Processing procedure optimization method and system based on process knowledge graph

Abstract

The invention relates to a processing procedure optimization method and system based on a process knowledge graph. The method comprises the following steps: performing preprocessing of standardized classification, data cleaning and normalization on entities of a process knowledge graph data set based on a structured process rule; classifying the structured process rules, converting the structured rules into text rules to form training data of a sequence-to-sequence model, and training the sequence-to-sequence model by using the training data, training a relational graph convolutional network model by using combined sample data of negative sample data output by the sequence-to-sequence model and positive sample data in the process knowledge graph; performing reasoning calculation on the received reasoning task based on the trained sequence-to-sequence model and the relational graph convolutional network model, and obtaining a comprehensive score, corresponding to the processing procedure, of the reasoning task; and optimizing the current processing technology based on the processing procedure obtained by reasoning.

Description

Technical Field

[0001] The present invention relates to the technical field of machining, and particularly to a machining process optimization method and system based on a process knowledge graph. Background Art

[0002] In the field of machining, process reasoning and planning involve using professional knowledge and a large amount of historical data to analyze, understand, and solve problems such as material processing, production processes, and quality control. Application scenarios of process reasoning and planning include recommending suitable machining processes for products or features, selecting appropriate manufacturing resources (such as tools, machine tools) for processes, and predicting equipment failures and performing maintenance, etc.

[0003] Traditional process reasoning methods mainly rely on rule bases or expert systems, which are based on predefined rule sets and logical reasoning, and the rules are manually written by experts. However, the machining process is complex and variable, involving numerous variables and uncertain factors, such as material properties, machining paths, equipment performance, etc., which makes rule definition extremely complex and difficult to comprehensively cover all possible situations. Once the rules are set, subsequent updates and expansions become difficult, resulting in insufficient flexibility of the system and difficulty in adapting to rapid process innovation and technological progress. In addition, traditional process reasoning methods highly rely on manually input rules and knowledge, lacking the ability of automatic learning and discovering new knowledge. Especially when facing a large amount of data, traditional process reasoning methods are difficult to effectively mine the potential patterns hidden therein.

[0004] The reasoning method based on a process knowledge graph combines machine learning and natural language processing technologies, can automatically extract knowledge from new data, and realize real-time update of knowledge. Different from traditional process reasoning methods, the knowledge graph does not rely on fixed rules, but can dynamically integrate new information, greatly improving the flexibility and scalability of the system. Through techniques such as graph embedding, the knowledge graph can make full use of a large amount of historical data, deeply analyze historical process data, and reveal hidden correlation patterns, which is difficult for traditional rule reasoning methods to achieve.

[0005] However, existing process reasoning methods based on process knowledge graphs still have some obvious limitations. First, the traditional node mapping-based reasoning method highly relies on node information in the existing knowledge graph, and for new parts or new processes not in the knowledge graph, this method cannot provide effective recommendations, limiting its applicability in dynamically changing scenarios. Second, node mapping reasoning usually relies on keyword matching but ignores the importance of node attributes. Even if the feature node names are the same, different attributes may correspond to completely different machining processes, thus affecting the accuracy of the derived process route.

[0006] More importantly, when dealing with process knowledge in current process reasoning research based on process knowledge graphs, only partial process content is involved, and the mining and utilization of text information and attribute information are not deep enough. In addition, the results of process reasoning must comply with the constraints in physical laws and engineering practices, but the existing research in this regard is not sufficient, resulting in certain limitations in the comprehensiveness and accuracy of process reasoning.

[0007] For example, CN119067028A discloses a method and system for optimizing integrated circuit process parameters based on machine learning. The method includes: obtaining process history data, constructing a process knowledge graph for the process to solve the semantic similarity between entities and relationships, extracting key process knowledge and constructing a prior knowledge base; determining the causal dependence relationship between process parameters and the yield rate, setting causal feature selection criteria, screening a subset of key causal features, expanding the process knowledge graph for the process, constructing a heterogeneous model integration and determining the dynamic changes of process parameters and the yield rate; constructing a multi-objective optimization model and setting optimization objectives, performing process parameter optimization to solve, generating optimized process parameters and performing transfer learning, comparing and mapping process nodes, constructing a virtual simulation system and performing causal intervention, generating process parameter optimization decisions and correcting the Bayesian causal model. This technical solution typically relies on historical data to optimize process parameters and cannot provide effective recommendations for new parts or new processes not in the graph; it relies on keyword matching but ignores the importance of node attributes; when dealing with process knowledge in the reasoning research, only partial content is involved, and the mining and utilization of text information and attribute information are not deep enough. This may lead to limitations in the comprehensiveness and accuracy of process reasoning.

[0008] CN118965970A discloses a method and system for simulating and optimizing casting process parameters based on a knowledge graph, including: constructing a simulation mathematical model; selecting variable parameters; constructing a process simulation database; constructing a casting process simulation knowledge graph; inputting the casting name and shrinkage hole diameter into the casting process simulation knowledge graph for retrieval to obtain simulation optimization data; using the simulation optimization data to re-perform casting simulation on the casting until the best casting process parameters are obtained, and storing the simulation optimization data corresponding to the best casting process parameters, as well as the corresponding casting name and shrinkage hole diameter, into the process simulation database to form a new casting process simulation knowledge graph.

[0009] This technical solution highly relies on the node information in the existing knowledge graph. When encountering new castings or shrinkage cavity diameters that have never been seen before, the existing knowledge graph may not be able to provide sufficient information for effective recommendation and optimization. Although the patent mentions storing new simulation optimization data into the process simulation database to form a new casting process simulation knowledge graph, this is still a process of post-update and cannot meet the requirements of new parts or new processes in real time. Therefore, when dealing with a dynamically changing production environment, the effectiveness and flexibility of this technical solution are limited to a certain extent.

[0010] When constructing the casting process simulation knowledge graph, this technical solution mainly relies on the casting name and shrinkage cavity diameter as the retrieval keywords. However, in the actual casting process, the attributes of the nodes (such as material properties, temperature, pressure, cooling rate, etc.) have an important impact on the process results. Even if the names of the nodes with different attributes are the same, they may correspond to completely different processing procedures. For example, even if the casting names are the same, if different materials are used or the pouring temperature is different, it may lead to completely different shrinkage cavity diameters and casting qualities. Therefore, the practice of relying only on keyword matching and ignoring attribute information may result in inaccurate derivation of the process route. To improve the accuracy of reasoning, the system should more deeply explore and utilize the attribute information of the nodes to ensure that the reasoning results can reflect the requirements of the actual process.

[0011] In summary, although the reasoning method based on the process knowledge graph has, to a certain extent, solved the limitations of the traditional method, it still needs to be further improved in dealing with new parts, new processes, and ensuring that the reasoning results meet the requirements of actual engineering. The present invention hopes to provide a processing procedure optimization method and system based on the process knowledge graph, which can further improve the current machining process.

