An AI power knowledge base retrieval method based on deep learning

Abstract

This invention discloses an AI-based power knowledge base retrieval method based on deep learning, comprising the following steps: acquiring power review knowledge data, processing the knowledge base construction to obtain a power knowledge base; performing clause segmentation and constraint element extraction to construct a two-layer knowledge unit structure; parsing power engineering design documents to obtain structured design data; constructing a query term sequence; encoding the query term sequence and the two-layer knowledge unit structure to construct a knowledge retrieval index; employing the ColBERT fine-grained delayed interactive semantic matching algorithm to perform term-level relevance calculation on the query-side term vector representation and the knowledge-side term vector representation to obtain a target knowledge clause set; performing hit evidence attribution on the target knowledge clause set and outputting the review results. This invention achieves accurate semantic retrieval and intelligent matching between power engineering design documents and knowledge clauses.

Description

Technical Field

[0001] This invention relates to the field of information technology in power engineering, and in particular to an AI-based power knowledge base retrieval method based on deep learning. Background Technology

[0002] With the continuous expansion of power grid construction, distribution network engineering, transmission and transformation engineering, and related design review work have placed higher demands on the efficiency of knowledge acquisition, the accuracy of clause retrieval, and the level of intelligence in the review process. In existing technologies, power engineering reviews typically rely on design specifications, industry standards, technical guidelines, review rules, and historical cases as the basis for judgment. Reviewers need to combine information such as textual descriptions, drawing content, and table parameters in design documents to verify the compliance, rationality, and completeness of the plan. This type of review work involves a wide range of knowledge sources, a large number of clauses, and complex data types, exhibiting obvious knowledge-intensive and experience-dependent characteristics.

[0003] Currently, most technical solutions for retrieving and accessing power industry knowledge clauses primarily rely on keyword matching, full-text search, or ordinary vector retrieval for knowledge location. Information extraction from design documents also largely focuses on processing single text content or only partially identifying drawings and tables, lacking a unified structured representation and cross-modal association capabilities. In practical applications, because power engineering design documents typically contain multiple data formats such as text, drawings, and tables, and there are often inconsistencies in expression, terminology, and constraints between design descriptions and specification clauses, existing technologies struggle to accurately establish a deep semantic correspondence between design content and knowledge clauses.

[0004] Therefore, how to provide an AI-based power knowledge base retrieval method based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-based power knowledge base retrieval method based on deep learning. This invention constructs a power knowledge base, a two-layer knowledge unit structure, and a knowledge retrieval index, and combines it with the ColBERT fine-grained delayed interactive semantic matching algorithm to achieve accurate semantic retrieval and intelligent matching between power engineering design documents and knowledge clauses. This can improve retrieval efficiency and matching accuracy, reduce the workload of manual review, reduce the risk of incorrect or missed reviews, and enhance the automation and intelligence level of power engineering review.

[0006] According to an embodiment of the present invention, an AI power knowledge base retrieval method based on deep learning includes the following steps: Acquire power review knowledge data, process the power review knowledge data to construct a knowledge base, and obtain the power knowledge base; Based on the power knowledge base, clause segmentation and constraint element extraction are performed to construct a two-layer knowledge unit structure including clause level and element level; Obtain the power engineering design documents to be reviewed, parse the power engineering design documents, and obtain structured design data; Construct query term sequences based on structured design data; A deep learning semantic coding model is used to encode the query word sequence and the two-layer knowledge unit structure respectively, so as to obtain the query word vector representation and the knowledge word vector representation, and a knowledge retrieval index is constructed based on the knowledge word vector representation. Based on the knowledge retrieval index, the ColBERT fine-grained delayed interaction semantic matching algorithm is used to calculate the word-level relevance between the query word vector representation and the knowledge word vector representation, and to perform grouped delayed interaction and matching score aggregation to obtain the target knowledge clause set. The system performs hit evidence attribution on the target knowledge clause set, constructs a clause matching result and rule constraint mapping structure, generates a review basis set, and performs clause comparison and compliance verification based on the review basis set, outputting the review results.

[0007] Optionally, obtaining the power knowledge base specifically includes: Acquire and aggregate power review knowledge data, and uniformly receive and collect it according to knowledge source category, business application category, and review scenario category to form an original knowledge data set; The original knowledge data set is preprocessed, and the knowledge content is segmented and organized according to the corresponding rule units to form a standardized set of knowledge items with clause identifiers, source identifiers, hierarchical identifiers and rule association identifiers; A knowledge base is constructed based on a standardized set of knowledge items to form a power knowledge base oriented towards power engineering design review scenarios.

[0008] Optionally, the construction of the two-layer knowledge unit structure is based on: The knowledge entries in the power knowledge base are segmented according to preset segmentation rules to form different independent and semantically complete clause-level knowledge units, and all clause-level knowledge units are summarized to form a clause-level knowledge unit set. Based on the set of clause-level knowledge units, constraint elements are extracted for each clause-level knowledge unit. The corresponding constraint elements of each clause-level knowledge unit are identified one by one, and element-level knowledge units corresponding to the clause-level knowledge units are generated, forming a set of element-level knowledge units that correspond one-to-one with each clause-level knowledge unit. Based on the set of knowledge units at the clause level and the set of knowledge units at the element level, the association between each knowledge unit at the clause level and the corresponding knowledge unit at the element level is established. At the same time, the association is organized and stored in a unified manner, thus constructing a two-layer knowledge unit structure including the clause level and the element level.

