the edges to be updated in the initial knowledge graph to obtain the set of nodes to be updated and the set of edges to be updated. Then, based on the attribute values in the verified candidate completion results, assign values to the attributes of each node in the set of nodes to be updated to obtain the set of nodes after attribute updates. S62. Based on the relation types in the candidate completion results that have passed the verification, update the relation types of each edge in the edge set to be updated to obtain the edge set after relation update. Based on the node set after attribute update and the edge set after relation update, determine the updated subgraph. Merge the updated subgraph into the initial knowledge graph to obtain the semantically completed knowledge graph. Mark the candidate knowledge that has not passed the verification as pending manual review. S63. Generate a list of candidate knowledge to be reviewed based on the candidate knowledge marked as pending manual review in the semantically completed knowledge graph, and establish a review tracking record according to the source identifier and generation time of each candidate knowledge in the list of pending review, and store the review tracking record in the metadata layer of the semantically completed knowledge graph.
[0020] In this solution, an updated subgraph is generated by updating node attributes and edge relationship types and then merged into the initial knowledge graph to complete semantic completion. Knowledge that fails verification is marked for manual review and an audit list is generated. Audit tracking records are established and stored in the metadata layer, thus achieving an effective connection between automatic knowledge graph updates and manual review, ensuring the accuracy of completion results and the traceability of operations.
[0021] Furthermore, S7 includes the following steps: S71. Based on the preset version identifier and valid time interval, perform version validity checks on entities and relations in the semantically completed knowledge graph to obtain the version validity status of each entity and the version validity status of each relation, and analyze the impact range of the data update event to obtain the impact range marker. S72. Based on the version validity status of each entity, the version validity status of each relation, and the scope of influence, match the entities and relations affected by the data update event in the semantically completed knowledge graph to obtain the set of affected entities and the set of affected relations. S73. Based on the new attribute values in the data update event, iterate the attributes of each entity in the affected entity set to obtain the iterated entity. Based on the new relation type in the data update event, iterate the relation type of each relation in the affected relation set to obtain the iterated relation. S74. Determine the initial value of source credibility based on the source type of the data update event, and determine the change range coefficient based on the change range of the data update event. Determine the confidence level of the association between each iterated entity in the affected entity set and each iterated relation in the affected relation set based on the initial value of source credibility and the change range coefficient. Mark the confidence level of the association between the iterated entities and relations based on the confidence level of the association, and obtain the updated knowledge graph.
[0022] In this solution, the entity and relation status are determined and the scope of impact of update events is analyzed through version validity checks. The entities and relations affected by the update events are located, and the version iteration of entity attributes and relation types is completed. The association confidence is marked by combining the source and scope of the update event, so as to achieve reliable updates of the knowledge graph and ensure that the graph data version is controllable and the association confidence results are accurate. Attached Figure Description
[0023] Figure 1 This is a flowchart illustrating a method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle, according to an embodiment of the present invention. Detailed Implementation
[0024] The following detailed description illustrates the specific implementation method: like Figure 1 As shown, a method for constructing and semantically completing an industrial knowledge graph throughout the product lifecycle includes the following steps: S1. Collect the operating status of industrial products to obtain full life cycle data of industrial products, and preprocess the full life cycle data of industrial products to obtain preprocessed data. S2. Based on the preprocessed data, determine the list of available feature extractors, extract features from each knowledge element in the preprocessed data, linearly transform the initial feature vectors of each modality to a unified target dimension, weight the feature vectors of the target dimension, and perform time-series encoding on the weighted feature vectors according to the generation time of each knowledge element to determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object. Then, perform weighted aggregation on the time-series enhanced feature vectors according to the cross-modal similarity matrix and the preset feature weight rules to obtain the similarity analysis results. S3. Based on the similarity analysis results, determine the set of candidate knowledge elements belonging to the same industrial object, perform conflict detection on the set of candidate knowledge elements, obtain the conflict detection results, and filter the set of candidate knowledge elements based on the conflict detection results to obtain the target set of knowledge elements. S4. Construct a basic entity layer based on the target knowledge element set, and construct a domain relationship layer according to the structural composition relationship and temporal sequence relationship between entities in the basic entity layer. Establish multidimensional semantic relationships between entities based on the basic entity layer and the domain relationship layer to obtain the initial knowledge graph. S5. Perform structural constraint reasoning on entities with missing attributes and entities with missing relationships using the initial knowledge graph. Determine the set of path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph. Predict the candidate relationship types for entity pairs with missing relationships. Determine the comprehensive confidence level based on the cross-modal matching score, graph structure consistency score, industrial rule compliance, historical case similarity, and data source credibility of each candidate completion result in the completion candidate set. Validate the completion candidate set based on the comprehensive confidence level and preset multiple constraint rules to obtain the candidate completion results that pass the validation and the candidate knowledge that fails the validation and is pending review. S6. Based on the candidate completion results that have passed the verification, perform semantic completion on the initial knowledge graph to obtain the semantically completed knowledge graph, and mark the candidate knowledge that has not passed the verification as pending manual review. S7. Iterate the semantically completed knowledge graph according to the preset version identifier and effective time interval to obtain the iterated entities and relations, and mark the confidence level of the correlation between the iterated entities and relations to obtain the updated knowledge graph.
[0025] Specifically, S1 includes the following steps: S11. Collect raw operating status signals from each monitoring node of the industrial product according to the preset acquisition frequency and preset acquisition channel, and perform analog-to-digital conversion on the raw operating status signals to obtain digital operating status signals. S12. Based on the time-domain and frequency-domain characteristics of each signal component in the digital operation status signal, noise component identification is performed on the digital operation status signal to obtain noise-marked signal components. Then, the digital operation status signal is filtered according to the noise-marked signal components to obtain the filtered operation status signal. S13. Divide the filtered running status signal into time windows according to the preset time window length and preset sliding step size to obtain window signal segments corresponding to each time window, and perform missing value detection on the window signal segments to obtain missing value detection results. Then, fill in the missing value segments with data according to the missing value detection results to obtain filled signal segments. S14. Perform outlier detection on the filled signal segment, obtain the outlier detection results, and perform outlier correction on the filled signal segment based on the outlier detection results to obtain preprocessed data.
[0026] Further, in step S11, a corresponding type of sensor, such as a vibration sensor, temperature sensor, or pressure sensor, is first installed at each monitoring node of the industrial product. Each sensor converts the physical quantity into a voltage or current analog signal through a signal conditioning circuit. The preset acquisition frequency is determined based on the Nyquist criterion of the highest effective signal frequency under the operating conditions of the industrial product. For example, if the highest effective signal frequency is 2 kHz, the preset acquisition frequency is set to 5 kHz to balance anti-aliasing filtering and data processing efficiency. The preset acquisition channels are determined based on the physical distribution of the monitoring nodes and the signal type. For example, for rotating machinery with 8 vibration measurement points and 4 temperature measurement points, the preset acquisition channels are set to 12 analog input channels, each corresponding to a sensor output. The data acquisition card synchronously samples each preset acquisition channel at the preset acquisition frequency, obtaining a set of multi-channel analog voltage values at each sampling moment. Analog-to-digital conversion is achieved through a successive approximation analog-to-digital converter (ADC). This ADC quantizes each analog voltage value into a 16-bit binary digital code. The quantized voltage range corresponds to the sensor's measurement range, for example, from -10V to +10V. The quantization step size is 20V divided by 65536, yielding the digital value for each sampling point. The continuously sampled digital values from all channels are arranged in chronological order to form a digital operating status signal. This digital operating status signal includes a timestamp and channel identifier.
