Knowledge graph-based complex product design knowledge service method and system

Abstract

The application provides a kind of complex product design knowledge service method and system based on knowledge graph, comprising: through text tree structure and improved TFIDF algorithm combination N-gram strategy, reference document is semantically annotated;Knowledge carding template and rule extraction method are used to extract design process knowledge, generate knowledge triple and build knowledge graph in Neo4j graph database;On this basis, product design knowledge super network is built, and the semantic matching degree between tasks and knowledge is calculated by combining super edge correlation degree and information entropy modified vector space model;Finally, the query push of process knowledge is realized based on Cypher language, and the sorting recommendation of reference knowledge is completed by multidimensional comprehensive evaluation method.The application realizes the task-oriented knowledge structured expression, semantic correlation calculation and intelligent push in complex product design, and improves the enterprise knowledge utilization efficiency and design decision quality.

Description

Technical Field

[0001] This invention relates to the field of knowledge service technology for complex product design, and more specifically, to a knowledge graph-based method and system for providing knowledge services for complex product design. Background Technology

[0002] As a knowledge-intensive activity, the success of complex product design is closely related to the ability to effectively manage and share engineering knowledge and experience related to product design throughout the development process. Due to the characteristics of complex processes and extensive knowledge involved in complex product design, the design tasks are usually distributed to designers in the form of work packages based on the decomposition of the work structure.

[0003] While most companies developing complex products have accumulated a large amount of business data, their knowledge utilization rate is low due to a lack of analysis and processing of this vast amount of unstructured textual knowledge resources. Furthermore, the lack of organization and summarization of design process knowledge leads to unclear design processes and frequent issues such as missing steps and rework.

[0004] This invention addresses the common problem in enterprises developing complex products: the time-consuming and labor-intensive passive knowledge retrieval and the lack of proactive knowledge delivery for design tasks. It researches a task-oriented semantic matching method for complex product design knowledge and a task-oriented knowledge delivery method for complex product design, providing methodological support for the implementation of information systems that facilitate task-oriented knowledge delivery for complex product design.

[0005] Patent application CN118839606A discloses a product intelligent design method and system based on a domain knowledge-enhanced large language model, including: step S1, acquiring a design task dataset within the product design domain; step S2, obtaining a large language model based on multi-task adaptation and knowledge enhancement based on the design task dataset and product design domain knowledge; step S3, obtaining a product design generation agent based on the large language model based on multi-task adaptation and knowledge enhancement through a product design task toolset; and step S4, the product design generation agent adjusting the product design scheme in real time based on user feedback. However, this patent cannot completely solve the existing technical problems, nor can it meet the needs of this invention. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a knowledge service method and system for complex product design based on knowledge graphs.

[0007] The knowledge graph-based knowledge service method for complex product design provided by the present invention includes:

[0008] Step 1: By standardizing the document structure, introducing a text tree model, and combining the improved TFIDF algorithm with the N-gram strategy, semantic annotation and structured representation of reference knowledge are achieved.

[0009] Step 2: By designing a knowledge organization template and combining it with a rule-based knowledge extraction method, we can achieve the systematic acquisition and triple generation of key knowledge in complex product design, and finally build a task-oriented knowledge graph based on Neo4j.

[0010] Step 3: By constructing a task-oriented complex product design knowledge hypernetwork model and combining a semantic matching method with multi-window co-occurrence and information entropy correction, the correlation between tasks and reference knowledge is calculated and semantic matching is performed.

[0011] Step 4: By constructing a task-oriented design knowledge push framework, a process knowledge retrieval method based on Cypher statements, and a reference knowledge push method based on multi-dimensional comprehensive evaluation, the acquisition and push of task-related knowledge in complex product design can be realized.

[0012] Preferably, step 1 includes:

[0013] Step 1.1: Standardize and represent reference knowledge;

[0014] Define the document knowledge element (DKU) as: DKU =<DP,DA,PC,DO,OT,BT,PT,MT> Wherein, DP represents the R&D stage of the design task to which the reference document belongs, DA represents the specific professional field involved in the reference document, PC represents the specific product model to which the reference document belongs, DO represents the specific design object involved in the reference document, OT represents the conceptual terms related to the design object in the reference document, BT represents the operational or behavioral terms related to the design task in the reference document, PT represents the design parameters or design features related to the design object and design task in the reference document, and MT represents the method or tool terms used to support the design task in the reference document.

[0015] Step 1.2: Design the semantic annotation process for reference knowledge;

[0016] Preprocessing and structured modeling of design documents, including document type identification, content extraction, title structure analysis, and text tree generation;

[0017] Combining word segmentation, stop word removal, and text tree analysis, we improved the TFIDF algorithm and used the N-gram strategy to extract keywords.

[0018] Based on the constructed ontology, a collection of semantic elements containing domain classes and attributes is generated;

[0019] Calculate the semantic similarity between keywords and semantic elements, complete the semantic matching and annotation of keywords, and record their correspondence in the ontology;

[0020] For unmatched keywords, their ontology dimension is determined with human assistance and semantic elements are expanded;

[0021] Step 1.3: Text content structuring based on text tree;

[0022] The document title and its direct content are used as the root node. Subheadings and their direct content at all levels form child nodes according to their heading level. The lowest level headings and their direct content form leaf nodes.

[0023] Direct content includes text paragraphs, table headings, and image titles;

[0024] The text tree of document d is d= <N0,N 1,1 ,…,N p,q ,…>, where N p,q This represents the q-th node of the p-th level subheading, and N p,q =(NT) p,q NC p,q ), where NT p,q Representative node N p,q Title information, NC p,q Representative node N p,q The direct content information, N0 is the initial node;

[0025] This is a Python-based regular expression toolkit that generates text trees using rule matching.

[0026] Step 1.4: Keyword extraction based on text tree and improved TFIDF;

[0027] Natural language processing tools are used for word segmentation and part-of-speech tagging, and a domain dictionary is loaded and stop words are removed;

[0028] Introducing an external corpus to calculate improved inverse document frequencies:

[0029]

[0030] Among them, S G and S E These represent a document set and an external corpus set, respectively, in the field of complex product design. G ∪S E | represents the total number of documents in the document set and external corpus set in the field of complex product design, |w i ∈S G ∪S E | This collection of documents and external corpora representing the field of complex product design contains the term w i The number of documents;

[0031] The formula for calculating word frequency is:

[0032]

[0033] Among them, |w i :w i ∈d and d∈S G |Representative term w i The frequency of occurrence of document d in the field of complex product design, |d:d∈S G | Represents the total number of words in document d of the domain;

[0034] The weighting formula based on title position is:

[0035]

[0036] In the formula, |w i :w i ∈NT pq |Representative term w i Appears in the text tree node title NT pq The number of times in, n p Let len(NT) be the number of nodes in the p-th level text tree. pq ) represents the title of a text tree node. pq The length, λ p This represents the weight of the p-th level text tree node, where m is the number of levels in the text tree.

