A knowledge graph-based method for recommending manufacturing resources

Abstract

This invention relates to a knowledge graph-based method for recommending manufacturing resources, comprising the following steps: establishing a supply and demand information model for the manufacturing domain; constructing a manufacturing domain ontology model to represent the concepts and relationships of demand information and manufacturing resources, serving as the schema layer of the manufacturing domain knowledge graph; performing knowledge extraction; utilizing the knowledge graph for visualization; implementing knowledge graph embedding, during training, obtaining erroneous triples (i.e., negative samples) by randomly replacing head entities; continuously optimizing the loss function using stochastic gradient descent to obtain qualified embedding vectors; after training the vector representations of entities using the TransE model, using vector value calculation to measure the degree of conformity of each resource, and employing cosine similarity calculation to calculate the similarity between vectors; and adding resource QoS service matching based on feature matching.

Description

Technical Field

[0001] This invention belongs to the field of digital manufacturing technology, specifically a manufacturing resource recommendation method based on knowledge graphs. Background Technology

[0002] Against the backdrop of economic globalization and the continuous development of intelligent manufacturing technologies, the manufacturing industry continues to expand in scale, but its manufacturing models and technological platforms lag behind. Many enterprises suffer from geographically dispersed manufacturing resources and weak resource connections, making it difficult to form efficient and collaborative product manufacturing methods. This leads to unreasonable resource utilization and waste of valuable resources during product manufacturing and processing. With the emergence of cloud manufacturing, manufacturing enterprises have begun to use technologies such as cloud computing and big data to serve and represent various manufacturing resources and help users obtain higher-quality resource services. However, the development of cloud manufacturing is still imperfect, with an incomplete technological system. Problems remain, including disorganized information models, incomplete resource summaries, and unreasonable resource recommendation methods. These issues result in manufacturing enterprises being slow to analyze and solve problems when responding to users' product manufacturing and processing needs, and producing low-quality resource recommendations.

[0003] This patent addresses the aforementioned problems by designing a knowledge graph-based manufacturing resource recommendation method. It expresses product manufacturing and processing requirements and resource information in the form of a knowledge graph, and uses this method to recommend the necessary manufacturing resources for product processing. This patent aims to explore methods for constructing knowledge graphs in the manufacturing field for product manufacturing and processing, and to recommend manufacturing resources based on the constructed knowledge graph, thereby improving enterprise service quality and resource utilization. Summary of the Invention

[0004] The purpose of this invention is to develop a knowledge graph-based manufacturing resource recommendation method. This method extracts enterprise manufacturing resource information and product processing demand information, represents supply and demand information in the manufacturing field using a knowledge graph, and utilizes knowledge graph embedding methods and resource service attribute information to recommend manufacturing resources, thereby achieving high-quality service responses to product manufacturing and processing needs. The technical solution is as follows:

[0005] A knowledge graph-based method for recommending manufacturing resources includes the following steps:

[0006] The first step is to establish a supply and demand information model for the manufacturing sector.

[0007] Based on the characteristics of the parts, the product manufacturing requirements are decomposed into different requirement units, and a product requirement information model is established:

[0008] Ru = {ID,RTy,RN,RFu,RTq,RC}

[0009] Ru represents the demand unit; ID represents the demand number; RTy represents the demand type, which includes site resources, equipment resources, and human resources; RN represents the demand name, which includes site resource requirements such as machining site and assembly site, equipment resource requirements such as turning equipment and milling equipment, and human resource requirements such as machining workers and assembly workers; RTq represents the time limit, also known as the machining time quota; RC represents the cost limit; and RCu represents the part demand characteristic unit.

[0010] In accordance with the specifications for establishing demand information models, construct the corresponding manufacturing resource information model:

[0011] MR={MRID,MRAt,MRF,MRAb,MRQos}

[0012] MR represents manufacturing resources, which includes five different attribute units: MRID is the manufacturing resource identifier, MRAt represents the basic attribute of the resource, MRF represents the functional attribute of the resource, MRAb represents the capability attribute of the resource, and MRQos represents the service quality attribute of the resource.