[0027] In step S12, empirical mode decomposition (EMD) is first performed on the digitized operating status signal, decomposing it into multiple intrinsic mode function (IMF) components, each of which is a signal component. Time-domain features are calculated for each signal component, including peak factor and kurtosis. The peak factor is the ratio of the maximum absolute value of the signal component to its root mean square (RMS) value, and the kurtosis is the ratio of the fourth central moment of the signal component to the square of its second central moment. A fast Fourier transform (FFT) is then performed on each signal component to obtain frequency-domain features, including the spectral centroid frequency and spectral bandwidth. The discrimination rule for noise component identification is as follows: when the peak factor of a signal component is less than a first peak threshold, the kurtosis is less than a first kurtosis threshold, and the spectral centroid frequency is higher than a first frequency threshold, the signal component is marked as a noise-labeled signal component. The first peak threshold is determined based on the statistical average of the peak factors of historical noise-free signal components under normal operating conditions of industrial products. For example, the first peak threshold is set to 2.5. The first kurtosis threshold is determined based on the theoretical value of normal distribution. For example, the first kurtosis threshold is set to 3.2. The theoretical value of normal distribution is calculated by the mathematical ratio of the fourth central moment to the square of the second central moment of the probability density function of normal distribution. For example, the theoretical value of normal distribution is set to 3.0. The first frequency threshold is determined based on the typical lower frequency limit of the inherent electrical noise of the acquisition system. For example, the first frequency threshold is set to 500 Hz. The filtering process employs a wavelet thresholding denoising method. For the time-domain waveform corresponding to the noise-marked signal component, a five-level wavelet decomposition is performed using the Daubechies wavelet basis function. The high-frequency detail coefficients of each level are then shrunk using a soft thresholding function. The threshold of the soft thresholding function is determined by multiplying the standard deviation estimate of the noise-marked signal component by a threshold scaling factor. The processed wavelet coefficients are then reconstructed to obtain the filtered running state signal. The threshold scaling factor is determined by adjusting the signal-to-noise ratio gain based on the experimental denoising effect; for example, the threshold scaling factor is set to 0.8.
[0028] In step S13, starting from the time starting point of the filtered operating status signal, all sampling points within the first time window are extracted according to the preset time window length to form the first window signal segment. Then, the starting position of the window is moved backward by a preset sliding step size, and sampling points of the same length are extracted again to form the second window signal segment. This process is repeated until the end of the window reaches the end of the filtered operating status signal, thus obtaining the window signal segments corresponding to all time windows. The preset time window length is determined based on the minimum stable cycle of the industrial product's operating status. For example, for a rotating device with a rotation speed of 3000 revolutions per minute and a rotation frequency of 50 Hz, one rotation cycle is 0.02 seconds. The preset time window length is set to 0.1 seconds, containing 5 complete rotation cycles. The preset sliding step size is determined based on the real-time requirements of online monitoring. For example, setting the preset sliding step size to 0.02 seconds ensures an 80% overlap rate between adjacent windows. For each window signal segment, missing value detection employs a point-by-point checking of timestamp intervals. If the time interval between two adjacent sampling points is greater than 1.5 times the preset sampling interval, a missing value is considered to exist within that interval. The missing value detection result includes the missing location and the missing duration. Data imputation uses a linear interpolation method. For each missing interval, the values of the two valid sampling points before and after the missing interval are linearly interpolated. The interpolated values are then filled into the missing locations according to equally spaced timestamps to obtain the imputed signal segment.
[0029] In step S14, for each filled signal segment, outlier detection employs a method based on the local outlier factor. First, the local reachability density of each sampling point within the filled signal segment is calculated. The local reachability density is the reciprocal of the average reachability distance between the sampling point and its N nearest neighbors. The value of N is determined based on the square root of the total number of sampling points within the window signal segment. For example, if the total number of sampling points within the window is 1024, then the value of N is set to 32. Then, the local outlier factor for each sampling point is calculated. The local outlier factor is the ratio of the local reachability density of the sampling point to the average local reachability density of its N nearest neighbors. The rule for determining the outlier detection result is: when the local outlier factor of a sampling point is greater than a preset outlier threshold, the sampling point is marked as an outlier. The preset outlier threshold is determined based on the empirical value of the outlier factor corresponding to the 99.7% confidence interval under the standard normal distribution. For example, the preset outlier threshold is set to 2.0. Outlier correction employs a median filtering method. For each sample point marked as an outlier, the median value of the L normal sample points before and after it is replaced. The value of L is determined based on one-hundredth of the total length of the window signal segment. For example, if the window signal segment contains 1024 sample points, the value of L is set to 10. During correction, the values of the L sample points before and after the outlier position are first sorted, and the value of the median position after sorting is taken as the correction value. This correction value is then written to the original outlier position, resulting in preprocessed data. For cases with multiple consecutive outliers, correction is performed sequentially from front to back. Each correction uses the median of the adjacent normal sample points (both corrected and uncorrected) to calculate the median.
[0030] Specifically, S2 includes the following steps: S21. Determine the feature extractor corresponding to each knowledge element based on the modality type of each knowledge element in the preprocessed data, obtain a candidate list of feature extractors, and filter the candidate list of feature extractors according to the available status identifier of each feature extractor in the candidate list to obtain a list of available feature extractors. S22. Based on the output dimensions of each feature extractor in the list of available feature extractors, perform feature extraction on each knowledge element in the preprocessed data to obtain the initial feature vector of each modality, and perform linear transformation on the initial feature vector of each modality to obtain the target dimension feature vector. S23. The target dimension feature vector is weighted according to the source identifier of each knowledge element in the preprocessed data to obtain a weighted feature vector. The weighted feature vector is then time-encoded according to the generation time of each knowledge element in the preprocessed data to obtain a time-enhanced feature vector. S24. Based on the time-series enhanced feature vectors, determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object, and perform weighted aggregation of the time-series enhanced feature vectors according to the cross-modal similarity matrix and preset feature weight rules to obtain the similarity analysis results. The mathematical expression of the cross-modal similarity matrix is: In the formula, This indicates the row number in the cross-modal similarity matrix. This represents the column index in the cross-modal similarity matrix. and These correspond to the serial numbers of different knowledge elements belonging to the same industrial object, and , Indicates the first Temporal enhanced feature vectors of knowledge elements, Indicates the first Temporal enhanced feature vectors of knowledge elements, Indicates the first The magnitude of the temporal augmentation feature vector of each knowledge element. Indicates the first The magnitude of the temporal augmentation feature vector of each knowledge element. Represents an exponential function. Indicates the first The moment when a knowledge element is generated. Indicates the first The moment when a knowledge element is generated. This represents the time decay coefficient.
[0031] Further, in step S21, each knowledge element in the preprocessed data is first labeled with its modal type, which includes vibration mode, temperature mode, pressure mode, and sound mode. For each modal type, all feature extractors bound to that modal type are searched in the feature extractor registry. The feature extractor registry pre-stores the mapping relationship between modal types and feature extractor names. All the found feature extractors are combined into a feature extractor candidate list. Each feature extractor in the candidate list carries an availability status flag, which is a Boolean value. A true Boolean value indicates that the corresponding algorithm library for the feature extractor has been loaded and its dependencies are complete, while a false Boolean value indicates that the feature extractor is unavailable. The availability status flag is determined based on the self-check results of the feature extractor's operating environment. For example, for a frequency domain feature extractor based on Fast Fourier Transform, if the Fast Fourier Transform algorithm library version is compatible and memory allocation is successful, the availability status flag is set to true. Iterate through the candidate list of feature extractors, check the availability status flag of each feature extractor, keep the feature extractors with the availability status flag being true, and delete the feature extractors with the availability status flag being false, to obtain the list of available feature extractors.
