A numerical control machine tool field knowledge graph and a construction method thereof

Abstract

The application discloses a kind of numerical control machine tool field knowledge graph and its construction method, including numerical control machine tool data acquisition;Based on a variety of word segmentation strategies to build entity corpus;Entity ontology library is constructed to entity corpus;Based on entity co-occurrence defines the relationship between entity and entity;Design auxiliary marking platform, mark entity and relationship;Active learning module carries out visualization to entity precision and relationship precision, and related marking sample is supplemented to low-precision category;Based on graph topology self-supervised model, text representation self-supervised model, graph topology mixed text representation self-supervised model multi-channel ablation data fusion;Based on graph database, large-scale knowledge graph node storage;Large-scale front-end graph visualization is realized based on front-end preloading technology;Based on knowledge graph representation model, graph topology self-supervised model, text representation self-supervised graph topology mixed text representation self-supervised model multi-channel node search implementation scheme carries out front-end visualization data focusing method.

Description

Technical Field

[0001] This invention belongs to the field of CNC machine tool knowledge graph construction technology, and relates to a knowledge graph in the field of CNC machine tools and its construction method, which is used to construct a high-precision knowledge graph in the field of CNC machine tools. The knowledge graph in the field of CNC machine tools constructed based on this method is larger in scale and richer in knowledge types. Background Technology

[0002] The rapid development of technologies such as the Internet of Things, artificial intelligence, and deep learning has brought new research methods to the manufacturing industry, driving its continuous development towards intelligence. Throughout the entire product lifecycle, a vast amount of data and knowledge is accumulated at different stages, containing a wealth of useful manufacturing information and patterns. Previously, this data and knowledge were primarily stored in traditional relational databases, resulting in high redundancy, scattered distribution, weak correlation, and relatively small storage capacity. Furthermore, there is a lack of semantic research into the relationships, cognition, understanding, and reasoning within data and knowledge.

[0003] Knowledge graphs (KG) originated from Google's next-generation intelligent semantic search engine technology. Their initial purpose was to optimize search engine results and enhance user search quality and experience. Essentially, a knowledge graph is a semantic network that reveals the relationships between entities. It describes concepts and their relationships in the physical world in symbolic form; it's a directed graph structure that describes entities, entity attributes, and entity relationships. Nodes in the graph represent entities or attributes, while edges represent relationships between entities. Complex relationships between things are described using multiple "entity-relationship-entity" triples.

[0004] Current research on knowledge graphs largely focuses on general-purpose knowledge graphs. Research on domain-specific knowledge graphs is somewhat insufficient. Regarding knowledge graphs in the machine tool field, previous research and existing technologies, such as the machine tool fault diagnosis method based on knowledge graphs disclosed in Chinese invention patent application CN202210948275.2, the knowledge graph relationship completion method for CNC machine tool equipment maintenance disclosed in CN202110450657.8, and the dynamic construction method of knowledge graphs for CNC machine tool fault diagnosis disclosed in CN201911116876.1, mainly concentrate on fault diagnosis knowledge graphs in the CNC machine tool / machining center field, process data knowledge graphs in the CNC machining field, or knowledge graphs in the CNC programming field. Further research is needed on knowledge graphs covering the entire CNC machine tool domain.

[0005] Enterprises manufacturing and using CNC machine tools, as well as academia, have accumulated a vast amount of data and knowledge. This includes, but is not limited to, CNC machine tool design data, machining process data, assembly data, data generated during the CNC machine tool production process, CNC machine tool maintenance and fault data, and books, papers, and patents related to CNC machine tools. These data types are complex and diverse in format, including structured, semi-structured, and unstructured data, making efficient utilization of this data and knowledge challenging. Efficiently extracting the necessary information and knowledge from this existing data and knowledge remains extremely difficult. Summary of the Invention

[0006] (I) Purpose of the Invention

[0007] The construction of existing machine tool domain knowledge graphs mainly focuses on fault diagnosis knowledge graphs in the CNC machine tool / machining center domain, process data knowledge graphs in the CNC machining domain, or CNC programming knowledge graphs. However, there is a lack of in-depth research on knowledge graphs for the entire CNC machine tool domain. As a result, although CNC machine tool manufacturers, users, and academia have accumulated a great deal of data and knowledge, it remains a very difficult technical problem to efficiently extract the required information and knowledge from this existing data and knowledge. This invention aims to propose a knowledge graph for the CNC machine tool domain and its construction method. This knowledge graph construction method collects unstructured, semi-structured, and structured data covering the entire CNC machine tool domain through various methods and means. It performs multi-channel word segmentation on the collected data through various algorithms and summarizes the segmentation results to form a summary weight. The results are then reviewed and modified by experts. A combination of manual and automatic methods is used to annotate entities and relationships. An active learning module is used to visualize the accuracy and supplement relevant labeled samples for low-precision categories. This invention constructs a high-precision CNC machine tool knowledge database using the above method. The knowledge graph in the field of CNC machine tools constructed based on this invention is larger in scale and more comprehensive in knowledge types.