[0009] Optionally, obtaining the structured design data specifically includes: The power engineering design documents to be reviewed are obtained, and the power engineering design documents are received, registered, verified for integrity and identified in terms of document type, forming a set of original design documents to be parsed. Based on the original design file set, text content reading, paragraph content segmentation, and semantic fragment extraction are performed on text-based design files; drawing content reading, drawing annotation recognition, equipment identification extraction, connection relationship extraction, and parameter label extraction are performed on drawing-based design files; and table area recognition, table header content recognition, field content extraction, and parameter item content extraction are performed on table-based design files. The text content fragments, drawing content fragments, and table content fragments obtained from parsing various types of design files are then uniformly summarized to form a set of design content parsing results. Based on the set of design content analysis results, the content fragments are structured, and content fragments belonging to the same review object or the same technical constraint are merged and organized to form structured design data.

[0010] Optionally, the construction of the query term sequence specifically includes: Each field record in the structured design data is processed to identify retrieval requirements. The content of the review object, attribute description, parameter description and source location related to the knowledge clause are extracted from each field record. The extracted content is then denoised, standardized and consistent to form a set of query elements. Based on the set of query elements, query elements belonging to the same review object or the same technical constraint are clustered and merged. Review intent reconstruction processing is performed on the clustered and merged query elements. Each group of query elements is arranged and organized in the order of review object priority, attribute description follow, parameter description supplement and source location association, forming query term fragments corresponding to each group of query elements. Based on query term fragments, sequential concatenation, duplicate term removal, and term boundary marking are performed on each query term fragment to generate a query term sequence corresponding to the structured design data.

[0011] Optionally, obtaining the knowledge retrieval index specifically includes: Obtain the query term sequence and two-layer knowledge unit structure, and use the clause-level knowledge unit and element-level knowledge unit as knowledge-end input data to construct a deep learning semantic coding model; The query term sequence is input into the deep learning semantic encoding model. Each query term in the query term sequence is sequentially processed by term embedding, position encoding, and context semantic encoding. Review role constraint injection and sparse term constraint processing are performed on each query term. The sparse term constraint processing activates and enhances professional term terms related to power design specifications, review rules, and engineering constraints, and sparsely suppresses ordinary descriptive terms that are not critical to review. The query term vector representation is output. The two-layer knowledge unit structure is input into the deep learning semantic coding model. The term-level knowledge units and element-level knowledge units in the two-layer knowledge unit structure are processed by word embedding, position encoding and context semantic encoding respectively. The hierarchical association injection and element affiliation preservation processing are performed on the knowledge words. The hierarchical semantic association information is jointly encoded and further residual compression processing is performed to compress and store the knowledge word vector representation. While maintaining the semantic distinguishability and hierarchical association relationship, the storage overhead of the knowledge word vector representation is reduced, forming a knowledge word vector set. Based on the knowledge term vector set, clause-level index items are established for knowledge term vector representations belonging to the same clause-level knowledge unit, and element-level index items are established for knowledge term vector representations belonging to the corresponding element-level knowledge unit. A hierarchical association index item is also established between the clause-level index items and the element-level index items. At the same time, a central representation is generated for the knowledge term vector representations under the same clause-level knowledge unit, and a central interactive pre-screening process is performed based on the central representation to form a knowledge retrieval index.

[0012] Optionally, obtaining the target knowledge clause set specifically includes: Based on the knowledge retrieval index, candidate knowledge unit location processing is performed. The knowledge retrieval index is jointly searched to determine the candidate clause-level knowledge units and candidate element-level knowledge units associated with the query term vector representation. The term vector representations at the knowledge end are then associated, extracted, and grouped to form a set of candidate knowledge end term vectors. The ColBERT fine-grained delayed interactive semantic matching algorithm is used to perform word-level relevance calculation on the query word vector representation and the candidate knowledge word vector set. Review role constraints are introduced. Each query word vector in the query word vector representation is constrained according to the review role. Only query word vectors with corresponding review role relationships are allowed to participate in the relevance calculation with the corresponding knowledge word vectors in the candidate knowledge word vector set. The result is a word-level relevance result set containing review role constraint information. Based on the word-level relevance result set, a grouped delayed interaction processing is performed. The query word vector representation is divided into different query word groups according to the review role. The candidate knowledge word vector set is divided into different knowledge word groups according to the clause-level affiliation and the element-level affiliation. For each query word group, delayed interaction matching is performed with the corresponding clause-level knowledge word group and element-level knowledge word group. Within each group, the knowledge word with the highest relevance is retained as the intra-group matching result of the current query word. The intra-group matching results of the clause-level knowledge word group and the element-level knowledge word group are collaboratively associated to form a grouped delayed interaction result that reflects both clause-level matching strength and element-level matching strength. Based on the grouped delayed interaction results, the matching score aggregation process is performed. Evidence consistency judgment is performed on the clause-level matching results and element-level matching results corresponding to each query tuple. Enhanced aggregation is performed on the matching results with consistent pointing relationship between clause-level matching results and element-level matching results, and suppression aggregation is performed on the matching results with inconsistent relationship between clause-level matching results and element-level matching results. The comprehensive matching score corresponding to each candidate knowledge unit is obtained, and all candidate knowledge units are sorted according to the comprehensive matching score to form the candidate knowledge unit ranking result. The target knowledge clause set is determined based on the ranking results of candidate knowledge units. Candidate clause-level knowledge units and their corresponding candidate element-level knowledge units that meet the preset retention conditions in terms of comprehensive matching scores are retained and then merged and organized according to hierarchical association and evidence consistency to generate the target knowledge clause set.