[0032] In step S22, the output dimension parameter of each feature extractor is read from the list of available feature extractors. The output dimension parameter represents the length of the feature vector generated by the feature extractor after transforming the input knowledge element. For each knowledge element in the preprocessed data, the corresponding feature extractor is selected according to the modality type of the knowledge element. The numerical sequence of the knowledge element is input into the corresponding feature extractor. The feature extractor calculates and outputs a fixed-length numerical sequence according to its internal algorithm. This numerical sequence is the initial feature vector. Since the output dimensions of feature extractors for different modal types are different, the length of the initial feature vector for each modality is inconsistent. For example, the initial feature vector dimension for the vibration mode is 128 dimensions, and the initial feature vector dimension for the temperature mode is 32 dimensions. A linear transformation is performed on the initial feature vector of each modality. The linear transformation method is to multiply the initial feature vector by a transformation matrix. The number of rows in the transformation matrix is equal to the target dimension, and the number of columns in the transformation matrix is equal to the dimension of the initial feature vector. Each element in the transformation matrix is a trainable weight parameter. The target dimension is determined based on the statistical distribution of the dimensions of all modal initial feature vectors. For example, the maximum value of 256 for all modal initial feature vector dimensions is set as the target dimension. The target dimension feature vector is obtained after a linear transformation.
[0033] In step S23, the source identifier of each knowledge element is read from the preprocessed data. The source identifier represents the sensor number or data acquisition channel number from which the knowledge element originates. A preset source weight table is consulted based on the source identifier. This table records the weight coefficient corresponding to each source identifier. The weight coefficient for each source identifier is determined based on the historical data reliability and measurement accuracy of that source. For example, for a calibrated high-precision pressure sensor, the weight coefficient corresponding to that source identifier is set to 1.2; for an uncalibrated vibration sensor, the weight coefficient is set to 0.8. Each element value in each target dimension feature vector is multiplied by the weight coefficient corresponding to that knowledge element to obtain a weighted feature vector. The generation time of each knowledge element is read from the preprocessed data; the generation time is the timestamp when the knowledge element was acquired. Each weighted feature vector is temporally encoded using a sinusoidal position encoding function, converting the generation time into a temporally encoded vector with the same dimension as the weighted feature vector. This temporally encoded vector is then added element-wise to the weighted feature vector to obtain a temporally enhanced feature vector.
[0034] In step S24, the time-series enhanced feature vectors of all knowledge elements belonging to the same industrial object are first obtained. The industrial object is identified by the equipment number field in the knowledge element. The time-series decay coefficient is determined according to the time scale of the industrial object's operating state change. For example, for an industrial object with slow temperature change, the time-series decay coefficient is set to 3600 seconds, so that the similarity between two knowledge elements spaced one hour apart decays to 0.37 times the original value. The preset feature weighting rule is determined according to the importance of modal type to the state representation of the industrial object. For example, the vibration mode weight is set to 0.4, the temperature mode weight is set to 0.3, the pressure mode weight is set to 0.2, and the sound mode weight is set to 0.1. The weighted aggregation method is as follows: for the time-series enhanced feature vector of each knowledge element, it is weighted and summed with the time-series enhanced feature vectors of other knowledge elements belonging to the same industrial object according to the corresponding element values in the cross-modal similarity matrix. During the weighted summation, each vector participating in the aggregation is also multiplied by the weight coefficient of the corresponding modal type in the preset feature weighting rule, and finally the aggregated feature vector of the knowledge element is obtained. The aggregated feature vectors of all knowledge elements are gathered to form the similarity analysis results.
[0035] Specifically, S3 includes the following steps: S31. Based on the similarity analysis results, perform entity type identification on each candidate knowledge element in the candidate knowledge element set belonging to the same industrial object, obtain the entity type label of each candidate knowledge element, and extract the entity attribute value of each candidate knowledge element based on the entity type label of each candidate knowledge element to obtain the entity attribute value set. S32. Classify the entity attribute values of each candidate knowledge element based on the semantic type corresponding to each entity attribute value in the entity attribute value set to obtain classified entity attribute values. Extract relation paths from the entity relation types in the candidate knowledge element set based on the classified entity attribute values to obtain a relation path set. S33. Based on the source identifier of each candidate knowledge element in the candidate knowledge element set, trace the source of the entity attribute value set and the relation path set to obtain the source-marked entity attribute value set and the source-marked relation path set. S34. Conflict detection is performed on the entity attribute values and entity relationship types in the candidate knowledge element set by using the data type of each entity attribute value in the source tag entity attribute value set and the semantic direction of each relationship path in the source tag relationship path set, and the conflict detection results are obtained. S35. Based on the conflict detection results, filter the entity attribute values and entity relationship types in the candidate knowledge element set to obtain the target knowledge element set.
[0036] Further, in step S31, all candidate knowledge elements belonging to the same industrial object are first read from the similarity analysis results. Each candidate knowledge element in the candidate knowledge element set is input into the entity type classifier. The entity type classifier adopts a hybrid classification architecture based on rule matching and statistical models. The rule matching layer contains a predefined entity type keyword dictionary, and the statistical model layer uses a conditional random field model for sequence labeling. The model parameters of the entity type classifier are determined by training based on an industrial domain labeled corpus. The industrial domain labeled corpus contains five categories of labels: equipment entities, parameter entities, event entities, time entities, and location entities, with a size of 100,000 labeled knowledge elements. The feature functions of the conditional random field model include part-of-speech features, context features, and word form features. The model training uses the maximum likelihood estimation method. After the candidate knowledge element is input into the classifier, the classifier outputs the entity type label to which the candidate knowledge element belongs. The system queries a preset attribute value extraction template based on the entity type label. This template binds a list of corresponding attribute field names to each entity type. For example, the attribute fields for the equipment entity type include equipment number, equipment name, and manufacturer; the attribute fields for the parameter entity type include parameter value, unit of measurement, and measurement precision. For each candidate knowledge element, the system locates and extracts the corresponding field content from the text or numerical sequence of the candidate knowledge element according to the attribute field names bound in the extraction template. All extracted field content constitutes the entity attribute value of the candidate knowledge element. The entity attribute values of all candidate knowledge elements are then collected into an entity attribute value set.
[0037] In step S32, for each entity attribute value in the entity attribute value set, the semantic type of the entity attribute value is first determined by a semantic type recognizer. The semantic type recognizer uses a text classifier based on a pre-trained language model, specifically the BERT model. The classification output layer of the semantic type recognizer includes four semantic categories: numerical type, time type, text type, and enumeration type. The classification threshold of the semantic type recognizer is determined based on the principle of maximizing the F1 score on the validation set; for example, the classification threshold is set to 0.75. Each entity attribute value is then categorized according to its determined semantic type, with entity attribute values of the same semantic type grouped together to obtain the categorized entity attribute values. For every two candidate knowledge elements in the candidate knowledge element set, it is determined whether there is an entity relationship type between these two candidate knowledge elements. The entity relationship type extraction method utilizes a relationship path extraction rule base for matching. This rule base stores multiple rules, each described using a subject-verb-object logical structure. The confidence score of a rule is determined based on its hit frequency in the historical knowledge base. For example, rules with a confidence score greater than a preset confidence score are considered valid. The preset confidence score is determined based on the principle of balancing precision and recall in selecting valid rules; for example, it might be set to 0.8. For two successfully matched candidate knowledge elements, the matched entity relationship type is added to the relationship path set, and the entity attribute values of these two candidate knowledge elements are also stored as endpoint attributes of the relationship path in the relationship path set.
