Knowledge-driven automatic generation method for part machining process and related product

Abstract

The invention relates to the field of intelligent manufacturing and digital process planning, in particular to a knowledge-driven automatic generation method for a part machining process and a related product, and the method comprises the steps: constructing a machining-process-oriented knowledge graph, and coupling the knowledge graph with a large knowledge model to form a feedback closed loop; extracting process knowledge units from the multi-source data, and performing quantitative scoring and credibility screening on the process knowledge units based on a preset knowledge evaluation algorithm to obtain a high-reliability process knowledge data set; performing fine-grained attribute relation extraction on the process knowledge in the process knowledge data set, structuring an extraction result, and dynamically updating the extraction result to the knowledge graph to realize incremental updating of the knowledge; through collaborative innovation of four core links of efficient and automatic knowledge acquisition, structured and refined process knowledge expression and dynamic extension and real-time updating capability generation, the technical bottleneck in traditional process planning is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent manufacturing and digital process planning, in particular to a knowledge-driven automatic generation method of part machining process and related products, which belongs to the cross-technology of computer-aided process planning (CAPP), knowledge engineering, artificial intelligence and manufacturing informatization, and is suitable for process planning, optimization and decision support in the machining process of complex parts. BACKGROUND

[0002] Under the background of intelligent transformation of manufacturing industry, part machining process automatic generation technology as the core support of production efficiency improvement and quality guarantee, its technical maturity directly affects the agile response ability of manufacturing system. However, the existing process knowledge sources are scattered and the formats are not unified, the data structures of different information sources are different, there are natural language description, structured report, sensor time series data and image data. The production personnel usually need to clean, label and arrange a large amount of manual text, which is time-consuming and easy to make mistakes, and it is difficult to form a unified knowledge base that can be directly calculated and utilized.

[0003] The existing means cannot effectively evaluate the quality and credibility of these knowledge units, resulting in the need for frequent manual review or repeated experiments in the downstream process decision-making process, affecting production efficiency and cost control. Although in recent years, process decision-making models based on rules or machine learning have tried to be applied to parameter recommendation, such models are usually disconnected from the domain ontology or expert system, and lack a deep understanding of process constraints and contextual semantics. On the other hand, their knowledge base is maintained statically and cannot absorb new materials, tools or process experimental results in real time, making it difficult to meet the changing production needs. Although some research has organized machining process information in the form of knowledge graph, the maintenance of nodes and relationships in the knowledge graph is mostly dependent on manual rules or periodic manual updates, making it difficult to reflect new situations in the production site in a timely manner. In terms of knowledge retrieval, traditional methods based on keyword or relationship queries are not efficient, and their matching ability for complex semantic combinations is also very limited. SUMMARY

[0004] The technical problem to be solved by the present application is that the existing part machining process knowledge management and decision support system generally has the problems of difficult knowledge acquisition, insufficient quality guarantee, weak dynamic evolution and low retrieval efficiency, etc. The purpose is to provide a knowledge-driven automatic generation method of part machining process and related products, which can automatically extract high-reliability knowledge from massive multi-source data, combine structured graph and deep semantic reasoning, and has the ability of dynamic expansion and real-time update of the whole process automation solution.

[0005] The present application is realized by the following technical scheme:

[0006] A knowledge-driven automatic generation method of part machining process, comprising:

[0007] construct a knowledge graph for the machining process, and couple the knowledge graph with a knowledge large model to form a feedback loop, wherein the knowledge graph provides structured process constraints and semantic background for the knowledge large model, the knowledge large model makes reasoning and decision-making according to the process constraints, and the decision-making results are fed back to update the knowledge graph;

[0008] extract process knowledge units from multi-source data, and quantitatively score and credibility filter the process knowledge units based on a preset knowledge evaluation algorithm to obtain a high-reliability process knowledge dataset;

[0009] extract fine-grained attribute relationships of process knowledge in the process knowledge dataset, and dynamically update the extracted results to the knowledge graph to realize incremental updating of knowledge.

[0010] Optionally, the step of constructing the knowledge graph for the machining process comprises:

[0011] establish a machining process method chain, and the step of establishing the machining process method chain comprises: introducing a process matching engine, establishing a bidirectional coupling mechanism driven by part requirements and enterprise resource capability, the bidirectional coupling mechanism generates a candidate set of machining methods from top to bottom according to part features, and filters and replaces the candidate set from bottom to top according to enterprise resource capability to generate a machining process method chain that meets the machining requirements and resource constraints;

[0012] map the machining process method chain to a graph database, wherein the machining parts, machining surfaces, processes, tools, fixtures and materials in the machining process method chain are taken as instance nodes in the graph, and the machining sequence, use relationship or application relationship between the nodes are taken as edges in the graph for semantic connection to construct the knowledge graph.

[0013] Optionally, the step of coupling the knowledge graph with the knowledge large model to form the feedback loop comprises:

[0014] vectorize the entities and relationships in the knowledge graph, and construct a vector index library;

[0015] when receiving a machining requirement, retrieve a subgraph related to the machining requirement in the vector index library;

[0016] input the subgraph and the machining requirement to the knowledge large model for reasoning to generate a decision-making suggestion with a confidence score;

[0017] evaluate the credibility of the newly generated knowledge in the reasoning process, and incrementally update the knowledge that meets a preset credibility threshold to the knowledge graph and the vector index library.