[0008] (II) Technical Solution

[0009] To achieve its inventive purpose and solve its technical problems, the technical solution adopted by this invention is as follows:

[0010] A method for constructing a knowledge graph in the field of CNC machine tools, characterized in that the knowledge graph construction method includes at least the following steps:

[0011] SS1. Knowledge data collection in the field of CNC machine tools: Collect all knowledge data directly or indirectly related to the field of CNC machine tools from publicly available knowledge data sources using automatic, semi-automatic and / or manual methods, and perform preliminary processing on the collected knowledge data to construct an unstructured text knowledge database for CNC machine tools;

[0012] SS2. Based on the unstructured text knowledge database of CNC machine tools constructed in step SS1, construct a CNC machine tool entity corpus based on multiple word segmentation strategies;

[0013] SS3. Define entity annotation standards. CNC machine tool experts annotate a portion of the entity corpus in the CNC machine tool entity corpus of step SS2 according to the entity annotation standards. A general algorithm is used to learn and train the results of expert annotation and complete the annotation of all entity categories, thereby constructing a CNC machine tool entity ontology library.

[0014] SS4. Define entity relationship category labeling standards. Based on the CNC machine tool entity ontology library constructed in step SS3, define the relationships between entities according to the entity relationship category labeling standards and through entity co-occurrence calculation. The entity relationship categories include at least master-slave relationships, configuration relationships, time relationships, causal relationships, state relationships and / or processing operation relationships between entities.

[0015] SS5. Design an auxiliary labeling platform to label entities and relationships from the unstructured text collected in step SS1. In this process, it is necessary to design a sliding labeling method, an entity category module, a relationship category module, a division of labor module, an active learning module, a remote supervision module, and a remote collaborative cloud labeling module for entity and relationship labeling scenarios.

[0016] SS6. Based on the active learning module in step SS5, visualize the entity accuracy and relation accuracy, and supplement relevant labeled samples for low-accuracy categories;

[0017] SS7. Multi-channel ablation data fusion based on graph topology self-supervised model, text representation self-supervised model, and graph topology hybrid text representation self-supervised model;

[0018] SS8. A large-scale knowledge graph node storage scheme based on graph databases;

[0019] SS9. Large-scale front-end graph visualization based on front-end preloading technology;

[0020] SS10. A front-end visualization data focusing method based on a multi-channel node search implementation scheme of knowledge graph representation model, graph topology self-supervised model, text representation self-supervised graph topology hybrid text representation self-supervised model.

[0021] In a preferred embodiment of the present invention, step SS1 above involves at least the following sub-steps: collecting all publicly available knowledge data directly or indirectly related to the field of CNC machine tools from relevant knowledge data sources, and performing preliminary processing on the collected relevant knowledge data:

[0022] SS1.1. Complete the collection of all knowledge data directly or indirectly related to the field of CNC machine tools that are publicly available from relevant knowledge data sources. The relevant knowledge data from these sources shall include at least the CNC machine tool process information generated during the enterprise's design, manufacturing, and assembly process; relevant books on the field of CNC machine tools published in libraries; relevant journal articles on the field of CNC machine tools published on academic websites; relevant patents on the field of CNC machine tools published by the Patent Office; and / or industry information on CNC machine tool industry websites. Furthermore, the relevant knowledge data shall include unstructured text data and / or image-based knowledge data.

[0023] SS1.2. For the image-based knowledge data in the knowledge data collected in step SS1.1, convert them into unstructured text data using optical character recognition technology and / or layout recognition technology;

[0024] SS1.3. Construct a CNC machine tool unstructured text knowledge database based on the unstructured text data collected and processed in steps SS1.1 and SS1.2. Based on the CNC machine tool unstructured text knowledge database, train a pre-trained language model that does not require manual data annotation. The pre-trained language model is an unsupervised pre-trained language model, a pre-trained language model based on semi-supervised learning, a pre-trained language model based on multi-task learning, and / or a pre-trained language model based on cross-language learning, to support the initial vectorization of text for downstream named entity recognition tasks and the initial vectorization of text for relation extraction tasks.