[0038] In step S33, the source identifier field of each candidate knowledge element is read from the candidate knowledge element set. The source identifier field includes the device number of the data acquisition device, the data acquisition channel number, and the data file number. A one-to-one correspondence is established between the source identifier and the candidate knowledge element, meaning each candidate knowledge element carries a unique source identifier. Each entity attribute value in the entity attribute value set is traversed, and the candidate knowledge element to which that entity attribute value belongs is found. A copy of the source identifier of that candidate knowledge element is appended to the metadata area of that entity attribute value, forming an entity attribute value carrying a source identifier. All entity attribute values carrying source identifiers constitute the source-marked entity attribute value set. Each relationship path in the relationship path set is traversed. Each relationship path includes a starting entity, an ending entity, and a relationship type. The source identifiers of the two candidate knowledge elements involved in the relationship path are found, and both source identifiers are appended to the metadata area of the relationship path, forming a relationship path carrying a source identifier. All relationship paths carrying source identifiers constitute the source-marked relationship path set.
[0039] In step S34, for multiple entity attribute values in the source marker entity attribute value set that have the same entity type label and the same attribute field name but different source identifiers, the data type field of each entity attribute value is obtained. The data type field includes integer, floating-point, string, and boolean types. The first type of conflict detection rule is data type consistency detection. When the data types of multiple entity attribute values corresponding to the same attribute field name are not completely consistent, the attribute field is marked as having a data type conflict. The second type of conflict detection rule is numerical range consistency detection. For floating-point or integer entity attribute values, a preset numerical range table stores the reasonable numerical range corresponding to each attribute field name. The reasonable numerical range is determined according to the industrial product design specifications. For example, the reasonable numerical range for temperature attributes is set to -20 degrees Celsius to 150 degrees Celsius. When an entity attribute value exceeds this reasonable numerical range, it is marked as a numerical range conflict. For each relationship path in the source marker relationship path set, the semantic direction field of each relationship path is read. The semantic direction field is represented by an ordered pair of starting entity type and ending entity type. For example, a device entity pointing to a parameter entity indicates the semantic direction of "the device has parameters". The third type of conflict detection rule is semantic direction consistency detection. For the same pair of starting and ending entity types, a pre-defined standard semantic direction table stores the allowed standard semantic directions between the entity types. The pre-defined standard semantic directions are determined based on the ontology of the industry domain. When the semantic direction of a relationship path is opposite to the allowed direction recorded in the standard semantic direction table, the relationship path is marked as having a semantic direction conflict. All conflict markers are collected into the conflict detection result.
[0040] In step S35, each conflict marker recorded in the conflict detection results and its corresponding entity attribute value or relation path's unique number are obtained. For entity attribute values marked as data type conflicts, a majority voting method is used for filtering. The frequency of occurrence of each data type under the same attribute field name is counted, and the entity attribute value corresponding to the data type with the highest frequency is retained, while entity attribute values of other data types are deleted. For entity attribute values marked as numerical range conflicts, entity attribute values exceeding a reasonable numerical range are directly deleted. For relation paths marked as semantic direction conflicts, a preset semantic direction priority rule is used for filtering. The preset semantic direction priority rule is determined according to the standard specifications of industrial knowledge graphs. For example, the semantic direction priority of device pointing to parameter is set to level 1, and the semantic direction priority of parameter pointing to device is set to level 2. The smaller the priority value, the higher the priority. When a direction conflict is detected between the same pair of entity types, the relation path corresponding to the semantic direction with the smaller priority value is retained, and the other secondary direction relation path with the opposite direction is deleted. After filtering all conflict types, entity attribute values that were not deleted are retained in the candidate knowledge element set, relationship types that were not deleted are retained in the candidate knowledge element set, and candidate knowledge elements that were not deleted and did not involve any conflicts are directly retained. Finally, all retained entity attribute values and entity relationship types and their respective candidate knowledge elements constitute the target knowledge element set.
[0041] Specifically, S4 includes the following steps: S41. Establish an entity index table based on the unique identifier of each entity in the target knowledge element set, and generate an initial entity embedding vector based on the entity type and entity attribute value of each entity in the entity index table to obtain a set of initial entity embedding vectors. S42. Cluster the entities in the target knowledge element set according to the semantic distance corresponding to the initial embedding vectors of each entity in the initial embedding vector set, obtain the entity clustering results, determine the structural composition relationship between each entity according to the entity clustering results, obtain the structural composition relationship set, determine the temporal sequence relationship between each entity according to the generation time of each knowledge element in the target knowledge element set, obtain the temporal sequence relationship set. S43. Construct a domain relation layer based on the set of structural composition relations and the set of temporal sequence relations, and establish multidimensional semantic relations between entities according to the set of initial entity embedding vectors and the relation types in the domain relation layer to obtain the initial knowledge graph.
[0042] Further, in step S41, a unique identifier field for each entity is obtained from the target knowledge element set. This unique identifier field includes the device number, attribute code, and timestamp, and each unique identifier is unique within the set. A hash map table is established as the entity index table, with the unique identifier as the key and the hash map table value as a pointer to the entity's storage location in the target knowledge element set. Each entity in the entity index table is traversed, and its entity type field and entity attribute value field are read. The entity type field includes device type, parameter type, and event type. Different embedding generation strategies are adopted for different entity types. For device type entities, the device number string is converted into a binary vector using character-level one-hot encoding, and the device name string is converted into a word vector using a pre-trained word embedding model. The binary vector and word vector are concatenated to form the initial entity embedding vector. For parameter type entities, the parameter values are normalized to obtain a floating-point vector, and the parameter unit string is converted into a unit vector using word embedding. The floating-point vector and unit vector are concatenated to form the initial entity embedding vector. For event-type entities, the event description text is converted into a 768-dimensional semantic vector using a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model, which serves as the initial entity embedding vector. The dimension of the initial entity embedding vector is determined based on the maximum dimension after encoding all entity attribute values; for example, all initial entity embedding vectors are uniformly padded or truncated to 256 dimensions. The initial entity embedding vectors corresponding to all entities are then aggregated into a set of initial entity embedding vectors.
[0043] In step S42, all initial embedding vectors of entities are read from the initial embedding vector set, and the Euclidean distance between any two initial embedding vectors is calculated as the semantic distance. A hierarchical clustering algorithm is used to cluster all entities. This algorithm employs a bottom-up aggregation strategy, initially treating each entity as a separate cluster. In each iteration, the two clusters with the closest semantic distance are merged into a new cluster until all entities are merged into a single cluster. The clustering cutoff threshold is determined based on the elbow inflection point of the semantic distance distribution curve. For example, a clustering cutoff threshold of 1.2 is set. Merging stops when the minimum semantic distance between all clusters exceeds the clustering cutoff threshold, and the current clusters represent the entity clustering results. Structural composition relationships are determined based on the entity clustering results. These relationships represent the spatial affiliation or component membership between entities within the same cluster. For each cluster, entities of type "device" are marked as parent nodes, and entities of type "parameter" or "event" are marked as child nodes. A structural composition relationship is established between the parent node and each child node, and all structural composition relationships constitute a set of structural composition relationships. Read the generation time field of each knowledge element from the target knowledge element set. The generation time is in timestamp format. Sort all entities according to the order of their generation time. For entities with earlier generation time and entities with later generation time, establish a time sequence relationship between the two entities. The direction of the time sequence relationship is from the earlier entity to the later entity. All time sequence relationships constitute a time sequence relationship set.
