A method for assisting generation of a garment process route based on knowledge graph embedding

Abstract

The application discloses a garment process route auxiliary generation method based on knowledge graph embedding, and a garment process knowledge graph containing a data layer, a construction layer and an application layer is hierarchically built. Style, part, fabric, process, equipment entities and various semantic relationships are defined through ontology modeling. Graph storage is completed by using Neo4j. Embedding learning is carried out on graph triples by using a TransD model. An entity mapping vector is trained by using a special distance scoring function and an interval loss function. After new garment styles, fabric and part features are input, a comprehensive matching score of historical process cases is calculated, and process routes are generated by reusing optimal case process and equipment information. The application can solve the defects of traditional process planning, such as dependence on manual work and knowledge fragmentation, is suitable for small-batch production of garments, and can effectively improve the intelligent efficiency of garment process planning.

Description

Technical Field

[0001] This invention relates to a method for generating garment process routes based on knowledge graph embedding, belonging to the fields of intelligent garment manufacturing, knowledge graph, natural language processing and deep learning. Background Technology

[0002] Garment process planning is a core aspect of garment production, directly determining production efficiency and costs. Currently, domestic garment companies largely rely on the personal experience of senior process engineers for process planning. This leads to fragmented process knowledge and difficulties in integrating heterogeneous data from multiple sources. Traditional retrieval methods can only search historical process sheets by style number, failing to achieve intelligent transfer and reuse of process experience.

[0003] Existing research largely focuses on building knowledge graphs for single equipment, single processes, or partial sewing techniques, lacking a systematic knowledge integration model for the entire process route. Furthermore, conventional deep learning-generated models are prone to issues such as missing processes and disordered sequences in small-sample scenarios of garment manufacturing, making them difficult to apply practically. To address these shortcomings, there is an urgent need for a technical solution that can integrate knowledge of all categories of garment manufacturing processes and combine semantic learning to achieve intelligent matching of similar processes and route reuse. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies, such as fragmented management of garment process knowledge, high dependence on manual process planning, low efficiency in generating new garment processes, and poor adaptability to small sample scenarios. This invention provides a garment process route generation method based on knowledge graph embedding, which realizes structured management of garment process knowledge, intelligent reuse of historical process experience, and automated output of new garment process routes and supporting equipment information, thereby promoting the intelligent transformation of garment production.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: This invention discloses a method for assisting in the generation of garment process routes based on knowledge graph embedding. The method includes the following steps: Step 1) Construct a knowledge graph for garment process planning: Divide garment process knowledge into three major areas: product and component information, equipment resource information, and process decision and flow information. Adopt a hybrid strategy of top-down as the main approach and bottom-up as the supplementary approach to complete ontology modeling, process data processing, triple construction, and realize knowledge graph storage and visualization through graph database. Step 2) Knowledge graph embedding learning: Using knowledge graph triples as training samples, negative samples are constructed by combining a negative sampling strategy, and low-dimensional vectors of entities and relations are obtained by training the TransD model. Step 3) Assisted generation of process route: Input the style, fabric and component features of the new garment, calculate the comprehensive matching score of each historical process route based on the trained TransD model, select the optimal candidate route according to the score, reuse its process and equipment information, and output the process route of the new garment.

[0006] Preferably, in step 1), the Stanford seven-step method and Protégé 5.5 tools are used to carry out ontology modeling. The ontology is divided into three subdomains: clothing product and component domain, manufacturing resource domain, and process decision and flow domain. The ontology defines object attributes and data attributes. The object attributes include hasComponent, isMadeOf, isPartOf, requiresProcess, usesEquipment, isPrerequisiteFor, hasCharacteristic, and hasQualitativeDescription. At the same time, process matching rules and process sequence rules are formally defined.

[0007] hasComponent: Represents the subordinate relationship between style and style component, and style component and basic component; isMadeOf: Represents the correspondence between style component and fabric; isPartOf: Represents the belonging relationship between component and component category; requiresProcess: Represents the association relationship between style component and required processing process; usesEquipment: Represents the usage relationship between process and processing equipment; isPrerequisiteFor: Represents the sequential constraint relationship between different processes.