[0018] Optionally, the step of extracting process knowledge units from multi-source data and performing quantitative scoring and credibility screening comprises:

[0019] Text data in the multi-source data is preprocessed, and an entity recognition model adopting a BiLSTM-CNN-CRF architecture is used to process and extract process knowledge entities;

[0020] The extracted knowledge entities are subjected to existence verification in a knowledge graph, and if they exist in the knowledge graph, multi-hop path retrieval is initiated for the corresponding knowledge entities; if they do not exist in the knowledge graph, they are stored in a temporary cache queue;

[0021] Based on the path comprehensive score, an adaptive screening threshold is determined according to the statistical distribution of the path comprehensive score to divide the reliability level of the knowledge entity.

[0022] Optionally, at the input layer of the entity recognition model, the Shannon entropy calculated according to the label distribution of the local context of the text is spliced with the embedding vector corresponding to the text to form an enhanced input feature, so that the convolutional layer of the model can adaptively focus on the high-entropy area during feature extraction.

[0023] Optionally, the method for calculating the path comprehensive score comprises:

[0024] A three-factor linear combination model is used to perform weighted scoring on the multi-hop path to obtain the path comprehensive score; the three factors include:

[0025] A path length factor calculated based on the logarithm of the number of entity nodes in the path;

[0026] A termination node authority factor quantified by comprehensively evaluating the global importance and local connection density of the end node of the path;

[0027] A relationship continuity factor obtained by calculating the semantic similarity between embedding vectors of adjacent relationships in the path.

[0028] Optionally, the step of dividing the reliability level of the knowledge entity comprises:

[0029] The path comprehensive scores of multiple paths associated with a knowledge entity are subjected to maximum-minimum normalization processing;

[0030] The multiple path scores after normalization processing are weighted and aggregated to calculate the final reliability score of the knowledge entity;

[0031] Based on the statistical distribution of the final reliability scores of all knowledge entities, an improved box plot method is used to determine an adaptive threshold;

[0032] According to the comparison result of the final reliability score and the third quartile and the adaptive threshold, the knowledge entity is divided into a high reliability, a low reliability or an unreliable level, wherein the knowledge entity with the final reliability score greater than the adaptive threshold is high reliability, the knowledge entity with the final reliability score less than or equal to the third quartile is unreliable, and the rest of the knowledge entity is low reliability.

[0033] Optionally, the multi-source data at least includes one or more of the following: process literature, standard specification, historical case, structured report and sensor data.

[0034] The step of preprocessing the text data in the multi-source data includes:

[0035] A binary classifier based on BERT model fine-tuning is adopted to identify and label the noise samples in the text data by a supervised learning method.

[0036] According to the identification and labeling results, the data identified as noise samples are filtered out to obtain cleaned text.

[0037] The cleaned text is segmented to generate sequence labeling samples for subsequent extraction of process knowledge units.

[0038] Optionally, the step of dynamically updating the knowledge graph includes:

[0039] For the knowledge entity divided into high reliability, a Span-Based framework and a deep learning classifier are adopted to perform fine-grained extraction of attributes and relationships of the knowledge entity.

[0040] The extracted attributes and relationships are structured and encapsulated, and are automatically mapped to the corresponding ontology structure of the knowledge graph to realize logical association of knowledge.

[0041] After the knowledge graph is mapped and updated, incremental vector embedding recalculation is triggered for the newly added knowledge entity and relationship, and the retrieval index is synchronously updated.

[0042] A computer program product includes a computer program / instruction, which, when executed by a processor, implements the knowledge-driven automatic generation method of the part machining process according to any one of the above.

[0043] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0044] The present application solves the technical bottleneck in traditional process planning through the collaborative innovation of four core links of efficient and automatic knowledge acquisition, structured and refined process knowledge expression and dynamic expansion and real-time updating capability.

[0045] The structured graph and the pre-training knowledge large model are deeply coupled, so that the system can utilize the ontology constraints and semantic context provided by the graph, and can also exert the deep understanding and reasoning ability of the large model in complex production scenarios; through the real-time feedback and incremental updating mechanism, the processing method chain is continuously verified and optimized, so as to ensure the feasibility and accuracy of the process scheme in the dynamic manufacturing environment.

[0046] Through the large-scale knowledge extraction and pre-screening module, the retrieval, preprocessing, filtering and information extraction of massive original data from multiple sources are automatically completed, and combined with the credibility quantitative scoring algorithm, the process knowledge unit with high reliability can be quickly output, which greatly reduces the workload of manual cleaning, labeling and sorting, and significantly improves the knowledge acquisition efficiency.

[0047] Based on the processing technology ontology, a field knowledge graph is constructed, and a Span-Based framework and a deep learning classifier are used on the knowledge graph to perform fine-grained attribute and relationship extraction on high-confidence entities, so as to realize accurate structured representation of core elements such as process parameters, tool characteristics and process logic, and fully meet the dual requirements of semantic richness and logical consistency. Relying on the incremental vector embedding recalculation and vector retrieval index updating, the system can immediately incorporate the newly added knowledge into the retrieval and reasoning process after each new knowledge is added, so as to ensure that the new knowledge is hit and called efficiently in the subsequent 'perception' stage, and realize the continuous self-evolution of the knowledge graph.

[0048] The system has modular and interfaced design, and can seamlessly access new data sources, algorithm models and ontology definitions; as the application scenarios are popularized and continuously fed back, the knowledge large model will be continuously self-improved, forming an efficient, scalable and continuously self-strengthening intelligent manufacturing knowledge management and decision support platform. BRIEF DESCRIPTION OF DRAWINGS

[0049] The accompanying drawings illustrate exemplary embodiments of the present application and, together with the general description of the application given above, and the detailed description of the embodiments given below, serve to explain the principles of the present application. These drawings should not be considered limiting; in fact, they illustrate the principles of the application in a more detailed fashion. In the drawings:

[0050] Figure 1 FIG. 1 is a flowchart of a knowledge-driven automatic generation method of a part processing technology according to the present application.