[0013] Optionally, obtaining the review results specifically includes: Extract and organize the target knowledge clause set, and associate the query terms involved in the matching in the query term vector representation with the matched knowledge terms in the target knowledge clause set to form a set of matching term correspondences; The attribution of hit evidence is performed based on the set of hit word correspondences. The correspondence of each hit word is recorded and associated. The hit word correspondences belonging to the same target knowledge clause are merged to form a set of hit evidence corresponding to each target knowledge clause. Based on the set of hit evidence, the hit evidence in each target knowledge clause is organized in layers according to clause-level content and element-level content. The hit evidence is mapped and associated with the corresponding rule constraints, parameter verification items and review requirements to form clause matching results. The correspondence between target knowledge clauses, hit evidence, rule constraints and parameter verification items is established to form a rule constraint mapping structure. Based on the clause matching results and rule constraint mapping structure, the clause matching results belonging to the same review object or the same review matter are merged and organized, and the corresponding target knowledge clause content, hit evidence content, rule constraint item content and parameter verification item content are uniformly organized to form a set of review basis; Based on the review basis set, clause comparison and compliance verification are performed. The structured design data in the power engineering design documents to be reviewed is compared item by item with the target knowledge clauses, rule constraints, and parameter verification items in the review basis set. At the same time, compliance judgment, missing item judgment, conflict item judgment, and abnormal item judgment are performed to form a verification result. The review result is output based on the verification result.

[0014] The beneficial effects of this invention are: This invention constructs a knowledge base for power review knowledge data and further builds a two-layer knowledge unit structure at the clause and element levels. This enables knowledge content such as power design specifications, industry standards, technical guidelines, review rules, and historical review cases to be organized and accessed in a more suitable way for retrieval and matching. This improves the structuring and reusability of power knowledge clauses and provides a stable knowledge foundation for subsequent accurate semantic retrieval and rule mapping.

[0015] This invention addresses the characteristic of power engineering design documents containing multimodal data such as text, drawings, and tables. By parsing the design documents and forming structured design data, and then constructing query term sequences based on the structured design data, the originally scattered, heterogeneous, and inconsistently expressed design content can be transformed into retrieval input that can be used for unified processing and semantic matching. This improves the completeness and consistency of content extraction during review and enhances the correspondence between design content and knowledge clauses.

[0016] This invention employs a deep learning semantic coding model and a knowledge retrieval index built based on knowledge-side lexical vector representations. Combined with the ColBERT fine-grained delayed interaction semantic matching algorithm, it performs lexical-level relevance calculation, grouped delayed interaction, and matching score aggregation on the query-side and knowledge-side lexical vector representations. This enables the retrieval process to move beyond traditional keyword matching or coarse-grained text similarity judgment, and instead identify the deep semantic connections between the design content and knowledge terms at the lexical level. This improves the accuracy of obtaining the target knowledge term set and the relevance of the retrieval results.

[0017] This invention forms a set of review basis by attributing the target knowledge clause set to the hit evidence, constructing clause matching results, and mapping rule constraints. Based on the set of review basis, clause comparison and compliance verification are carried out, so that the search results can be further transformed into review output results with evidence support and rule association. This improves the interpretability and traceability of the review process and reduces the risk of incorrect or missed reviews in manual review. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Fig. 1 This is a flowchart of an AI power knowledge base retrieval method based on deep learning proposed in this invention; Fig. 2 This is a schematic diagram of the ColBERT fine-grained delayed interactive semantic matching structure for an AI power knowledge base retrieval method based on deep learning proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figs. 1-2 A deep learning-based AI power knowledge base retrieval method includes the following steps: Acquire power review knowledge data, process the power review knowledge data to construct a knowledge base, and obtain the power knowledge base; Based on the power knowledge base, clause segmentation and constraint element extraction are performed to construct a two-layer knowledge unit structure including clause level and element level; Obtain the power engineering design documents to be reviewed, parse the power engineering design documents, and obtain structured design data; Construct query term sequences based on structured design data; A deep learning semantic coding model is used to encode the query word sequence and the two-layer knowledge unit structure respectively, so as to obtain the query word vector representation and the knowledge word vector representation, and a knowledge retrieval index is constructed based on the knowledge word vector representation. Based on the knowledge retrieval index, the ColBERT fine-grained delayed interaction semantic matching algorithm is used to calculate the word-level relevance between the query word vector representation and the knowledge word vector representation, and to perform grouped delayed interaction and matching score aggregation to obtain the target knowledge clause set. The system performs hit evidence attribution on the target knowledge clause set, constructs a clause matching result and rule constraint mapping structure, generates a review basis set, and performs clause comparison and compliance verification based on the review basis set, outputting the review results.

[0021] In this embodiment, obtaining the power knowledge base specifically includes: Acquire and aggregate power review knowledge data. The power review knowledge data consists of knowledge content from power design specifications, industry standards, technical guidelines, review rules, and historical review cases. It is uniformly received and aggregated according to knowledge source category, business application category, and review scenario category to form the original knowledge data set. Data preprocessing is performed on the original knowledge data set. Data preprocessing includes identifying and processing invalid content, duplicate content, content with abnormal format, and non-review-related content in the original knowledge data. The knowledge text is standardized in terms of format, terminology, numbering, and hierarchy. The knowledge content is then segmented and organized according to the corresponding rule units to form a standardized set of knowledge entries with clause identifiers, source identifiers, hierarchy identifiers, and rule association identifiers. The knowledge base is constructed based on a standardized set of knowledge items. The construction of the knowledge base includes the structured organization of standardized knowledge items, the establishment of item associations, the establishment of rule constraint relationships, the merging of similar items, and cross-source knowledge mapping, forming a power knowledge base oriented towards power engineering design review scenarios.