[0037] The weight correction formula based on special locations is:

[0038]

[0039] In the formula, λ special For special position weighting coefficients, |w i :w i ∈NC pq |Representative term w i NC appears in the text tree node content pq The number of times in, len(NC) pq ) represents the content of a text tree node (NC) pq Length, S H It is a set of title trigger words and their synonyms, I(NT) pq ∈S H ) represents the title NT pq Is it included in set S? H Indicator functions within;

[0040] Introducing term w i Professional term weighting coefficient The calculation formula is as follows:

[0041]

[0042] Among them, S S For the domain dictionary set, γ S Weighting coefficients for domain-specific terms;

[0043] The comprehensive formula for calculating keyword weight is:

[0044]

[0045] in, It is the term w i Part-of-speech weighting factor;

[0046] The tanh function is used to normalize the importance weights of candidate keywords to achieve a mapping of weights to the interval (0,1). The calculation formula is as follows:

[0047]

[0048] By combining the above equations, we obtain the weight values ​​of all terms in document d, sort them from highest to lowest, and extract candidate words based on the set threshold to obtain the candidate keyword set dd. shortlist =(w1,w2,…,w r ) and their weights The threshold adopts a dynamic change strategy, and the expression is:

[0049]

[0050] In the formula, n is the total number of words in document d. Let be the weight value of the i-th term;

[0051] Step 1.5: Semantic element matching and semantic annotation;

[0052] Generate a set of semantic elements for design tasks, design objects, design phases, professional fields, and contextual terms;

[0053] Calculate candidate keywords w i and ontology instance I j The similarity is calculated using the following formula:

[0054]

[0055] Among them, P(w i ,I j ) is a candidate keyword obtained based on the pre-trained word vector model Word2Vec. i and ontology instance I j Similarity between them;

[0056] The similarity between two word vectors is calculated using cosine similarity, expressed as follows:

[0057]

[0058] Where W is the Word2Vec Chinese pre-trained model;

[0059] If semantic similarity Sim ij Greater than the matching threshold γ match In this case, semantic element I is used. j Replacement candidate keyword w i And perform semantic annotation of the text based on the type of semantic elements;

[0060] If the semantic similarity is less than the matching threshold γ match But it is greater than the threshold γ for new elements. new In such cases, domain experts will make the judgment and add new elements.

[0061] Preferably, step 2 includes:

[0062] Step 2.1: By introducing the IDEF0 functional modeling method, a structured analysis of complex product design tasks is conducted to identify their functional modules and related information, and then a knowledge organization template containing four categories of content, including basic attributes, control elements, process elements and resource elements, is designed.

[0063] Step 2.2: Extract process knowledge related to the design task using a method based on preset rules;

[0064] Step 2.3: Perform data layer fusion on the extracted knowledge;

[0065] Step 2.4: Import the knowledge triples in batches into Neo4j for node and relationship modeling and storage.

[0066] Preferably, step 3 includes:

[0067] Step 3.1: Define the task-oriented complex product design knowledge hypernetwork as H = (V, E), where V is the set of subnetwork nodes and E is the set of hyperedges; the subnetwork nodes are derived from the knowledge graph, and the relevance is calculated on the domain corpus;

[0068] Step 3.2: Obtain the set of hyperedge feature terms for design task T1 and reference document K2; calculate the semantic relevance of design task T1 and reference document K2 from the 8 dimensions of DKU based on subnet relevance and hyperedge relevance, and obtain the relevance vector Sim(T1,K2);

[0069] The semantic space vector for constructing the design task is VT1, and the semantic space vector for the reference document is VK2;

[0070] Step 3.3: Calculate semantic matching similarity using weighted Euclidean distance. The normalized calculation formula is as follows:

[0071]

[0072] In the formula, ω r_coe To determine the knowledge correlation between the final design task T1 and reference document K2, α k ω is the feature term class weight correction coefficient. 1k ω 2k These represent the feature class weights for the design task and the reference document, respectively.

[0073] Preferably, step 4 includes:

[0074] Step 4.1: Based on the WBS decomposition process, after parsing the task features, generate Cypher statement retrieval process knowledge, and combine hypernetwork and reconstructed VSM matching reference knowledge;

[0075] Step 4.2: Cypher-based process knowledge push: Levenshtein distance and cosine similarity are used to locate graph task nodes, retrieve relevant process knowledge and push it;

[0076] Step 4.3: Multidimensional comprehensive evaluation;

[0077] Knowledge Importance ω i_coe calculate:

[0078]

[0079] in, D, representing the designers involved in this design task j Regarding the evaluation of the importance of this reference knowledge Representative designer D j Importance evaluation weight;

[0080] Knowledge Support ω s_coe calculate:

[0081]

[0082] in, D, representing the designers involved in this design task j Evaluation of the support level for this reference knowledge Representative designer D j Support evaluation weights;

[0083] Knowledge recommendation rating ω final calculate:

[0084] ω final =γ r_coe·ω r_coe +γ i_coe ·ω i_coe +γ s_coe ·ω s_coe

[0085] In the formula, γ r_coe γ i_coe γ s_coe These are the empirical weight hyperparameters for the relevance of design reference knowledge, the importance of knowledge, and the support of knowledge, respectively, and the sum of the three is 1.

[0086] The knowledge graph-based complex product design knowledge service system provided by the present invention includes:

[0087] Module M1: By standardizing the document structure, introducing a text tree model, and combining an improved TFIDF algorithm with an N-gram strategy, semantic annotation and structured representation of reference knowledge are achieved.

[0088] Module M2: By designing knowledge organization templates and combining them with rule-based knowledge extraction methods, it realizes the systematic acquisition and triple generation of key knowledge in complex product design, and finally builds a task-oriented knowledge graph based on Neo4j.

[0089] Module M3: By constructing a task-oriented complex product design knowledge hypernetwork model and combining a semantic matching method with multi-window co-occurrence and information entropy correction, it realizes the calculation and semantic matching of the correlation between tasks and reference knowledge.

[0090] Module M4: By constructing a task-oriented design knowledge push framework, a process knowledge retrieval method based on Cypher statements, and a reference knowledge push method based on multi-dimensional comprehensive evaluation, it realizes the acquisition and push of task-related knowledge in complex product design.

[0091] Preferably, the module M1 includes:

[0092] Module M1.1: Standardized representation of reference knowledge;

[0093] Define the document knowledge element (DKU) as: DKU =<DP,DA,PC,DO,OT,BT,PT,MT> Wherein, DP represents the R&D stage of the design task to which the reference document belongs, DA represents the specific professional field involved in the reference document, PC represents the specific product model to which the reference document belongs, DO represents the specific design object involved in the reference document, OT represents the conceptual terms related to the design object in the reference document, BT represents the operational or behavioral terms related to the design task in the reference document, PT represents the design parameters or design features related to the design object and design task in the reference document, and MT represents the method or tool terms used to support the design task in the reference document.

[0094] Module M1.2: Design reference knowledge semantic annotation process;

[0095] Preprocessing and structured modeling of design documents, including document type identification, content extraction, title structure analysis, and text tree generation;

[0096] Combining word segmentation, stop word removal, and text tree analysis, we improved the TFIDF algorithm and used the N-gram strategy to extract keywords.

[0097] Based on the constructed ontology, a collection of semantic elements containing domain classes and attributes is generated;

[0098] Calculate the semantic similarity between keywords and semantic elements, complete the semantic matching and annotation of keywords, and record their correspondence in the ontology;

[0099] For unmatched keywords, their ontology dimension is determined with human assistance and semantic elements are expanded;

[0100] Module M1.3: Text content structuring based on text tree;

[0101] The document title and its direct content are used as the root node. Subheadings and their direct content at all levels form child nodes according to their heading level. The lowest level headings and their direct content form leaf nodes.