[0025] In a preferred embodiment of the present invention, step SS2 above involves at least the following sub-steps to complete the construction of the CNC machine tool entity corpus:

[0026] SS2.1. Combining different word segmenters and word segmentation algorithms to form several word segmentation strategies, performing multi-channel text word segmentation on the CNC machine tool knowledge data in the unstructured text knowledge database constructed in step SS1, wherein the word segmenter includes the jieba word segmenter and / or the lac word segmenter, and the word segmentation algorithm includes the ngram word segmentation algorithm, the dag word segmentation algorithm and / or the ac automaton word segmentation algorithm;

[0027] SS2.2. Classify and summarize the text segmentation results generated by different segmentation strategies in step SS2.1. Assign different weights to the text segmentation results of different categories during the summary, and sort the text segmentation results based on the weights.

[0028] SS2.3. Experts in the field of CNC machine tools define entity categories for the sorted text segmentation results from step SS2.2;

[0029] SS2.4. Construct a corpus of CNC machine tool entities.

[0030] In a preferred embodiment of the present invention, step SS3 above involves at least the following sub-steps to complete the construction of the CNC machine tool entity library:

[0031] SS3.1. Define entity annotation standards one by one based on entity attributes, and build an entity annotation standard library;

[0032] SS3.2. Based on the aforementioned entity annotation standard library, experts in the CNC machine tool field perform entity category annotation on a portion of the entity corpus in the CNC machine tool entity corpus;

[0033] SS3.3. A general algorithm is used to learn and train the results of expert annotation. When the learning accuracy reaches the preset condition, the annotation of all other entity categories is completed. Each word annotated here corresponds to the meaning of the word in the original text.

[0034] SS3.4. Based on the entity category annotation results of CNC machine tool experts and general algorithms, complete the construction of the CNC machine tool entity ontology library.

[0035] In a preferred embodiment of the present invention, step SS4 above involves defining the relationships between entities in the CNC machine tool entity body library based on at least the following sub-steps:

[0036] SS4.1. Define entity relationship category annotation standards based on entity relationships between entities;

[0037] SS4.2. Based on the entity categories annotated by experts in the CNC machine tool entity ontology library, firstly, the named entity recognition dataset is formed by matching using the AC automaton. Then, an entity discovery model and a text segmentation model for the CNC machine tool domain are established based on the Globle Pointer algorithm.

[0038] SS4.3. After identifying the unstructured text data collected in step SS1 based on the domain entity discovery model, the entities in the already labeled entity corpus are filtered out. Experts review the new entities discovered by the domain entity discovery algorithm and add them to the entity corpus. Then, the unstructured text data collected in step SS1 is segmented based on the domain text segmentation model.

[0039] SS4.4. After segmenting unstructured text data, entity probability correlation is calculated by calculating word frequency co-occurrence and mutual information co-occurrence, and text, entity, and entity similarity data are extracted to construct an entity relationship auxiliary labeling dataset.

[0040] SS4.5. Based on the entity relationship auxiliary labeling dataset constructed in step SS4.4, experts annotate the relationships between entities and between entities within the original text. This approach enables the labeling of different categories of relationship data when two entities appear in different texts. Furthermore, by using entity word frequency co-occurrence probability and entity mutual information co-occurrence probability, the overall data can be focused on.

[0041] In a preferred embodiment of the present invention, step SS6 above includes at least the following sub-steps:

[0042] SS6.1. Entity recognition is performed based on Keras tuner and Ray [tune] automatic parameter tuning technology and named entity recognition models BERT CRF and BERT Global Pointer.

[0043] SS6.2 performs relation extraction based on the BERT semi-pointer semi-labeling model, global pointer linker, and UIE model.

[0044] In a preferred embodiment of the present invention, step SS7 above includes at least the following sub-steps:

[0045] SS7.1 employs text representation similarity models, graph topology similarity models (deepwalk, node2vec, graphsage), and graph topology hybrid text representation self-supervised models (roberta graph sage, ernie graph sage) to support data fusion.

[0046] SS7.2. Implement knowledge graph relationship completion suggestion function based on link prediction algorithm;

[0047] SS7.3. Experts review and verify the completion of knowledge graph relationships. Expert verification confirms the knowledge graph data by passing the knowledge graph relationship completion review, thereby achieving the ability to ensure the quality of knowledge graph data.

[0048] In a preferred embodiment of the present invention, in step SS8 above, the triplet data reviewed by the current expert is stored in a knowledge graph database, using a graph database as the storage medium.

[0049] In a preferred embodiment of the present invention, step SS9 above includes at least the following sub-steps:

[0050] SS9.1. Large-scale front-end graph visualization based on front-end preloading technology;

[0051] SS9.2 implements a graph topology node relationship preloading method based on pyvis to achieve visualization and rendering of node relationships in a large-scale CNC machine tool knowledge graph. During the rendering process, a node dimensionality reduction approach is used to reduce data latency issues caused by front-end rendering operations.