[0051] Figure 2 FIG. 2 is a flowchart of constructing a knowledge graph for processing technology according to the present application.

[0052] Figure 3 FIG. 3 is a flowchart of quantitative scoring and credibility screening according to the present application. DETAILED DESCRIPTION

[0053] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0054] It should also be noted that, for ease of description, only the parts relevant to the present invention are shown in the accompanying drawings.

[0055] Where there is no conflict, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] Example 1

[0057] like Figure 1 As shown, this embodiment provides a knowledge-driven automated generation system for the entire process of parts processing technology, aiming to solve the problems of scattered sources of knowledge for parts processing technology, inconsistent quality, low efficiency of manual processing, and the inability of isolated model decisions to evolve dynamically in the existing technology; specifically, it includes the following three steps:

[0058] Step 1: Construct a large-scale process knowledge model oriented towards the processing method chain.

[0059] A knowledge graph oriented towards the manufacturing process is constructed, and the knowledge graph is coupled with a large knowledge model to form a feedback loop. The knowledge graph provides the large knowledge model with structured process constraints and semantic background. The large knowledge model makes inference decisions based on process constraints and feeds back the decision results to update the knowledge graph.

[0060] A domain knowledge graph is constructed based on the manufacturing process ontology and deeply coupled with a large knowledge model. In this closed loop, the knowledge graph provides the model with structured process constraints and semantic background, while the pre-trained model undertakes deep understanding and reasoning decisions regarding the manufacturing context. The model's decision results are continuously updated through a feedback mechanism to update the knowledge graph, enabling feasibility verification, accuracy assurance, and continuous iterative optimization of the process method chain in a dynamic manufacturing environment.

[0061] The knowledge graph provides structured process constraints and semantic context, supporting the modeling of logical connections in the manufacturing process; the large knowledge model is responsible for deep understanding and intelligent reasoning of process requirements and the production environment, providing support for process decision-making. The system optimizes and verifies the processing method chain through a continuous feedback mechanism, ensuring the feasibility and accuracy of the process plan under dynamic manufacturing conditions.

[0062] Knowledge graphs store various entities (specific tools, processes, materials) in the field of machining processes and their precise relationships (applicable to, preceding processes, etc.) in the form of a network.

[0063] Large knowledge model refers to a large-scale artificial intelligence model with strong natural language understanding and logical reasoning capabilities, which can understand complex and unstructured process requirements and make decisions based on the constraints provided by the knowledge graph, such as planning process routes or recommending process parameters.

[0064] Feedback loop refers to the decisions made by the large knowledge model, after being verified, its results (such as a new process combination proven to be efficient) will be used to update and enrich the knowledge graph.

[0065] Step two, build an automatic extraction and screening mechanism for high-credibility process knowledge.

[0066] Extract process knowledge units from multiple sources of data, and based on the preset knowledge evaluation algorithm, quantitatively score and screen the process knowledge units for reliability, to obtain a high-reliability process knowledge dataset.

[0067] Large-scale retrieval is implemented on the target information source, and massive raw data is collected. The pre-processed raw data is processed and extracted based on knowledge filtering and information extraction technology to extract knowledge units with practical application value. Based on the preset knowledge evaluation algorithm, the extracted knowledge units are quantitatively scored and screened for reliability to obtain a high-reliability knowledge dataset.

[0068] The multi-source data includes at least one or more of the following: process literature, standard specifications, historical cases, structured reports, and sensor data. From multiple sources of information, automatically retrieve and pre-process massive raw data, extract knowledge units with practical application value through filtering and information extraction technology, and screen out high-reliability process knowledge by combining credibility quantitative scoring algorithm, thereby providing solid data support for subsequent process decision-making.

[0069] This step addresses the difficulties in obtaining process knowledge in manufacturing processes, as well as the uneven quality of the obtained knowledge. A new method is proposed that combines multi-source heterogeneous information processing and reliable knowledge identification. This mechanism can automatically retrieve and efficiently pre-process massive raw data from process literature, standard specifications, historical cases, and sensor data. Using filtering and information extraction technology based on natural language processing and semantic understanding, it accurately extracts knowledge units with practical application value. By introducing a credibility quantitative scoring algorithm, the extracted knowledge units are evaluated for reliability and prioritized, and high-reliability process knowledge is selected, significantly improving knowledge quality and usability.

[0070] Step three, build a fine-grained attribute relationship extraction and graph dynamic updating mechanism.

[0071] Fine-grained attribute relation extraction is performed on the process knowledge in the process knowledge dataset, and the extracted results are structured and dynamically updated to the knowledge graph to realize incremental updating of knowledge.

[0072] The fine-grained attribute relation extraction and graph dynamic updating mechanism is constructed. The high-credibility knowledge unit entity original text is located in step one, and the Span-Based attribute and relation extraction is performed. A fixed-length representation is constructed for each span, and an attribute classifier and a relation classifier are used to identify whether the span corresponds to an entity attribute value and a semantic relation between different entities, respectively. Structured packaging and domain graph mapping are performed to map the new nodes and edges to the corresponding ontology positions of the domain knowledge graph, complete semantic structuring and logical association, and perform incremental updating to ensure that the newly supplemented knowledge can be retrieved and called immediately in the next round of closed loop.