[0102] Direct content includes text paragraphs, table headings, and image titles;

[0103] The text tree of document d is d= <N0,N 1,1 ,…,N p,q ,…>, where N p,q This represents the q-th node of the p-th level subheading, and N p,q =(NT) p,q NC p,q ), where NT p,q Representative node N p,q Title information, NC p,q Representative node N p,q The direct content information, N0 is the initial node;

[0104] This is a Python-based regular expression toolkit that generates text trees using rule matching.

[0105] Module M1.4: Text Tree-Based Improved TFIDF Keyword Extraction;

[0106] Natural language processing tools are used for word segmentation and part-of-speech tagging, and a domain dictionary is loaded and stop words are removed;

[0107] Introducing an external corpus to calculate improved inverse document frequencies:

[0108]

[0109] Among them, S G and S E These represent a document set and an external corpus set, respectively, in the field of complex product design. G ∪S E | represents the total number of documents in the document set and external corpus set in the field of complex product design, |w i ∈S G ∪S E | This collection of documents and external corpora representing the field of complex product design contains the term w i The number of documents;

[0110] The formula for calculating word frequency is:

[0111]

[0112] Among them, |w i :w i ∈d and d∈S G |Representative term w i The frequency of occurrence of document d in the field of complex product design, |d:d∈S G | Represents the total number of words in document d of the domain;

[0113] The weighting formula based on title position is:

[0114]

[0115] In the formula, |w i :w i ∈NT pq |Representative term w i Appears in the text tree node title NT pq The number of times in, n p Let len(NT) be the number of nodes in the p-th level text tree. pq ) represents the title of a text tree node. pq The length, λ p This represents the weight of the p-th level text tree node, where m is the number of levels in the text tree.

[0116] The weight correction formula based on special locations is:

[0117]

[0118] In the formula, λ special For special position weighting coefficients, |w i :w i ∈NC pq |Representative term w iNC appears in the text tree node content pq The number of times in, len(NC) pq ) represents the content of a text tree node (NC) pq Length, S H It is a set of title trigger words and their synonyms, I(NT) pq ∈S H ) represents the title NT pq Is it included in set S? H Indicator functions within;

[0119] Introducing term w i Professional term weighting coefficient The calculation formula is as follows:

[0120]

[0121] Among them, S S For the domain dictionary set, γ S Weighting coefficients for domain-specific terms;

[0122] The comprehensive formula for calculating keyword weight is:

[0123]

[0124] in, It is the term w i Part-of-speech weighting factor;

[0125] The tanh function is used to normalize the importance weights of candidate keywords to achieve a mapping of weights to the interval (0,1). The calculation formula is as follows:

[0126]

[0127] By combining the above equations, we obtain the weight values ​​of all terms in document d, sort them from highest to lowest, and extract candidate words based on the set threshold to obtain the candidate keyword set dd. shortlist =(w1,w2,…,w r ) and their weights The threshold adopts a dynamic change strategy, and the expression is:

[0128]

[0129] In the formula, n is the total number of words in document d. Let be the weight value of the i-th term;

[0130] Module M1.5: Semantic element matching and semantic annotation;

[0131] Generate a set of semantic elements for design tasks, design objects, design phases, professional fields, and contextual terms;

[0132] Calculate candidate keywords w i and ontology instance I j The similarity is calculated using the following formula:

[0133]

[0134] Among them, P(w i ,I j ) is a candidate keyword obtained based on the pre-trained word vector model Word2Vec. i and ontology instance I j Similarity between them;

[0135] The similarity between two word vectors is calculated using cosine similarity, expressed as follows:

[0136]

[0137] Where W is the Word2Vec Chinese pre-trained model;

[0138] If semantic similarity Sim ij Greater than the matching threshold γ match In this case, semantic element I is used. j Replacement candidate keyword w i And perform semantic annotation of the text based on the type of semantic elements;

[0139] If the semantic similarity is less than the matching threshold γ match But it is greater than the threshold γ for new elements. new In such cases, domain experts will make the judgment and add new elements.

[0140] Preferably, the module M2 includes:

[0141] Module M2.1: By introducing the IDEF0 functional modeling method, a structured analysis of complex product design tasks is conducted to identify their functional modules and related information, and then a knowledge organization template containing four categories of content: basic attributes, control elements, process elements and resource elements is designed.

[0142] Module M2.2: Employs a method based on preset rules to extract process knowledge related to the design task;

[0143] Module M2.3: Performs data layer fusion on the extracted knowledge;

[0144] Module M2.4: Imports knowledge triples in batches into Neo4j for node and relation modeling and storage.

[0145] Preferably, the module M3 includes:

[0146] Module M3.1: Defines a task-oriented knowledge hypernetwork for complex product design as H = (V, E), where V is the set of subnetwork nodes and E is the set of hyperedges; the subnetwork nodes are derived from the knowledge graph, and the relevance is calculated on the domain corpus;

[0147] Module M3.2: Obtain the set of hyperedge feature terms for design task T1 and reference document K2; calculate the semantic relevance of design task T1 and reference document K2 from the 8 dimensions of DKU based on subnet relevance and hyperedge relevance, and obtain the relevance vector Sim(T1,K2);

[0148] The semantic space vector for constructing the design task is VT1, and the semantic space vector for the reference document is VK2;

[0149] Module M3.3: Uses weighted Euclidean distance to calculate semantic matching similarity. The normalized calculation formula is as follows:

[0150]

[0151] In the formula, ω r_coe To determine the knowledge correlation between the final design task T1 and reference document K2, α k ω is the feature term class weight correction coefficient. 1k ω 2k These represent the feature class weights for the design task and the reference document, respectively.

[0152] Preferably, the module M4 includes:

[0153] Module M4.1: Based on the WBS decomposition process, it generates Cypher statement retrieval process knowledge after parsing task features, and combines hypernetwork and reconstructed VSM matching reference knowledge;

[0154] Module M4.2: Cypher-based process knowledge push: Levenshtein distance and cosine similarity are used to locate graph task nodes, retrieve relevant process knowledge and push it;

[0155] Module M4.3: Multidimensional Comprehensive Evaluation;

[0156] Knowledge Importance ω i_coe calculate:

[0157]

[0158] in, D, representing the designers involved in this design task j Regarding the evaluation of the importance of this reference knowledge Representative designer D j Importance evaluation weight;

[0159] Knowledge Support ω s_coe calculate:

[0160]

[0161] in, D, representing the designers involved in this design task j Evaluation of the support level for this reference knowledge Representative designer D j Support evaluation weights;

[0162] Knowledge recommendation rating ω final calculate:

[0163] ω final =γ r_coe ·ω r_coe +γ i_coe ·ω i_coe +γ s_coe ·ω s_coe

[0164] In the formula, γ r_coe γ i_coe γ s_coe These are the empirical weight hyperparameters for the relevance of design reference knowledge, the importance of knowledge, and the support of knowledge, respectively, and the sum of the three is 1.

[0165] Compared with the prior art, the present invention has the following beneficial effects:

[0166] (1) The task-oriented design reference knowledge semantic annotation method proposed in this invention comprehensively considers text features such as text structure features, content meaning, and text length in the keyword extraction stage, improves the TFIDF algorithm, and introduces the N-gram strategy to generate keywords. It effectively solves the problems of dataset skewness and fragmentation of professional vocabulary in semantic annotation in the field of product design, and realizes semantic annotation of a large number of unstructured document resources in the field.