[0012] In addition, on the one hand, there are differences in the understanding of those skilled in the art; on the other hand, although the applicant has studied a large number of documents and patents when making this invention, due to space limitations, all details and contents are not listed in detail. However, this does not mean that this invention does not possess the features of these prior arts. On the contrary, this invention already possesses all the features of the prior arts, and the applicant reserves the right to add relevant prior arts in the background art. Summary of the Invention

[0013] Currently, there are several deficiencies in the technical solutions for process reasoning using process knowledge graphs. The initial problem is that traditional node mapping reasoning techniques highly rely on the node data existing in the knowledge graph. This means that for new components or new processes that are not included, it is difficult for the system to generate effective suggestions, thus limiting the application scope of this method in a rapidly changing environment. In addition, such reasoning methods often focus on keyword matching while ignoring the crucial role of node attributes. Even if the names of feature nodes are the same, if their attributes are different, they may be associated with completely different manufacturing processes, which undoubtedly affects the accuracy of the derived process paths.

[0014] Further analysis shows that in existing research, the process reasoning work based on process knowledge graphs does not comprehensively handle process content and fails to fully exploit the value of text information and attribute information. At the same time, to ensure the practical feasibility of process reasoning results, physical principles and engineering operation specifications must be strictly followed, but the current research still needs to be strengthened in this regard, which directly restricts the integrity and reliability of process reasoning results.

[0015] In view of the deficiencies of the existing technology, the present invention provides, from the first aspect, a method for optimizing processing procedures based on a process knowledge graph. The method includes: performing preprocessing of standard classification, data cleaning, and normalization on the entities of the process knowledge graph data set based on structured process rules; classifying the structured process rules and converting the structured rules into text rules to form training data for a sequence-to-sequence model, training the sequence-to-sequence model using the training data, and training a relational graph convolutional network model using a combined sample data of negative sample data output by the sequence-to-sequence model and positive sample data in the process knowledge graph; performing inference calculations on the received inference task based on the trained sequence-to-sequence model and relational graph convolutional network model, and obtaining a comprehensive score corresponding to the processing procedure of the inference task; and optimizing the current processing technology based on the inferred processing procedure.

[0016] By performing preprocessing of standard classification, data cleaning, and normalization on the entities of the process knowledge graph data set based on structured process rules, this method can effectively handle new parts or new processes that are not in the existing knowledge graph. This method not only enhances the model's learning ability for unknown entities but also improves its ability to adapt to dynamically changing scenarios. In addition, by using the combination of a sequence-to-sequence model and a relational graph convolutional network model for inference calculations, it can more accurately capture the complex relationships between different nodes, thereby improving the accuracy of the derived processing procedures.

[0017] According to a preferred embodiment, the steps of preprocessing the entities in the process knowledge graph dataset include: dividing the numerical range of the entities based on structured process rules to achieve standardized classification; traversing the triples in the process knowledge graph dataset, identifying and deleting the entities with missing data items to achieve data cleaning; normalizing the numerical attributes of the process knowledge graph dataset to a predefined value range, and constructing a numerical identification table corresponding to the value range.

[0018] Dividing the numerical range of the entities in the process knowledge graph dataset to achieve standardized classification, and identifying and deleting the entities with missing data items by traversing the triples in the process knowledge graph dataset, which ensures the quality and consistency of the data. The normalization process of numerical attributes enables data of different scales to be compared and analyzed in a unified framework, further improving the generalization ability and robustness of the model. At the same time, constructing a numerical identification table corresponding to the value range helps to improve the efficiency and accuracy of subsequent processing steps.

[0019] According to a preferred embodiment, the steps of converting structured rules into text rules to form the training data of the sequence-to-sequence model include: after classifying the structured process rules, traversing the dataset of structured process rules and extracting the attribute information of the structured process rules; replacing the value range with identifiers according to the numerical identification table, thereby converting the structured process rules into text rules and completing the attribute annotation of the structured process rules at the same time; calculating the Jaccard similarity coefficient according to the attributes of the process knowledge graph nodes and the attributes of the rule categories corresponding to the inference tasks. Converting the structured process rules into text rules and enhancing the understanding of node attributes through Jaccard similarity coefficient calculation significantly improves the depth and breadth of information mining in the process knowledge graph. This conversion not only retains the core content of the original structured rules, but also makes them easier to be understood and applied by machine learning algorithms. More importantly, this method can better consider the importance of node attributes, thus avoiding misjudgment problems caused by keyword matching.

[0020] According to a preferred embodiment, the steps of preprocessing the entities in the process knowledge graph dataset further include: screening the features and processing procedures of the inference tasks that do not exist in the process knowledge graph, and importing the information of the screened processing procedures into the process knowledge graph to form new process knowledge graph nodes; this process greatly expands the coverage of the knowledge graph and makes it more comprehensive.

[0021] Convert the attributes of entities in the process knowledge graph into node representations in the process knowledge graph; initialize the feature vectors of the nodes in the process knowledge graph based on a pre-trained word embedding model, and construct an adjacency matrix of the feature vectors. This provides strong support for the subsequent relational graph convolutional network, enabling the network to more accurately simulate the physical laws and constraints in actual engineering, thus ensuring the authenticity and reliability of the inference results. According to a preferred embodiment, the steps of performing inference calculation on the received inference task based on the trained sequence-to-sequence model and the relational graph convolutional network model include: inputting the entities in the form of coherent sentences formed by converting the inference task into the sequence-to-sequence model; the sequence-to-sequence model assigns probability scores to the decoded entities, and forms candidate triples by combining the entities with the top-k probability scores with the original triples, and the sequence-to-sequence model delivers the candidate triples to the relational graph convolutional network model, and the relational graph convolutional network model outputs scores corresponding to the candidate triples; calculate the comprehensive score corresponding to the processing operation based on the probability scores output by the sequence-to-sequence model and the scores output by the relational graph convolutional network model.

[0022] The present invention inputs the inference task in the form of coherent sentences into the sequence-to-sequence model and combines the relational graph convolutional network model to output scores. This method can not only effectively capture the relationships between entities, but also comprehensively consider various factors to give the optimal solution. The comprehensive score calculated based on the probability scores and the scores provides a clear and intuitive operation guide for users to help optimize the current processing process.

[0023] The present invention provides a processing operation optimization system based on a process knowledge graph from a second aspect. The system includes a processor, and the processor includes a preprocessing unit, an inference unit, and an optimization unit.

[0024] The preprocessing unit includes a standardization classification module, a rule stratification module, and a conversion module. The standardization classification module performs preprocessing of standardization classification, data cleaning, and normalization on the entities in the process knowledge graph data set based on structured process rules. The rule stratification module classifies the structured process rules and converts the structured rules into text rules to form the training data of the sequence-to-sequence model. The conversion module converts the triples in the process knowledge graph into entities in the form of coherent sentences based on a conversion template, and the entities in the form of coherent sentences are the training data of the sequence-to-sequence model.

[0025] The preprocessing unit performs standardized classification, data cleaning, and normalization on the entities in the dataset through the standardized classification module. This process utilizes the powerful computing capabilities of the processor, enabling a large amount of data to be effectively organized in a short period of time and ensuring the data quality for subsequent model training. The rule layering module converts structured process rules into text form and generates training data. This step achieves high-throughput data conversion with the help of a high-performance processor, reducing the latency time. The conversion module is responsible for converting triple information into coherent sentence form, further enhancing the consistency and usability of the data.