[0073] The Span-based information extraction framework is used to accurately locate the text context of high-credibility entities, and deep learning classifiers are used to efficiently identify the semantic relation between entities and their attribute values, thereby realizing fine-grained structured expression of process knowledge. Meanwhile, the extracted knowledge triples are automatically mapped to the ontology nodes and edges of the domain knowledge graph, and the complete organization and expression of knowledge are realized through semantic association. Combined with the incremental vector embedding recalculation and retrieval index updating mechanism, the system can immediately hit and efficiently call new knowledge after each new knowledge supplement.

[0074] Through the cooperative realization of the above three targets, the embodiment can significantly improve the generation efficiency and reliability of the part machining process scheme, reduce the cost of manual intervention, and continuously self-iterate and optimize in actual production, thereby providing an efficient, scalable, and continuously self-strengthening process knowledge automatic generation solution for intelligent manufacturing.

[0075] Embodiment Two

[0076] As shown in Figure 2 , the embodiment provides a specific implementation process of step one, which can be divided into two major links: first, the construction of the knowledge graph itself, and second, the operation mechanism of the feedback closed loop between the knowledge graph and the knowledge big model. That is, a high-quality process knowledge graph that conforms to the theoretical logic and is consistent with the actual production capacity is constructed through a bidirectional verification mechanism; then, based on the graph, a closed loop workflow of "perception-decision-updating" is established, so that the system can respond to external demands, make intelligent decisions, and feed back the new knowledge generated in the decision-making process to the system in real time, thereby realizing self-iteration.

[0077] Step 1.1, establish a machining process method chain.

[0078] The method for establishing the machining process method chain comprises the following steps: introducing a process matching engine, establishing a two-way coupling mechanism driven by part requirements and enterprise resource capability, deriving a candidate set of machining methods from top to bottom according to part features, and screening and replacing the candidate set from bottom to top according to enterprise resource capability, to generate a machining process method chain that meets the machining requirements and resource constraints.

[0079] To ensure the rationality and executability of the process method chain, factors such as machining process rules, experience constraints and device feasibility need to be comprehensively embedded in the modeling process, including matching rules between machining features and machining methods, sequence logic rules between machining methods, usage restrictions of equipment / clamps / tools, and compatibility constraints between process parameters. By defining part machining requirement functions and enterprise resource capability functions , the type of the main machining surface belongs to (such as hole, groove, plane, inclined plane, etc.); geometric and topological information of the part (such as shape features, adjacent surface connection relationship, size coordinates, etc.); machining information related to the machining surface (such as the position, direction, accessibility of the surface, etc.); user-defined machining technical requirements (such as accuracy, tolerance, surface roughness, heat treatment requirements, etc.).

[0080] The process matching engine is introduced to establish a two-way coupling mechanism driven by part requirements and combined with actual resource capability. This mechanism starts from part features from top to bottom, and sequentially derives the required machining features, processes, process parameters, and generates a candidate set of machining methods. At the same time, the candidate machining methods are screened and replaced from bottom to top according to the current resource capability of the enterprise, to ensure that the selected process path can be executed. After the above two-way matching and multiple constraint screening are completed, the system finally outputs a set of machining method sequences that meet the machining requirements and are executable , where each machining method is matched with a specific process, machining surface and tool, process parameter setting (determined by ), executable equipment and tool resources (limited by ), and the tool and process parameters are one-to-one corresponding. The final machining method chain is the optimization result under the joint coupling of and .

[0081] Step 1.2, map the machining process method chain to the graph database.

[0082] The machining process method chain is mapped to a graph database, wherein the machining parts, machining surfaces, processes, tools, fixtures and materials in the machining process method chain are taken as instance nodes in the graph, and the machining sequence, use relationship or applicable relationship between the nodes are taken as edges in the graph for semantic connection, to construct a knowledge graph.

[0083] After the machining method chain is constructed, the generated machining method chain is mapped to a graph database, each machining entity, attribute and relationship therebetween are represented in a graph structure, a standardized machining process knowledge graph is constructed, and the knowledge graph is set as The nodes in the graph are composed of each core element in the machining method chain, including machining parts, machining surfaces, processes, tools, fixtures, materials and process parameters; each machining method unit is taken as an instance node in the graph, carrying its associated multi-dimensional attribute information such as geometric characteristics, machining requirements and equipment resources. Each process unit in the machining method chain is taken as an instance node in the graph, and the constraint relationship (such as "machining sequence", "use fixture", "use tool") between the nodes is taken as an edge in the graph for semantic connection, to form a structured and queryable knowledge network.

[0084] Step 1.3, feedback loop operation of knowledge graph and knowledge large model.

[0085] The entities and relationships in the knowledge graph are vectorized, abstract process concepts are converted into digital vectors that can be mathematically operated by computers, and a vector index library is constructed; when a machining requirement is received, first, a quick search is performed in the vector index library to find the most relevant existing knowledge to the current requirement, i.e. a subgraph; then, the system inputs the subgraph and the machining requirement to the knowledge large model for reasoning, to generate decision suggestions with confidence score.

[0086] The newly generated knowledge in the reasoning process is evaluated for credibility, and the knowledge that meets the preset credibility threshold after evaluation is incrementally updated to the knowledge graph and the vector index library.