[0008] Preferably, in step 1), using the semi-structured process documents of the garment enterprise as the data source, Python is used to complete data cleaning, text normalization, and component category merging. The data is mapped to Style entity, Style component instance entity StylePart, Component entity Part, Fabric entity Fabric, Process entity Process, and Equipment entity Equipment, and triples are generated. Finally, 6668 process records, 197 components, 15 component categories, 24 types of equipment, 4 types of fabric, 282 component-level process route cases, and 8382 knowledge graph triples are obtained. Neo4j is used to store the triples, forming a process topology network containing 4006 nodes.

[0009] Preferably, step 2) uses 14,621 triples as training samples; the triple distance scoring function of the TransD model is: ;in, For triples Distance rating For the head entity, For tail entities, For relation vectors, For vector 2 norm operations, The head entity mapping vector; The tail entity is mapped to a vector; the lower the score, the higher the semantic rationality of the triple.

[0010] Preferably, the interval loss function used in step 2) is: In the formula: This represents the total loss value of the model. For positive sample distance scores, For negative sample distance scores, The interval parameter is used; model parameters are set as follows: embedding dimension 64, number of training epochs 80, batch size 512, learning rate 0.005. The training uses the Adam optimizer, and the entity and relation vectors are normalized after each training round. The dataset is divided by style number, with the training set containing 54 styles and the test set containing 14 styles.

[0011] Preferably, the comprehensive score function for candidate routes in step 3) is: In the formula: The overall score for candidate route c. , , The scores are based on style, components, and fabric matching, in that order. These are the business preference items for component name, component category, and fabric; in the formula, 1.6, 1.1, and 1.3 are weight values.

[0012] Preferred style matching score The fabric matching score is calculated based on the hasComponent relationship between the target style and candidate style component instances. The component matching score is calculated based on the isMadeOf relationship between candidate style component instances and input fabric. It is calculated from the hasComponent relationship between candidate style component instances and input components.

[0013] Preferably, in step 3), the candidate route with the highest comprehensive score is selected as the optimal reference route, and the process number, process description and processing equipment information in the route are directly reused without generating new process nodes.

[0014] Preferably, the method also includes an effect verification step: selecting a new garment that did not participate in the model training as the verification object, inputting information on fabric, component structure, and process features, and statistically analyzing route overlap rate, longest common subsequence similarity, and complete matching rate to evaluate the process route generation effect.

[0015] This invention discloses a knowledge graph-based garment process route generation system, comprising a knowledge graph construction module, an embedding learning module, a process matching and recommendation module, and a visualization module, with communication connections between the modules. The knowledge graph construction module is used for ontology modeling, data processing, and graph storage; the embedding learning module is used for TransD model training and vector extraction; the process matching and recommendation module is used for feature input, score calculation, and process route output; and the visualization module is used to display the knowledge graph, recommended process information, and similar historical cases. The knowledge graph construction module outputs structured triple data to the embedding learning module, and the embedding learning module outputs entity and relation vectors to the process matching and recommendation module. The process matching and recommendation module connects to the knowledge graph construction module and the visualization module, respectively. The visualization module receives the original knowledge graph data and the process recommendation results and presents them externally.

[0016] Beneficial effects: This invention constructs a full-dimensional knowledge graph of garment manufacturing processes, integrating knowledge across the entire chain of styles, components, fabrics, processes, and equipment. This solves the problems of fragmented traditional process knowledge and inefficient retrieval, enabling structured and computable management of process knowledge. By combining the TransD knowledge embedding model to mine implicit semantic relationships between entities, it achieves higher matching accuracy compared to traditional rule matching, enabling the transfer of process experience across style numbers and similar structures. The invention adopts a scheme that reuses historical process cases, rather than generating processes from scratch, adapting to the small-sample production scenarios of the garment industry and effectively avoiding the defects of chaotic process sequences and missing content in generative models. The system's modules have clear division of labor, smooth data flow, and a high degree of automation, significantly reducing the workload of process personnel, improving the efficiency of new garment process planning, and providing technical support for the digital and intelligent transformation of garment enterprises. Attached Figure Description

[0017] Figure 1 This is a diagram illustrating the construction and application architecture of the garment technology knowledge graph for this invention.