[0013] The second step is to construct an ontology model for the manufacturing domain, representing the concepts and relationships between demand information and manufacturing resources, which serves as the schema layer of the manufacturing domain knowledge graph.

[0014] The third step is knowledge extraction: the BIO annotation method is used to annotate the data, and the BiLSTM-CRF model is used to complete the entity recognition task. The input of BiLSTM is the trained word vector, which reflects each word in a sentence. After the model is trained, the output of BiLSTM is the score of each category corresponding to a certain word.

[0015] The fourth step is to use knowledge graphs to achieve visual representation: the Neo4j graph database is used to store and represent the obtained data. The obtained entity and relation data are stored in a structured CSV file. A Neo4j program is written to import nodes and relations, import the data, and generate a knowledge graph for the manufacturing field.

[0016] The fifth step is to embed the knowledge graph: The TransE model is used to embed the knowledge graph, which maps entities and relations to the same space. By training the existing triples T = (h, r, t) in the manufacturing domain knowledge graph, the vectors of entities and relations satisfy h + rt = 0.

[0017] During training, incorrect triples, i.e. negative samples, are obtained by randomly replacing the head entity; the loss function is continuously optimized using stochastic gradient descent to obtain qualified embedding vectors.

[0018] The sixth step involves training the vector representation of entities using the TransE model, then using the vector values ​​to measure the degree of conformity of each resource, and employing the cosine similarity calculation method to calculate the similarity between vectors.

[0019] The knowledge graph stores product manufacturing and processing demand information and manufacturing resource information, and uses nodes and relationships to represent demand units in the product demand information model and resource units in the manufacturing resource information model. By using vectorization and similarity queries of different nodes, manufacturing resource features are matched according to the manufacturing and processing features in the demand unit, thereby obtaining the resource unit that matches the demand unit.

[0020] The seventh step is to add resource QoS service matching on the basis of feature matching. Select the best service quality candidate set from the manufacturing resources that meet the similarity requirements of basic features, so that the final candidate resources have good performance in terms of functional attributes and service quality.

[0021] Resource QoS service is defined by four indicators: processing time (PT), cost level (CL), processing quality (PQ), and customer satisfaction (S).

[0022] Resource processing time is expressed as:

[0023]

[0024] Among them, T a T is the average time T takes for the selected resource to complete its processing task. ave The average time to complete the processing task for all similar processing resources;

[0025] Cost levels are expressed as:

[0026]

[0027] Among them, C a C is the cost of the selected resource to complete its processing task. ave The average cost of completing the processing task using all similar processing resources within the enterprise;

[0028] Processing quality is expressed as:

[0029]

[0030] Among them, Q a Q represents the number of qualified tasks out of all processing tasks for the selected processing resource. e To complete the total number of processing tasks;

[0031] Satisfaction is expressed as:

[0032]

[0033] Among them, S i This refers to the user satisfaction rating obtained from each previous product manufacturing and processing.

[0034] The weights of the four indicators are calculated using the coefficient of variation method, and the Qos score of the manufacturing resources is calculated using these four indicators. The manufacturing resources are then ranked according to their scores, and a resource selection score is established. The group or several groups of manufacturing resources with the highest ranking are selected as the result of the knowledge graph-based manufacturing resource recommendation method.

[0035] Furthermore, in the first step,

[0036] The basic attributes of manufacturing resources include the following three types of information:

[0037] MRAt = {MRn, MRL, MRw}

[0038] Where MRn represents the resource name, which is a static attribute of the resource; MRl represents the resource location, which is dynamic information; and MRw represents the working status of the resource.

[0039] The functional attributes of manufacturing resources include the following three types of information.

[0040] MRF = {MRt, MRe, MRp}

[0041] Wherein, MRt represents the machining feature type, MRe represents the machining process type, and MRp represents the applicable machining parts;

[0042] The ability attributes of manufacturing resources include the following three types of information.

[0043] MRAb = {MRz, MRm, MRpow}

[0044] Wherein, MRz represents the maximum machining size, MRm represents the machinable material, and MRpow represents the machining power;

[0045] The service quality attributes of manufacturing resources include the following three types of information.