[0167] (2) The task-oriented complex product design knowledge hypernetwork model proposed in this invention realizes the modeling and quantitative description of complex network relationships between task-oriented knowledge elements from the aspects of design task attributes and design task content context. In specific details, it proposes improvement measures such as the correlation calculation method based on multi-window co-occurrence degree and the term importance correction based on information entropy, which effectively improves the correlation measurement between task-related knowledge elements.

[0168] (3) The semantic matching method proposed in this invention combines knowledge hypernetwork features and vector space model theory, and takes into account the statistical features of prior knowledge in the engineering field and large-scale domain knowledge resources, thus achieving more accurate engineering semantic matching. Attached Figure Description

[0169] Other features, objects, and advantages of the present invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0170] Figure 1 For the text tree generation process;

[0171] Figure 2 Design a knowledge graph construction process for complex, task-oriented products;

[0172] Figure 3 Design a knowledge hypernetwork construction process for complex, task-oriented products;

[0173] Figure 4 Design a knowledge delivery framework for complex, task-oriented products. Detailed Implementation

[0174] The present invention will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present invention, but do not limit the invention in any way. It should be noted that those skilled in the art can make several changes and improvements without departing from the concept of the present invention. These all fall within the scope of protection of the present invention.

[0175] Example 1

[0176] This invention provides a knowledge service method for complex product design based on knowledge graphs, comprising the following steps:

[0177] Step 1: By standardizing the document structure, introducing a text tree model, and combining the improved TFIDF algorithm with the N-gram strategy, efficient semantic annotation and structured representation of reference knowledge are achieved.

[0178] Step 2: By designing a knowledge organization template and combining it with a rule-based knowledge extraction method, we can achieve the systematic acquisition and triple generation of key knowledge in complex product design, and finally build a task-oriented knowledge graph based on Neo4j.

[0179] Step 3: By constructing a task-oriented complex product design knowledge hypernetwork model, and proposing a semantic matching method that combines multi-window co-occurrence and information entropy correction, the accurate calculation and semantic matching of the correlation between tasks and reference knowledge are achieved.

[0180] Step 4: By constructing a task-oriented design knowledge push framework, we propose a process knowledge retrieval method based on Cypher statements and a reference knowledge push method based on multi-dimensional comprehensive evaluation, so as to realize the intelligent acquisition and accurate push of task-related knowledge in complex product design.

[0181] Step 1 includes the following steps:

[0182] (1) Reference knowledge standardization representation and semantic annotation process

[0183] Step 1.1.1: Task-Oriented Standardized Representation of Reference Knowledge

[0184] Document Knowledge Units (DKUs) are defined to achieve standardized representation of knowledge resources such as various complex product design documents. A design document knowledge unit can be represented as DKU =<DP,DA,PC,DO,OT,BT,PT,MT> In this document, element DP represents the R&D stage of the design task to which the reference document belongs; element DA represents the specific professional field involved in the reference document; element PC represents the specific product model to which the reference document belongs; element DO represents the specific design object involved in the reference document; element OT represents the specific conceptual terms related to the design object involved in the reference document; element BT represents the specific operational or behavioral terms related to the design task involved in the reference document; element PT represents the specific design parameters or design features related to the design object and design task involved in the reference document; and element MT represents the relevant methods or tools used to support the design task involved in the reference document.

[0185] Step 1.1.2: Task-Oriented Design Reference Knowledge Semantic Annotation Process

[0186] First, the design document is preprocessed and structurally modeled through document type identification, content extraction, title structure analysis, and text tree generation. Second, keywords are extracted using an N-gram strategy, combining word segmentation, stop word removal, and a text tree-improved TFIDF algorithm. Third, based on the constructed ontology, a set of semantic elements containing domain classes, attributes, and other elements is parsed and generated. Fourth, semantic similarity between keywords and semantic elements is calculated to complete semantic matching and annotation of keywords, and their correspondence in the ontology is recorded. Finally, for keywords that fail to match, their ontology dimensions are determined with manual assistance, and semantic elements are expanded, thereby achieving continuous improvement of the ontology and providing semantic foundation support for knowledge graph construction.

[0187] Based on this process, specific document knowledge resources are semantically annotated in eight dimensions, such as product model and design stage, in DKU, providing a semantic annotation foundation for task-oriented knowledge services.

[0188] (2) Text content structuring based on text tree

[0189] Before structuring the text, text extraction and noise reduction are performed first. Based on the acquired document data resources, various open-source tools or commercial software can be used to extract the text from the documents. A text tree structure is constructed using the document title and its direct content as the root node, with each level of subtitles and their direct content forming sub-nodes according to their title level, and the bottom-level titles and their direct content forming leaf nodes.

[0190] Direct content refers to text content contained below the current heading and above the next heading, including three types of information: text paragraphs, table headings, and image titles. Let the text tree of document d be d= <N0,N 1,1 ,…,N p,q ,…>, where N p,q This represents the q-th node of the p-th level subheading, and N p,q =(NT) p,q NC p,q ), where NT p,q Representative node N p,q Title information, NC p,q Representative node N p,q Directly subordinate content information.

[0191] Finally, based on Python's regular expression toolkit, a text tree is generated using rule matching, such as... Figure 1 .

[0192] (3) Keyword extraction based on text tree-improved TFIDF

[0193] Step 1.3.1: Word segmentation, part-of-speech tagging, and stop word removal

[0194] First, the preprocessed document needs to undergo word segmentation and stop word removal. After comprehensively comparing the segmentation effects of various commonly used Chinese word segmentation tools, this invention uses the PKUSEG natural language processing tool for text segmentation and part-of-speech tagging, and improves the accuracy of domain-specific word segmentation by loading a user dictionary (a dictionary for complex product design). Based on this, regular expressions are used to write matching rules to remove stop words from a custom stop word dictionary.

[0195] Step 1.3.2: Calculation of term frequency and inverse document frequency

[0196] The TFIDF algorithm suffers from problems such as dataset skewness and inter-class / intra-class distribution bias. To address these issues, this invention introduces an external corpus to mitigate the unreasonable calculation of IDF. The improved IDF calculation formula is as follows:

[0197]

[0198] Among them, S G and SE These represent a document set and an external corpus set, respectively, in the field of complex product design. G ∪S E | represents the total number of documents in the document set and external corpus set in the field of complex product design, |w i ∈S G ∪S E | This collection of documents and external corpora representing the field of complex product design contains the term w i The number of documents. The formula for calculating term frequency (TF) is as follows:

[0199]

[0200] Among them, |w i :w i ∈d and d∈S G |Representative term w i The frequency of occurrence of document d in the field of complex product design, |d:d∈S G | Represents the total number of words in document d of the domain.

[0201] Step 1.3.3: Improved TFIDF weight calculation based on text tree

[0202] This invention improves the traditional TFIDF weight calculation method by introducing text tree information and domain dictionary, and by comprehensively considering information such as text length and part of speech, and extracts a weight calculation method based on text tree-improved TFIDF (T-TFIDF).

[0203] A weighting formula based on title position is proposed:

[0204]

[0205] In the formula, |w i :w i ∈NT pq |Representative term w i Appears in the text tree node title NT pq The number of times in, n p Let len(NT) be the number of nodes in the p-th level text tree. pq ) represents the title of a text tree node. pq The length, λ p represents the weight of the p-th level text tree node, and m is the number of levels in the text tree.

[0206] Weight correction formula based on specific locations:

[0207]

[0208] In the formula, λspecial For special position weighting coefficients, |w i :w i ∈NC pq |Representative term w i NC appears in the text tree node content pq The number of times in, len(NC) pq ) represents the content of a text tree node (NC) pq Length, S H It is a collection of important title trigger words such as "Preface" and "Conclusion" and their synonyms, I(NT) pq ∈S H ) represents the title NT pq Is it included in set S? H Indicator Function.