[0026] The inference unit includes a sequence-to-sequence model and a relational graph convolutional network model. The sequence-to-sequence model is trained based on the training data and outputs negative sample data. The relational graph convolutional network model is formed in a way that it is trained using a combined sample data of the negative sample data output by the sequence-to-sequence model and the positive sample data in the process knowledge graph. Among them, the trained sequence-to-sequence model and relational graph convolutional network model perform inference calculations on the received inference task and obtain a comprehensive score corresponding to the processing procedure of the inference task.

[0027] The inference unit contains a sequence-to-sequence model and a relational graph convolutional network model, which jointly act on the received inference task. The parallel computing ability and cache mechanism of the processor enable the sequence-to-sequence model to quickly complete the training of large-scale data and output negative sample data. The relational graph convolutional network model uses these negative sample data in combination with positive sample data for deeper learning, thereby improving the model's ability to capture complex relationships. Such a design not only speeds up the system's response speed but also enhances the model's prediction accuracy.

[0028] The optimization unit optimizes the current processing technology based on the processing procedure obtained from the inference.

[0029] The optimization unit uses the high-efficiency computing performance of the processor to evaluate multiple processing schemes and select the optimal solution. This method not only improves production efficiency but also reduces resource consumption. In addition, the optimization unit can dynamically adjust the strategy according to changes in production requirements, maintaining the flexibility of production.

[0030] According to a preferred embodiment, the standardized classification module is configured to: divide the numerical range of entities based on structured process rules to achieve standardized classification; traverse the triples in the process knowledge graph dataset, identify and delete entities with missing data items to achieve data cleaning; normalize the numerical attributes of the process knowledge graph dataset to a predefined value range, and construct a numerical identification table corresponding to the value range.

[0031] The standardized classification module in this configuration can effectively improve the consistency and accuracy of data. Through the division of numerical ranges, it can ensure that data from different sources are compared and analyzed under the same standard. In addition, the data cleaning process can effectively remove incomplete or incorrect information, avoiding interference from this information to the inference unit. The normalization step ensures that all numerical attributes are on the same scale, facilitating model learning and prediction. At the same time, the constructed numerical identification table provides the ability for quick query, improving the processing efficiency.

[0032] According to a preferred embodiment, the rule layering module is configured to: after classifying the structured process rules, traverse the dataset of the structured process rules and extract the attribute information of the structural process rules; replace the value range with an identifier according to the numerical identification table, thereby converting the structured process rules into text rules and completing the attribute annotation of the structured process rules at the same time; calculate the Jaccard similarity coefficient according to the attributes of the process knowledge graph nodes and the attributes of the rule categories corresponding to the inference tasks.

[0033] The design of the rule layering module enables complex process rules to be systematically organized and parsed. The converted text rules are easier to understand and process, especially in tasks involving natural language processing. The attribute annotation function enhances the interpretability of the rules, enabling users to better understand the application scenarios of each rule. The process of calculating the Jaccard similarity coefficient helps to identify and associate process knowledge graph nodes with similar characteristics, thereby improving the working efficiency and accuracy of the inference unit.

[0034] According to a preferred embodiment, the conversion module is configured to: screen the features and processing procedures of the inference tasks that do not exist in the process knowledge graph, and import the information of the screened processing procedures into the process knowledge graph to form new process knowledge graph nodes; convert the attributes of the entities in the process knowledge graph into node representations in the process knowledge graph; initialize the feature vectors of the process knowledge graph nodes based on a pre-trained word embedding model, and construct an adjacency matrix of the feature vectors.

[0035] The function of the conversion module is to dynamically expand the content of the process knowledge graph, enabling it to adapt to new processing requirements and technological developments. By screening and importing new information, the knowledge base can be continuously enriched and improved, enhancing the flexibility and adaptability of the system. Initializing the node feature vectors using a pre-trained word embedding model can capture the semantic relationships between entities, while the construction of the adjacency matrix provides the necessary input form for the relational graph convolutional network (R-GCN) model, facilitating the effective application of deep learning algorithms.

[0036] According to a preferred embodiment, the inference unit is configured as follows: the sequence-to-sequence model receives entities in the form of coherent sentences transformed from the inference task, assigns probability scores to the decoded entities, forms candidate triples by combining the entities with the top-k probability scores with the original triples; the sequence-to-sequence model delivers the candidate triples to the relational graph convolutional network model; the relational graph convolutional network model outputs scores corresponding to the candidate triples; and calculates a comprehensive score corresponding to the processing procedure based on the probability scores output by the sequence-to-sequence model and the scores output by the relational graph convolutional network model.

[0037] This configuration of the inference unit allows it to handle highly complex and variable processing procedure optimization problems. The sequence-to-sequence model is good at capturing the mapping relationship between input and output, especially performing well when dealing with process rules in natural language form. The present invention combined with the R-GCN model can deeply explore the relationships between entities, thereby providing more accurate scoring results. This method not only improves the accuracy and robustness of inference, but also enhances the generalization ability of the system for unseen data. The finally obtained comprehensive score can help enterprises make more scientific and reasonable decisions, improving production efficiency and product quality. Brief Description of the Drawings

[0038] Figure 1 is a schematic diagram of the simplified module connection relationship of the processing procedure optimization system based on the process knowledge graph provided by the present invention;

[0039] Figure 2 is a partial rule schematic diagram of a hole processing method provided by the present invention;

[0040] Figure 3 is a schematic diagram of the inference process of a tapered hole process provided by the present invention;

[0041] Figure 4 is a schematic diagram of the process of the processing procedure optimization method based on the process knowledge graph provided by the present invention.

[0042] List of Reference Numerals

[0043] 100: Processor; 110: Preprocessing Unit; 111: Standardization Classification Module; 112: Rule Hierarchy Module; 113: Conversion Module; 120: Inference Unit; 121: Sequence-to-Sequence Model; 122: Relational Graph Convolutional Network Model; 130: Optimization Unit. Detailed Embodiment

[0044] The following is a detailed description with reference to the drawings.

[0045] The present invention explains some noun terms.

[0046] Process Knowledge Graph: In the field of machining technology, it is a structured model used to describe and manage various types of knowledge involved in the machining process. It organizes and expresses the entities (such as machine tools, cutting tools, workpieces, materials, etc.) in machining and the relationships between them in the form of a graph, forming a comprehensive knowledge network. The Process Knowledge Graph not only records specific machining steps and technical parameters but also includes related engineering principles, operation experiences, fault diagnosis methods, etc.

[0047] Process Knowledge Graph Dataset: It is to extract the entities and relationships of the Process Knowledge Graph from various specific machining process manuals through knowledge extraction means. Specifically, knowledge extraction tools are used to identify various entities (such as features, processes, cutting tools, machine tools, etc.) in the manuals and the relationships between them (such as "has process", "uses cutting tool"), thereby forming structured triples. Based on these entities and the relationships between them, the Process Knowledge Graph is constructed.

[0048] Process Rule Data: Its main sources include industry standards related to manufacturing processes and expert experiences. Part of the data of industry standards usually comes from specification documents issued by authoritative organizations such as ISO, ANSI, ASTM, etc. These documents specify in detail elements such as material selection, machining parameters, quality inspection standards, etc., providing a scientific basis and consistency guidelines for process design and implementation. The rules covered in the content of industry standards often include key parameters such as specific temperature, pressure, speed, etc., as well as strict requirements for process quality, which are the core sources of the process rule system. At the same time, process rule data also comes from the rich experiences accumulated by experts in actual production. The rules in industry standards and expert experiences are manually collected, sorted, and stored to form the process rule database used in the present invention.