[0052] The second objective of this invention is to provide a knowledge graph in the field of CNC machine tools constructed according to the above method.

[0053] (III) Technical Effects

[0054] Compared with existing technologies, the method for constructing a knowledge graph in the field of CNC machine tools of the present invention has significant technical advantages:

[0055] 1. This invention provides a top-down method for constructing a knowledge graph for the CNC machine tool professional field. It collects unstructured, semi-structured, and structured data covering the entire CNC machine tool field using various methods and techniques. The collected data is segmented into words using multiple algorithms, and the segmentation results are aggregated to form a weighted summary, which is then reviewed and modified by experts. A combination of manual and automatic methods is used to label entities and relationships, and an active learning module is used to visualize the accuracy, supplementing low-precision categories with relevant labeled samples. This method constructs a high-precision CNC machine tool knowledge database. The CNC machine tool field knowledge graph constructed based on this invention is larger in scale and more comprehensive in its knowledge categories.

[0056] 2. This invention provides an auxiliary annotation platform, featuring modules for designing entity categories, relationship categories, division of labor, active learning, remote supervision, and remote collaborative cloud annotation. Cloud deployment reduces the requirements for local hardware. The remote collaborative cloud annotation module supports simultaneous online annotation by multiple users and provides real-time statistics on annotation details, significantly improving annotation efficiency.

[0057] 3. This invention achieves rapid visualization of knowledge graphs. It utilizes front-end preloading technology to visualize large-scale front-end graphs, and employs a pyvis-based graph topology node relationship preloading method to render the visualization of node relationships in a large-scale CNC machine tool knowledge graph. During rendering, node dimensionality reduction reduces data latency caused by front-end rendering operations. Furthermore, it employs a multi-channel node search implementation scheme based on knowledge graph representation models, graph topology self-supervised models, and a hybrid text representation self-supervised graph topology and text representation self-supervised models to achieve front-end visualization data focusing. Attached Figure Description

[0058] Figure 1 This is a schematic diagram illustrating the method for constructing a knowledge graph in the field of CNC machine tools according to the present invention.

[0059] Figure 2This is a schematic diagram of the CNC machine tool domain knowledge data collection in step SS1.

[0060] Figure 3 This is a schematic diagram illustrating the construction of a CNC machine tool entity corpus based on multiple word segmentation strategies in step SS2.

[0061] Figure 4 This is a schematic diagram of constructing the CNC machine tool entity body library in step SS3.

[0062] Figure 5 This is a diagram illustrating the relationships between entities in step SS4.

[0063] Figure 6 This is a schematic diagram of the auxiliary marking platform in step SS5. Detailed Implementation

[0064] To better understand the present invention, the following embodiments further illustrate the content of the invention, so that the advantages and features of the invention can be more easily understood by those skilled in the art. It should be noted that the following descriptions are merely preferred embodiments of the present invention, but the content of the invention is not limited to the following embodiments. In fact, various modifications and variations can be made to the present invention without departing from the scope or spirit of the invention, which will be apparent to those skilled in the art. For example, features shown or described as part of one embodiment can be used with another embodiment to produce yet another embodiment. Therefore, it is intended that such modifications and variations be included within the scope of the appended claims and their equivalents.

[0065] This invention discloses a knowledge graph construction method in the field of CNC machine tools, comprising: CNC machine tool data acquisition; construction of a CNC machine tool entity corpus based on multiple word segmentation strategies; construction of a CNC machine tool entity ontology library by CNC machine tool experts based on the entity corpus; definition of relationships between entities based on entity co-occurrence; design of an auxiliary labeling platform to label entities and relationships from unstructured text; visualization of entity accuracy and relationship accuracy by an active learning module, supplementing relevant labeled samples for low-accuracy categories; multi-channel ablation data fusion based on graph topology self-supervised models, text representation self-supervised models, and graph topology hybrid text representation self-supervised models; a large-scale knowledge graph node storage scheme based on the graph database Hugegraph; large-scale front-end graph visualization based on front-end preloading technology; and a front-end visualization data focusing method based on a multi-channel node search implementation scheme of knowledge graph representation models, graph topology self-supervised models, text representation self-supervised graph topology hybrid text representation self-supervised models.