[0046] MRQos = {MRtim, MRc, MRa}

[0047] Where MRtim represents processing time, MRc represents processing cost, and MRa represents processing precision.

[0048] In this invention, the manufacturing domain supply and demand information model construction module can dominate the description method of product manufacturing and processing demand information and manufacturing resource information; the manufacturing domain ontology modeling module and knowledge extraction module can fill the knowledge graph from both conceptual and data aspects; the knowledge graph visualization expression module and knowledge graph embedding module can visualize entities and relationships and form embedded representations; the feature matching module based on entity similarity calculation and the manufacturing resource recommendation generation module based on resource services can achieve the recommendation of optimal manufacturing resources from both basic features and resource service quality aspects.

[0049] Existing manufacturing resource matching models often fail to fully integrate internal enterprise manufacturing resources, and the disorganized description of resource information significantly increases the difficulty of matching manufacturing resources, making it difficult to provide users with accurate manufacturing resource recommendations. Furthermore, the recommendation indicators are often singular, failing to consider the actual user experience of manufacturing resource services. Therefore, compared with existing technologies, this invention not only fully mines and standardizes all internal enterprise data, but also achieves higher-quality manufacturing resource recommendations based on manufacturing requirements and service quality. Attached Figure Description

[0050] Figure 1 System architecture diagram of a knowledge graph-based manufacturing resource recommendation method

[0051] Figure 2 Schematic diagram of supply and demand information model in the manufacturing sector

[0052] Figure 3 Basic framework diagram of BiLSTM-CRF model

[0053] Figure 4 Knowledge graph diagram of manufacturing and processing requirements for a certain product

[0054] Figure 5 Partial diagram of the manufacturing resource knowledge graph

[0055] Figure 6 TransE model diagram

[0056] Figure 7 Schematic diagram of similar units for manufacturing and processing requirements of a certain product Detailed Implementation

[0057] The following describes specific embodiments of the present invention. Figure 1 This is a system architecture diagram of the knowledge graph-based manufacturing resource recommendation method of the present invention.

[0058] like Figure 1As shown, the knowledge graph-based manufacturing resource recommendation method (hereinafter referred to as the recommendation method) of the present invention mainly includes a manufacturing domain supply and demand information model construction module, a manufacturing domain ontology modeling module, a knowledge extraction module, a knowledge graph visualization expression module, a knowledge graph embedding module, a feature matching module based on entity similarity calculation, and a manufacturing resource recommendation generation module based on resource services.

[0059] This invention's recommendation system primarily targets the product manufacturing process, aiming to address the challenges of fully utilizing enterprise resources due to their fragmented nature and the low quality of resource recommendations resulting from a single recommendation method. This invention constructs a formalized and unified demand information model and manufacturing resource model, and then builds an ontology model as the schema layer of a manufacturing domain knowledge graph. Knowledge extraction technology is then used to extract resource information as the data layer of the domain knowledge graph, completing the construction of the manufacturing domain knowledge graph. Finally, Neo4j is used to visualize the knowledge graph.

[0060] After constructing the manufacturing domain knowledge graph, the TransE model is used to learn the embedded representation of entities in the manufacturing domain knowledge graph using knowledge graph embedding technology. Resource units that match the demand units are obtained through entity similarity calculation. Finally, the manufacturing resource with the highest service quality is selected through resource service QoS calculation.

[0061] This invention establishes a standardized information expression method by constructing a unified supply and demand model in the manufacturing field. This enables the efficient integration of product manufacturing and processing demand information as well as internal enterprise manufacturing resource information, facilitating resource management and mobilization. The construction of a manufacturing field knowledge graph allows for the visual representation of manufacturing knowledge, and the embedding of entities and relationships within the knowledge graph provides a more direct, faster, and more accurate representation. Furthermore, the manufacturing resource recommendation method, which integrates both feature and resource service information, is more comprehensive than general recommendation methods, meeting users' requirements for manufacturing resource service quality in an intelligent manufacturing environment.

[0062] The structure and working principle of each module will be explained in turn below.