[0049] Sequence-to-Sequence Model 121 (abbreviated as Seq2Seq model): It usually consists of two parts: an encoder and a decoder. The encoder is responsible for receiving the input sequence and converting it into one or more context vectors, which contain all the information of the input sequence. The decoder then generates the output sequence step by step based on this context vector. Specifically, the encoder reads the elements in the input sequence one by one, represents them as one or more hidden state vectors through an RNN or other types of recurrent networks, and finally uses these hidden state vectors as the initial state of the decoder. The decoder predicts the next output element step by step based on these hidden state vectors, combined with the previously generated output elements, until the complete output sequence is generated.

[0050] Relational Graph Convolutional Networks (R-GCN) model 122: A graph neural network model for processing multi-relational data. The core idea of the R-GCN model is to aggregate neighbor information according to different types of edges between nodes in each layer, thereby generating richer node representations. Specifically, for a given node v i , the R-GCN model will consider all different types of relationships r connected to it, and perform a weighted sum on the set of neighbor nodes under each relationship, and then aggregate these results as the new representation of the node. This mechanism allows the R-GCN model to not only consider directly connected neighbors when updating node features, but also distinguish the impacts brought by different types of edges.

[0051] Example 1

[0052] Existing process inference methods based on process knowledge graphs still have obvious limitations when applied to the machining process.

[0053] First of all, traditional node mapping inference methods highly rely on node information in the existing graph. It is difficult to provide effective suggestions for new components or processes not in the graph, which limits their applicability in a dynamically changing environment. Specifically, when encountering new components or new machining processes that are not recorded, these methods cannot make reasonable inferences based on existing data, resulting in insufficient flexibility in practical applications.

[0054] Secondly, node mapping inference usually relies on keyword matching and ignores the importance of node attributes. Even if two nodes have the same name, the differences in their attributes may lead to completely different machining steps, thus affecting the accuracy of the process route. For example, the same "drilling" operation may require different tools and parameter settings due to differences in material hardness, size, or precision requirements. Therefore, inference methods relying solely on name matching may lead to incorrect process decisions.

[0055] More importantly, current process inference research often only covers part of the process knowledge when dealing with process knowledge, and the mining and utilization of text and attribute information are not deep enough. Many methods focus on extracting information from structured data, while paying less attention to unstructured text descriptions and implicit knowledge in technical documents. This limitation makes it difficult for process inference systems to comprehensively understand complex process flows, thereby affecting the accuracy and reliability of their inference results.

[0056] In addition, the results of process reasoning must comply with physical laws and various constraints in engineering practice, such as material properties, equipment capabilities, safety standards, etc. However, existing methods are insufficient in this regard and lack effective modeling and verification mechanisms for these constraints. This not only affects the comprehensiveness of process reasoning but also reduces its application value in actual production. For example, some reasoning results may be theoretically feasible but violate physical laws or engineering specifications in actual operation, leading to unforeseen problems.

[0057] Aiming at the deficiencies of the prior art, the present invention provides a processing procedure reasoning method and system based on a process knowledge graph. Further, the present invention also provides a processing procedure optimization method and system based on a process knowledge graph, and the present invention can also provide a processing procedure anomaly detection method and system based on a process knowledge graph.

[0058] The system of the present invention includes a processor 100 that executes the method of the present invention. Preferably, the processor 100 may be an application-specific integrated chip, a server, a cloud server, or a combination thereof.

[0059] Preferably, as Figure 1 shown, a preprocessing unit 110, an inference unit 120, and an optimization unit 130 are provided inside the processor 100 of the present invention. A connection is established between the preprocessing unit 110 and the inference unit 120 to achieve data transmission. A connection is established between the inference unit 120 and the optimization unit 130 to achieve data transmission. When the preprocessing unit 110, the inference unit 120, and the optimization unit 130 are integrated inside the processor 100, the preprocessing unit 110, the inference unit 120, and the optimization unit 130 are connected through an internal bus.

[0060] Preferably, the preprocessing unit 110 is used for preprocessing data.

[0061] The preprocessing unit 110 includes a standardization and classification module 111, a rule stratification module 112, and a conversion module 113. A connection is established between the standardization and classification module 111 and the rule stratification module 112 to achieve data transmission. A connection is established between the rule stratification module 112 and the conversion module 113 to achieve data transmission.

[0062] The standardization and classification module 111 is used for processing the attribute information of entities in the process knowledge graph based on process rules. Preferably, the standardization and classification module 111 is mainly used for standardizing and classifying numerical information.

[0063] The rule stratification module 112 is used for performing rule stratification processing on process rules based on the inference task and the attribute information involved in the process rules.

[0064] The conversion module 113 is used to preset a conversion template, incorporate text information such as entity attributes, entity types, and related rules into the conversion template, and convert the structured knowledge graph into a text representation. The text representation is used as the input of the inference unit 120 for subsequent training and inference.

[0065] The inference unit 120 is used to train a sequence-to-sequence model 121 and a relational graph convolutional network model 122 on the basis of data preprocessing, and output an inference result by synthesizing the scores of the two.

[0066] The inference unit 120 is provided with a sequence-to-sequence model 121 and a relational graph convolutional network model 122. The sequence-to-sequence model 121 and the relational graph convolutional network model 122 input the processed negative sample data into the relational graph convolutional network model 122.

[0067] The optimization unit 130 is used to optimize and update the current processing procedure.

[0068] The connection port of the processor 100 of the present invention is connected to a terminal or a third-party system for receiving process knowledge. The connection port of the processor 100 is also connected to a machining system, and is used for the optimization unit 130 to communicate with the control unit of the machining system and update the processing procedure.

[0069] The machining process optimization method based on the process knowledge graph executed by the processor 100 of the present invention is as follows.

[0070] S100: Perform preprocessing of data.

[0071] The steps executed by the standardization and classification module 111 are as follows.

[0072] S110: Perform preprocessing of standardizing classification, data cleaning, and normalization on the entities of the process knowledge graph dataset based on structured process rules.

[0073] S111: Divide the numerical range of entities based on structured process rules to achieve standardization classification.

[0074] Figure 2 Shows a part of the structured process rules for the machining method of holes.

[0075] Figure 2 The table in details the selection rules for the hole machining method, that is, different machining procedures are selected based on different diameters (D), machining accuracies (IT), and surface roughnesses (Ra). The first column on the left of the table represents the value ranges of the machining accuracy and surface roughness of the hole, the second column represents the diameter range of the hole, and the first and second rows of the third and fourth columns represent whether there are reserved holes. The third to seventh rows of the third and fourth columns list the machining procedures under different attributes of the hole.

[0076] According to the above Figure 2 structured process rules, numerical ranges are divided for numerical attributes such as machining accuracy (IT), surface roughness (Ra), diameter (D), etc.

[0077] For example, the machining accuracy (IT) can be divided into three numerical ranges: IT≥10, 9≤IT<10, 7≤IT<8. The surface roughness (Ra) can be divided into three numerical ranges: Ra≥6.3, 3.2≤Ra<6.3, 0.8≤Ra<1.6. Since there is an overlap in the regular ranges of the diameter attribute, a more refined division is required to ensure mutually exclusive ranges when dividing the ranges. The diameter (D) is divided into five value ranges: D<6, 6≤D<10, 10≤D<12, 12≤D<20, 20≤D<30. This more refined division can avoid the ambiguity caused by overlap and ensure that each range of the numerical attribute does not overlap.