[0022] In this embodiment, the construction of the two-layer knowledge unit structure is based on the following: The knowledge entries in the power knowledge base are segmented according to preset segmentation rules to form different independent and semantically complete clause-level knowledge units, and all clause-level knowledge units are summarized to form a clause-level knowledge unit set. Based on the set of clause-level knowledge units, constraint elements are extracted for each clause-level knowledge unit. The constraint elements include review object elements, constraint attribute elements, parameter condition elements, and scenario limitation elements. The corresponding constraint elements of each clause-level knowledge unit are identified one by one, and element-level knowledge units corresponding to the clause-level knowledge units are generated, forming a set of element-level knowledge units that correspond one-to-one with each clause-level knowledge unit. Based on the set of knowledge units at the clause level and the set of knowledge units at the element level, the association between each knowledge unit at the clause level and the corresponding knowledge unit at the element level is established. At the same time, the association is organized and stored in a unified manner, thus constructing a two-layer knowledge unit structure including the clause level and the element level.

[0023] In this embodiment, obtaining structured design data specifically includes: The process involves acquiring power engineering design documents to be reviewed, receiving, registering, verifying the integrity of the documents, and identifying their document types. The power engineering design documents are then categorized into text-based, drawing-based, and tabular-based design documents. Corresponding document identifiers, source identifiers, and content identifiers are established for each type of design document, forming a set of original design documents to be parsed. Based on the original design file set, text content reading, paragraph content segmentation, and semantic fragment extraction are performed on text-based design files; drawing content reading, drawing annotation recognition, equipment identification extraction, connection relationship extraction, and parameter label extraction are performed on drawing-based design files; and table area recognition, table header content recognition, field content extraction, and parameter item content extraction are performed on table-based design files. The text content fragments, drawing content fragments, and table content fragments obtained from parsing various types of design files are then uniformly summarized to form a set of design content parsing results. Based on the set of design content analysis results, each content fragment is processed in a structured manner. The structured processing includes identifying, extracting, classifying, associating, and uniformly storing the review object content, attribute description content, parameter description content, and source location content in each content fragment. Content fragments belonging to the same review object or the same technical constraint are merged and organized to form structured design data.

[0024] In this embodiment, the construction of the query term sequence specifically includes: Each field record in the structured design data is processed to identify retrieval requirements. The content of the review object, attribute description, parameter description and source location related to the knowledge clause are extracted from each field record. The extracted content is then denoised, standardized and consistent to form a set of query elements. Based on the set of query elements, query elements belonging to the same review object or the same technical constraint are clustered and merged. Review intent reconstruction processing is performed on the clustered and merged query elements. Each group of query elements is arranged and organized in the order of review object priority, attribute description follow, parameter description supplement and source location association, forming query term fragments corresponding to each group of query elements. Based on query term fragments, sequential concatenation, duplicate term removal, and term boundary marking are performed on each query term fragment to generate a query term sequence corresponding to the structured design data.

[0025] In this embodiment, obtaining the knowledge retrieval index specifically includes: The query term sequence and two-layer knowledge unit structure are obtained. The clause-level knowledge unit and the element-level knowledge unit are used as the knowledge-end input data to construct a deep learning semantic coding model. The deep learning semantic coding model includes a term embedding layer, a positional coding layer, a contextual semantic coding layer, and a vector output layer. The term embedding layer is used to convert query terms and knowledge terms into initial semantic representations. The positional coding layer is used to inject term order information into the initial semantic representations. The contextual semantic coding layer is used to semantically model the contextual relationships between terms. The vector output layer is used to output the term vector representations corresponding to the query terms and knowledge terms. The query term sequence is input into a deep learning semantic encoding model. Each query term in the query term sequence is sequentially processed by term embedding, positional encoding, and contextual semantic encoding. Review role constraint injection and sparse term constraint processing are then performed on each query term. The sparse term constraint processing activates and enhances professional term terms related to power design specifications, review rules, and engineering constraints, and sparsely suppresses ordinary descriptive terms that are not critical to review. The output is a query term vector representation. The two-layer knowledge unit structure is input into a deep learning semantic coding model. Lexical embedding, positional encoding, and contextual semantic encoding are performed on the clause-level knowledge units and element-level knowledge units in the two-layer knowledge unit structure, respectively. Hierarchical association injection and element affiliation preservation processing are performed on the knowledge lexical units. The hierarchical semantic association information is jointly encoded and further residual compression processing is performed to compress and store the knowledge-end lexical vector representation. While maintaining semantic distinguishability and hierarchical association, the storage overhead of the knowledge-end lexical vector representation is reduced, forming a knowledge-end lexical vector set. The hierarchical semantic association information includes the overall semantic information of the clause-level knowledge unit, the constraint semantic information of the element-level knowledge unit, and the hierarchical association information between the clause-level knowledge unit and the corresponding element-level knowledge unit. Based on the knowledge-end term vector set, clause-level index items are established for knowledge-end term vector representations belonging to the same clause-level knowledge unit, and element-level index items are established for knowledge-end term vector representations belonging to the corresponding element-level knowledge unit. A hierarchical association index item is also established between the clause-level index items and the element-level index items. At the same time, a central representation is generated for the knowledge-end term vector representations under the same clause-level knowledge unit, and a central interactive pre-screening process is performed based on the central representation. The central interactive pre-screening process adopts a candidate pre-screening mechanism based on the central representation to pre-position candidate clause-level knowledge units and candidate element-level knowledge units that have a high relevance to the query term vector sequence, thus forming a knowledge retrieval index.