[0018] Figure 2 This is a partial topology graph of the knowledge graph corresponding to a certain type of trousers.

[0019] Figure 3 A schematic diagram of the interface for assisting in the planning of garment manufacturing processes. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Figure 1 The architecture is divided into three levels from top to bottom: application layer, construction layer and data layer, which fully reflects the whole process logic of this invention from the input of original process data source, knowledge graph construction to the implementation of process reuse: Data layer: is the source of original data input for this invention, including three types of basic data sources: clothing process information, clothing BOM material list and equipment resource information, covering the basic original data of the entire clothing production process, and providing original materials for the construction of the upper-level graph. Construction layer: is the core execution layer for knowledge graph generation, divided into three processing units: (1) Knowledge sorting and classification unit: completes information collection, concept classification, entity attribute definition and entity relationship definition in sequence, realizing the standardization of original unstructured / semi-structured process data; (2) Ontology construction unit: based on the sorted knowledge, defines entity type and relationship type, and finally generates a standardized ontology model in the field of clothing process, specifying the specifications of all entities and relationships in the graph; (3) Knowledge storage unit: uses Neo4j graph database, transforms standardized knowledge into triple network based on ontology model, and completes the instantiation and storage of clothing process knowledge graph. Application Layer: This is the final business implementation module of this invention. Based on the knowledge graph built from the bottom layer, it uses the TransD embedding model to calculate process similarity, enabling the reuse of garment processes based on process similarity, i.e., the function of assisting in the generation of new garment process routes. The three-layer hierarchical data flow is unidirectionally transmitted from bottom to top, with the upper-layer application inversely dependent on the lower-layer graph data.

[0022] Figure 2 This is a partial visualization of the knowledge graph corresponding to the trousers with style number S26118126K00 in the embodiment. It intuitively shows the actual storage structure of entities, object relationships, and triples within the graph of this invention: Entity type differentiation: Different colored circular nodes in the graph represent different categories of entities. Red nodes are style component instances (StylePart) (trousers - waistband - 2 pieces, single placket - lining 1 piece, double-embedded faux pockets); purple nodes are fabric entities (Fabric) (fine wool-like fabric); nodes are process entities (Process) (hand-stitched waistband hooks, concealed back waistband label, three-stitched front pocket fabric, etc.); light blue nodes are style entities (S26118126K00). The lines between nodes indicate the standard object relationships defined in this invention: isMadeOf: connects the style component entity and the fabric entity, indicating that the component uses the corresponding fabric; hasComponent: connects the style entity and all its subordinate style component instances, indicating that a garment contains multiple components; requiresProcess: connects the style component entity and all corresponding processing process entities, which is the core association relationship for process matching in this invention.

[0023] Figure 3This is the test interface for the effect verification step. The verification object is a new wide-leg trousers (style number S26318502K30) that did not participate in model training. The fabric category is worsted wool-like fabric. It is used to quantitatively prove the generation effect of the process route of this invention. It is divided into three main display areas: The outer new product summary area on the left displays the basic information of the new product to be tested, including style number, fabric category, total number of component-level route cases participating in the verification (8 groups in total), and lists four global quantitative evaluation indicators: average overlap rate 0.6822, average LCS similarity 0.646. 5. Weighted similarity rate: 0.6591; Weighted LCS similarity: 0.6364; Complete match rate: 0.25. The upper right overall verification index area summarizes the global statistical results of 8 process modules, intuitively reflecting the overall matching performance of this method. The lower right module-level verification result table area breaks down the independent matching effect of each process module by major component categories such as placket, waist loop, front piece, back piece, waistband, and pockets, including the actual number of process steps, process overlap rate, longest common subsequence (LCS) similarity, whether a complete match is achieved, and the best similar historical style number. The placket module has an overlap rate and LCS similarity of 1, achieving 100% complete process matching; the waist loop module has an overlap rate of 0.8571, with a high degree of matching in the core processes; complex modules such as multi-process assembly, back piece, waistband, and pockets have some process differences, but the core processing processes can be effectively matched.