[0078] Figure 2 The specific rules are as follows:

[0079] When IT≥10 and Ra≥6.3:

[0080] If the diameter D<30: With pre - drilled hole: 1 - reaming; Without pre - drilled hole: 1 - drilling.

[0081] If the diameter D≥30: With pre - drilled hole: 1 - rough boring (turning); Without pre - drilled hole: 1 - drilling - rough boring (turning); 2 - drilling - reaming.

[0082] When 9≤IT<10 and 3.2≤Ra<6.3:

[0083] If the diameter D<10: With pre - drilled hole: 1 - reaming - broaching; Without pre - drilled hole: 1 - drilling - broaching.

[0084] If the diameter 10≤D<30: With pre - drilled hole: 1 - reaming - enlarging; Without pre - drilled hole: 1 - drilling - enlarging.

[0085] If the diameter D≥30: With pre - drilled hole: 1 - rough boring (turning); Without pre - drilled hole: 1 - drilling - rough boring (turning); 2 - drilling - reaming.

[0086] When 7≤IT<8 and 0.8≤Ra<1.6:

[0087] If the diameter D<6: With pre - drilled hole: 1 - reaming - finish broaching; Without pre - drilled hole: 1 - drilling - finish broaching.

[0088] If the diameter 6≤D<12: With pre - drilled hole: 1 - reaming - rough broaching - finish broaching; Without pre - drilled hole: 1 - drilling - rough broaching - finish broaching.

[0089] If 12 ≤ D < 20 in diameter: there is a reserved hole: 1 - drill and ream - ream - rough ream - finish ream; without a reserved hole: 1 - drill - ream - rough ream - finish ream.

[0090] S112: Traverse the triples in the process knowledge graph dataset, identify and delete entities with missing data items to achieve data cleaning.

[0091] Specifically, traverse the triples in the process knowledge graph dataset one by one, identify data items with missing head entities, relationships, or tail entities, and delete them from the dataset to ensure data integrity. Based on the cleaned data, obtain the training triple set from the knowledge graph. Each training triple includes an entity pair consisting of a source entity and a target entity corresponding to the source entity, and the relationship between the source entity and the target entity.

[0092] S113: Normalize the numerical attributes of the process knowledge graph dataset to a predefined value range, and construct a numerical identification table corresponding to the value range.

[0093] Perform standardization processing on the numerical attributes of the process knowledge graph nodes, and normalize them to a predefined value range. Each value range is mapped through a unique identifier. For example, map the three value ranges of 9 ≤ IT < 10, 3.2 ≤ Ra < 6.3, and 0.8 ≤ Ra < 1.6 to [IT1], [Ra1], and [Ra2] respectively to form a numerical identifier table based on the numerical range.

[0094] S120: Classify the structured process rules and convert the structured rules into text rules to form the training data of the sequence-to-sequence model 121.

[0095] S121: Perform classification of structured process rules.

[0096] Classify the rules in the process rule knowledge base according to the goal of the reasoning task and the type of the node to be reasoned. In a possible embodiment, according to the two reasoning tasks targeted by the present invention: reasoning the corresponding processing procedures for process features and reasoning suitable manufacturing resources (such as tools and machine tools) for the processing procedures, classify the rules at multiple levels.

[0097] ① Process reasoning task

[0098] The query form is: (feature, has process,?). Define process reasoning class rules in this task, and subdivide them according to different subclasses of features (such as holes, grooves, threads, etc.):

[0099] Hole process reasoning rules: Used for reasoning of hole features (such as drilling, reaming, finishing reaming, etc.);

[0100] Groove process reasoning rules: Applicable to reasoning of groove features (such as grooving, milling grooves, etc.);

[0101] Thread process inference rules: used for process inference of thread features (such as thread machining).

[0102] In the process inference task, according to the specific category of features (such as holes, grooves or threads), the corresponding class of rules is dynamically selected and added to the query to improve the accuracy and applicability of the inference.

[0103] ② Manufacturing resource inference task

[0104] The query form is: (process, using tool,?) or (process, using machine tool,?). In this task, the tool inference rules and machine tool inference rules are defined respectively, and further subdivided according to different subclasses of the process (such as drilling, reaming, boring, etc.). According to the category of the process, the tool inference rules are subdivided into multiple subclasses: drill tool inference rules, reaming tool inference rules, boring tool inference rules, etc.; the machine tool inference rules are subdivided according to the process type, including drill machine tool inference rules, reaming machine tool inference rules, boring machine tool inference rules, etc.

[0105] S122: After classifying the structured process rules, traverse the dataset of the structured process rules and extract the attribute information of the structural process rules.

[0106] In a specific embodiment, for the rule: when the hole machining accuracy is [IT 1 , the diameter is [D 1 ， the surface roughness is [Ra 1 , and there is no reserved hole, the machining process is drill-ream. Mark its relevant attributes as {machining accuracy, diameter, surface roughness, whether there is a reserved hole}.

[0107] S123: Replace the value range with an identifier according to the numerical identification table, so as to convert the structured process rules into text rules, and at the same time complete the attribute annotation of the structured process rules.

[0108] In step S111, the Figure 2 machining accuracy (IT), surface roughness (Ra), and diameter (D) attributes are divided, and the attributes are converted into [IT 1 , [IT 2 , [IT 3 , [Ra 1 , [Ra 2 , [Ra 3 , [D 1 , [D 2 , [D 3 , [D 4 , [D 5 in the numerical identification table. Figure 2The structured process rules extracted are converted into text rules. For example, when the hole has 7 ≤ IT < 8 and 0.8 ≤ Ra < 1.6, if the hole diameter D < 6, there is a reserved hole, and the processing procedures are: drill and ream; according to the content of the text rules, the attributes related to the rules can be manually marked as {processing accuracy, diameter, surface roughness, whether there is a reserved hole}. According to the numerical identification table, the value range is replaced with identifiers, and the text rules are obtained after conversion. When the hole processing accuracy is [IT 3 , the surface roughness is [Ra 3 , the diameter is [D 1 , and there is a reserved hole, the processing procedures are drill and ream.

[0109] S124: Calculate the Jaccard similarity coefficient according to the attributes of the process knowledge graph nodes and the attributes of the rule categories corresponding to the reasoning tasks.

[0110] After completing the attribute marking of the process structure rules, judge the reasoning task category according to the type of the node to be reasoned, and select the corresponding category of structured process rules. For example, for the reasoning task (tapered hole, with procedures,?). According to the definition in step S121, this query is classified as a "procedure reasoning task". Since the tapered hole belongs to the hole type feature node, according to the subdivision of the "procedure reasoning task" in step S121, the "hole procedure reasoning rule category" should be selected as the most matching rule type for the current reasoning task. After determining the most relevant rule category, use the Jaccard similarity coefficient to further screen these rules.

[0111] The calculation formula of the Jaccard similarity coefficient is:

[0112]

[0113] Among them, N a represents the attribute set of the node, R a represents the attribute set related to the category of the structured process rules, |N a ∩R a | represents the size of the intersection of the two sets, |N a ∪R a | represents the size of the union of the two sets.