[0087] The processed graph in step 1.2 is embedded into the graph neural coding module of nano-graphrag to generate low-dimensional vectors for each entity and relationship, and a vector index library is simultaneously constructed to support millisecond-level neighbor search. Then, when the system receives processing requirements and real-time features from users or devices, it uses the search(query_vector, top_k) interface of a large language model (such as DeepSeek) to retrieve the most relevant subgraph in the graph vector space for the current scenario, achieving accurate "perception" of historical process experience and parameters. Next, the system inputs the retrieved subgraph together with the original query into the reasoning engine of nano-graphrag, which uses a graph-text hybrid model to perform deep reasoning and decision-making on processing parameters, tool selection, or process optimization schemes, and outputs the optimal recommendations with confidence scores. Finally, for newly identified or verified entities and relationships during the reasoning process, the verify(entity_vector, threshold) API of DeepSeek is called again to evaluate their credibility. Only newly added knowledge that meets the preset threshold is written back to the main graph through the incremental update interface of nano-graphrag, and the vector index is updated in real time, thus completing the entire "perception-decision-update" closed loop and ensuring that the processing process knowledge base can dynamically evolve and continuously provide the latest and most reliable support for downstream applications.

[0088] On the basis of building a processing method chain and knowledge large model fusion framework, a systematic knowledge extraction and screening mechanism needs to be established to ensure the quality and practicality of knowledge input. This mechanism not only ensures the accuracy and reliability of the injected knowledge, but also provides a solid foundation for subsequent graph updating and model reasoning.

[0089] Based on the processing process ontology, a structured domain knowledge graph is constructed to systematically express key process elements and their relationships in the processing process. Through systematic organization and semantic modeling of process engineer domain knowledge, the knowledge graph ensures the accuracy, professionalism, and scalability of the graph content. A pre-trained general large language model is deeply integrated with the process domain knowledge graph. The knowledge graph provides process rules, constraint boundaries, and context information for the large model, enabling the model to follow professional process logic when reasoning and generating process schemes, thereby improving its explainability and controllability. During the generation of process schemes, the system automatically references the entity and attribute relationships in the graph to verify and correct the selected processing path.

[0090] Embodiment Three

[0091] As Figure 3As shown, the embodiment provides a specific implementation process of step two, first, valuable information is quickly identified and extracted from massive literature; then, the information is preliminarily checked in the background; finally, an objective score is given to the reliability of each piece of information through a quantitative scoring system, and classification is carried out accordingly, so as to ensure that only the highest quality knowledge can be adopted by the system.

[0092] Step 2.1, multi-source data processing.

[0093] A binary classifier based on BERT model fine-tuning is used to identify and label noise samples in text data through supervised learning method; according to the identification and labeling results, the data identified as noise samples are filtered out to obtain cleaned text; the cleaned text is segmented to generate sequence labeling samples for subsequent extraction of process knowledge units.

[0094] For the large-scale process text data obtained, first, preprocessing is performed to remove special characters, garbled characters and irrelevant layout markers. Then, noise samples are labeled through supervised learning method: a binary classifier is fine-tuned on BERT model, noise samples are labeled as "Noise", and the rest of normal text is labeled as "Text". The labeled text is uniformly processed, and all of them are converted to UTF-8 encoding. Based on the category label of each token, the tokens identified as noise are filtered out, and the remaining tokens are reorganized in the original order to obtain the cleaned text. Then, the text is segmented into WordPiece fragments by using the tokenizer provided by BERT, and the label misalignment problem caused by segmentation is adjusted, and finally the sequence labeling sample in the form of is generated. In this way, the standardization and consistency of the text are guaranteed, and high-quality data input is provided for subsequent downstream tasks.

[0095] Step 2.2, knowledge entity extraction.

[0096] The text data in the multi-source data is preprocessed, and a BiLSTM-CNN-CRF architecture entity recognition model is used for process knowledge entity extraction to obtain process knowledge entities. BiLSTM-CNN-CRF is a model combination used for sequence labeling tasks (such as named entity recognition) in the field of natural language processing, which can effectively combine the context information of the text (through BiLSTM), local key features (through CNN) and internal constraints of the label sequence (through CRF), so as to accurately identify the pre-defined entity categories in the text.

[0097] The extraction processing technology knowledge entity algorithm is constructed, after the preprocessing of the processing technology text (including denoising, unified coding and word segmentation) is completed, the BiLSTM-CNN-CRF architecture is used for processing technology knowledge entity extraction. The entropy reflects the uncertainty of the model in judging the category of each token in the local context, which is input into the CNN as a prior feature. It can be seen that which position is prone to ambiguity, so as to give additional attention and inhibition when extracting features.

[0098] Step 2.2.1, model optimization.

[0099] The traditional CNN module only relies on fixed size convolution kernel to extract features of local context at character or subword level, and it is difficult to distinguish the importance of different positions. In order to further improve the accuracy of entity extraction, this step introduces Shannon entropy, which is used to measure the "uncertainty" of the model in identifying the category of a word (or character) in the text. The higher the entropy value, the more uncertain the model's judgment of this position, and there may be ambiguity at this position.

[0100] In the input layer of the entity recognition model, the Shannon entropy calculated according to the label distribution of the local context of the text is spliced with the embedding vector corresponding to the text to form the enhanced input feature, so that the convolution layer of the model can adaptively pay attention to the high entropy area when extracting features.