[0063] 1. Manufacturing Sector Supply and Demand Information Model Building Module

[0064] The supply and demand information model in the manufacturing field of this invention includes a demand information model and a manufacturing resource information model. To facilitate the implementation of the recommendation method, it is necessary to construct two models with a unified form, and the manufacturing and processing information contained therein should correspond to each other. Figure 2 A schematic diagram of supply and demand information models in the manufacturing sector.

[0065] Based on the characteristics of the parts, the product manufacturing requirements are decomposed into different requirement units, and a product requirement information model is established:

[0066] Ru = {ID,RTy,RN,RFu,RTq,RC}

[0067] Ru represents the demand unit, ID represents the demand number, RTy represents the demand type, such as site resources, equipment resources, human resources, etc., RN represents the demand name, such as machining site and assembly site included in site resources, turning equipment and milling equipment included in equipment resources, or machining workers and assembly workers included in human resources, RTq represents the time limit, also called the machining time quota, such as the time for turning the outer diameter and inner diameter, RC represents the cost limit, and RFU represents the part demand characteristic unit. The element form is different for different resources. Taking equipment resource demand as an example, it can be represented as:

[0068] RFu = {RFc, RFm, RFma, RFs, RFa}

[0069] In this context, RFc represents the machining feature category, such as external circle, hole, and shaft; RFm represents the machining method, such as turning and milling; RFma represents the part material feature, such as 45 steel and 40Cr; RFs represents the part size feature, such as the diameter of hole machining or the length and width of plane machining; and RFa represents the required machining accuracy feature, such as IT8 and Ra6.3.

[0070] In accordance with the specifications for establishing demand information models, construct the corresponding manufacturing resource information model:

[0071] MR={MRID,MRAt,MRF,MRAb,MRQos}

[0072] Here, MR represents manufacturing resources, which includes five different attribute units: MRID is the manufacturing resource identifier, MRAt represents the basic attribute of the resource, MRF represents the functional attribute of the resource, MRAb represents the capability attribute of the resource, and MRQos represents the service quality attribute of the resource.

[0073] The basic attributes mainly include the following three types of information:

[0074] MRAt = {MRn, MRL, MRw}

[0075] Wherein, MRn represents the resource name, which is a static attribute of the resource; MRl represents the resource location, which is dynamic information; and MRw represents the working status of the resource.

[0076] Functional attributes mainly include the following three types of information.

[0077] MRF = {MRt, MRe, MRp}

[0078] Among them, MRt represents the machining feature type, such as external cylindrical machining; MRe represents the machining process type, such as turning; and MRp represents the applicable machining parts.

[0079] Ability attributes mainly include the following three types of information.

[0080] MRAb = {MRz, MRm, MRpow}

[0081] Where MRz represents the maximum machining size, MRm represents the machinable materials, and MRpow represents the machining power.

[0082] Service quality attributes mainly include the following three types of information.

[0083] MRQos = {MRtim, MRc, MRa}

[0084] Where MRtim represents processing time, MRc represents processing cost, and MRa represents processing precision.

[0085] Once the aforementioned supply and demand information model for the manufacturing sector is completed, product manufacturing and processing needs and manufacturing resource information can be represented using this model.

[0086] 2. Ontology Modeling Module for Manufacturing Domain

[0087] Ontology is a knowledge representation method. Based on the demand information model and manufacturing resource information model constructed using the manufacturing domain supply and demand information model construction module, it guides the construction of the manufacturing domain ontology model. A seven-step ontology construction method is employed to represent the concepts and relationships of demand information and manufacturing resources, serving as the schema layer of the manufacturing domain knowledge graph. The construction scope includes the parts that need to be processed and all manufacturing resources of the enterprise, considering the demand information of past products and the manufacturing resource information already retained in the enterprise's long-term manufacturing and processing tasks. A top-down hierarchical structure is adopted. Based on the relevant needs of product manufacturing and processing, manufacturing domain knowledge is further subdivided into two concepts: manufacturing resource knowledge and product manufacturing and processing knowledge. These two concepts are further subdivided according to the manufacturing domain supply and demand information model. Manufacturing resource knowledge is further divided into human resources, equipment resources, technical resources, site resources, and external resources. Different resources are further divided into different equipment, technologies, etc. Product manufacturing and processing knowledge is divided into different demand units. Establish the hierarchical relationship and interrelationship of four concepts: part-of (the relationship between the part and the whole), kind-of (the relationship between the parent class and the child class), instance-of (the relationship between the class and the instance), and attribute-of (the attributes of the class, including object attributes and data attributes).