[0114] For example, the tapered hole node in this reasoning task includes four attributes: {processing accuracy, surface roughness, diameter, whether there is a reserved hole}. Considering two candidate rules in the subclass, that is, the first rule "when the hole processing accuracy is [IT 1 , the diameter is [D 1 , the surface roughness is [Ra 1 , and there is no reserved hole, the processing procedures are drill and expand" and the second rule "when the hole surface roughness is [Ra 2 , the radius is [HD1 , with a depth of [Dp 1 is a drilling process. According to the method described in S122, the two sets of relevant attributes of the rules are {processing accuracy, diameter, surface roughness, whether there is a reserved hole} and {surface roughness, radius, depth} respectively.

[0115] The relevance between the rule and the current inference task can be judged by the Jaccard similarity coefficient. For the conical hole node, its attributes are exactly the same as those of the first rule (both the intersection and the union are 4), so the Jaccard similarity coefficient between the two is 1. Compared with the second rule, there is only one common point, surface roughness (the size of the intersection is 1, while the size of the union is 6), so the Jaccard similarity coefficient is 1 / 6. Thus, it can be seen that the first rule is highly relevant to the current task, while the second rule has a lower relevance to the current task.

[0116] To accurately match the rule most relevant to the current inference task, through experimental adjustment and verification, a threshold w = 0.7 is set. This threshold is not too loose to introduce too much noise, nor too strict to miss important rules. Any rule with a calculated Jaccard similarity exceeding this threshold will be considered relevant to the current inference task. Finally, all the rules identified as relevant will be integrated and used as the rule information Rules in the subsequent triple represented by coherent sentences and added to the input information of the sequence-to-sequence model 121.

[0117] S130: The conversion module 113 converts the triples in the process knowledge graph into entities in the form of coherent sentences based on the conversion template. The entities in the form of coherent sentences are the training data of the sequence-to-sequence model 121.

[0118] The preset conversion template is:

[0119] (Head,Relation,?) → Head name is a label,

[0120] Its attributes include

[0121] P 1name P 1value , …, P iname ∶P ivalue ;

[0122] Relationship: Relation name ;

[0123] Rule: Rules.

[0124] Among them, Head represents the head entity, and Relation name represents the relationship, and Head namerepresents the head entity name, label represents the label to which the entity belongs, P iname represents the name of the i-th attribute of the head entity, P ivalue represents the value of the i-th attribute of the head entity, i = 1, 2, …. Rules represents the rules corresponding to the query.

[0125] In this transformation template, an entity is represented by connecting the entity name and attribute information.

[0126] For the sequence-to-sequence model 121, after numerical attribute processing and rule hierarchical processing, the transformation template obtains the training text set from the process knowledge graph. The input of each training text consists of the head entity name, head entity type, head entity attributes, relationships, and related rules of the triple, and the output is the tail entity name. The sequence-to-sequence model 121 uses the cross-entropy loss function as the loss function for training.

[0127] Preferably, the triples of the training data transformed by the transformation template come from the process knowledge graph. The data in the process knowledge graph is in the form of triples.

[0128] S131: Screen the features and processing procedures of the inference tasks that do not exist in the process knowledge graph, and import the information of the screened features and processing procedures into the process knowledge graph to form new process knowledge graph nodes.

[0129] Check whether the features or procedures involved in the inference task exist in the knowledge graph. If not, import the information of the features or procedures to be recommended into the knowledge graph. The numerical attribute information of the new node is converted into a value range according to the aforementioned processing method, and the relationship between the new feature or procedure node and the process knowledge node in the graph is established. Since the relationship graph convolutional network model 122 needs to infer the relationship between the head entity of the inference task and the existing nodes in the knowledge graph. This step imports the head entity information involved in the inference task into the knowledge graph, and uses the relationship established with the existing nodes in the knowledge graph as the training data of the relationship graph convolutional network model 122.

[0130] S132: Convert the attributes of the entities in the process knowledge graph into node representations in the process knowledge graph.

[0131] Convert the attributes of each entity in the knowledge graph into the corresponding graph node representation, and generate an independent node structure called an "attribute node".

[0132] S133: Initialize the feature vectors of the process knowledge graph nodes based on the pre-trained word embedding model, and construct the adjacency matrix of the feature vectors.

[0133] Preferably, a word embedding model is set in the preprocessing unit 110. The word embedding model is used to initialize the feature vectors of the nodes in the process knowledge graph. The word embedding model constructs a feature vector adjacency matrix based on each relationship, and transmits the feature vector adjacency matrix to the relational graph convolutional network model as training data.

[0134] Specifically, a pre-trained word embedding model is used to initialize the feature vectors of each node. For each relationship type, an independent adjacency matrix is constructed. This adjacency matrix comes from the knowledge graph processed by steps S131 and S132, and serves as the most direct training data for the relational graph convolutional network.

[0135] S134: Use the triple (h, r, t) in the process knowledge graph as positive samples. After converting (h, r,?) into text through a conversion template, input it into the trained sequence-to-sequence model 121. The error entity in the output of the sequence-to-sequence model 121 replaces the tail entity t to generate negative sample triples (h, r, t') for training the relational graph convolutional network model 122. The cross-entropy loss function is used as the loss function for training.

[0136] S200: Based on the data preprocessing, train the sequence-to-sequence model 121 and the relational graph convolutional network model 122, and comprehensively output the inference result based on the scores of both.

[0137] The inference unit 120 includes a sequence-to-sequence model 121 and a relational graph convolutional network model 122.

[0138] The sequence-to-sequence model 121 is trained based on the training data and outputs negative sample data. The steps performed by the sequence-to-sequence model 121 are as follows.

[0139] S210: Receive the entities in the form of coherent sentences transformed from the inference task, and assign probability scores to the decoded entities.

[0140] Convert the received inference task into a text representation according to the conversion template and input it into the trained sequence-to-sequence model 121.

[0141] Randomly sample n output sequences from the decoder of the sequence-to-sequence model 121. For each decoded entity, assign a score equal to the (log) probability of decoding its sequence.

[0142] The formula for this step is:

[0143]

[0144]

[0145] P(Y|X) represents the conditional probability, which is the probability of the output sequence Y given the input sequence X. Y consists of a series of elements y 1 , y 2 ,..., y m composed. y t represents t elements in the output sequence. P(y t |y 1 , u 2 ,…, u t-1 , X) represents the probability of generating the next output y t under the condition of the given input sequence X and all previous outputs. This is how the probability is calculated at each step in the decoding process, reflecting the dependency relationship in the sequence generation process. Score seq represents the score of the output sequence Y, which is obtained by taking the logarithm of P(Y|X).

[0146] The probability score assigned to entities not encountered is -∞, and entities with the top-k scores are selected to form candidate triples with the original triples.

[0147] S220: Entities with the top-k probability scores are combined with the original triples to form candidate triples.

[0148] The relational graph convolutional network model 122 is trained using the combined sample data of the negative sample data output by the sequence-to-sequence model 121 and the positive sample data in the process knowledge graph.

[0149] The steps performed by the relational graph convolutional network model 122 are as follows.