[0101] First, the label distribution of each token in the sliding window is counted and the Shannon entropy is calculated as entropy prior. Then the entropy value is attached to the corresponding embedding vector to form the enhanced input: ;

[0102] wherein, represents the original embedding vector of the th token, with a dimension of , represents that the scalar is expanded to a one-dimensional vector after the vector splicing operation, and then it is concatenated with the original dimensional vector embedding to obtain the expanded dimensional vector representation , represents the Shannon information entropy of the label distribution of the token in the sliding window, and the calculation formula is: ;

[0103] wherein, is the probability of predicting the th class in the local window. The high entropy value of The position indicates that the context label distribution is not concentrated, and the prediction is uncertain. A low entropy value means that the model is very confident in the local context. After this processing, the convolutional layer can adaptively focus on high-entropy areas when extracting local features.

[0104] The word-level embedding and the character features derived by CNN are concatenated and sent to a bidirectional LSTM to extract the entity of the processing technology. BiLSTM can utilize both forward and backward contexts to generate hidden state sequences that take into account left and right dependencies. Based on the BiLSTM hidden state sequence, the CRF layer performs joint decoding on the entire sentence to ensure the consistency of the boundary and label sequence of the processing technology entity.

[0105] Step 2.3, existence verification.

[0106] For the extracted knowledge entity, existence verification is performed in the knowledge graph. If it exists in the knowledge graph, a multi-hop path retrieval is initiated for the corresponding knowledge entity. If it does not exist in the knowledge graph, it is stored in a temporary cache queue.

[0107] After extracting the processing technology knowledge unit in step 2.2, existence verification of the knowledge unit graph is performed. The SPARQL query statement is used to retrieve the type definition library of the knowledge graph to check whether the entity belongs to the predefined process concept category. If the entity exists in the graph, it is marked as "registered trusted entity" and enters the subsequent analysis process. If it does not exist, it is temporarily classified as "entity to be verified" and needs to be manually reviewed or supplemented with knowledge. This step ensures that the screening process is based on a verified knowledge system, avoiding the interference of invalid entities.

[0108] Step 2.4, path comprehensive score.

[0109] Based on the multi-hop path, the path comprehensive score is calculated, and the adaptive screening threshold is determined according to the statistical distribution of the path comprehensive score to classify the reliability level of the knowledge entity. For entities that have been confirmed to exist, a multi-hop path discovery mechanism based on breadth-first search (BFS) is started, the order of traversed entity nodes is saved, the semantic relationship type between nodes is recorded, and the path length, node importance, and relationship density are calculated in real time.

[0110] The method for calculating the path comprehensive score includes: using a three-factor linear combination model to weight the multi-hop path and obtain the path comprehensive score; the three factors include:

[0111] The path length factor is calculated based on the logarithm of the number of entity nodes in the path, and the calculation formula is: wherein, is the number of entity nodes in the path (e.g., from to ).

[0112] The termination node authority refers to the core position and influence of the end node of a certain association path in the entire knowledge network in the knowledge graph. The index is quantified by comprehensively evaluating the global importance and local connection density of the node. The termination node authority factor is quantified by comprehensively evaluating the global importance and local connection density of the end node of the path, and the calculation formula is: , wherein, is the global importance of the node, is the local connection density of the node.

[0113] The relationship continuity refers to calculating the semantic similarity of adjacent relationships through relationship type embedding vectors, ensuring that the path has process logic consistency. The relationship continuity factor is obtained by calculating the semantic similarity between the embedding vectors of adjacent relationships on the path, and the calculation formula is: wherein, is the similarity of the adjacent relationship pair, is the attenuation factor.

[0114] A path comprehensive scoring model is designed, and a three-factor linear combination model is used to form a weighted scoring system that can be explained: wherein .

[0115] The path comprehensive score of multiple paths associated with a knowledge entity is subjected to maximum-minimum normalization processing; wherein, P is the set of entity associated paths.

[0116] The contribution of each path is quantified, and the final reliability score of the knowledge entity is obtained by aggregation calculation; .

[0117] Step 2.5, reliability level division.

[0118] Based on the statistical distribution of the final reliability scores of all knowledge entities, an improved box plot method is used to determine an adaptive threshold; wherein, the filtering boundary is adaptively determined by statistical methods, is the first quartile, is the third quartile, is the interquartile range, and 1.5 is an empirical coefficient to control the deviation degree of the threshold and the data main body distribution.

[0119] According to the comparison results of the final reliability score and the third quartile and the adaptive threshold, the knowledge entity is divided into high reliability, low reliability or unreliable levels by using a multi-level threshold design, wherein the final reliability score is greater than the adaptive threshold (Q3+1.5*IQR) knowledge entities with final reliability scores less than or equal to the third quartile are unreliable, and the rest knowledge entities are low reliability.

[0120] This embodiment constructs a process knowledge extraction and pre-screening module based on information filtering and knowledge extraction. The module uniformly preprocesses massive raw data such as texts, reports, and sensor time series, and relies on a preset quantitative scoring algorithm to score and screen the extracted knowledge units, thereby automatically generating a high-reliability process knowledge set. This technology breaks through the low-efficiency mode of traditional manual cleaning and semi-automatic labeling, and significantly improves the efficiency and quality of cross-source process knowledge acquisition.

[0121] Embodiment Four

[0122] This embodiment provides a specific process of step three, that is, after screening out high-reliability knowledge, the system performs in-depth analysis on the knowledge and seamlessly and dynamically integrates the knowledge into the existing knowledge system to realize self-evolution and capability enhancement of the system.