[0209] Considering the importance of different parts of speech for keyword extraction (generally, nouns and verbs are more likely to be keywords than adjectives and adverbs), the term w is introduced. i Part-of-speech weighting factor Because the keyword distribution differs across corpora from different domains, The values ​​should also vary, requiring parameter tuning based on the specific corpus of complex product design domains. However, generally speaking, the part-of-speech weights of nouns and verbs are greater than those of adjectives and adverbs. Considering the importance of specialized domain vocabulary, the term w is introduced based on a domain dictionary. i Professional term weighting coefficient The calculation formula is as follows:

[0210]

[0211] Where S S For the domain dictionary set, γ S This refers to the weighting coefficients for domain-specific terms.

[0212] The comprehensive formula for calculating keyword weight is:

[0213]

[0214] Because the introduction of weighting coefficients often results in weight ranges not satisfying [0,1], this invention uses the tanh function to normalize the importance weights of candidate keywords to achieve a mapping of weights to the interval (0,1). The specific calculation formula is as follows:

[0215]

[0216] By combining the above equations, we obtain the weight values ​​of all terms in document d, sort them from highest to lowest, and extract candidate words based on the set threshold to obtain the candidate keyword set dd. shortlist =(w1,w2,…,w r ) and their weights The threshold in this paper adopts a dynamic change strategy, namely:

[0217]

[0218] In the formula, n is the total number of words in document d. Let be the weight value of the i-th term.

[0219] Step 1.3.4: Keyword generation based on N-grams

[0220] The keyword generation algorithm based on N-gram is shown in the table below.

[0221]

[0222] Among them, prob co_word The calculation formula is as follows:

[0223]

[0224] In the formula, p(w) c ,w s ) is the term w c and w s The co-occurrence window length l in the domain corpus window The frequency of co-occurrence within, p(w s ) is the term w c Frequency of occurrence in the domain corpus.

[0225] Finally, we obtain the keyword set for document d. and their weights

[0226] (4) Semantic element matching and semantic annotation

[0227] Step 1.4.1: Generation of Domain Semantic Element Set

[0228] This invention utilizes the Apache Jena library to decompose the ontology framework and parse the ontology elements, generating a set of semantic elements for the design task. Considering that different attributes in the ontology contain different data structures such as atomic classes, Boolean expressions, constraints, and quantifiers, additional decomposition processing is required. The original ontology parsing element set can be represented as:

[0229] Class = {C1, C2, ..., C} n};

[0230] Individual = {I1, I2, ..., I} o};

[0231] ObjectProperty={OP1,OP2,…,OP p};

[0232] DataProperty={DP1,DP2,…,DP q};

[0233] Description = {D1,D2,…,D} r};

[0234] Wherein: Class, Individual, ObjectProperty, DataProperty, and Description represent the ontology class, instance, object property, data property, and description collection of the design domain, respectively. By retaining key elements such as classes and instances, a set of semantic elements representing the design task, design object, design phase, professional domain, and contextual terminology is formed.

[0235] Step 1.4.2: Semantic element matching and semantic annotation

[0236] Assume candidate keyword set d shortlist =(w1,w2,…,w r The ontology instance set for the design task is Individual = {I1, I2, ..., I...} o The elements in the candidate keyword set are compared one by one with those in the ontology instance set, and the similarity between the two elements is calculated to form an r×o similarity matrix S as follows:

[0237]

[0238] Among them, Sim ij For candidate keywords w i and ontology instance I j The similarity is calculated using the following formula:

[0239]

[0240] Among them, P(w i ,I j ) is a candidate keyword obtained based on the pre-trained word vector model Word2Vec. i and ontology instance I j The similarity between word vectors. Assuming the Word2Vec Chinese pre-trained model is W, the cosine similarity is used to calculate the similarity between two word vectors, that is:

[0241]

[0242] The semantic element corresponding to the maximum value in each row of matrix S is selected as the candidate keyword w. i Let's assume the latent semantic matching element is I. j Then when the semantic similarity between the two is Sim ij Greater than the matching threshold γ match At that time, semantic element I can be used j Replacement candidate keyword w i The text is semantically annotated based on the type of semantic elements. If the semantic similarity between the two is less than the matching threshold γ, then... match But it is greater than the threshold γ for new elements. new When this is the case, it indicates that it may be a new semantic element. Domain experts then judge and add new elements, so that the domain ontology can be continuously expanded and improved.

[0243] Assume the final semantic annotation result is d final =(w1,w2,…,w s The corresponding semantic weight vector is d. k =(k1,k2,…,k s ), whose corresponding semantic type set is d type =(type1,type2,…,type s ).

[0244] Step 2 includes the following steps:

[0245] (1) Template design for reviewing knowledge of complex product design

[0246] By introducing the IDEF0 functional modeling method, a structured analysis of complex product design tasks is conducted to identify their functional modules and related information, and then a knowledge organization template containing four categories of content: basic attributes, control elements, process elements, and resource elements is designed.

[0247] (2) Rule-based knowledge extraction for complex product design

[0248] Based on the design process knowledge template compiled by designers, a rule-based method is used to extract design task-related process knowledge, taking into account the characteristics of the knowledge table. The extraction process and rules shown in the table below can be used to complete the extraction of design process knowledge from the complex product design knowledge table.

[0249]

[0250]

[0251] (3) Knowledge triple generation and knowledge fusion

[0252] After acquiring knowledge related to complex product design, the knowledge triples generated in the above steps are fused at the data layer to obtain the final data used to construct a task-oriented knowledge graph for complex product design.

[0253] (4) Knowledge graph construction based on Neo4j

[0254] Using extracted knowledge triples as foundational data, and employing the Neo4j graph database, various knowledge elements in the design process are modeled, visualized, and stored in a visual manner. Simultaneously, tools such as neo4j-admin import are used to achieve batch data import, enabling the construction of a task-oriented knowledge graph for complex product design. Figure 2 .

[0255] Step 3 includes the following steps:

[0256] (1) Task-oriented reference knowledge semantic matching process

[0257] First, relevant node information is extracted from the constructed knowledge graph, and a task-oriented complex product design knowledge hypernetwork is built through subnet correlation degree calculation and hyperedge correlation degree calculation. Based on the hypernetwork, the hyperedges of specific design tasks and reference documents to be matched are obtained, and the semantic matching degree is calculated by combining the feature information of the hypernetwork with the reconstructed vector space model. Finally, the semantic matching correlation degree between each reference document and the design task is obtained.

[0258] (2) Construction of a task-oriented product design knowledge hypernetwork model

[0259] like Figure 3 Define the task-oriented complex product design knowledge hypernetwork as H=(V,E), where V=(V T V P V C V K ), E = (E T E P E C E K (HE), where each sub-network node mainly comes from the complex product design domain knowledge graph constructed in steps 1 and 2, and the correlation degree and hyperedge correlation degree between sub-network nodes are calculated on the domain knowledge graph and domain corpus.