[0150] S230: The sequence-to-sequence model 121 delivers the candidate triples to the relational graph convolutional network model 122. The relational graph convolutional network model 122 outputs the scores corresponding to the candidate triples.

[0151] The top-k candidate triples generated by the sequence-to-sequence model 121 are input into the relational graph convolutional network model 122. The relational graph convolutional network model 122 outputs a triple score for each triple. Using DistMult, the output vectors of two nodes on an edge are mapped to a scoring function to represent the confidence of this edge as the triple score. The formula is as follows:

[0152]

[0153] where d represents the dimension of the vector, h i , r i , t i represent the components of the vectors h, r, and t in the i-th dimension respectively.

[0154] S240: Calculate the comprehensive score corresponding to the processing procedure based on the probability score output by the sequence-to-sequence model 121 and the score of its own output.

[0155] Integrate the scores of the sequence-to-sequence model 121 and the relational graph convolutional network model 122 to obtain the final top-m result (m < k). The scoring formula is:

[0156] Score = W seq Score seq + W R-GCN Score R-GCN .

[0157] Among them, W seq and W R-GCN represent the scores of the sequence-to-sequence model 121 and the relational graph convolutional network model 122 respectively; Score seq and Score R-GCN represent the weights of the sequence-to-sequence model 121 and the relational graph convolutional network model 122 respectively.

[0158] S300: The trained sequence-to-sequence model 121 and relational graph convolutional network model 122 perform inference calculations on the received inference task and obtain the comprehensive score corresponding to the processing procedure of the inference task.

[0159] As Figure 3 shown, when the sequence-to-sequence model 121 receives a question, for example, when receiving an inference task about the processing procedure of "tapered hole", this task is represented in the form of a triple (tapered hole, has process,?). First, check whether the entities involved in this task exist in the process knowledge graph. If not, import the relevant entity information into the knowledge graph and create a new node for "tapered hole". The numerical attributes of the new node are converted into a value range according to the established processing method and are connected to other process knowledge nodes in the graph. Then retrain the relational graph convolutional network model 122.

[0160] Then, according to the aforementioned conversion template, convert the target triple into a text representation: "The tapered hole is a feature, and its attributes are, feature category: hole;... whether it is a reserved hole: no; relationship: has process. Rule:..." Input this description into the sequence-to-sequence model 121, and randomly sample (generate) n sequence probability scores in the decoder stage. Each decoded entity is assigned a probability score, and its value is equal to the (log) probability of the decoded sequence. For example, process entities such as finish turning and rough grinding of the inner hole obtain corresponding scores according to their probabilities.

[0161] Next, the top-k candidate triples with the highest scores generated by the sequence-to-sequence model 121 are input into the relational graph convolutional network model 122. For each triple, the relational graph convolutional network model 122 gives a triple score, and the scoring function of the node vectors at both ends of the edge is calculated using the DistMult method to evaluate the confidence of this edge. Finally, combining the scores of the sequence-to-sequence model 121 and the relational graph convolutional network model 122, the top m results are obtained as the final recommended processing procedures, such as finish turning, rough grinding the inner hole, and finish grinding the inner hole. For example, the relational graph convolutional network model 122 outputs the top-m results of the comprehensive score; m = 3.

[0162] If the above-mentioned reasoning and optimization of the processing procedures are not carried out, the following problems may be encountered: If there is no entity of "tapered hole" in the knowledge graph, traditional node mapping-based methods may wrongly recommend processing procedures applicable to double tapered holes, which are not suitable for the case of tapered holes.

[0163] In addition, even if there is a node named "tapered hole" in the knowledge graph, if its attributes are different from those of the "tapered hole" in the current reasoning task, the processing procedure suggestions provided by traditional methods may also be inappropriate. For example, the tapered hole in the knowledge graph has specific attributes {IT≥10, Ra≥6.3, D<30, with reserved holes}, and the corresponding processing procedure is reaming; while the tapered hole in the reasoning task has different attributes {9≤IT<10, 3.2≤Ra<6.3, D<10, without reserved holes}, and the corresponding processing procedure should be drilling-reaming. Therefore, accurate reasoning and optimization are the key to ensuring the correct recommended processing procedures. If the above-mentioned reasoning and optimization of the processing procedures are not carried out, the following situations may occur: If there is no "tapered hole" entity in the knowledge graph, traditional node mapping-based methods may wrongly recommend processing procedures applicable to double tapered holes. In another case, even if there is a node named "tapered hole", if its attributes are different from those of the "tapered hole" in the current reasoning task, traditional methods may also provide inaccurate processing procedure suggestions. For example, if the tapered hole in the knowledge graph has specific attributes {IT≥10, Ra≥6.3, D<30, with reserved holes}, and the corresponding processing procedure is reaming; while the tapered hole in the reasoning task has different attributes {9≤IT<10, 3.2≤Ra<6.3, D<10, without reserved holes}, and the corresponding processing procedure should be drilling-reaming. Therefore, accurate reasoning and optimization are necessary to ensure the correct recommended processing procedures.

[0164] S400: The optimization unit 130 optimizes the current processing technology based on the inferred processing procedures.

[0165] First, the optimization unit 130 needs to receive the results from the inference unit 120, which includes the comprehensive scores obtained through the sequence-to-sequence model 121 and the relational graph convolutional network model 122. These scores reflect the performance in various aspects such as the quality and efficiency of different processing procedures.

[0166] Next, the optimization unit 130 will conduct an in-depth analysis of these scores to identify the bottlenecks or inefficient links in the current processing technology. This step may involve complex calculations, such as comparing the performance indicators of different processes, or simulating different production scenarios to predict possible results.

[0167] Based on the above analysis, the optimization unit 130 will generate a series of improvement strategies. These strategies are aimed at solving the identified problems and improving the overall production efficiency. For example, if it is found that the material removal rate of a certain process is low, the optimization unit 130 may suggest adjusting the tool selection or processing parameters.

[0168] After generating multiple potential optimization solutions, the optimization unit 130 will further screen and verify these solutions. This process may involve using the inference unit 120 again to evaluate the effects of each solution, ensuring that the selected solution can indeed bring the expected improvements.

[0169] After determining the optimal solution, the optimization unit 130 will guide the adjustment of the actual production process. At the same time, the optimization unit 130 will continuously monitor the implementation effects and collect feedback data to make necessary fine-tuning in a timely manner. This step emphasizes the importance of dynamic adjustment to ensure that the production process is always in the best state.

[0170] Preferably, the optimization unit 130 should have the ability of self-learning, that is, continuously updating its algorithms and models according to the new data in actual operation. This means that with the accumulation of more data, the optimization unit 130 can provide more accurate and effective optimization suggestions.

[0171] Based on the reasoning of the above steps, it is possible to comprehensively integrate information such as rules and knowledge graphs, and provide accurate process recommendations for features or processes that do not exist in the knowledge graph. The obtained processing procedures or tools / machine tools are first evaluated by the processing personnel. Once the effectiveness of the processing procedure or tool / machine tool is verified, the processing procedure (tool / machine tool) will be used in practical production applications, and the existing process knowledge graph will be checked to determine whether there is already a link between the corresponding features and processes (processes and tools / machine tools). In this example, if the best processing procedure for the tapered hole obtained through reasoning is {drill - ream}, and there is no link established between the tapered hole node and the drill process node in the process knowledge graph, then a triple (tapered hole, has process, drill) will be added to the process knowledge graph to complete the link. Through this method, the knowledge in the process knowledge graph can be updated to provide more comprehensive and accurate process recommendations in actual production.