[0024] Technical solution / content: A method for assisting in the generation of garment process routes based on knowledge graph embedding is divided into three core stages: garment process knowledge graph construction, TransD knowledge graph embedding learning, and process route assistance generation. The specific implementation steps are as follows.

[0025] I. Construction of Knowledge Graph for Garment Process Planning This stage adopts a hybrid construction strategy, primarily top-down and secondarily bottom-up, to sequentially complete ontology model construction, raw process data processing, triple generation, and graph visualization and storage.

[0026] 1.1 Ontology Model Construction This invention employs the seven-step ontology construction method proposed by Stanford University, combined with the Protégé 5.5 ontology editing tool, to complete the formal modeling of the ontology in the field of garment technology. Based on actual garment production operations, the technology knowledge is divided into three core subdomains: garment product and component domain, manufacturing resource domain, and process decision and workflow domain.

[0027] The Apparel Products and Components domain describes entities such as apparel styles, components, component categories, and fabrics; the Manufacturing Resources domain describes various processing equipment and the usage relationships between equipment and processes; and the Process Decisions and Flow domain describes processing procedures, complete process routes, and the sequential constraints between different processes.

[0028] In addition to defining entities and data attributes, this ontology model also formally defines process matching rules and process sequence rules, transforming traditional manual experience into a computable semantic structure. It also defines multiple object attributes to represent the semantic relationships between different entities. Specific object attributes and their meanings are as follows: `hasComponent`: represents the subordinate relationship between style and style components, and between style components and basic components; `isMadeOf`: represents the correspondence between style components and fabric; `isPartOf`: represents the belonging relationship between components and component categories; `requiresProcess`: represents the association between style components and required processing processes; `usesEquipment`: represents the usage relationship between processes and processing equipment; `isPrerequisiteFor`: represents the sequential constraint relationship between different processes.

[0029] 1.2 Data Layer Construction This invention uses semi-structured data from existing garment enterprises, such as style and process sheets, component process records, and fabric mapping documents, as its data source. A Python program is used to batch process the raw data. First, data cleaning and text standardization are performed, removing redundant spaces, special markers, quantity suffixes, and other non-standard content from process descriptions, and standardizing process names. Second, based on keyword rules, subdivided components are grouped into corresponding component categories, improving the comparability between different process examples.

[0030] The processed tabular data is mapped to standard entities, including Style, Style Part instance, Part, Fabric, Process, and Equipment entities. Combined with the object attributes defined in the ontology, standardized entity-relationship-entity triples are generated. After data processing and triple instantiation, 6668 process records, 197 parts, 15 part categories, 24 types of processing equipment, and 4 types of fabric information are obtained, forming 282 part-level process route cases, totaling 8382 knowledge graph triples.

[0031] Table 1. Definition of Core Object Attributes in the Clothing Technology Knowledge Graph hasComponent Style StylePart This indicates that the style contains instances of a certain style component. hasComponent StylePart Part Indicates the specific component type corresponding to the style component instance. isMadeOf StylePart Fabric Indicates the fabric type corresponding to the style component instance. isPartOf Part PartGroup This indicates the major component category to which a specific component belongs. requiresProcess StylePart Process This indicates the operations that need to be performed on a specific style of component. usesEquipment Process Equipment This indicates the processing equipment required for a specific process. isPrerequisiteFor Process Process This indicates the sequential dependency between processes within the same technological route. hasCharacteristic Style or Fabric Characteristic This indicates the characteristic information of the style or fabric. hasQualitativeDescription Style or Fabric Qualitative Description Text description information corresponding to style or fabric 1.3 Visualization and Storage of Atlases All standardized triples were imported into the Neo4j graph database to instantiate and store the knowledge graph, ultimately forming a garment process topology network with 4006 nodes. This topology network can intuitively display the complete processing chain of "style—component—fabric—process—equipment" and the sequential dependencies between processes, realizing both the visualization of process knowledge and providing standardized structured data for subsequent TransD model embedding learning.