[0260] (3) Semantic matching method based on hypernetwork and reconstructed VSM

[0261] Step 3.3.1: Obtain the hyperedge and generate the feature term set: Assume that the feature term set obtained from the hyperedge of design task T1 and reference document K2 are as follows:

[0262] Step 3.3.2: Calculate the semantic similarity of each feature class: Based on the subnet correlation and hyperedge correlation, calculate the semantic correlation between design task T1 and reference document K2 from the 8 dimensions (PC, DP, DO, DA, OT, BT, PT, MT) of DKU, and obtain the correlation vector Sim(T1,K2)=(ω PC ,ω DP ,…,ω MT );

[0263] Step 3.3.3: Generate the semantic space vector of the design task and reference documents: Construct the semantic space vector of the design task as VT1 = (ω) around the 8 dimensions of DKU. 11 ,ω 12 ,…,ω 18 The semantic space vector of the reference document is VK2=(ω) 21 ,ω 22 ,…,ω 28 ), where ω 11 =ω 12 =…=ω 18 =1, ω 21 =ω PC ,ω 22 =ω DP ,…,ω 28 =ω MT ;

[0264] Step 3.3.4: Calculate semantic matching similarity using weighted Euclidean distance. The normalized calculation formula is as follows:

[0265]

[0266] In the formula, ω r_coe To determine the knowledge correlation between the final design task T1 and reference document K2, α k ω is the feature term class weight correction coefficient. 1k ω 2k These represent the feature class weights for the design task and the reference document, respectively.

[0267] Step 4 includes the following steps:

[0268] (1) Design of a task-oriented knowledge delivery framework and process

[0269] like Figure 4By constructing a knowledge push process based on WBS decomposition, this paper clarifies how the system parses the design task characteristics and analyzes the knowledge requirements after task allocation for complex product models. It uses rules to generate Cypher statements to achieve the retrieval and push of process knowledge. At the same time, it combines hypernetwork and reconstructed VSM algorithm to match reference knowledge, and uses a multi-dimensional evaluation mechanism to select the most valuable reference knowledge to push to the client. Client designers can obtain relevant knowledge resources while receiving tasks, and provide feedback on the use of knowledge after completing tasks.

[0270] (2) Cypher-based knowledge retrieval and process knowledge push

[0271] This invention constructs a design process knowledge push workflow based on the Neo4j graph database and the Cypher query language, enabling precise location of design tasks within the knowledge graph and automatic push of relevant knowledge. Specifically, the system first uses Levenshtein distance and cosine similarity calculation methods from the vector space model to measure the matching correlation between the design task's name and description and known tasks in the knowledge graph, thus determining the task's position within the graph. Subsequently, the system retrieves design process knowledge related to the task from the graph database using generated Cypher statements and pushes it to the designer as a work package. Simultaneously, the matched nodes provide input data support for subsequent design reference knowledge matching. The entire process can be implemented using the Python py2neo tool.

[0272] (3) Push of design reference knowledge based on multidimensional comprehensive evaluation

[0273] Knowledge Importance ω i_coe :

[0274]

[0275] in, D, representing the designers involved in this design task j Regarding the evaluation of the importance of this reference knowledge Representative designer D j The importance evaluation weight depends on the designer's qualifications (work experience); the higher the qualifications (the more work experience), the greater the weight.

[0276] Knowledge Support ω s_coe :

[0277]

[0278] in, D, representing the designers involved in this design task j Evaluation of the support level for this reference knowledge Representative designer D j The weight of the support evaluation depends on the designer's qualifications (work experience); the higher the qualifications (the more work experience), the greater the weight.

[0279] Knowledge recommendation rating ω final :

[0280] ω final =γ r_coe ·ω r_coe +γ i_coe ·ω i_coe +γ s_coe ·ω s_coe (16)

[0281] In the formula, γ r_coe γ i_coe γ s_coe The system uses empirical weight hyperparameters to measure the relevance, importance, and support of reference knowledge, with a sum of 1 for each. Finally, reference knowledge is pushed based on the multi-dimensional comprehensive evaluation results.

[0282] Example 2

[0283] Example 2 is a preferred example of Example 1.

[0284] This invention takes the design task of a gas turbine aerodynamic scheme as an example. The data source includes WBS information and PBS information related to gas turbine design, as well as a knowledge summary table of turbine aerodynamic scheme design, expert information in the field of turbine aerodynamics, keywords, and semantic annotation results.

[0285] In terms of specific implementation, knowledge extraction is first performed based on a knowledge summary table for turbine aerodynamic design. This table includes four tasks: one-dimensional turbine aerodynamic design, turbine blade styling, three-dimensional turbine aerodynamic design calculation, and turbine transition section design and optimization. Taking turbine blade styling as an example, the content of the knowledge summary table for complex product design is illustrated. Based on the relevant rules and methods in step 2, knowledge extraction is performed on the design knowledge summary table to obtain the design process knowledge required for turbine aerodynamic design.

[0286] Secondly, the extracted design knowledge was supplemented. In the design experience extraction stage, relevant content from Chapter 2, Turbine Aerodynamic Scheme Design, and Chapter 4, Blade Forming, of the commonly used reference manual for gas turbine aerodynamic design, "Aircraft Engine Design Manual - Volume 10," was selected. After text recognition, noise reduction, and text structuring, the relevant rules and methods from step 2 were used to extract design experience, and the extraction results were evaluated using precision (P), recall (R), and F1 score.

[0287]

[0288] The results obtained from testing on the two data sources are shown in the table below:

[0289]

[0290] The specific design experience extracted is shown in the table below:

[0291]

[0292] Based on the knowledge extraction of gas turbine aerodynamic design, knowledge triples are generated and knowledge is fused from the extraction results and data source content. The Levenshtein distance similarity threshold hyperparameter in the knowledge fusion step is set to 2. Finally, a graph is constructed using the Neo4j graph database to obtain a knowledge graph for gas turbine aerodynamic design.

[0293] With the participation of experts from a gas turbine manufacturing company, knowledge about the turbine blade styling design process was obtained using the retrieval method proposed in step 4 in terms of knowledge dissemination during the design process.

[0294] Regarding the dissemination of design reference knowledge, experts from gas turbine manufacturing companies scored the top 7 documents based on their relevance to the "T2 turbine blade styling" task content, assigning them scores based on knowledge importance and support, as shown in the table below:

[0295]

[0296] The empirical hyperparameters for the design reference knowledge relevance, importance, and support are determined and set separately by domain experts.

[0297] Analysis of the above knowledge recommendation results from the document content level shows that document 6 is a paper, which contains a section on blade styling that introduces "blade styling method based on HIT turbine blade editor". Therefore, this document has great reference value for designers of turbine blade styling tasks and has obtained a high degree of knowledge support. After weighting the knowledge support, it has a high overall recommendation degree, which is in line with the understanding in the field of gas turbine blade styling.

[0298] Those skilled in the art will understand that, in addition to implementing the system, apparatus, and their modules provided by this invention in purely computer-readable program code, the same program can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, the system, apparatus, and their modules provided by this invention can be considered a hardware component, and the modules included therein for implementing various programs can also be considered structures within the hardware component; alternatively, modules for implementing various functions can be considered both software programs implementing the method and structures within the hardware component.

[0299] Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essence of the present invention. Unless otherwise specified, the embodiments and features described in this application can be arbitrarily combined with each other.