[0172] It should be noted that the above specific embodiments are exemplary. Those skilled in the art can come up with various solutions inspired by the disclosed content of the present invention, and these solutions also belong to the scope of the disclosure of the present invention and fall within the protection scope of the present invention. Those skilled in the art should understand that the description and drawings of the present invention are illustrative and do not constitute a limitation on the claims. The protection scope of the present invention is defined by the claims and their equivalents. The description of the present invention contains multiple inventive concepts. Phrases such as "preferably" and "according to a preferred embodiment" indicate that the corresponding paragraphs disclose an independent concept. The applicant reserves the right to file divisional applications based on each inventive concept.

Claims

1. A method for optimizing a processing procedure based on a process knowledge graph, characterized in that: The method comprises: Based on structured process rules, the entities in the process knowledge graph dataset are preprocessed by standardized classification, data cleaning and normalization; classifying the structured process rules and converting the structured rules into text rules to form training data for a sequence-to-sequence model (121), The sequence-to-sequence model (121) is trained using the training data, and a relational graph convolutional network model (122) is trained using combined sample data of negative sample data output by the sequence-to-sequence model (121) and positive sample data in the process knowledge graph; Performing reasoning calculation on the received reasoning task based on the trained sequence-to-sequence model (121) and the relationship graph convolution network model (122), and obtaining a comprehensive score of the reasoning task corresponding to the processing procedure; The current machining process is optimized based on the machining process obtained by reasoning.

2. The method according to claim 1, characterized in that The steps for preprocessing entities of the process knowledge graph dataset include: Dividing entities into numerical ranges based on the structured process rules to achieve standardized classification; Traversing the triples in the process knowledge graph dataset, identifying and deleting entities with missing data items to achieve data cleaning; The numerical attributes of the process knowledge graph dataset are normalized to a predefined value range, and a numerical identification table corresponding to the value range is constructed.

3. The method according to claim 1 or 2, characterized in that: The step of converting structured rules into textual rules to form training data for a sequence-to-sequence model (121) includes: After the structured process rules are classified, the data set of the structured process rules is traversed and the attribute information of the structured process rules is extracted; Replacing the value range with an identifier according to the numerical identification table, thereby converting the structured process rule into a text rule, and completing the attribute labeling of the structured process rule; The Jaccard similarity coefficient is calculated based on the attributes of the process knowledge graph nodes and the attributes of the rule categories corresponding to the reasoning tasks.

4. The method according to any one of claims 1 to 3, characterized in that: The steps of preprocessing the entities of the process knowledge graph dataset also include: Screening out features and processing procedures of the reasoning task that do not exist in the process knowledge graph, and importing the information of the screened features and processing procedures into the process knowledge graph to form a new process knowledge graph node; Converting the attributes of entities in the process knowledge graph into node representations in the process knowledge graph; The feature vectors of the process knowledge graph nodes are initialized based on the pre-trained word embedding model, and the adjacency matrix of the feature vectors is constructed.

5. The method according to any one of claims 1 to 4, characterized in that: The step of performing inference calculation on the received inference task based on the trained sequence-to-sequence model (121) and the relationship graph convolution network model (122) comprises: The entities in the form of coherent sentences transformed from the reasoning task are input into the sequence-to-sequence model (121); the sequence-to-sequence model (121) assigns probability scores to the decoded entities, and forms candidate triples with entities having top-k probability scores and the original triples. The sequence-to-sequence model (121) transmits the candidate triples to the relational graph convolutional network model (122), The relationship graph convolutional network model (122) outputs a score corresponding to the candidate triple; A comprehensive score corresponding to the processing step is calculated based on the probability score output by the sequence-to-sequence model (121) and the score output by the relationship graph convolutional network model (122).

6. A processing procedure optimization system based on process knowledge graph, characterized in that: The system comprises a processor (100), wherein the processor (100) comprises a preprocessing unit (110), an inference unit (120) and an optimization unit (130); The preprocessing unit (110) includes a standardized classification module (111), a rule stratification module (112) and a conversion module (113); The standardized classification module (111) performs standardized classification, data cleaning and normalization preprocessing on entities of the process knowledge graph data set based on structured process rules; The rule stratification module (112) classifies the structured process rules and converts the structured rules into text rules to form training data for a sequence-to-sequence model (121); The conversion module (113) converts the triples in the process knowledge graph into entities in the form of coherent sentences based on the conversion template, and the entities in the form of coherent sentences are training data of the sequence-to-sequence model (121); The reasoning unit (120) includes a sequence-to-sequence model (121) and a relational graph convolutional network model (122), The sequence-to-sequence model (121) is formed by training based on the training data, and outputs negative sample data; The relationship graph convolution network model (122) is trained by using the negative sample data output by the sequence-to-sequence model (121) and the combined sample data of the positive sample data in the process knowledge graph; The trained sequence-to-sequence model (121) and the relationship graph convolution network model (122) perform reasoning calculations on the received reasoning tasks, and obtain a comprehensive score of the reasoning tasks corresponding to the processing steps; The optimization unit (130) optimizes the current machining process based on the machining process obtained by inference.

7. The system according to claim 6, characterized in that The standardized classification module (111) is configured to: Dividing entities into numerical ranges based on the structured process rules to achieve standardized classification; Traversing the triples in the process knowledge graph dataset, identifying and deleting entities with missing data items to achieve data cleaning; The numerical attributes of the process knowledge graph dataset are normalized to a predefined value range, and a numerical identification table corresponding to the value range is constructed.

8. The system according to claim 6 or 7, characterized in that: The rule layering module (112) is configured to: After the structured process rules are classified, the data set of the structured process rules is traversed and the attribute information of the structured process rules is extracted; Replacing the value range with an identifier according to the numerical identification table, thereby converting the structured process rule into a text rule, and completing the attribute labeling of the structured process rule; The Jaccard similarity coefficient is calculated based on the attributes of the process knowledge graph nodes and the attributes of the rule categories corresponding to the reasoning tasks.

9. The system according to any one of claims 6 to 8, characterized in that: The conversion module (113) is configured to: Screening out features and processing procedures of the reasoning task that do not exist in the process knowledge graph, and importing the information of the screened features and processing procedures into the process knowledge graph to form a new process knowledge graph node; Converting the attributes of entities in the process knowledge graph into node representations in the process knowledge graph; The feature vectors of the process knowledge graph nodes are initialized based on the pre-trained word embedding model, and the adjacency matrix of the feature vectors is constructed.

10. The system according to any one of claims 6 to 9, characterized in that: The reasoning unit (120) is configured to: The sequence-to-sequence model (121) receives entities in the form of coherent sentences transformed by the reasoning task, assigns probability scores to the decoded entities, and forms candidate triples with entities with top-k probability scores and the original triples; The sequence-to-sequence model (121) transmits the candidate triples to the relational graph convolutional network model (122); The relational graph convolutional network model (122) outputs a score corresponding to the candidate triplet; and a comprehensive score corresponding to the processing step is calculated based on the probability score output by the sequence-to-sequence model (121) and the score output by the relational graph convolutional network model (122).

Images