[0044] In step S43, each structural composition relation in the set of structural composition relations is treated as a type of relation edge in the domain relation layer, and the type of the relation edge is labeled as structural composition type. Each temporal sequence relation in the set of temporal sequence relations is treated as another type of relation edge in the domain relation layer, and the type of the relation edge is labeled as temporal sequence type. The domain relation layer is represented using a directed graph data structure, which includes a node set and an edge set. The node set consists of unique identifiers for all entities in the target knowledge element set, and the edge set consists of all relations in the set of structural composition relations and the set of temporal sequence relations. The starting and ending nodes of each edge in the domain relation layer are referenced by unique identifiers. Each node in the domain relation layer is assigned a node feature vector based on the initial entity embedding vector set; the node feature vector is the initial entity embedding vector of the entity corresponding to that node. Each edge is assigned a relation type label based on the relation type in the domain relation layer. The multidimensional semantic association is established by performing a neighborhood aggregation operation on each node based on the directed graph of the domain relation layer. This operation weights and sums the node's initial entity embedding vector with the initial entity embedding vectors of all its neighboring nodes. The weighting coefficients are determined based on the reciprocal of the semantic distance between the current node and its neighboring nodes. The resulting aggregation yields a multidimensional semantic representation vector for each node. All multidimensional semantic representation vectors of all nodes, all relation edges, and their relation type labels are then stored together as an initial knowledge graph.
[0045] Specifically, S5 includes the following steps: S51. Determine the attribute statistical features of known entities based on the attribute distribution of known entities in the initial knowledge graph, and generate attribute candidate values for entities with missing attributes in the initial knowledge graph based on the attribute statistical feature vectors to obtain an attribute completion candidate set. Determine the path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph to obtain a path feature vector set. S52. Based on the path feature vector set, predict the candidate types of entity pairs with missing relations in the initial knowledge graph to obtain the candidate set of relation completion. Determine the candidate set of completion based on the candidate set of attribute completion and the candidate set of relation completion. S53. Determine the overall confidence level of each candidate completion result based on its cross-modal matching score, graph structure consistency score, industry rule compliance, historical case similarity, and data source credibility. The mathematical expression for the overall confidence level of each candidate completion result is as follows: In the formula, Indicates the first The overall confidence level of the candidate completion results The weighting coefficients represent the cross-modal matching scores. Indicates the first Cross-modal matching scores of candidate completion results The weighting coefficients represent the graph structure consistency score. Indicates the first The graph structure consistency score of each candidate completion result. Weighting coefficients representing the degree of compliance with industrial regulations. Indicates the first The degree of compliance with industry rules for each candidate completion result. The weighting coefficients representing the similarity of historical cases. Indicates the first Historical case similarity of each candidate completion result Weighting coefficients representing the credibility of data sources. Indicates the first The credibility of the data sources for each candidate completion result. ; S54. The candidate completion set is initially screened based on the comprehensive confidence level and the preset confidence threshold to obtain the initial candidate set. The candidate completion results in the initial candidate set are then constrained and verified according to the preset multiple constraint rules. Candidate completion results whose comprehensive confidence level is not less than the preset confidence threshold and which pass all constraint verifications are marked as verified candidate completion results. Candidate completion results whose comprehensive confidence level is less than the preset confidence threshold or which fail any constraint verification are marked as unverified candidate knowledge to be reviewed.
[0046] Further, in step S51, all entities in the initial knowledge graph are traversed, and entities with complete attribute fields are selected as known entities. Statistical calculations are performed on each attribute field of the known entities, including calculating the mean, variance, minimum, and maximum values of all values for that attribute field, to obtain the attribute statistical feature vector for that attribute field. The attribute field name of each entity with missing attributes is located, and the corresponding attribute statistical feature vector is found. A set of candidate values is generated based on the mean and variance in the attribute statistical feature vector. The candidate values are generated by sampling at equal intervals within the range of the mean plus or minus twice the variance. The sampling step size is determined based on one-tenth of the variance value. For example, for the temperature attribute field, if the mean is 80 degrees Celsius and the variance is 10 degrees Celsius, the sampling step size is set to 1 degree Celsius, generating a candidate value sequence from 60 degrees Celsius to 100 degrees Celsius. This candidate value sequence is used as the attribute completion candidate set. Extract all known entity pairs from the initial knowledge graph. Each known entity pair consists of two entities connected by at least one relation path. For each known entity pair, obtain all relation paths between the two entities. Each relation path is a sequence of relation edges connected sequentially. Concatenate the type labels of all relation edges in each relation path into a string in order. Then, convert the string into a 768-dimensional vector through a pre-trained BERT model to obtain the path feature vector corresponding to the relation path. Collect all path feature vectors of all known entity pairs into a path feature vector set.
[0047] In step S52, all entity pairs that do not have direct relationship edges but belong to the same industrial object and have different entity types are identified from the initial knowledge graph. These entity pairs are marked as entity pairs with missing relationships. The adjacent entity sets of each entity pair with a missing relationship in the initial knowledge graph are obtained for each pair. The number of identical adjacent entities in the two adjacent entity sets is calculated. If the number of identical adjacent entities is greater than a preset adjacency threshold, it is determined that there is a potential relationship between the entity pairs with missing relationships. The preset adjacency threshold is determined based on the average number of adjacent entity intersections of known entity pairs in the initial knowledge graph; for example, the preset adjacency threshold is set to 2. The path feature vector corresponding to the shortest relation path between the two entities of the entity pair with the missing relation is retrieved from the path feature vector set. This path feature vector is then input into a relation type classifier. The relation type classifier adopts a multilayer perceptron structure, consisting of an input layer, a hidden layer, and an output layer. The input layer has a dimension of 768, the hidden layer has a dimension of 256, and the output layer has a dimension equal to the total number of relation types. The model parameters of the relation type classifier are determined through training based on the relation type annotation data of known entity pairs in the initial knowledge graph. Each dimension of the relation type classifier's output layer corresponds to a predicted probability of a relation type. The relation type corresponding to the maximum predicted probability is selected as the relation candidate type for the entity pair with the missing relation. The entity pair identifiers are paired with the relation candidate types and stored as a relation completion candidate set. All entries in the attribute completion candidate set are merged with all entries in the relation completion candidate set; the merged set is the completion candidate set.