[0066] As a preferred example of the present invention, such as Figure 1As shown, the method for constructing a knowledge graph in the field of CNC machine tools provided by this invention mainly includes 10 steps in its specific implementation, such as CNC machine tool data acquisition, constructing an entity corpus, constructing an entity ontology library, etc. The specific implementation steps of each step are as follows:

[0067] Step SS1. Knowledge Data Collection in the CNC Machine Tool Field: This involves collecting all publicly available knowledge data directly or indirectly related to the CNC machine tool field using automatic, semi-automatic, and / or manual methods. Preliminary processing of the collected knowledge data is then performed to construct an unstructured text knowledge database for CNC machine tools. This is specifically accomplished through the following sub-steps:

[0068] like Figure 2 As shown, this specifically includes the collection of process information, PDF data, search engine crawler data, CNC machine tool papers, and CNC machine tool patents. OCR and layout recognition technologies are used to convert image information from which text cannot be directly obtained into text information, thereby supplementing the unstructured text data required for the knowledge collection section. Based on massive amounts of unstructured text data, pre-trained language models for the CNC machine tool field are trained (BERT, Roberta, and other pre-trained language models support the initial vectorization of text for downstream named entity recognition tasks and relation extraction tasks).

[0069] CNC machine tool data acquisition refers to the collection of all knowledge and data that directly or indirectly affect CNC machine tools using various collection methods, including automatic, semi-automatic, and manual methods. Specifically, it includes the following:

[0070] SS1.1. Process Information Collection: Process information data generated during the enterprise's design, manufacturing, and assembly processes.

[0071] Data collection in SS1.2.pdf format: PDF electronic documents related to CNC machine tools were obtained from electronic libraries, PDF papers related to CNC machine tools were obtained from academic websites such as CNKI, and PDF patent reports related to CNC machine tools were obtained from patent websites.

[0072] SS1.3. Search Engine Crawler Collection: Use web crawlers to collect relevant knowledge about CNC machine tools from search engines.

[0073] SS1.4. Collection of CNC Machine Tool Papers: Obtain relevant CNC machine tool papers through academic websites such as CNKI;

[0074] SS1.5. CNC Machine Tool Patent Collection: Obtain patent reports related to the field of CNC machine tools through patent websites.

[0075] SS1.6 uses OCR and layout recognition technology to convert image information from which text cannot be directly obtained into text information, thereby supplementing the unstructured text data required for the knowledge collection section.

[0076] SS1.7. Training pre-trained language models for the CNC machine tool domain based on massive amounts of unstructured text data (pre-trained language models such as BERT and Roberta support initial vectorization of text for downstream named entity recognition tasks and initial vectorization of text for relation extraction tasks).

[0077] Step SS2. Based on the unstructured text knowledge database for CNC machine tools constructed in Step SS1, such as... Figure 3 As shown, a CNC machine tool entity corpus is constructed based on multiple word segmentation strategies, and the construction of the CNC machine tool entity corpus is completed based on at least the following sub-steps:

[0078] SS2.1. Multi-channel text segmentation is performed on the collected CNC machine tool data using jieba segmenter, lac segmenter, ngram segmentation algorithm, dag segmentation algorithm, and ac automaton segmentation algorithm respectively.

[0079] SS2.2. After word segmentation, the segmentation results corresponding to different word segmentation methods are summarized to form a summary weight of the word segmentation results;

[0080] SS2.3. Experts redefine the entity categories of the sorted word segmentation results and construct an entity corpus for CNC machine tools.

[0081] Step SS3. CNC machine tool experts construct a CNC machine tool entity ontology library based on the entity corpus. Experts annotate entity categories based on the entity ranking results implemented by a general algorithm. Here, each annotation corresponds to a word and its meaning in the original text; entity annotation standards are defined. In step SS3 above, the construction of the CNC machine tool entity ontology library is completed based on at least the following sub-steps, such as... Figure 4 As shown:

[0082] SS3.1. Define entity annotation standards one by one based on entity attributes, and build an entity annotation standard library;

[0083] SS3.2. Based on the aforementioned entity annotation standard library, experts in the CNC machine tool field perform entity category annotation on a portion of the entity corpus in the CNC machine tool entity corpus;

[0084] SS3.3. A general algorithm is used to learn and train the results of expert annotation. When the learning accuracy reaches the preset condition, the annotation of all other entity categories is completed. Each word annotated here corresponds to the meaning of the word in the original text.

[0085] SS3.4. Based on the entity category annotation results of CNC machine tool experts and general algorithms, complete the construction of the CNC machine tool entity ontology library.