Claims

1. A knowledge graph-based method for providing knowledge services for complex product design, characterized in that, include: Step 1: By standardizing the document structure, introducing a text tree model, and combining the improved TFIDF algorithm with the N-gram strategy, semantic annotation and structured representation of reference knowledge are achieved. Step 2: By designing a knowledge organization template and combining it with a rule-based knowledge extraction method, we can achieve the systematic acquisition and triple generation of key knowledge in complex product design, and finally build a task-oriented knowledge graph based on Neo4j. Step 3: By constructing a task-oriented complex product design knowledge hypernetwork model and combining a semantic matching method with multi-window co-occurrence and information entropy correction, the correlation between tasks and reference knowledge is calculated and semantic matching is performed. Step 4: By constructing a task-oriented design knowledge push framework, a process knowledge retrieval method based on Cypher statements, and a reference knowledge push method based on multi-dimensional comprehensive evaluation, the acquisition and push of task-related knowledge in complex product design can be realized. Step 1 includes: Step 1.1: Standardize and represent reference knowledge; Define the document knowledge element (DKU) as: DKU=<DP,DA,PC,DO,OT,BT,PT,MT> Wherein, DP represents the R&D stage of the design task to which the reference document belongs, DA represents the specific professional field involved in the reference document, PC represents the specific product model to which the reference document belongs, DO represents the specific design object involved in the reference document, OT represents the conceptual terms related to the design object in the reference document, BT represents the operational or behavioral terms related to the design task in the reference document, PT represents the design parameters or design features related to the design object and design task in the reference document, and MT represents the method or tool terms used to support the design task in the reference document. Step 1.2: Design the semantic annotation process for reference knowledge; Preprocessing and structured modeling of design documents, including document type identification, content extraction, title structure analysis, and text tree generation; Combining word segmentation, stop word removal, and text tree analysis, we improved the TFIDF algorithm and used the N-gram strategy to extract keywords. Based on the constructed ontology, a collection of semantic elements containing domain classes and attributes is generated; Calculate the semantic similarity between keywords and semantic elements, complete the semantic matching and annotation of keywords, and record their correspondence in the ontology; For unmatched keywords, their ontology dimension is determined with human assistance and semantic elements are expanded; Step 1.3: Text content structuring based on text tree; The document title and its direct content are used as the root node. Subheadings and their direct content at all levels form child nodes according to their heading level. The lowest level headings and their direct content form leaf nodes. Direct content includes text paragraphs, table headings, and image titles; document The text tree is ,in This represents the q-th node of the p-th level subheading, and ,in Representative node Title information, Representative node Direct content information, It is the initial node; This is a Python-based regular expression toolkit that generates text trees using rule matching. Step 1.4: Keyword extraction based on text tree and improved TFIDF; Natural language processing tools are used for word segmentation and part-of-speech tagging, and a domain dictionary is loaded and stop words are removed; Introducing an external corpus to calculate improved inverse document frequencies: in, and These represent a document set and an external corpus set, respectively, in the field of complex product design. The total number of documents representing the document set and external corpus sets in the field of complex product design. The document collection and external corpus representing the field of complex product design contain terms The number of documents; The formula for calculating word frequency is: in, Representative terms Documentation in the field of complex product design Frequency of occurrence in Representative domain documents Total word count; The weighting formula based on title position is: In the formula, Representative terms Appears in the text tree node title The number of times in, Let be the number of nodes in the p-th level text tree. Represents the title of a text tree node. Length, This represents the weight of the p-th level text tree node, where m is the number of levels in the text tree. The weight correction formula based on special locations is: In the formula, For special position weighting coefficients, Representative terms Content appearing in text tree nodes The number of times in, Represents the content of the text tree node. Length, It is a collection of title trigger words and their synonyms. Representative title Is it included in the set? Indicator functions within; Introducing terms Professional term weighting coefficient The calculation formula is as follows: in, For a set of domain dictionaries, Weighting coefficients for domain-specific terms; The comprehensive formula for calculating keyword weight is: in, It is a term Part-of-speech weighting factor; The tanh function is used to normalize the importance weights of candidate keywords to achieve weight-to-interval distribution. The mapping is calculated using the following formula: By combining the above formulas, we can obtain the document. The weight values ​​of all terms in the text are sorted from high to low, and candidate terms are extracted based on a set threshold. Candidate keyword set and their weights The threshold adopts a dynamic change strategy, and the expression is: In the formula, For document Total word count For the first The weight value of each term; Step 1.5: Semantic element matching and semantic annotation; Generate a set of semantic elements for design tasks, design objects, design phases, professional fields, and contextual terms; Calculate candidate keywords and ontology instance The similarity is calculated using the following formula: in, Candidate keywords obtained based on the pre-trained word vector model Word2Vec and ontology instance Similarity between them; The similarity between two word vectors is calculated using cosine similarity, expressed as follows: in, It is a Word2Vec Chinese pre-trained model; If semantic similarity Greater than the matching threshold When that happens, use semantic elements. Alternative candidate keywords And perform semantic annotation of the text based on the type of semantic elements; If the semantic similarity is less than the matching threshold But greater than the threshold for new elements In such cases, domain experts will determine and add new elements accordingly. Step 3 includes: Step 3.1: Define the task-oriented complex product design knowledge hypernetwork as... Where V is the set of subnet nodes and E is the set of hyperedges; the subnet nodes are derived from the knowledge graph, and the relevance is calculated on the domain corpus. Step 3.2: Obtain the design task and reference documents The set of hyperedge feature terms; the design task is calculated from 8 dimensions of DKU based on subnet correlation degree and hyperedge correlation degree respectively. and reference documents The semantic relevance is used to obtain the relevance vector. ; The semantic space vector for constructing the design task is The semantic space vector of the reference document is ; Step 3.3: Calculate semantic matching similarity using weighted Euclidean distance. The normalized calculation formula is as follows: （13） In the formula, For the final design task and reference documents The degree of knowledge relevance, For feature term class weight correction coefficients, , These represent the feature class weights for the design task and the reference document, respectively.

2. The knowledge graph-based knowledge service method for complex product design according to claim 1, characterized in that, Step 2 includes: Step 2.1: By introducing the IDEF0 functional modeling method, a structured analysis of complex product design tasks is conducted to identify their functional modules and related information, and then a knowledge organization template containing four categories of content: basic attributes, control elements, process elements and resource elements is designed. Step 2.2: Extract process knowledge related to the design task using a method based on preset rules; Step 2.3: Perform data layer fusion on the extracted knowledge; Step 2.4: Import the knowledge triples in batches into Neo4j for node and relationship modeling and storage.

3. The knowledge graph-based knowledge service method for complex product design according to claim 1, characterized in that, Step 4 includes: Step 4.1: Based on the WBS decomposition process, after parsing the task features, generate Cypher statement retrieval process knowledge, and combine hypernetwork and reconstructed VSM matching reference knowledge; Step 4.2: Cypher-based process knowledge push: Levenshtein distance and cosine similarity are used to locate graph task nodes, retrieve relevant process knowledge and push it; Step 4.3: Multidimensional comprehensive evaluation; Knowledge Importance calculate: in, Representatives of the designers involved in this design task Regarding the evaluation of the importance of this reference knowledge Representative designers Importance evaluation weight; Knowledge support calculate: in, Representatives of the designers involved in this design task Evaluation of the support level for this reference knowledge Representative designers Support evaluation weights; Knowledge Recommendation calculate: In the formula, , , These are the empirical weight hyperparameters for the relevance of design reference knowledge, the importance of knowledge, and the support of knowledge, respectively, and the sum of the three is 1.