[0123] A fine-grained knowledge extraction and graph dynamic updating mechanism is introduced. This mechanism revolves around high-confidence entities, accurately locates their context information in the original text, and extracts their attributes and relationships through a Span-Based model. The extracted structured knowledge units can be automatically mapped to the ontology structure of the existing graph to form semantic connectivity and logical expansion. At the same time, after the graph is updated, it can also be fed back to the large model in real time to realize the synchronous improvement and self-enhancement closed loop of the knowledge base and the model capability.

[0124] The specific steps include:

[0125] For knowledge entities classified as high reliability, a Span-Based framework and a deep learning classifier are used to extract the attributes and relationships of the knowledge entities in a fine-grained manner. The context is located by backtracking the original text, and a fixed-length representation is constructed based on the Span-Based framework for candidate word segments. A deep attribute classifier and a relationship classifier are used to identify entity attribute values and semantic relationships between entities, respectively.

[0126] The extracted attributes and relationships are structured and encapsulated, and are automatically mapped to the corresponding ontology structure of the knowledge graph to realize logical association of knowledge. The obtained triples (subject-predicate-object) are encapsulated according to the specification and are automatically mapped to the corresponding ontology nodes and edges of the domain knowledge graph to realize fine-grained structuring and ontology logical association of process knowledge.

[0127] ​After the knowledge graph completes the mapping update, the newly added knowledge entities and relationships trigger incremental vector embedding recalculation, and the retrieval index is updated synchronously. After the incremental writing of the graph nodes and edges is completed, the vectorization engine automatically triggers the recalculation of the low-dimensional embedding of the newly added entities and relationships, and performs incremental update in the vector retrieval index, to ensure that the newly supplemented process knowledge is retrieved and called efficiently in real time, realizing the dynamic expansion and continuous self-strengthening of the knowledge graph.

[0128] Embodiment five

[0129] This embodiment provides a specific example.

[0130] 1. Automated knowledge learning and updating.

[0131] The system discovers a new type of drill named "ABC-123" from a newly published technical article, which performs excellently in machining TC4 titanium alloy.

[0132] The system performs internal procedures: extracts the knowledge entity "ABC-123 drill", evaluates it as a high-reliability level by analyzing its association path with existing knowledge, then performs fine-grained extraction on the entity to clarify its attributes and relationships, such as <ABC-123, applicable material, TC4 titanium alloy>, and finally dynamically updates these new knowledge to the knowledge graph and its retrieval index.

[0133] 2. Intelligent process task execution.

[0134] After the engineer submits the process planning request for "TC4 titanium alloy parts", the system retrieves "TABC-123" as a highly relevant knowledge when searching for solutions.

[0135] The system finally generates an optimized process plan, which clearly recommends "using ABC-123 drill for finishing process", and gives a confidence score.

[0136] Embodiment six

[0137] A knowledge-driven automatic generation terminal for a part machining process, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-mentioned knowledge-driven automatic generation method for a part machining process.

[0138] The memory can be used to store software programs and modules, and the processor can execute various functional applications and data processing of the terminal by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one execution program required for a function, etc.

[0139] The storage data area can store data created according to the use of the terminal, etc. In addition, the storage can include a high-speed random access memory, and can also include a non-volatile memory such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state memory device.

[0140] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the knowledge-driven automatic generation method of the part machining process.

[0141] Without loss of generality, the computer readable medium can include computer storage media and communication media. Computer storage media include volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions data structures, program modules or other data. Computer storage media include RAM, ROM, EPROM, EEPROM, flash memory or other solid-state storage technology, CD-ROM, DVD or other optical storage, magnetic cassette, magnetic tape, disk storage or other magnetic storage device. Of course, those skilled in the art can know that computer storage media are not limited to the above. The system memory and mass storage device described above can be collectively referred to as memory.

[0142] A computer program product includes computer programs / instructions that are executed by a processor to implement the knowledge-driven automatic generation method of the part machining process described above.

[0143] A computer program product includes computer programs or instruction sets for performing specific tasks or implementing specific functions. These programs or instructions are designed to be executed by a processor to implement a series of predefined steps or operations. The program product can be stored in various forms of computer storage media, such as memory, hard disk, solid state drive, optical disk or other forms of digital storage device. It can exist in the form of compiled binary code, or in the form of script or bytecode executable by an interpreter. The program product, through carefully designed algorithms and logical instructions, enables the processor to process data in a specific order and manner, complete various functions such as data analysis, user interaction, device control, etc.

[0144] In the description of the specification, the description of the terms "one embodiment / way", "some embodiments / ways", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment / way or example are included in at least one embodiment / way or example of the present application. In the description of the specification, the illustrative description of the above terms does not necessarily refer to the same embodiment / way or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any appropriate manner in any one or more embodiments / ways or examples. In addition, the person skilled in the art can combine and combine the different embodiments / ways or examples described in the specification and the features of the different embodiments / ways or examples, without contradiction.

[0145] In addition, the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the present application, the meaning of "a plurality of" is at least two, for example, two, three, etc., unless otherwise specifically limited.

[0146] The person skilled in the art should understand that the above-mentioned embodiments are only for the purpose of clearly illustrating the present application, and are not intended to limit the scope of the present application. For those skilled in the art, other changes or modifications can be made on the basis of the above-mentioned application, and these changes or modifications are still within the scope of the present application.