[0088] After defining the concepts, relationships, and attributes of the ontology, the Protégé ontology modeling method is used to construct an ontology model for the manufacturing domain.

[0089] 3. Knowledge Extraction Module

[0090] Knowledge graphs are essentially network knowledge representations based on graph models, with the core being triples consisting of "entity-relationship-entity". Knowledge extraction methods are used to extract entities and relations from the data, forming the data layer of the knowledge graph. Much of the knowledge in the manufacturing domain is unstructured or semi-structured data that cannot be directly accessed; the role of knowledge extraction is to extract and store key information from this unstructured data. Knowledge extraction mainly includes entity extraction, relation extraction, and attribute extraction.

[0091] Before knowledge extraction, data preprocessing and text annotation are required. The BIO annotation method is used for data annotation, where B, I, and O represent the beginning, middle, or end characters of an entity, as well as other content that is not an entity. For example, a requirement in product manufacturing is: "The internal turning base is installed on the vertical lathe turntable, with the small end facing down and the large end facing up. It is calibrated and fixed to ensure that the outer circle runout of the large end is within 0.05 mm. The internal surface is semi-finished, with at least 5 layers, and the rotation speed is not less than 80 r / min." Here, "circular runout" and "rotation speed" are entities that need to be extracted.

[0092] Entity extraction is the process of extracting named entities from text. Commonly used entity extraction methods include rule-based and dictionary-based methods, machine learning methods, and deep learning methods. To ensure the effectiveness of entity extraction, the BiLSTM-CRF model is used to complete the entity extraction task.

[0093] Figure 3 This is a schematic diagram of the basic framework of the BiLSTM-CRF model. BiLSTM, or Bidirectional Long Short-Term Memory Network, adds a backward computation process compared to LSTM. Unlike LSTM, which only passes information forward to backward, BiLSTM utilizes the contextual information of the input sequence by simultaneously feeding the input sequence into both the forward and backward LSTM layers and then connecting both layers to the output layer. CRF (Conditional Random Field) is a decision model that analyzes the output to find relationships and predict sequences. In the BiLSTM-CRF model, the input is word vectors, and the output is a sequence of labels. Specifically, the input to BiLSTM is trained word vectors reflecting each word in a sentence; after training, the output of BiLSTM is the score of a word for each category.

[0094] Recognition Performance Analysis: This invention utilizes the manufacturing resources and actual product manufacturing processes of a company to verify its content. Internal data and text resources were integrated, resulting in 10,353 data entries. These were preprocessed using the BIO annotation method, and the dataset was divided into training, validation, and test sets in a 3:1:1 ratio. To evaluate whether the training effect met expectations, the model evaluation metrics were set as accuracy, recall, and F1 score. After 30 rounds of training, the model reached a stable state with an accuracy of 96.12%, a recall of 95.64%, and an F1 score of 95.51%, meeting the entity recognition requirements.

[0095] An ontology-based relation extraction method is adopted. An ontology is an expression of a concept, while an entity is an instantiation of the ontology. Therefore, the concepts and relationships between concepts in the ontology can be inherited by the entities. After the entity extraction is completed, a large number of entities will be obtained. At the same time, the entities will be labeled. Based on the labels, the entities will be paired to obtain triples. For example, entity A "lathe" and entity B "JIMK460" can be paired to obtain the triple (A, part-of, B).

[0096] 4. Knowledge Graph Visualization Module

[0097] To address the problems of inconvenient data representation and retrieval, unclear data structure, and difficulty in uncovering data relationships, Neo4j graph database is used to store and represent the data obtained from the above three modules.

[0098] Neo4j is a widely used graph database that uses the Cypher language. It can not only visualize knowledge graphs but also perform tasks such as querying and node matching. The obtained entity and relationship data are stored in a structured CSV file. A Neo4j program to import nodes and relationships is shown in Table 1, automatically importing the data to generate a knowledge graph for the manufacturing domain.