4. A knowledge graph-based knowledge service system for complex product design, characterized in that, include: Module M1: By standardizing the document structure, introducing a text tree model, and combining an improved TFIDF algorithm with an N-gram strategy, semantic annotation and structured representation of reference knowledge are achieved. Module M2: By designing knowledge organization templates and combining them with rule-based knowledge extraction methods, it realizes the systematic acquisition and triple generation of key knowledge in complex product design, and finally builds a task-oriented knowledge graph based on Neo4j. Module M3: By constructing a task-oriented complex product design knowledge hypernetwork model and combining a semantic matching method with multi-window co-occurrence and information entropy correction, it realizes the calculation and semantic matching of the correlation between tasks and reference knowledge. Module M4: By constructing a task-oriented design knowledge push framework, a process knowledge retrieval method based on Cypher statements, and a reference knowledge push method based on multi-dimensional comprehensive evaluation, it realizes the acquisition and push of task-related knowledge in complex product design. The module M1 includes: Module M1.1: Standardized representation of reference knowledge; Define the document knowledge element (DKU) as: DKU=<DP,DA,PC,DO,OT,BT,PT,MT> Wherein, DP represents the R&D stage of the design task to which the reference document belongs, DA represents the specific professional field involved in the reference document, PC represents the specific product model to which the reference document belongs, DO represents the specific design object involved in the reference document, OT represents the conceptual terms related to the design object in the reference document, BT represents the operational or behavioral terms related to the design task in the reference document, PT represents the design parameters or design features related to the design object and design task in the reference document, and MT represents the method or tool terms used to support the design task in the reference document. Module M1.2: Design reference knowledge semantic annotation process; Preprocessing and structured modeling of design documents, including document type identification, content extraction, title structure analysis, and text tree generation; Combining word segmentation, stop word removal, and text tree analysis, we improved the TFIDF algorithm and used the N-gram strategy to extract keywords. Based on the constructed ontology, a collection of semantic elements containing domain classes and attributes is generated; Calculate the semantic similarity between keywords and semantic elements, complete the semantic matching and annotation of keywords, and record their correspondence in the ontology; For unmatched keywords, their ontology dimension is determined with human assistance and semantic elements are expanded; Module M1.3: Text content structuring based on text tree; The document title and its direct content are used as the root node. Subheadings and their direct content at all levels form child nodes according to their heading level. The lowest level headings and their direct content form leaf nodes. Direct content includes text paragraphs, table headings, and image titles; document The text tree is ,in This represents the q-th node of the p-th level subheading, and ,in Representative node Title information, Representative node Direct content information, It is the initial node; This is a Python-based regular expression toolkit that generates text trees using rule matching. Module M1.4: Text Tree-Based Improved TFIDF Keyword Extraction; Natural language processing tools are used for word segmentation and part-of-speech tagging, and a domain dictionary is loaded and stop words are removed; Introducing an external corpus to calculate improved inverse document frequencies: in, and These represent a document set and an external corpus set, respectively, in the field of complex product design. The total number of documents representing the document set and external corpus sets in the field of complex product design. The document collection and external corpus representing the field of complex product design contain terms The number of documents; The formula for calculating word frequency is: in, Representative terms Documentation in the field of complex product design Frequency of occurrence in Representative domain documents Total word count; The weighting formula based on title position is: In the formula, Representative terms Appears in the text tree node title The number of times in, Let be the number of nodes in the p-th level text tree. Represents the title of a text tree node. Length, This represents the weight of the p-th level text tree node, where m is the number of levels in the text tree. The weight correction formula based on special locations is: In the formula, For special position weighting coefficients, Representative terms Content appearing in text tree nodes The number of times in, Represents the content of the text tree node. Length, It is a collection of title trigger words and their synonyms. Representative title Is it included in the set? Indicator functions within; Introducing terms Professional term weighting coefficient The calculation formula is as follows: in, For a set of domain dictionaries, Weighting coefficients for domain-specific terms; The comprehensive formula for calculating keyword weight is: in, It is a term Part-of-speech weighting factor; The tanh function is used to normalize the importance weights of candidate keywords to achieve weight-to-interval distribution. The mapping is calculated using the following formula: By combining the above formulas, we can obtain the document. The weight values ​​of all terms in the text are sorted from high to low, and candidate terms are extracted based on a set threshold. Candidate keyword set and their weights The threshold adopts a dynamic change strategy, and the expression is: In the formula, For document Total word count For the first The weight value of each term; Module M1.5: Semantic element matching and semantic annotation; Generate a set of semantic elements for design tasks, design objects, design phases, professional fields, and contextual terms; Calculate candidate keywords and ontology instance The similarity is calculated using the following formula: in, Candidate keywords obtained based on the pre-trained word vector model Word2Vec and ontology instance Similarity between them; The similarity between two word vectors is calculated using cosine similarity, expressed as follows: in, It is a Word2Vec Chinese pre-trained model; If semantic similarity Greater than the matching threshold When that happens, use semantic elements. Alternative candidate keywords And perform semantic annotation of the text based on the type of semantic elements; If the semantic similarity is less than the matching threshold But greater than the threshold for new elements In such cases, domain experts will determine and add new elements accordingly. The module M3 includes: Module M3.1: Defines a task-oriented knowledge hypernetwork for complex product design. Where V is the set of subnet nodes and E is the set of hyperedges; the subnet nodes are derived from the knowledge graph, and the relevance is calculated on the domain corpus. Module M3.2: Obtaining Design Tasks and reference documents The set of hyperedge feature terms; the design task is calculated from 8 dimensions of DKU based on subnet correlation degree and hyperedge correlation degree respectively. and reference documents The semantic relevance is used to obtain the relevance vector. ; The semantic space vector for constructing the design task is The semantic space vector of the reference document is ; Module M3.3: Uses weighted Euclidean distance to calculate semantic matching similarity. The normalized calculation formula is as follows: （13） In the formula, For the final design task and reference documents The degree of knowledge relevance, For feature term class weight correction coefficients, , These represent the feature class weights for the design task and the reference document, respectively.

5. The knowledge graph-based knowledge service system for complex product design according to claim 4, characterized in that, The module M2 includes: Module M2.1: By introducing the IDEF0 functional modeling method, a structured analysis of complex product design tasks is conducted to identify their functional modules and related information, and then a knowledge organization template containing four categories of content: basic attributes, control elements, process elements and resource elements is designed. Module M2.2: Employs a method based on preset rules to extract process knowledge related to the design task; Module M2.3: Performs data layer fusion on the extracted knowledge; Module M2.4: Imports knowledge triples in batches into Neo4j for node and relation modeling and storage.

6. The knowledge graph-based knowledge service system for complex product design according to claim 4, characterized in that, The module M4 includes: Module M4.1: Based on the WBS decomposition process, it generates Cypher statement retrieval process knowledge after parsing task features, and combines hypernetwork and reconstructed VSM matching reference knowledge; Module M4.2: Cypher-based process knowledge push: Levenshtein distance and cosine similarity are used to locate graph task nodes, retrieve relevant process knowledge and push it; Module M4.3: Multidimensional Comprehensive Evaluation; Knowledge Importance calculate: in, Representatives of the designers involved in this design task Regarding the evaluation of the importance of this reference knowledge Representative designers Importance evaluation weight; Knowledge support calculate: in, Representatives of the designers involved in this design task Evaluation of the support level for this reference knowledge Representative designers Support evaluation weights; Knowledge Recommendation calculate: In the formula, , , These are the empirical weight hyperparameters for the relevance of design reference knowledge, the importance of knowledge, and the support of knowledge, respectively, and the sum of the three is 1.

Images