Claims

1. A knowledge-driven automatic generation method for part processing technology, characterized in that, include: A knowledge graph oriented towards processing technology is constructed, and the knowledge graph is coupled with a large knowledge model to form a feedback loop. The knowledge graph provides the large knowledge model with structured process constraints and semantic background. The large knowledge model makes inference decisions based on process constraints and feeds back the decision results to update the knowledge graph. Process knowledge units are extracted from multi-source data, and quantitative scoring and credibility screening of process knowledge units are performed based on a preset knowledge evaluation algorithm to obtain a highly reliable process knowledge dataset. Fine-grained attribute relationships are extracted from the process knowledge dataset, and the extracted results are structured and dynamically updated to the knowledge graph to achieve incremental knowledge updates.

2. The knowledge-driven automatic generation method for part processing technology according to claim 1, characterized in that, The steps to construct a knowledge graph oriented towards manufacturing processes include: The steps to establish a machining process method chain include: introducing a process matching engine, establishing a two-way coupling mechanism driven by part requirements and enterprise resource capabilities, the two-way coupling mechanism derives a candidate set of machining methods from top to bottom based on part characteristics, and filters and replaces the candidate set from bottom to top based on enterprise resource capabilities to generate a machining process method chain that meets machining requirements and resource constraints. The machining process method chain is mapped to a graph database. The machining parts, machining surfaces, operations, tools, fixtures and materials in the machining process method chain are used as instance nodes in the graph. The machining sequence, usage relationship or applicability relationship between nodes are used as edges in the graph for semantic connection to construct a knowledge graph.

3. The knowledge-driven automatic generation method for part processing technology according to claim 1, characterized in that, The steps involved in coupling knowledge graphs with large knowledge models to form a feedback loop include: Vectorize the entities and relations in the knowledge graph and build a vector index library; When a processing requirement is received, the subgraph related to the processing requirement is retrieved from the vector index library; The subgraph and processing requirements are input into the knowledge model for reasoning, generating decision recommendations with confidence scores. The credibility of newly generated knowledge during the reasoning process is evaluated, and knowledge that meets the preset credibility threshold after evaluation is incrementally updated to the knowledge graph and vector index library.

4. The knowledge-driven automatic generation method for part processing technology according to claim 1, characterized in that, The steps for extracting process knowledge units from multi-source data and performing quantitative scoring and credibility screening include: Text data from multi-source data is preprocessed, and a BiLSTM-CNN-CRF architecture entity recognition model is used to extract process knowledge entities to obtain process knowledge entities. For the extracted knowledge entities, an existence verification is performed in the knowledge graph. If the entity exists in the knowledge graph, a multi-hop path retrieval is initiated for the corresponding knowledge entity. If the entity does not exist in the knowledge graph, it is stored in a temporary cache queue. A comprehensive path score is calculated based on multi-hop paths, and an adaptive screening threshold is determined based on the statistical distribution of the comprehensive path score to classify the reliability level of knowledge entities.

5. The knowledge-driven automatic generation method for part processing technology according to claim 4, characterized in that, In the input layer of the entity recognition model, the Shannon entropy, calculated based on the label distribution of the local text context, is concatenated with the embedding vector corresponding to the text to form enhanced input features, enabling the model's convolutional layers to adaptively focus on high-entropy regions during feature extraction.

6. The knowledge-driven automatic generation method for part processing technology according to claim 4, characterized in that, Methods for calculating the overall path score include: A three-factor linear combination model is used to weight and score multi-hop paths to obtain a comprehensive path score; the three factors include: The path length factor is calculated based on the logarithm of the number of entity nodes in the path. Termination node authority factor is quantified by comprehensively evaluating the global importance and local connectivity density of the terminal node in the path. The relational coherence factor is obtained by calculating the semantic similarity between the embedding vectors of adjacent relationships on the path.

7. The knowledge-driven automatic generation method for part processing technology according to claim 4, characterized in that, The steps for classifying the reliability levels of knowledge entities include: The path score of multiple paths associated with a knowledge entity is subjected to minimization normalization. The weighted aggregation of multiple path scores after normalization is performed to calculate the final reliability score of the knowledge entity. Based on the statistical distribution of the final reliability scores of all knowledge entities, an improved box plot method is used to determine the adaptive threshold. Knowledge entities are classified into high reliability, low reliability, or unreliable levels based on the comparison results of the final reliability score with the third quartile and the adaptive threshold. Knowledge entities with a final reliability score greater than the adaptive threshold are considered high reliable, knowledge entities with a final reliability score less than or equal to the third quartile are considered unreliable, and the remaining knowledge entities are considered low reliable.

8. The knowledge-driven automatic generation method for part processing technology according to claim 1, characterized in that, Multi-source data includes at least one or more of the following: process documents, standards and specifications, historical cases, structured reports, and sensor data; The steps for preprocessing text data from multi-source datasets include: A binary classifier based on BERT model fine-tuning is used to identify and label noisy samples in text data through supervised learning methods; Based on the identification and annotation results, data identified as noise samples are filtered out to obtain cleaned text. The cleaned text is segmented into words to generate sequence labeling samples for subsequent extraction of process knowledge units.

9. The knowledge-driven automatic generation method for part processing technology according to claim 7, characterized in that, The steps for dynamically updating a knowledge graph include: For knowledge entities classified as highly reliable, a Span-Based framework and a deep learning classifier are used to extract the attributes and relationships of the knowledge entities in a fine-grained manner. The extracted attributes and relationships are encapsulated in a structured manner and automatically mapped to the corresponding ontology structure of the knowledge graph to realize the logical association of knowledge. After the knowledge graph completes the mapping update, incremental vector embedding recalculation is triggered for newly added knowledge entities and relationships, and the retrieval index is updated synchronously.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the knowledge-driven automatic generation method for part processing technology as described in any one of claims 1-9.

Images