[0048] In step S53, the cross-modal matching score represents the cosine similarity between multiple modal feature vectors involved in the candidate completion result. For example, for attribute completion results, the cross-modal matching score is the cosine similarity between the semantic vector corresponding to the attribute value and the typical semantic vector corresponding to the modality type of the entity, with a calculated value ranging from 0 to 1. The graph structure consistency score represents the structural similarity between the newly formed triples after the candidate completion result is added to the initial knowledge graph and the known triples in the initial knowledge graph. The structural similarity is determined by calculating the proportion of common adjacent nodes of the head and tail entities of the new triples in the initial knowledge graph. The industrial rule compliance score represents whether the candidate completion result conforms to the rule constraints in the preset industrial rule base. The preset industrial rule base stores the legal correspondence between different entity types and attribute types. For example, for voltage entities, the industrial rule base stipulates that the voltage value attribute must be positive and not exceed 1000 volts. If the candidate completion result satisfies all relevant rules, the industrial rule compliance score is 1; otherwise, it is 0. Historical case similarity represents the average similarity between the candidate completion result and the attribute values or relationship types of completed cases in the historical knowledge graph completion records. The historical case database contains 100,000 approved completion records, and the similarity is calculated using edit distance or cosine similarity methods. Data source credibility represents the comprehensive score of the sensor calibration level and signal-to-noise ratio of the acquisition channel of the original data source on which the candidate completion result is based. The sensor calibration level is divided into three levels: A, B, and C, corresponding to baseline scores of 1.0, 0.8, and 0.6, respectively. The signal-to-noise ratio score is normalized to the range of 0 to 1 by the ratio of signal power to noise power. The data source credibility is obtained by multiplying the baseline score by the signal-to-noise ratio score. The five weight coefficients were determined using the Analytic Hierarchy Process (AHP). In the AHP, the pairwise comparison matrix was constructed by three industrial knowledge graph experts who scored the importance of the final completion quality according to each dimension. After consistency testing, the weight coefficients were calculated. For example, the weight coefficient for cross-modal matching score was set to 0.25, the weight coefficient for graph structure consistency score was set to 0.20, the weight coefficient for industrial rule compliance was set to 0.25, the weight coefficient for historical case similarity was set to 0.15, and the weight coefficient for data source credibility was set to 0.15.
[0049] In step S54, the comprehensive confidence score of each candidate completion result is read from the comprehensive confidence score, and the comprehensive confidence score is compared with a preset confidence threshold. The preset confidence threshold is determined based on the balance requirements of accuracy and recall in industrial knowledge graph completion application scenarios. For example, the preset confidence threshold is set to 0.75. The preset multiple constraint rules include data type consistency rules, entity relationship uniqueness rules, and attribute value mutual exclusion rules. The data type consistency rule requires that the data type of the attribute completion result must be consistent with the data type definition of the attribute field in the initial knowledge graph. The entity relationship uniqueness rule requires that the new triple formed by the relationship completion result does not have duplicate records in the initial knowledge graph. The attribute value mutual exclusion rule requires that if the same attribute field of the same type of entity already has a definite value in the initial knowledge graph, the completion value cannot be mutually exclusive with the definite value. The exclusion relationship is defined by a preset exclusion table. For each candidate completion result in the candidate set after the initial screening, it is checked against the above three constraint rules in turn. If all three checks pass, the candidate completion result is marked as a candidate completion result that has passed the check. If any check fails, the candidate completion result is marked as candidate knowledge that has failed the check and is pending review.
[0050] Specifically, S6 includes the following steps: S61. Based on the entity identifier and relation identifier in the verified candidate completion results, locate the nodes to be updated and the edges to be updated in the initial knowledge graph to obtain the set of nodes to be updated and the set of edges to be updated. Then, based on the attribute values in the verified candidate completion results, assign values to the attributes of each node in the set of nodes to be updated to obtain the set of nodes after attribute updates. S62. Based on the relation types in the candidate completion results that have passed the verification, update the relation types of each edge in the edge set to be updated to obtain the edge set after relation update. Based on the node set after attribute update and the edge set after relation update, determine the updated subgraph. Merge the updated subgraph into the initial knowledge graph to obtain the semantically completed knowledge graph. Mark the candidate knowledge that has not passed the verification as pending manual review. S63. Generate a list of candidate knowledge to be reviewed based on the candidate knowledge marked as pending manual review in the semantically completed knowledge graph, and establish a review tracking record according to the source identifier and generation time of each candidate knowledge in the list of pending review, and store the review tracking record in the metadata layer of the semantically completed knowledge graph.
[0051] Further, in step S61, the entity identifier field and relation identifier field carried by each candidate completion result are read from the verified candidate completion results. The entity identifier field is the unique number of the node to be updated in the initial knowledge graph node set, and the relation identifier field is the unique number of the edge to be updated in the initial knowledge graph edge set. All initial knowledge graph nodes corresponding to the read entity identifiers are marked as nodes to be updated and collected into a set of nodes to be updated. Similarly, all initial knowledge graph edges corresponding to the read relation identifiers are marked as edges to be updated and collected into a set of edges to be updated. For each node in the set of nodes to be updated, an attribute value entry matching the node's entity identifier is searched from the verified candidate completion results. The attribute value entry contains an attribute name and an attribute value. The original value of the corresponding attribute field in the initial knowledge graph for that node is replaced with the attribute value from the attribute value entry. After the replacement, the attribute-updated node is obtained. All attribute-updated nodes form the attribute-updated node set.
[0052] In step S62, for each edge in the set of edges to be updated, a relation type entry matching the relation identifier of the edge is found from the verified candidate completion results. The relation type entry contains a new relation type name. The original value of the relation type field of the edge in the initial knowledge graph is replaced with the new relation type name. After the replacement, the updated relation edge is obtained. All updated relation edges form the updated relation edge set. All nodes in the attribute-updated node set and all edges in the updated relation edge set are combined into a directed graph structure, which is the updated subgraph. Each node in the updated subgraph is merged with the node corresponding to the entity identifier in the initial knowledge graph through attribute coverage. Each edge in the updated subgraph is merged with the edge corresponding to the relation identifier in the initial knowledge graph through relation type coverage. After merging, the semantically completed knowledge graph is obtained. All candidate knowledge entries that failed verification are traversed, and the status field of each candidate knowledge entry is changed from the default value to the status indicator of pending manual review.
[0053] In step S63, the system queries the semantically completed knowledge graph for all candidate knowledge items whose status field is marked as "awaiting manual review." All retrieved items are then sorted according to their creation time, and the sorted list constitutes the review list. The source identifier field and creation time field of each candidate knowledge item in the review list are obtained. The source identifier field includes the sensor number and channel number. A review tracking record is created for each candidate knowledge item, containing the source identifier, creation time, current status, and creation time. The current status field is initially set to "pending processing," and the creation time field is set to the system's current timestamp. All created review tracking records are written as key-value pairs into the metadata layer of the semantically completed knowledge graph. The metadata layer is an additional data area stored independently of entity nodes and relation edges. In the metadata layer, the key in the key-value pairs is the unique identifier of the candidate knowledge item, and the value is the review tracking record structure, completing the associated storage.
[0054] Specifically, S7 includes the following steps: S71. Based on the preset version identifier and valid time interval, perform version validity checks on entities and relations in the semantically completed knowledge graph to obtain the version validity status of each entity and the version validity status of each relation, and analyze the impact range of the data update event to obtain the impact range marker. S72. Based on the version validity status of each entity, the version validity status of each relation, and the scope of influence, match the entities and relations affected by the data update event in the semantically completed knowledge graph to obtain the set of affected entities and the set of affected relations. S73. Based on the new attribute values in the data update event, iterate the attributes of each entity in the affected entity set to obtain the iterated entity. Based on the new relation type in the data update event, iterate the relation type of each relation in the affected relation set to obtain the iterated relation. S74. Determine the initial source confidence value based on the source type of the data update event, and determine the change range coefficient based on the change range of the data update event. Based on the initial source confidence value and the change range coefficient, determine the association confidence between each iterated entity in the affected entity set and each iterated relation in the affected relation set. Then, label the association confidence between the iterated entities and relations based on the association confidence to obtain the updated knowledge graph. The mathematical expression for the association confidence is: In the formula, Indicates the first After the iteration, the entity and the... The confidence level of the correlation between relationships after each iteration Indicates the first The initial value of the source credibility of the entity after each iteration. Indicates the first After the iteration, the entity and the... The range of changes between relationships after each iteration Indicates the first After the iteration, the entity and the... The coefficient of change range between relationships after each iteration Represents an exponential function. Indicates the first After the iteration, the entity and the... The shortest path distance between each relation in the semantically completed knowledge graph after each iteration. This represents the distance attenuation constant.