[0099] Table 1. Neo4j Node and Relationship Import Procedure

[0100]

[0101] Figure 4 This is a schematic diagram of the knowledge graph of the manufacturing and processing requirements of a certain product according to the present invention. Figure 5 This is a partial schematic diagram of the manufacturing resource knowledge graph of the present invention.

[0102] 5. Knowledge Graph Embedding Module

[0103] Knowledge graphs consist of entities and relations. To achieve a structured representation of product manufacturing and processing demand information and manufacturing resource information in the knowledge graph, the edges of the knowledge graph are expressed in the form of fact triples:

[0104] T = (h, r, t)

[0105] Where h represents the head entity, t represents the tail entity, and r represents the relationship between the head and tail entities.

[0106] Knowledge graph embedding represents entities and relations in a knowledge graph as low-dimensional vectors, serving as a method for entity alignment. Knowledge graph embedding proposes a scoring function to represent the validity of triples, and then learns and trains the embedded representations of entities and relations through model learning, continuously improving the scores of triples. The TransE model is used to implement knowledge graph embedding, mapping entities and relations to the same space.

[0107] Figure 6 The TransE model is illustrated in the diagram. Relation r is regarded as a translation operation from head entity h to tail entity t. The training objective of this model is to train the existing triples in the manufacturing domain knowledge graph so that the vectors of entities and relations satisfy h + rt = 0. That is, the closer the new vector after summing the vectors of the head entity and relation in the triple is to the vector of the tail entity, the higher the correctness of the triple.

[0108] For correct triples, lower potential energy indicates a closer relationship between the two entities; conversely, for incorrect triples, higher potential energy is desirable. The experiment involves selecting as many entity data points as possible from the knowledge graph and extracting triples from them. Relationship indexes are also constructed for each entity to build the training set required for the experiment. Based on the open-source machine learning library TensorFlow, the data is first preprocessed, with triples stored in a txt file. The parameters of the TransE model are set, and the input training set consists of the defined triples, entity sets, and relation sets. The marginal margin hyperparameter γ is set to 1, the dimension k of the embedding vector is 100, the learning rate is 0.001, and the number of iterations is 100.

[0109] Since all triples in the knowledge graph are correct, a replacement method is used to obtain incorrect triples for model training. During training, the head and tail entities are randomly replaced to obtain incorrect triples, i.e., negative samples. Then, the loss function is continuously optimized through stochastic gradient descent to obtain qualified embedding vectors.

[0110] 6. Feature matching module based on entity similarity calculation

[0111] After training the vector representations of entities using the TransE model, the similarity between resources is measured by calculating the vector values. The cosine similarity method is used to calculate the similarity between vectors, and the formula is as follows:

[0112]

[0113] Among them, a i ∈A=[a1,a2,...,an ], b i ∈B=[b1,b2,...,b n ], where n is the vector dimension value.

[0114] The knowledge graph stores product manufacturing and processing demand information and manufacturing resource information. It uses nodes and relationships to represent demand units in the product demand information model and resource units in the manufacturing resource information model. By using vectorization and similarity queries of different nodes, manufacturing resource features are matched according to the manufacturing and processing features in the demand unit, thereby obtaining the resource unit that matches the demand unit.

[0115] Figure 7 This diagram illustrates similar units related to product manufacturing and processing needs. It includes a partial knowledge graph of a specific need unit within the enterprise's knowledge graph of manufacturing and processing needs for a certain outer circular product, and partial knowledge graphs of two resource units within the manufacturing resource knowledge graph. Through knowledge graph embedding, the need resource units are treated as target features. A 100-dimensional vector is obtained through entity training. Cosine similarity is used to find similar feature vectors. After initial screening of the candidate set in the manufacturing resource knowledge graph, many similar manufacturing resource units are obtained, and these units are sorted from highest to lowest similarity. Table 2 shows the sorting information for some similar units.

[0116] Table 2 shows the sorting information for some similar units.