[0086] Step SS4. Define the relationship between entities based on entity co-occurrence: Implement a text segmentation model based on expert-annotated entity categories and the global pointer algorithm. The unstructured data collected in step (1) is segmented using the text segmentation model algorithm. After segmentation, entity probability correlation is constructed using word frequency co-occurrence calculation and mutual information co-occurrence calculation to build text, entity, entity, and entity similarity data. Experts define the relationship between two entities and their corresponding relationships in the original text using word frequency co-occurrence calculation and mutual information co-occurrence calculation. Define relationship annotation standards. Relationship types include: master-slave relationship, configuration relationship, temporal relationship, causal relationship, state relationship, and processing operation relationship.

[0087] In step SS4 above, the definition of the relationships between entities in the CNC machine tool entity body library should be completed based on at least the following sub-steps, such as... Figure 5 As shown:

[0088] SS4.1. Based on the expert-annotated entity corpus of CNC machine tools, firstly, the named entity recognition dataset is formed by matching using the AC automaton, and then the domain entity discovery model is implemented based on the Globe Pointer algorithm.

[0089] SS4.2. Based on the domain entity discovery algorithm, the unstructured data collected in step SS1 is identified and entities in the already labeled entity corpus are filtered out.

[0090] SS4.3. Experts review and supplement new entities discovered by the domain entity discovery algorithm into the entity corpus;

[0091] SS4.4. An auxiliary labeled dataset is constructed by calculating co-occurrence of entity frequencies and co-occurrence of mutual information to extract the following structural relationships: (text, entity A, entity B, entity co-occurrence probability of word frequencies, entity co-occurrence probability of mutual information). Experts then label the relationships between entity A and entity B on this dataset. This enables the labeling of different categories of relationship data when entity A and entity B appear in different texts. Based on the entity co-occurrence probability of word frequencies and the entity co-occurrence probability of mutual information, a focused approach to the overall data can be achieved.

[0092] SS4.5 defines relationship labeling standards. Relationship types include: master-slave relationship, configuration relationship, temporal relationship, causal relationship, state relationship, and processing operation relationship.

[0093] Step SS5. Design an auxiliary tagging platform to annotate entities and relationships from unstructured text, such as... Figure 6As shown, this paper presents a sliding labeling method for entity and relation labeling scenarios, including modules for entity categories, relation categories, task allocation, active learning, remote supervision, and remote collaborative cloud labeling. An auxiliary labeling platform is used to label unstructured designs manually, semi-automatically, and fully automatically. To improve labeling efficiency and operability, the task allocation module and the remote collaborative cloud labeling module were designed.

[0094] Step SS6. The active learning module visualizes entity and relation accuracy and supplements relevant labeled samples for low-precision categories; it performs entity recognition based on Keras tuner and Ray [tune] automatic parameter tuning techniques and named entity recognition models BERT CRF and BERT Global Pointer, and relation extraction based on BERT semi-pointer / semi-labeling models, Global Pointer Linker, and UIE models. Step SS6 includes at least the following sub-steps:

[0095] SS6.1. The Active Learning module visualizes entity accuracy and relation accuracy;

[0096] SS6.2. Supplement relevant labeled samples for low-precision categories;

[0097] SS6.3 uses Keras tuner and Ray [tune] automatic parameter tuning technology, BERT CRF and BERT Global Pointer models for entity recognition, and BERT semi-pointer / semi-labeling model, Global Pointer Linker and UIE models for relation extraction.

[0098] Step SS7. Multi-channel ablation data fusion based on graph topology self-supervised models, text representation self-supervised models, and graph topology hybrid text representation self-supervised models: In the direction of data fusion, text representation similarity models, graph topology similarity models (deepwalk, node2vec, graphsage), and graph topology hybrid text representation self-supervised models (roberta graphsage, ernie graphsage) can currently be used to support data fusion. During experiments, it was found that different representations of the same entity follow certain predictable methods. A knowledge graph relationship completion prompt function is implemented based on the link prediction algorithm. Expert verification confirms the knowledge graph data through knowledge graph relationship completion review, thereby achieving the ability to ensure the quality of knowledge graph data. This step respectively realizes the knowledge graph fusion based on self-supervised methods, the knowledge graph fusion based on expert annotation supervision, and the knowledge graph processing capabilities based on link prediction. The final data quality is controlled by experts. Step SS7 above includes at least the following sub-steps:

[0099] SS7.1 employs text representation similarity models, graph topology similarity models (deepwalk, node2vec, graphsage), and graph topology hybrid text representation self-supervised models (roberta graph sage, ernie graph sage) to support data fusion.

[0100] SS7.2. Implement knowledge graph relationship completion suggestion function based on link prediction algorithm;

[0101] SS7.3. Experts review and verify the completion of knowledge graph relationships. Expert verification confirms the knowledge graph data by passing the knowledge graph relationship completion review, thereby achieving the ability to ensure the quality of knowledge graph data.