[0055] Further, in step S71, the preset version identifier field and valid time interval field corresponding to each entity and each relation are read from the metadata layer of the semantically completed knowledge graph. The preset version identifier is an incrementing integer, and the valid time interval is defined by the start timestamp and the end timestamp. The version validity check rule is as follows: if the current system time is between the start timestamp and the end timestamp of the valid time interval, and the preset version identifier is equal to the current global version number, then the version validity status of the entity or relation is marked as valid; if the current system time exceeds the valid time interval or the preset version identifier is less than the current global version number, then it is marked as expired. The current global version number is determined based on the number of updates to the semantically completed knowledge graph. For example, the initial global version number is set to 1, and the global version number is incremented by 1 after each data update event. The scope of influence analysis adopts a graph traversal method, starting from the entity nodes involved in the data update event and spreading outward along the relation edges in the semantically completed knowledge graph. The diffusion depth is determined according to the intensity of the change in the data update event. For example, if the diffusion depth is set to 3 layers, all entities and relations accessed within the diffusion depth range will be marked as scope of influence markers.
[0056] In step S72, all entity nodes in the semantically completed knowledge graph are traversed, and the version validity status and scope of influence of each entity are checked. Entities marked as valid in version validity status and marked as affected in scope of influence are added to the affected entity set. Entities marked as expired in version validity status but marked as affected in scope of influence are also added to the affected entity set, because expired entities need to be iteratively updated in data update events. All relation edges in the semantically completed knowledge graph are traversed, and the version validity status and scope of influence of each relation are checked. Relations marked as valid in version validity status and marked as affected in scope of influence are added to the affected relation set, and relations marked as expired in version validity status and marked as affected in scope of influence are also added to the affected relation set. After completing all traversals, the complete set of affected entities and the set of affected relations are obtained.
[0057] In step S73, a list of new attribute values and a list of new relationship types are read from the data packet of the data update event. Each record in the list of new attribute values contains an entity identifier, an attribute name, and an attribute value. Each record in the list of new relationship types contains a relationship identifier and a new relationship type name. The system checks if there is a matching record in the new attribute value list for the entity identifier of each entity in the affected entity set. If a matching record exists, the original value of the corresponding attribute field of that entity is replaced with the attribute value from the new attribute value list. Simultaneously, the version number field of that entity is incremented by 1, indicating that one version iteration has been completed. After the replacement, the iterated entity is obtained. The system checks if there is a matching record in the new relationship type list for the relationship identifier of each relationship in the affected relationship set. If a matching record exists, the original value of the relationship type field of that relationship is replaced with the new relationship type name. Simultaneously, the version number field of that relationship is incremented by 1. After the version iteration is completed, the iterated relationship is obtained.
[0058] In step S74, the source type is read from the source type field of the data update event. Source types include manually labeled, automatically collected by sensors, and imported from external systems. Each source type corresponds to a preset initial source confidence value, which is determined based on historical accuracy statistics for each source type. For example, the initial source confidence value for manually labeled types is set to 0.95, for automatically collected by sensors to 0.85, and for imported from external systems to 0.75. The change range coefficient is determined based on the number of entities and relations involved in the data update event. The change range coefficient is calculated as the sum of the number of entities and relations involved divided by the ratio of the total number of entities to the total number of relations in the semantically completed knowledge graph. A larger ratio indicates a larger change range. The change range coefficient is linearly mapped between 0.5 and 1.0. For example, if the ratio is 0.2, the change range coefficient is set to 0.6. The distance decay constant is determined based on the average shortest path length of the semantically completed knowledge graph; for example, it can be set to 3.0. For each iterated entity in the affected entity set and each iterated relation in the affected relation set, the shortest path distance between the two nodes in the semantically completed knowledge graph is calculated using a breadth-first search algorithm. The calculated association confidence is then appended as a label to the semantic association between the corresponding iterated entity and iterated relation. After completing the confidence labeling, the updated knowledge graph is obtained.
Claims
1. A method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle, characterized in that: Includes the following steps: S1. Collect the operating status of industrial products to obtain full life cycle data of industrial products, and preprocess the full life cycle data of industrial products to obtain preprocessed data. S2. Based on the preprocessed data, determine the list of available feature extractors, extract features from each knowledge element in the preprocessed data, linearly transform the initial feature vectors of each modality to a unified target dimension, weight the feature vectors of the target dimension, and perform time-series encoding on the weighted feature vectors according to the generation time of each knowledge element to determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object. Then, perform weighted aggregation on the time-series enhanced feature vectors according to the cross-modal similarity matrix and the preset feature weight rules to obtain the similarity analysis results. S3. Based on the similarity analysis results, determine the set of candidate knowledge elements belonging to the same industrial object, perform conflict detection on the set of candidate knowledge elements, obtain the conflict detection results, and filter the set of candidate knowledge elements based on the conflict detection results to obtain the target set of knowledge elements. S4. Construct a basic entity layer based on the target knowledge element set, and construct a domain relationship layer according to the structural composition relationship and temporal sequence relationship between entities in the basic entity layer. Establish multidimensional semantic relationships between entities based on the basic entity layer and the domain relationship layer to obtain the initial knowledge graph. S5. Perform structural constraint reasoning on entities with missing attributes and entities with missing relationships using the initial knowledge graph. Determine the set of path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph. Predict the candidate relationship types for entity pairs with missing relationships. Determine the comprehensive confidence level based on the cross-modal matching score, graph structure consistency score, industrial rule compliance, historical case similarity, and data source credibility of each candidate completion result in the completion candidate set. Validate the completion candidate set based on the comprehensive confidence level and preset multiple constraint rules to obtain the candidate completion results that pass the validation and the candidate knowledge that fails the validation and is pending review. S6. Based on the candidate completion results that have passed the verification, perform semantic completion on the initial knowledge graph to obtain the semantically completed knowledge graph, and mark the candidate knowledge that has not passed the verification as pending manual review. S7. Iterate the semantically completed knowledge graph according to the preset version identifier and effective time interval to obtain the iterated entities and relations, and mark the confidence level of the correlation between the iterated entities and relations to obtain the updated knowledge graph.
2. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S1 includes the following steps: S11. Collect raw operating status signals from each monitoring node of the industrial product according to the preset acquisition frequency and preset acquisition channel, and perform analog-to-digital conversion on the raw operating status signals to obtain digital operating status signals. S12. Based on the time-domain and frequency-domain characteristics of each signal component in the digital operation status signal, noise component identification is performed on the digital operation status signal to obtain noise-marked signal components. Then, the digital operation status signal is filtered according to the noise-marked signal components to obtain the filtered operation status signal. S13. Divide the filtered running status signal into time windows according to the preset time window length and preset sliding step size to obtain window signal segments corresponding to each time window, and perform missing value detection on the window signal segments to obtain missing value detection results. Then, fill in the missing value segments with data according to the missing value detection results to obtain filled signal segments. S14. Perform outlier detection on the filled signal segment, obtain the outlier detection results, and perform outlier correction on the filled signal segment based on the outlier detection results to obtain preprocessed data.
3. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S2 includes the following steps: S21. Determine the feature extractor corresponding to each knowledge element based on the modality type of each knowledge element in the preprocessed data, obtain a candidate list of feature extractors, and filter the candidate list of feature extractors according to the available status identifier of each feature extractor in the candidate list to obtain a list of available feature extractors. S22. Based on the output dimensions of each feature extractor in the list of available feature extractors, perform feature extraction on each knowledge element in the preprocessed data to obtain the initial feature vector of each modality, and perform linear transformation on the initial feature vector of each modality to obtain the target dimension feature vector. S23. The target dimension feature vector is weighted according to the source identifier of each knowledge element in the preprocessed data to obtain a weighted feature vector. The weighted feature vector is then time-encoded according to the generation time of each knowledge element in the preprocessed data to obtain a time-enhanced feature vector. S24. Based on the time-series enhanced feature vector, determine the cross-modal similarity matrix between different knowledge elements belonging to the same industrial object, and perform weighted aggregation of the time-series enhanced feature vector according to the cross-modal similarity matrix and the preset feature weight rules to obtain the similarity analysis results.
4. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S3 includes the following steps: S31. Based on the similarity analysis results, perform entity type identification on each candidate knowledge element in the candidate knowledge element set belonging to the same industrial object, obtain the entity type label of each candidate knowledge element, and extract the entity attribute value of each candidate knowledge element based on the entity type label of each candidate knowledge element to obtain the entity attribute value set. S32. Classify the entity attribute values of each candidate knowledge element based on the semantic type corresponding to each entity attribute value in the entity attribute value set to obtain classified entity attribute values. Extract relation paths from the entity relation types in the candidate knowledge element set based on the classified entity attribute values to obtain a relation path set. S33. Based on the source identifier of each candidate knowledge element in the candidate knowledge element set, trace the source of the entity attribute value set and the relation path set to obtain the source-marked entity attribute value set and the source-marked relation path set. S34. Conflict detection is performed on the entity attribute values and entity relationship types in the candidate knowledge element set by using the data type of each entity attribute value in the source tag entity attribute value set and the semantic direction of each relationship path in the source tag relationship path set, and the conflict detection results are obtained. S35. Based on the conflict detection results, filter the entity attribute values and entity relationship types in the candidate knowledge element set to obtain the target knowledge element set.
5. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S4 includes the following steps: S41. Establish an entity index table based on the unique identifier of each entity in the target knowledge element set, and generate an initial entity embedding vector based on the entity type and entity attribute value of each entity in the entity index table to obtain a set of initial entity embedding vectors. S42. Cluster the entities in the target knowledge element set according to the semantic distance corresponding to the initial embedding vectors of each entity in the initial embedding vector set, obtain the entity clustering results, determine the structural composition relationship between each entity according to the entity clustering results, obtain the structural composition relationship set, determine the temporal sequence relationship between each entity according to the generation time of each knowledge element in the target knowledge element set, obtain the temporal sequence relationship set. S43. Construct a domain relation layer based on the set of structural composition relations and the set of temporal sequence relations, and establish multidimensional semantic relations between entities according to the set of initial entity embedding vectors and the relation types in the domain relation layer to obtain the initial knowledge graph.
6. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S5 includes the following steps: S51. Determine the attribute statistical features of known entities based on the attribute distribution of known entities in the initial knowledge graph, and generate attribute candidate values for entities with missing attributes in the initial knowledge graph based on the attribute statistical feature vectors to obtain an attribute completion candidate set. Determine the path feature vectors between known entity pairs based on the relationship paths between known entities in the initial knowledge graph to obtain a path feature vector set. S52. Based on the path feature vector set, predict the candidate types of entity pairs with missing relations in the initial knowledge graph to obtain the candidate set of relation completion. Determine the candidate set of completion based on the candidate set of attribute completion and the candidate set of relation completion. S53. Determine the overall confidence level of each candidate completion result based on its cross-modal matching score, graph structure consistency score, industry rule compliance, historical case similarity, and data source credibility. The mathematical expression for the overall confidence level of each candidate completion result is as follows: In the formula, Indicates the first The overall confidence level of the candidate completion results The weighting coefficients represent the cross-modal matching scores. Indicates the first Cross-modal matching scores of candidate completion results The weighting coefficients represent the graph structure consistency score. Indicates the first The graph structure consistency score of each candidate completion result. Weighting coefficients representing the degree of compliance with industry regulations. Indicates the first The degree of compliance with industry rules for each candidate completion result. The weighting coefficients representing the similarity of historical cases. Indicates the first Historical case similarity of each candidate completion result Weighting coefficients representing the credibility of data sources. Indicates the first The credibility of the data sources for each candidate completion result. ; S54. The candidate completion set is initially screened based on the comprehensive confidence level and the preset confidence threshold to obtain the initial candidate set. The candidate completion results in the initial candidate set are then constrained and verified according to the preset multiple constraint rules. Candidate completion results whose comprehensive confidence level is not less than the preset confidence threshold and which pass all constraint verifications are marked as verified candidate completion results. Candidate completion results whose comprehensive confidence level is less than the preset confidence threshold or which fail any constraint verification are marked as unverified candidate knowledge to be reviewed.
7. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S6 includes the following steps: S61. Based on the entity identifier and relation identifier in the verified candidate completion results, locate the nodes to be updated and the edges to be updated in the initial knowledge graph to obtain the set of nodes to be updated and the set of edges to be updated. Then, based on the attribute values in the verified candidate completion results, assign values to the attributes of each node in the set of nodes to be updated to obtain the set of nodes after attribute updates. S62. Based on the relation types in the candidate completion results that have passed the verification, update the relation types of each edge in the edge set to be updated to obtain the edge set after relation update. Based on the node set after attribute update and the edge set after relation update, determine the updated subgraph. Merge the updated subgraph into the initial knowledge graph to obtain the semantically completed knowledge graph. Mark the candidate knowledge that has not passed the verification as pending manual review. S63. Generate a list of candidate knowledge to be reviewed based on the candidate knowledge marked as pending manual review in the semantically completed knowledge graph, and establish a review tracking record according to the source identifier and generation time of each candidate knowledge in the list of pending review, and store the review tracking record in the metadata layer of the semantically completed knowledge graph.
8. The method for constructing and semantically completing an industrial knowledge graph throughout the entire product lifecycle as described in claim 1, characterized in that: S7 includes the following steps: S71. Based on the preset version identifier and valid time interval, perform version validity checks on entities and relations in the semantically completed knowledge graph to obtain the version validity status of each entity and the version validity status of each relation, and analyze the impact range of the data update event to obtain the impact range marker. S72. Based on the version validity status of each entity, the version validity status of each relation, and the scope of influence, match the entities and relations affected by the data update event in the semantically completed knowledge graph to obtain the set of affected entities and the set of affected relations. S73. Based on the new attribute values in the data update event, iterate the attributes of each entity in the affected entity set to obtain the iterated entity. Based on the new relation type in the data update event, iterate the relation type of each relation in the affected relation set to obtain the iterated relation. S74. Determine the initial value of source credibility based on the source type of the data update event, and determine the change range coefficient based on the change range of the data update event. Determine the confidence level of the association between each iterated entity in the affected entity set and each iterated relation in the affected relation set based on the initial value of source credibility and the change range coefficient. Mark the confidence level of the association between the iterated entities and relations based on the confidence level of the association, and obtain the updated knowledge graph.
Citation Information
Patent Citations
Geological disaster knowledge graph construction method and system based on large language model enhancement
CN122713387A