[0117]

[0118] 7. Manufacturing resource recommendation and generation module based on resource services

[0119] Traditional manufacturing resource recommendations only consider the fitness of features and whether the processing attributes of resources meet manufacturing requirements. However, in the context of intelligent manufacturing, users are paying more and more attention to the service evaluation of manufacturing resources, and the service quality of manufacturing resources is gradually becoming an important evaluation factor.

[0120] Based on feature matching, this invention adds a resource QoS service matching method, which aims to select the resource with the best service quality from manufacturing resources that meet the similarity requirements of basic features, so that the final candidate resources have good performance in terms of functional attributes and service quality.

[0121] QoS service is defined by four indicators: processing timeliness (PT), cost level (CL), processing quality (PQ), and satisfaction level (S). Resource processing timeliness is expressed as...

[0122]

[0123] Among them, T a T is the average time T takes for the selected resource to complete its processing task. aveThe average time to complete the processing task for all similar processing resources.

[0124] Cost level expressed as

[0125]

[0126] Among them, C a C is the cost of the selected resource to complete its processing task. ave The average cost of completing the processing task for all similar processing resources within the enterprise.

[0127] Processing quality is expressed as

[0128]

[0129] Among them, Q a Q represents the number of qualified tasks out of all processing tasks for the selected processing resource. e To complete the total number of processing tasks.

[0130] Satisfaction is expressed as

[0131]

[0132] Among them, S i This refers to the user satisfaction rating obtained from each previous product manufacturing and processing.

[0133] An evaluation matrix Q can be constructed using the Qos parameters of each resource. Q contains four evaluation indicators for n manufacturing resources.

[0134]

[0135] The values ​​of processing quality PQ and satisfaction S are both within [0,1] and can be directly used in the calculation. However, the resource processing time PT and cost level index CL need to be normalized. The calculation method is as follows.

[0136]

[0137] Wherein, PT' and CL' are the normalized resource processing timeliness and cost level indicators, P(i) are all resource individuals with processing time records in the historical database, and Q(i) are all resource individuals with cost records.

[0138] The weighting coefficients of each indicator are determined using the coefficient of variation method.

[0139]

[0140] Among them, V i σ is the coefficient of variation. i Standard deviation, This is the average of all indicators.

[0141] The weight ω can be obtained from the coefficient of variation. i for

[0142]

[0143] Using the above method, the candidate manufacturing resources obtained by feature matching are weighted and summed according to the four indicators to obtain the Qos score. The manufacturing resources are then sorted according to the scores. Table 3 shows the service matching and sorting information of the manufacturing resources.

[0144] Table 3 Manufacturing Resource Service Matching Ranking Information

[0145]

[0146] Based on the ranking of manufacturing resource services, a resource selection score is established, and the group or several groups of manufacturing resources with the highest ranking are selected as the result of the knowledge graph-based manufacturing resource recommendation method.