[0102] Step SS8. Large-scale knowledge graph node storage scheme based on graph database; The triple data after expert review is stored in the knowledge graph database, using a graph database as the storage medium.

[0103] Step SS9. Large-scale front-end graph visualization based on front-end preloading technology; large-scale CNC machine tool knowledge graph node relationship visualization rendering is achieved using the pyvis graph topology node relationship preloading method. During the rendering process, node dimensionality reduction reduces data latency caused by front-end rendering operations. Step SS9 includes at least the following sub-steps:

[0104] SS9.1. Large-scale front-end graph visualization based on front-end preloading technology;

[0105] SS9.2 implements a graph topology node relationship preloading method based on pyvis to achieve visualization and rendering of node relationships in a large-scale CNC machine tool knowledge graph. During the rendering process, a node dimensionality reduction approach is used to reduce data latency issues caused by front-end rendering operations.

[0106] Step SS10. A front-end visualization data focusing method is implemented based on a multi-channel node search scheme that combines knowledge graph representation models, graph topology self-supervised models, and text representation self-supervised graph topology hybrid text representation self-supervised models.

[0107] The objectives of this invention have been fully and effectively achieved through the above embodiments. All equivalent or simple variations made to the structures, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art can make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not deviate from the structure of this invention or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

Claims

1. A method for constructing a knowledge graph in the field of CNC machine tools, characterized in that, The method for constructing the knowledge graph includes at least the following steps: SS1. Knowledge data collection in the field of CNC machine tools: Collect all knowledge data directly or indirectly related to the field of CNC machine tools from publicly available knowledge data sources using automatic, semi-automatic and / or manual methods, and perform preliminary processing on the collected knowledge data to construct an unstructured text knowledge database for CNC machine tools; SS2. Based on the unstructured text knowledge database of CNC machine tools constructed in step SS1, construct a CNC machine tool entity corpus based on multiple word segmentation strategies. The construction of the CNC machine tool entity corpus is completed based on at least the following sub-steps: Several word segmentation strategies are formed by combining different word segmenters and word segmentation algorithms to perform multi-channel text word segmentation on CNC machine tool knowledge data in unstructured text knowledge databases. The word segmenters include jieba word segmenter and / or lac word segmenter, and the word segmentation algorithms include ngram word segmentation algorithm, dag word segmentation algorithm and / or ac automaton word segmentation algorithm. The text segmentation results generated by different segmentation strategies are classified and summarized. Different weights are assigned to different categories of text segmentation results during the summarization, and the text segmentation results are sorted based on the weights. Experts in the field of CNC machine tools define entity categories for the sorted text segmentation results, and finally construct a CNC machine tool entity corpus. SS3. Define entity annotation standards. CNC machine tool experts annotate a portion of the entity corpus in the CNC machine tool entity corpus of step SS2 according to the entity annotation standards. A general algorithm is used to learn and train the results of expert annotation and complete the annotation of all entity categories, thereby constructing a CNC machine tool entity ontology library. SS4. Define entity relationship category labeling standards. Based on the CNC machine tool entity ontology library constructed in step SS3, define the relationships between entities according to the entity relationship category labeling standards and through entity co-occurrence calculation. The entity relationship categories include at least master-slave relationships, configuration relationships, time relationships, causal relationships, state relationships and / or processing operation relationships between entities. SS5. Design an auxiliary labeling platform to label entities and relationships from the unstructured text collected in step SS1. In this process, it is necessary to design a sliding labeling method, an entity category module, a relationship category module, a division of labor module, an active learning module, a remote supervision module, and a remote collaborative cloud labeling module for entity and relationship labeling scenarios. SS6. Based on the active learning module in step SS5, visualize the entity accuracy and relation accuracy, and supplement relevant labeled samples for low-accuracy categories; SS7. Multi-channel ablation data fusion based on graph topology self-supervised models, text representation self-supervised models, and graph topology hybrid text representation self-supervised models, including: Data fusion is supported by text representation similarity model, graph topology similarity model, and graph topology hybrid text representation self-supervised model; knowledge graph relationship completion prompt function is implemented based on link prediction algorithm; experts review the knowledge graph relationship completion, and the knowledge graph data is confirmed by the expert verification of the knowledge graph relationship completion review, thereby achieving the ability to ensure the quality of knowledge graph data; SS8. A large-scale knowledge graph node storage scheme based on graph databases; SS9. Large-scale front-end graph visualization based on front-end preloading technology; SS10. A front-end visualization data focusing method based on a multi-channel node search implementation scheme of knowledge graph representation model, graph topology self-supervised model, text representation self-supervised graph topology hybrid text representation self-supervised model.

2. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, In step SS1 above, at least the following sub-steps are required to complete the collection of all publicly available knowledge data directly or indirectly related to the field of CNC machine tools from relevant knowledge data sources, and to perform preliminary processing on the collected relevant knowledge data: SS1.

1. Complete the collection of all knowledge data directly or indirectly related to the field of CNC machine tools that are publicly available from relevant knowledge data sources. The relevant knowledge data from these sources shall include at least the CNC machine tool process information generated during the enterprise's design, manufacturing, and assembly process; relevant books on the field of CNC machine tools published in libraries; relevant journal articles on the field of CNC machine tools published on academic websites; relevant patents on the field of CNC machine tools published by the Patent Office; and / or industry information on CNC machine tool industry websites. Furthermore, the relevant knowledge data shall include unstructured text data and / or image-based knowledge data. SS1.

2. For the image-based knowledge data in the knowledge data collected in step SS1.1, convert them into unstructured text data using optical character recognition technology and / or layout recognition technology; SS1.

3. Construct a CNC machine tool unstructured text knowledge database based on the unstructured text data collected and processed in steps SS1.1 and SS1.

2. Based on the CNC machine tool unstructured text knowledge database, train a pre-trained language model that does not require manual data annotation. The pre-trained language model is an unsupervised pre-trained language model, a pre-trained language model based on semi-supervised learning, a pre-trained language model based on multi-task learning, and / or a pre-trained language model based on cross-language learning, to support the initial vectorization of text for downstream named entity recognition tasks and the initial vectorization of text for relation extraction tasks.

3. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, In step SS3 above, the construction of the CNC machine tool entity library is completed based on at least the following sub-steps: SS3.

1. Define entity annotation standards one by one based on entity attributes, and build an entity annotation standard library; SS3.

2. Based on the aforementioned entity annotation standard library, CNC machine tool experts performed entity category annotation on a portion of the entity corpus in the CNC machine tool entity corpus; SS3.

3. A general algorithm is used to learn and train the results of expert annotation. When the learning accuracy reaches the preset condition, the annotation of all other entity categories is completed. Each word annotated here corresponds to the meaning of the word in the original text. SS3.

4. Based on the entity category annotation results of CNC machine tool experts and general algorithms, complete the construction of the CNC machine tool entity ontology library.

4. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, In step SS4 above, the definition of the relationships between entities in the CNC machine tool entity body library is completed based on at least the following sub-steps: SS4.

1. Define entity relationship category annotation standards based on entity relationships between entities; SS4.

2. Based on the entity categories annotated by experts in the CNC machine tool entity ontology library, firstly, the named entity recognition dataset is formed by matching using the AC automaton. Then, an entity discovery model and a text segmentation model for the CNC machine tool domain are established based on the Globle Pointer algorithm. SS4.

3. After identifying the unstructured text data collected in step SS1 based on the domain entity discovery model, the entities in the already labeled entity corpus are filtered out. Experts review the new entities discovered by the domain entity discovery algorithm and add them to the entity corpus. Then, the unstructured text data collected in step SS1 is segmented based on the domain text segmentation model. SS4.

4. After segmenting unstructured text data, entity probability correlation is calculated by calculating word frequency co-occurrence and mutual information co-occurrence, and text, entity, and entity similarity data are extracted to construct an entity relationship auxiliary labeling dataset. SS4.

5. Based on the entity relationship auxiliary labeling dataset constructed in step SS4.4, experts annotate the relationships between entities and between entities in the original text.

5. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, Step SS6 above includes at least the following sub-steps: SS6.

1. Entity recognition is performed based on Keras tuner, Ray [tune] automatic parameter tuning technology, named entity recognition model, BERT CRF, and BERT GlobalPointer model. SS6.2 performs relation extraction based on the BERT semi-pointer semi-labeling model, global pointer linker, and UIE model.

6. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, In step SS8 above, the triplet data reviewed by the current experts is stored in a knowledge graph database, using a graph database as the storage medium.

7. The method for constructing a knowledge graph in the field of CNC machine tools according to claim 1, characterized in that, Step SS9 above includes at least the following sub-steps: SS9.

1. Large-scale front-end graph visualization based on front-end preloading technology; SS9.

2. Based on the pyvis graph topology node relationship preloading method, the visualization and rendering of node relationships of knowledge graphs for large-scale CNC machine tools is realized.

8. A knowledge graph in the field of CNC machine tools, characterized in that, The CNC machine tool domain knowledge graph is constructed according to the construction method of the CNC machine tool domain knowledge graph according to any one of claims 1 to 7.

Images