[0032] II. TransD Knowledge Graph Embedding Representation Learning In this stage, the TransD model was selected to embed the knowledge graph. The TransD model can set independent projection vectors for entities and relations, which can adapt to the characteristics of multiple entities and multiple types of relations in the clothing technology knowledge graph, and achieve a good balance between the number of model parameters and semantic representation capabilities.

[0033] 2.1 Training Sample Construction All 14,621 triples in the knowledge graph were extracted as a positive sample set. During training, a negative sampling strategy was adopted. For each positive sample triple, its head or tail entity was randomly replaced to generate a negative sample triple that did not hold semantic logic. The positive and negative samples were combined to participate in the model training.

[0034] 2.2 Triple Distance Scoring Function The TransD model first maps the head and tail entities to the vector space corresponding to the relation, and then uses a distance scoring function to determine the semantic rationality of the triples. The formula for the distance scoring function is as follows: ;in, For triples Distance rating For the head entity, For tail entities, For relation vectors, For vector 2 norm operations, The head entity mapping vector; The tail entity mapping vector; the smaller the score value, the more reasonable the semantic logic of the corresponding triple; in the process route prediction stage, the negative number of the score is taken as the confidence score, which is used for sorting and screening candidate process routes.

[0035] 2.3 Interval Loss Function To constrain model convergence and ensure that the distance score of positive sample triples is less than that of negative sample triples, this invention employs an interval loss function. The formula for the single-sample loss function is as follows: In the formula: This represents the total loss value of the model. For positive sample distance scores, For negative sample distance scores, The interval parameter is used; model parameters are set as follows: embedding dimension 64, number of training epochs 80, batch size 512, learning rate 0.005. The training uses the Adam optimizer, and the entity and relation vectors are normalized after each training round. The dataset is divided by style number, with the training set containing 54 styles and the test set containing 14 styles.

[0036] 2.4 Model Training Parameters and Training Process The specific training parameters for the model in this invention are set as follows: embedding dimension 64, number of training epochs 80, batch size 512, learning rate 0.005, and interval parameter. The Adam optimizer was used during training. After each training round, all entity mapping vectors and relation vectors were normalized to improve the stability of model training.

[0037] The dataset was divided into two parts based on clothing style numbers: a training set containing 54 styles and a test set containing 14 styles. After model training, low-dimensional mapping vectors corresponding to all entities and relationships were saved to provide data support for subsequent process matching.

[0038] III. Similar Case Matching and Process Route Generation This stage uses the style, fabric, and component structural features of the new garment as input. A trained TransD model is used to calculate the comprehensive matching score of historical process cases, selecting the optimal case and reusing its process information to automatically generate the process route. This invention does not generate entirely new processes from scratch, but rather reuses mature historical process cases, avoiding the problems of process errors and omissions in small sample scenarios.

[0039] 3.1 Calculation of the overall score of candidate routes Construct a comprehensive score function for candidate process routes to complete multi-dimensional matching evaluation. The formula is as follows: In the formula: The overall score for candidate route c. , , The scores are, in order, for style, component, and fabric matching. In the formula, 1.6, 1.1, and 1.3 correspond to... This belongs to weight. This includes business preferences for component name, component category, and fabric.

[0040] 3.2 Process Route Output All historical candidate process routes are sorted in descending order based on their comprehensive scores, and the candidate route with the highest score is selected as the optimal reference route. All information contained in this route, including process numbers, process descriptions, and processing equipment, is directly reused to finally output a complete process route adapted to the new garment.

[0041] IV. Effect Verification New garments not used in model training were selected as validation subjects. Information such as fabric category, component structure, and special process features of the new garments were entered, and the proposed method was run to generate recommended process routes. The recommended routes were compared with the actual process routes of the new garments, and three indicators—route overlap rate, longest common subsequence similarity, and complete matching rate—were used to quantitatively evaluate the process route generation effect of the proposed method.