Claims

1. A knowledge graph-based method for recommending manufacturing resources, comprising the following steps: The first step is to establish a supply and demand information model for the manufacturing sector. Based on the characteristics of the parts, the product manufacturing requirements are decomposed into different requirement units, and a product requirement information model is established: Ru = {ID,RTy,RN,RFu,RTq,RC} Ru represents the demand unit; ID represents the demand number; RTy represents the demand type, which includes site resources, equipment resources, and human resources; RN represents the demand name, which includes site resource requirements such as machining site and assembly site, equipment resource requirements such as turning equipment and milling equipment, and human resource requirements such as machining workers and assembly workers; RTq represents the time limit, also known as the machining time quota; RC represents the cost limit; and RCu represents the part demand characteristic unit. In accordance with the specifications for establishing demand information models, construct the corresponding manufacturing resource information model: MR={MRID,MRAt,MRF,MRAb,MRQos} in, MR stands for Manufacturing Resource, which contains five different attribute units: MRID is the Manufacturing Resource Identifier, MRAt represents the basic attribute of the resource, MRF represents the functional attribute of the resource, MRAb represents the capability attribute of the resource, and MRQos represents the service quality attribute of the resource. The second step is to construct an ontology model for the manufacturing domain, representing the concepts and relationships between demand information and manufacturing resources, which serves as the schema layer of the manufacturing domain knowledge graph. The third step is knowledge extraction: the BIO annotation method is used to annotate the data, and the BiLSTM-CRF model is used to complete the entity recognition task. The input of BiLSTM is the trained word vector, which reflects each word in a sentence. After the model is trained, the output of BiLSTM is the score of each category corresponding to a certain word. The fourth step is to use knowledge graphs to achieve visual representation: the Neo4j graph database is used to store and represent the obtained data. The obtained entity and relation data are stored in a structured CSV file. A Neo4j program is written to import nodes and relations, import the data, and generate a knowledge graph for the manufacturing field. The fifth step is to embed the knowledge graph: The TransE model is used to embed the knowledge graph, mapping entities and relations to the same space. By training the existing triples T = (h, r, t) in the manufacturing domain knowledge graph, relation r is regarded as a translation operation from head entity h to tail entity t. The training goal is to train the existing triples in the manufacturing domain knowledge graph so that the vectors of entities and relations satisfy h + rt = 0. That is, the closer the new vector after summing the vectors of the head entity and relation in the triple is to the vector of the tail entity, the higher the correctness of the triple. During training, incorrect triples, i.e. negative samples, are obtained by randomly replacing the head entity; the loss function is continuously optimized using stochastic gradient descent to obtain qualified embedding vectors. The sixth step involves training the vector representation of entities using the TransE model, then using the vector values ​​to measure the degree of conformity of each resource, and employing the cosine similarity calculation method to calculate the similarity between vectors. The knowledge graph stores product manufacturing and processing demand information and manufacturing resource information, and uses nodes and relationships to represent demand units in the product demand information model and resource units in the manufacturing resource information model. By using vectorization and similarity queries of different nodes, manufacturing resource features are matched according to the manufacturing and processing features in the demand unit, thereby obtaining the resource unit that matches the demand unit. The seventh step is to add resource QoS service matching on the basis of feature matching. Select the best service quality candidate set from the manufacturing resources that meet the similarity requirements of basic features, so that the final candidate resources have good performance in terms of functional attributes and service quality. Resource QoS service is defined by four indicators: processing time (PT), cost level (CL), processing quality (PQ), and customer satisfaction (S). Resource processing time is expressed as: Among them, T a T is the average time T takes for the selected resource to complete its processing task. ave The average time to complete the processing task for all similar processing resources; Cost levels are expressed as: Among them, C a C is the cost of the selected resource to complete its processing task. ave The average cost of completing the processing task using all similar processing resources within the enterprise; Processing quality is expressed as: Among them, Q a Q represents the number of qualified tasks out of all processing tasks for the selected processing resource. e To complete the total number of processing tasks; Satisfaction is expressed as: Among them, S i This refers to the user satisfaction rating obtained from each previous product manufacturing and processing. The weights of the four indicators are calculated using the coefficient of variation method, and the Qos score of the manufacturing resources is calculated using these four indicators. The manufacturing resources are then ranked according to their scores, and a resource selection score is established. The group or several groups of manufacturing resources with the highest ranking are selected as the result of the knowledge graph-based manufacturing resource recommendation method.

2. The knowledge graph-based manufacturing resource recommendation method according to claim 1, characterized in that, In the first step, The basic attributes of manufacturing resources include the following three types of information: MRAt = {MRn, MRL, MRw} Where MRn represents the resource name, which is a static attribute of the resource; MRl represents the resource location, which is dynamic information; and MRw represents the working status of the resource. The functional attributes of manufacturing resources include the following three types of information. MRF = {MRt, MRe, MRp} Wherein, MRt represents the machining feature type, MRe represents the machining process type, and MRp represents the applicable machining parts; The ability attributes of manufacturing resources include the following three types of information. MRAb = {MRz, MRm, MRpow} Wherein, MRz represents the maximum machining size, MRm represents the machinable material, and MRpow represents the machining power; The service quality attributes of manufacturing resources include the following three types of information. MRQos = {MRtim, MRc, MRa} Where MRtim represents processing time, MRc represents processing cost, and MRa represents processing precision.

Images