[0026] This invention constructs a deep learning semantic encoding model for both the query and knowledge ends, and combines it with review role constraints, sparse term constraints, hierarchical semantic association encoding, residual compression, and a candidate pre-screening mechanism based on center representation to form a knowledge retrieval index suitable for power knowledge clause retrieval. This improves the consistency of semantic representation between query terms and knowledge terms, enhances index organization efficiency, reduces vector storage overhead on the knowledge end, and improves the accuracy and retrieval efficiency of candidate knowledge unit positioning. It also provides a stable and reliable retrieval foundation for subsequent fine-grained semantic matching and target knowledge clause acquisition.

[0027] In this embodiment, obtaining the target knowledge clause set specifically includes: Based on the knowledge retrieval index, candidate knowledge unit location processing is performed. The knowledge retrieval index is jointly searched to determine the candidate clause-level knowledge units and candidate element-level knowledge units associated with the query term vector representation. The term vector representations at the knowledge end are then associated, extracted, and grouped to form a set of candidate knowledge end term vectors. The ColBERT fine-grained delayed interactive semantic matching algorithm is used to perform word-level relevance calculation on the query word vector representation and the candidate knowledge word vector set. Review role constraints are introduced. Each query word vector in the query word vector representation is constrained according to the review role. Only query word vectors with corresponding review role relationships are allowed to participate in the relevance calculation with the corresponding knowledge word vectors in the candidate knowledge word vector set. The result is a word-level relevance result set containing review role constraint information. Based on the word-level relevance result set, a grouped delayed interaction processing is performed. The query word vector representation is divided into different query word groups according to the review role. The candidate knowledge word vector set is divided into different knowledge word groups according to the clause-level affiliation and the element-level affiliation. For each query word group, delayed interaction matching is performed with the corresponding clause-level knowledge word group and element-level knowledge word group. Within each group, the knowledge word with the highest relevance is retained as the intra-group matching result of the current query word. The intra-group matching results of the clause-level knowledge word group and the element-level knowledge word group are collaboratively associated to form a grouped delayed interaction result that reflects both clause-level matching strength and element-level matching strength. Based on the grouped delayed interaction results, the matching score aggregation process is performed. Evidence consistency judgment is performed on the clause-level matching results and element-level matching results corresponding to each query tuple. Enhanced aggregation is performed on the matching results with consistent pointing relationship between clause-level matching results and element-level matching results, and suppression aggregation is performed on the matching results with inconsistent relationship between clause-level matching results and element-level matching results. The comprehensive matching score corresponding to each candidate knowledge unit is obtained, and all candidate knowledge units are sorted according to the comprehensive matching score to form the candidate knowledge unit ranking result. The target knowledge clause set is determined based on the ranking results of candidate knowledge units. Candidate clause-level knowledge units and their corresponding candidate element-level knowledge units that meet the preset retention conditions in terms of comprehensive matching scores are retained and then merged and organized according to hierarchical association and evidence consistency to generate the target knowledge clause set.

[0028] This invention achieves precise retrieval and deep semantic matching between power engineering design content and knowledge clauses at both the clause and element levels by locating candidate knowledge units based on knowledge retrieval indexes, introducing ColBERT fine-grained delayed interactive semantic matching with review role constraints, and aggregating matching scores based on evidence consistency. This improves the accuracy of obtaining the target knowledge clause set, the relevance and interpretability of the retrieval results, enhances the completeness and stability of knowledge clause retrieval in complex power review scenarios, and provides a more reliable retrieval foundation for subsequent rule mapping, compliance verification, and review opinion generation.

[0029] In this embodiment, obtaining the review result specifically includes: Extract and organize the target knowledge clause set, and associate the query terms involved in the matching in the query term vector representation with the matched knowledge terms in the target knowledge clause set to form a set of matching term correspondences; The attribution of hit evidence is performed based on the set of hit word correspondences. The correspondence of each hit word is recorded and associated. The hit word correspondences belonging to the same target knowledge clause are merged to form a set of hit evidence corresponding to each target knowledge clause. Based on the set of hit evidence, the hit evidence in each target knowledge clause is organized in layers according to clause-level content and element-level content. The hit evidence is mapped and associated with the corresponding rule constraints, parameter verification items and review requirements to form clause matching results. The correspondence between target knowledge clauses, hit evidence, rule constraints and parameter verification items is established to form a rule constraint mapping structure. Based on the clause matching results and rule constraint mapping structure, the clause matching results belonging to the same review object or the same review matter are merged and organized, and the corresponding target knowledge clause content, hit evidence content, rule constraint item content and parameter verification item content are uniformly organized to form a set of review basis; Based on the review criteria set, clause comparison and compliance verification are performed. The structured design data in the power engineering design documents to be reviewed is compared item by item with the target knowledge clauses, rule constraints, and parameter verification items in the review criteria set. Simultaneously, compliance judgment, missing item judgment, conflict item judgment, and anomaly item judgment are performed to form a verification result. Based on the verification result, the review result is output. The review result includes the target knowledge clause matching result, rule constraint verification result, and review opinion result. The review opinion result includes the clause identification information found in the power knowledge base retrieval, the evidence fragment information found in the retrieval, the semantic matching relationship information between query terms and knowledge clauses, the clause source identification information, and the compliance explanation information generated based on the knowledge clause matching result.