[0042] V. System Architecture and Module Connection Relationships To implement the above method, this invention also provides a knowledge graph-based garment process route generation system. The system consists of four main functional modules: a knowledge graph construction module, an embedding learning module, a process matching and recommendation module, and a visualization module. The functions, connections, and data flow relationships of each module are as follows: 5.1 Module Function Description (1) Knowledge graph construction module: responsible for completing ontology modeling, cleaning of raw process data, entity relationship mapping, triple generation, and graph storage and topology network construction based on Neo4j; (2) Embedded learning module: responsible for loading knowledge graph triple samples, performing negative sampling, TransD model training, distance scoring and loss calculation, and outputting entity and relation low-dimensional mapping vectors; (3) Process matching and recommendation module: responsible for receiving the feature input of new clothing, calling the mapping vector to complete the comprehensive score calculation, screening the optimal process route and outputting the planning results; (4) Visualization module: responsible for displaying the knowledge graph topology, recommended process sequences, processing equipment information, and matched similar historical process cases.

[0043] 5.2 Module Connection and Data Flow (1) The knowledge graph construction module serves as the system's data source, outputting standardized triplet data to the embedding learning module, while simultaneously pushing the original graph topology data to the visualization module. (2) After the embedded learning module completes the model training, it outputs the entity and relation mapping vectors to the process matching and recommendation module; (3) The process matching and recommendation module and the knowledge graph construction module interact bidirectionally to retrieve historical process case data in real time; at the same time, the final process recommendation results are pushed to the visualization display module. (4) The visualization module summarizes the map data, process recommendation data, and similar case data, and presents them to the outside world in a unified way.

[0044] Example 1: Construction and Implementation of a Knowledge Graph for Garment Technology Ontology Modeling: Using Protégé 5.5, an ontology for the garment technology domain was built, dividing the domain into three subdomains: garment products and components, manufacturing resources, and process decisions and procedures. Eight core object attributes were defined, and constraints such as process sequence, components and fabrics, and processes and equipment were solidified, transforming industry process experience into formal semantic rules.

[0045] Data processing: Collect actual production process documents from enterprises, totaling 6,668 process records, 197 components, 24 types of processing equipment, and 4 types of fabrics. Use Python scripts to complete data cleaning, text normalization, and component classification, ultimately generating 8,382 standard triples and 282 component-level process route cases.

[0046] Graph storage and visualization: Import all triples into the Neo4j graph database to generate a process topology network with 4006 nodes, which can intuitively display the complete process chain and process constraints of categories such as trousers and tops.

[0047] Example 2: Implementation of TransD Model Embedding Learning Sample preparation: 14,621 triples were extracted from the knowledge graph as positive samples. For each positive sample, the head entity or tail entity was randomly replaced to generate corresponding negative samples, and a training dataset was constructed.

[0048] Parameter configuration: Set the embedding dimension to 64, the number of training rounds to 80, the batch size to 512, the learning rate to 0.005, the interval parameter γ=1.0, and select the Adam optimizer.

[0049] Model training: The model underwent 80 rounds of iterative training. After each round, the entity mapping vectors and relation vectors were normalized. The dataset was divided according to style number, with 54 styles used as the training set and 14 styles as the test set. After training, a 64-dimensional mapping vector file of all entities and relations was saved.

[0050] Example 3: Generating and Verifying the Process Route of New Clothing. A new clothing item with style number S26318502K30 was selected as the verification object. This style was not used in the model training.

[0051] Input feature information: The fabric category is worsted wool-like fabric, including a 6-piece waistband, single placket, inner placket, J-shaped placket, back pocket, patch pocket, etc., with special processing features such as topstitching, zip stitching, and serger stitching. Feature analysis: The system extracts the input fabric, component, structural, and processing features to form structured matching conditions; Case matching and score calculation: Traverse 282 component-level historical process route cases and call TransD mapping vectors to calculate the comprehensive matching score of each candidate route; Process route output: Select the best case with the highest score and output the corresponding process sequence, process description and supporting processing equipment; Statistical Analysis and Results: This verification included 8 component / process stage route examples. The overall average route overlap rate was 0.6822, the average longest common subsequence similarity was 0.6465, the weighted route overlap rate was 0.6591, the weighted longest common subsequence similarity was 0.6364, and the complete matching rate was 0.25. The placket module achieved 100% complete matching; the waist loop module had a route overlap rate of 0.8571 and a longest common subsequence similarity of 0.7143, with highly consistent core processes; the multi-process assembly module, affected by process complexity, had an overlap rate of 0.5385, but still fully covered the core processing steps. The experimental results demonstrate that this invention can effectively assist in the generation of process routes for new garments, meeting the needs of actual production.

[0052] Finally, it should be noted that the present invention is not limited to the above embodiments, and many variations are possible. All variations that can be directly derived or conceived by those skilled in the art from the disclosure of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for assisting in the generation of garment process routes based on knowledge graph embedding, characterized in that, The method includes the following steps: Step 1) Construct a knowledge graph for garment process planning: Divide garment process knowledge into three major areas: product and component information, equipment resource information, and process decision and flow information. Adopt a hybrid strategy of top-down as the main approach and bottom-up as the supplementary approach to complete ontology modeling, process data processing, triple construction, and realize knowledge graph storage and visualization through graph database. Step 2) Knowledge graph embedding learning: Using knowledge graph triples as training samples, negative samples are constructed by combining a negative sampling strategy, and low-dimensional vectors of entities and relations are obtained by training the TransD model. Step 3) Assisted generation of process route: Input the style, fabric and component features of the new garment, calculate the comprehensive matching score of each historical process route based on the trained TransD model, select the optimal candidate route according to the score, reuse its process and equipment information, and output the process route of the new garment.

2. The method for assisting in the generation of garment process routes based on knowledge graph embedding according to claim 1, characterized in that, Step 2) Use 14,621 triples as training samples; the triple distance scoring function of the TransD model is: ;in, For triples Distance rating For the head entity, For tail entities, For relation vectors, For vector 2 norm operations, The head entity mapping vector; The tail entity is mapped to a vector; the lower the score, the higher the semantic rationality of the triple.

3. The method for assisting in the generation of garment process routes based on knowledge graph embedding according to claim 1 or 2, characterized in that, Step 2) uses the following interval loss function: In the formula: This represents the total loss value of the model. For positive sample distance scores, For negative sample distance scores, The interval parameter is used; model parameters are set as follows: embedding dimension 64, number of training epochs 80, batch size 512, learning rate 0.

005. ; The Adam optimizer was used for training, and the entity and relation vectors were normalized after each training round. The dataset was divided by style number, with the training set containing 54 styles and the test set containing 14 styles.

4. The method for assisting in the generation of garment process routes based on knowledge graph embedding according to claim 1, characterized in that, The comprehensive score function for candidate routes in step 3) is: In the formula: The overall score for candidate route c. , , The scores are based on style, components, and fabric matching, in that order. These are the business preference items for component name, component category, and fabric; in the formula, 1.6, 1.1, and 1.3 are weight values.

5. A method for assisting in the generation of garment process routes based on knowledge graph embedding, as described in claim 1 or 4, characterized in that, In step 3), the candidate route with the highest comprehensive score is selected as the optimal reference route. The process number, process description, and processing equipment information within this route are directly reused without generating new process nodes.

6. The method for assisting in the generation of garment process routes based on knowledge graph embedding according to claim 1, characterized in that, It also includes an effect verification step: select a new garment that did not participate in the model training as the verification object, input the fabric, component structure, and process feature information, and count the route overlap rate, the longest common subsequence similarity, and the complete matching rate to evaluate the process route generation effect.

7. A knowledge graph-embedded garment process route auxiliary generation system, characterized in that, The method for implementing any one of claims 1 to 6 includes a knowledge graph construction module, an embedding learning module, a process matching and recommendation module, and a visualization module, with communication connections between the modules; the knowledge graph construction module is used to complete ontology modeling, data processing, and graph storage; the embedding learning module is used to complete TransD model training and vector extraction; the process matching and recommendation module is used for feature input, score calculation, and process route output; the visualization module is used to display the knowledge graph, recommended process information, and similar historical cases; The knowledge graph construction module outputs structured triple data to the embedding learning module, and the embedding learning module outputs entity and relation vectors to the process matching and recommendation module. The process matching and recommendation module connects to the knowledge graph construction module and the visualization module, respectively. The visualization module receives the original knowledge graph data and the process recommendation results and presents them externally.

Images