[0030] This invention performs hit evidence attribution, clause matching result construction, and rule constraint mapping on the target knowledge clause set. It structurally associates the semantic matching relationship between query terms and knowledge clauses, forming a traceable set of review basis. Based on this, it systematically compares and verifies the compliance of structured design data, knowledge clause content, rule constraint items, and parameter verification items in power engineering design documents. This enables the power knowledge base retrieval results to directly support rule judgment and review opinion generation in the design review process, thereby improving the accuracy of semantic matching between power engineering design documents and knowledge clauses and the interpretability of retrieval results. At the same time, it improves the automation, review efficiency, and result consistency of the review work, and reduces the risk of incorrect and missed reviews in manual review.

[0031] Example 1: To verify the feasibility of this invention in practice, it was applied to the distribution network engineering design review business scenario of a power design company. Deployed within the company's integrated project intelligent management platform, it serves as the core processing module for knowledge retrieval and intelligent verification. In its daily review work, the company needs to examine a large number of feasibility study documents, preliminary design documents, and supporting drawings and tables. The review objects include not only the power supply scheme, equipment selection, parameter configuration, and project boundaries in the text descriptions, but also the wiring relationships, equipment annotations, drawing parameters, and equipment lists, capacity data, and constraints in the tables. Traditionally, reviewers rely on manually consulting design specifications, industry standards, technical guidelines, review rules, and historical cases, comparing each item between the design documents and the knowledge clauses. This approach has obvious shortcomings. On the one hand, the sources of knowledge are scattered and the number of relevant clauses is huge, so reviewers often need to search back and forth between multiple standard texts, which is time-consuming. On the other hand, the expression in the design documents is not completely consistent with the wording of the standard clauses, and there are differences in terminology, abbreviations, and scattered information, which can easily lead to the inaccurate location of relevant clauses, resulting in mis-review, omissions, or incomplete citation of review basis.

[0032] In this embodiment, the power design specifications, industry standards, technical guidelines, review rules, and historical review cases accumulated by the enterprise over a long period of time are first uniformly collected to form power review knowledge data. The knowledge data is then processed to build a knowledge base. During the processing, knowledge entries with inconsistent formats, terminology, numbering, and duplicate content are cleaned and standardized. Standardized knowledge entries are formed by segmenting them according to rule units and review constraint boundaries. Furthermore, a two-level knowledge unit structure between clause-level knowledge units and element-level knowledge units is established. In this way, the knowledge text that originally required manual page-by-page searching is organized into a structured knowledge base that can directly participate in semantic encoding and matching, providing a foundation for subsequent retrieval.

[0033] Subsequently, a batch of power distribution network engineering design review documents were imported into the system. This batch of design documents included text files, drawing files, and table files. The system first uniformly received, classified, and parsed the design documents. It extracted text and split paragraphs from the text content, identified equipment identifiers, connection relationships, drawing parameters, and constraint annotations in the drawings, and extracted equipment names, models, parameter values, quantities, and related fields from the tables. The parsing results were then uniformly converted into structured design data. Then, based on the structured design data, a query term sequence was constructed. The review object, attribute description, parameter conditions, and source location were reorganized into an input format more suitable for semantic retrieval. Unlike the traditional method of directly retrieving raw sentences, this method can reduce the interference of invalid descriptions and background expressions on the search results, making the design content more accurately mapped to knowledge base clauses.

[0034] During the retrieval phase, the system uses a deep learning semantic encoding model to encode the query term sequence and the two-layer knowledge unit structure, respectively, to obtain query-side term vector representations and knowledge-side term vector representations, and further constructs a knowledge retrieval index. Based on this, the ColBERT fine-grained delayed interaction semantic matching algorithm is used to calculate the term-level relevance of the query-side and knowledge-side term vector representations. This algorithm does not simply compare the overall similarity of the entire text, but performs fine-grained matching of query terms and knowledge terms one by one. It also incorporates review role constraints, the two-layer knowledge unit structure, and an evidence consistency aggregation mechanism, performing delayed interaction at both the clause and element levels to ultimately obtain the set of target knowledge clauses most relevant to the current design content. Subsequently, the system performs hit evidence attribution on the target knowledge clause set, associating the participating query terms, knowledge clauses, and corresponding rule constraints to form clause matching results, rule constraint mapping structures, and a set of review basis. This completes the determination of the target knowledge clause set, hit evidence attribution, clause matching result construction, and rule constraint mapping. After the review basis set is generated, the review basis set, target knowledge clause matching results, rule constraint verification results, and structured design data are input into the large language model processing module. This processing module is pre-configured with a retrieval enhancement generation process adapted to power engineering review business. After receiving the review basis set, it first performs retrieval enhancement context organization on the target knowledge clause content, rule constraint item content, parameter verification item content, and hit evidence content in the review basis set. Then, the organized context content, along with the corresponding structured design data, clause matching results, and rule constraint verification results, is input into the large language model to generate review opinion content and review report content for review business.

[0035] To verify the effectiveness of this invention, design review documents were selected as test objects from the same batch of business. The method of this invention was compared with the traditional manual review method and the ordinary vector retrieval method. The traditional manual review was completed independently by three reviewers with many years of experience. The ordinary vector retrieval method used a general text vector retrieval model for clause recall. The specific comparison data is shown in Table 1: Table 1 Performance Comparison Statistics in Power Engineering Design Document Review Tasks

[0036] As shown in Table 1, when handling power engineering design review tasks of the same scale, the method of this invention significantly outperforms traditional manual review methods and ordinary vector retrieval-assisted methods in both review efficiency and retrieval accuracy. Under test data consisting of 120 design documents, 360 drawings, and 240 tables, the method of this invention completed the review of all documents in an average of 5.8 hours, a significant reduction compared to the 48.6 hours of traditional manual review. Simultaneously, the average review time per document decreased from 24.3 minutes to 2.5 minutes. Regarding retrieval accuracy, the method of this invention achieved a 93.8% correct hit rate for knowledge clauses, significantly higher than the 81.3% of the ordinary vector retrieval method, while reducing the non-compliance omission rate to 4.1%. Furthermore, the method of this invention provides an average of 3.4 valid supporting clauses for each review question, higher than the 1.8 of the traditional method and the 2.1 of the ordinary vector retrieval method. This indicates that the invention not only improves the retrieval efficiency of the power knowledge base but also enhances the completeness and interpretability of the review basis, thereby effectively improving the automation and intelligence level of the power engineering design review process.

[0037] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based AI power knowledge base retrieval method, characterized in that, Includes the following steps: Acquire power review knowledge data, process the power review knowledge data to construct a knowledge base, and obtain the power knowledge base; Based on the power knowledge base, clause segmentation and constraint element extraction are performed to construct a two-layer knowledge unit structure including clause level and element level; Obtain the power engineering design documents to be reviewed, parse the power engineering design documents, and obtain structured design data; Construct query term sequences based on structured design data; A deep learning semantic coding model is used to encode the query word sequence and the two-layer knowledge unit structure respectively, so as to obtain the query word vector representation and the knowledge word vector representation, and a knowledge retrieval index is constructed based on the knowledge word vector representation. Based on the knowledge retrieval index, the ColBERT fine-grained delayed interaction semantic matching algorithm is used to calculate the word-level relevance between the query word vector representation and the knowledge word vector representation, and to perform grouped delayed interaction and matching score aggregation to obtain the target knowledge clause set. The system performs hit evidence attribution on the target knowledge clause set, constructs a clause matching result and rule constraint mapping structure, generates a review basis set, and performs clause comparison and compliance verification based on the review basis set, outputting the review results.

2. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The acquisition of the power knowledge base specifically includes: Acquire and aggregate power review knowledge data, and uniformly receive and collect it according to knowledge source category, business application category, and review scenario category to form an original knowledge data set; The original knowledge data set is preprocessed, and the knowledge content is segmented and organized according to the corresponding rule units to form a standardized set of knowledge items with clause identifiers, source identifiers, hierarchical identifiers and rule association identifiers; A knowledge base is constructed based on a standardized set of knowledge items to form a power knowledge base oriented towards power engineering design review scenarios.

3. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The construction of the two-layer knowledge unit structure is based on: The knowledge entries in the power knowledge base are segmented according to preset segmentation rules to form different independent and semantically complete clause-level knowledge units, and all clause-level knowledge units are summarized to form a clause-level knowledge unit set. Based on the set of clause-level knowledge units, constraint elements are extracted for each clause-level knowledge unit. The corresponding constraint elements of each clause-level knowledge unit are identified one by one, and element-level knowledge units corresponding to the clause-level knowledge units are generated, forming a set of element-level knowledge units that correspond one-to-one with each clause-level knowledge unit. Based on the set of knowledge units at the clause level and the set of knowledge units at the element level, the association between each knowledge unit at the clause level and the corresponding knowledge unit at the element level is established. At the same time, the association is organized and stored in a unified manner, thus constructing a two-layer knowledge unit structure including the clause level and the element level.

4. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The acquisition of the structured design data specifically includes: The power engineering design documents to be reviewed are obtained, and the power engineering design documents are received, registered, verified for integrity and identified in terms of document type, forming a set of original design documents to be parsed. Based on the original design file set, text content reading, paragraph content segmentation, and semantic fragment extraction are performed on text-based design files; drawing content reading, drawing annotation recognition, equipment identification extraction, connection relationship extraction, and parameter label extraction are performed on drawing-based design files; and table area recognition, table header content recognition, field content extraction, and parameter item content extraction are performed on table-based design files. The text content fragments, drawing content fragments, and table content fragments obtained from parsing various types of design files are then uniformly summarized to form a set of design content parsing results. Based on the set of design content analysis results, the content fragments are structured, and content fragments belonging to the same review object or the same technical constraint are merged and organized to form structured design data.

5. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The construction of the query term sequence specifically includes: Each field record in the structured design data is processed to identify retrieval requirements. The content of the review object, attribute description, parameter description and source location related to the knowledge clause are extracted from each field record. The extracted content is then denoised, standardized and consistent to form a set of query elements. Based on the set of query elements, query elements belonging to the same review object or the same technical constraint are clustered and merged. Review intent reconstruction processing is performed on the clustered and merged query elements. Each group of query elements is arranged and organized in the order of review object priority, attribute description follow, parameter description supplement and source location association, forming query term fragments corresponding to each group of query elements. Based on query term fragments, sequential concatenation, duplicate term removal, and term boundary marking are performed on each query term fragment to generate a query term sequence corresponding to the structured design data.

6. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The acquisition of the knowledge retrieval index specifically includes: Obtain the query term sequence and two-layer knowledge unit structure, and use the clause-level knowledge unit and element-level knowledge unit as knowledge-end input data to construct a deep learning semantic coding model; The query term sequence is input into the deep learning semantic encoding model. Each query term in the query term sequence is sequentially processed by term embedding, position encoding, and context semantic encoding. Review role constraint injection and sparse term constraint processing are performed on each query term. The sparse term constraint processing activates and enhances professional term terms related to power design specifications, review rules, and engineering constraints, and sparsely suppresses ordinary descriptive terms that are not critical to review. The query term vector representation is output. The two-layer knowledge unit structure is input into the deep learning semantic coding model. The term-level knowledge units and element-level knowledge units in the two-layer knowledge unit structure are processed by word embedding, position encoding and context semantic encoding respectively. The hierarchical association injection and element affiliation preservation processing are performed on the knowledge words. The hierarchical semantic association information is jointly encoded and further residual compression processing is performed to compress and store the knowledge word vector representation. While maintaining the semantic distinguishability and hierarchical association relationship, the storage overhead of the knowledge word vector representation is reduced, forming a knowledge word vector set. Based on the knowledge term vector set, clause-level index items are established for knowledge term vector representations belonging to the same clause-level knowledge unit, and element-level index items are established for knowledge term vector representations belonging to the corresponding element-level knowledge unit. A hierarchical association index item is also established between the clause-level index items and the element-level index items. At the same time, a central representation is generated for the knowledge term vector representations under the same clause-level knowledge unit, and a central interactive pre-screening process is performed based on the central representation to form a knowledge retrieval index.

7. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The specific process of obtaining the target knowledge clause set includes: Based on the knowledge retrieval index, candidate knowledge unit location processing is performed. The knowledge retrieval index is jointly searched to determine the candidate clause-level knowledge units and candidate element-level knowledge units associated with the query term vector representation. The term vector representations at the knowledge end are then associated, extracted, and grouped to form a set of candidate knowledge end term vectors. The ColBERT fine-grained delayed interactive semantic matching algorithm is used to perform word-level relevance calculation on the query word vector representation and the candidate knowledge word vector set. Review role constraints are introduced. Each query word vector in the query word vector representation is constrained according to the review role. Only query word vectors with corresponding review role relationships are allowed to participate in the relevance calculation with the corresponding knowledge word vectors in the candidate knowledge word vector set. The result is a word-level relevance result set containing review role constraint information. Based on the word-level relevance result set, a grouped delayed interaction processing is performed. The query word vector representation is divided into different query word groups according to the review role. The candidate knowledge word vector set is divided into different knowledge word groups according to the clause-level affiliation and the element-level affiliation. For each query word group, delayed interaction matching is performed with the corresponding clause-level knowledge word group and element-level knowledge word group. Within each group, the knowledge word with the highest relevance is retained as the intra-group matching result of the current query word. The intra-group matching results of the clause-level knowledge word group and the element-level knowledge word group are collaboratively associated to form a grouped delayed interaction result that reflects both clause-level matching strength and element-level matching strength. Based on the grouped delayed interaction results, the matching score aggregation process is performed. Evidence consistency judgment is performed on the clause-level matching results and element-level matching results corresponding to each query tuple. Enhanced aggregation is performed on the matching results with consistent pointing relationship between clause-level matching results and element-level matching results, and suppression aggregation is performed on the matching results with inconsistent relationship between clause-level matching results and element-level matching results. The comprehensive matching score corresponding to each candidate knowledge unit is obtained, and all candidate knowledge units are sorted according to the comprehensive matching score to form the candidate knowledge unit ranking result. The target knowledge clause set is determined based on the ranking results of candidate knowledge units. Candidate clause-level knowledge units and their corresponding candidate element-level knowledge units that meet the preset retention conditions in terms of comprehensive matching scores are retained and then merged and organized according to hierarchical association and evidence consistency to generate the target knowledge clause set.

8. The AI ​​power knowledge base retrieval method based on deep learning according to claim 1, characterized in that, The results of the review are obtained in the following specific ways: Extract and organize the target knowledge clause set, and associate the query terms involved in the matching in the query term vector representation with the matched knowledge terms in the target knowledge clause set to form a set of matching term correspondences; The attribution of hit evidence is performed based on the set of hit word correspondences. The correspondence of each hit word is recorded and associated. The hit word correspondences belonging to the same target knowledge clause are merged to form a set of hit evidence corresponding to each target knowledge clause. Based on the set of hit evidence, the hit evidence in each target knowledge clause is organized in layers according to clause-level content and element-level content. The hit evidence is mapped and associated with the corresponding rule constraints, parameter verification items and review requirements to form clause matching results. The correspondence between target knowledge clauses, hit evidence, rule constraints and parameter verification items is established to form a rule constraint mapping structure. Based on the clause matching results and rule constraint mapping structure, the clause matching results belonging to the same review object or the same review matter are merged and organized, and the corresponding target knowledge clause content, hit evidence content, rule constraint item content and parameter verification item content are uniformly organized to form a set of review basis; Based on the review basis set, clause comparison and compliance verification are performed. The structured design data in the power engineering design documents to be reviewed is compared item by item with the target knowledge clauses, rule constraints, and parameter verification items in the review basis set. At the same time, compliance judgment, missing item judgment, conflict item judgment, and abnormal item judgment are performed to form a verification result. The review result is output based on the verification result.

Images