Grinding and polishing process parameter question-answering system design method based on knowledge graph

Abstract

The invention discloses a design method of a knowledge graph process parameter question-answering system for a grinding and polishing process. The invention provides a mode layer general construction method SDPKM of an industrial field processing technology, and provides a specific processing technology characteristic design by taking a grinding and polishing technology as an example, so that construction of a field knowledge graph is facilitated; according to the method, a multi-dimensional expression method MDPKA of a domain knowledge graph data layer is provided, the coupling and query efficiency of different types of data is improved, a double-domain knowledge graph construction method is provided for numerical data, a longitudinal graph is constructed, and the query efficiency is improved from a bottom-layer storage structure; according to the method, the large language model and the pre-training language model are adopted for knowledge extraction and graph completion, so that the training cost is reduced, and the usability of the method is improved; according to the semi-automatic construction method of the grinding and polishing process parameter question-answering system based on the knowledge graph, the technical threshold of users is greatly reduced, and landing of an actual production scene is facilitated.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic processing of grinding and polishing process, and particularly relates to a design method of a grinding and polishing process parameter question and answer system based on a knowledge graph. BACKGROUND

[0002] Grinding and polishing processing refers to a process of grinding, polishing and modifying the surface of a workpiece by mechanical, chemical or ultrasonic methods to achieve the required surface quality and appearance effect, which can meet various processing requirements such as cutting processing, surface treatment and mirror manufacturing. Due to the limitation of process parameter recommendation, the grinding and polishing robot cannot currently realize the customization and offline programming tasks of non-standard products independently.

[0003] Traditional process parameter recommendation and processing flow usually rely on manual teaching methods, that is, workers actually operate or input parameters to control the robot and then perform batch production according to a fixed process. The manual teaching method is easily affected by the difficulties such as a large number of grinding and polishing sub-processes, complex grinding and polishing mechanisms, large industrial production uncertainties, high worker training threshold and non-generalization and migration of production processes.

[0004] To promote the automation of the grinding and polishing process, the offline programming method has attracted widespread attention. However, this method faces difficulties such as difficult process parameter recommendation and high program preparation cost, resulting in poor accuracy and general applicability of the current industrial robot offline programming processing method and high application cost.

[0005] For the difficulty of process parameter recommendation in the grinding and polishing process, the current process parameter scheme design mainly relies on methods such as process manuals and worker experience, which are difficult for grinding and polishing industrial robots to accurately understand. Therefore, in terms of data processing, grinding and polishing process data needs to be cleaned and arranged as processing knowledge and stored as a knowledge graph. However, there is no mature framework for the entire data processing flow to be applied, the labor cost is high, the reliability of the extracted knowledge is poor, and thus the grinding and polishing process knowledge graph is difficult to be applied in specific production scenarios. SUMMARY

[0006] To solve the above problems of data processing and industrial application methods in the technical field of automatic grinding and polishing process, the present application provides a design method of a grinding and polishing process knowledge graph process parameter question and answer system.

[0007] A design method of a grinding and polishing process parameter question and answer system based on a knowledge graph, comprising: Step S1, a knowledge graph mode layer with six dimensions is designed, including six dimensions of process principle, workpiece, processing equipment, processing parameter, sub-process and post-processing; Step S2: For the constructed knowledge graph pattern layer, design a knowledge graph data layer with cognitive dimension, perception dimension and instance dimension; wherein, the cognitive dimension stores theoretical knowledge, the perception dimension stores pure numerical data in case data, and the instance dimension has timeliness and is used to store product information that conforms to the node attributes of the perception dimension. Step S3: Extract knowledge from the cognitive dimension, perceptual dimension, and instance dimension based on the large language model; Step S4: Complete the knowledge graph based on the pre-trained model; Step S5: Based on the improved grinding and polishing process knowledge graph, a knowledge graph question-answering system is constructed using a retrieval enhancement algorithm combined with semantic analysis. Step S6: Using the constructed grinding and polishing verification platform, conduct experiments based on the processing parameters recommended by the knowledge question and answer system, and return the quality parameters to the atlas for updating. If the quality of the workpiece after processing does not meet the standards, the parameters are re-recommended and the experiment is repeated.

[0008] Ideally, the basic principles of pattern layer design include conforming to natural language expression habits and human cognitive learning habits, so as to meet the interpretability and credibility of knowledge graphs. Ideally, ontology language is used for modeling, and Protégé software is used for visualization of the schema layer skeleton.

[0009] Preferably, in step 2, the cognitive dimension stores theoretical knowledge and classifies grinding and polishing processes based on the underlying physical and chemical principles of different processes. For grinding and polishing methods with different underlying principles, each process is decomposed step by step, and the union of relevant parameters is taken to ensure the completeness of the pattern layer. The decomposition rule is that the lower-level categories are mutually exclusive and do not overlap, and the mapping from the upper level to the lower level satisfies the injective condition. Using the inclusion relation in Protégé, the inference rule is that if an entity belongs to the tail category of the inclusion relation, then it must belong to the head category of the inclusion relation, and the mapping between parameters at the same level is bijective. The cognitive dimension node further provides sub-labels. The perceptual dimension's infrastructure can store data of the smallest unit of process processing, design the implementation layer to couple different unit actions, and provide evaluation parameters for the overall solution; the perceptual dimension proposes grinding and polishing data categories, which are related to all process parameters, so as to realize the connection and integration of multiple influencing factors.

[0010] Ideally, a dual-domain knowledge graph is constructed for knowledge storage; the core graph is arranged according to the process ordering pattern layer, and the core graph is used as the node attribute to construct the perception data node; the domain graph is named according to the SDPKM method to label the nodes, and the sensor output parameters in the actual industrial production process are stored as node attributes.

[0011] Preferably, step S3 specifically includes: Step S3-1, data preprocessing: the conversion of data format selects formats such as Markdown or JSON with clear semantic specifications, and gives prompts to improve the understanding accuracy of the large model for such format documents; a prompt template is designed and introduced to clearly instruct the model to output results in a specified format; for table data, OCR technology is used in combination with table structure recognition algorithms for automatic extraction; for picture data, screenshots are stored; Step S3-2, knowledge extraction: triple extraction is divided into three sub-tasks, which respectively process the relationships between entities, the relationships between entities and events, and the relationships between events; Step S3-3, label completion: after extraction, the distribution abstracts, clusters and generalizes entity, relationship and event types, merges synonymous expressions, associates with the mode layer skeleton, and stores as labels; Step S3-4, graph visualization: the extracted triples and labels are imported into Neo4j software for interactive visualization. Cypher language is used to set visualization parameters.

[0012] Preferably, in step S4, a graph completion algorithm that combines semantic representation and graph structure information is used to complete the missing triples (h, r, t) in the knowledge graph, and the task form includes link prediction and triple classification.

[0013] Preferably, in step S4, semantic embedding representation is realized based on the pre-trained language model BERT, and the ability of the model to distinguish between positive and negative samples is improved through a structured training strategy: specifically, in the training phase, a training set including positive and negative samples is constructed, and the loss function is optimized to minimize the loss of positive samples and maximize the loss of negative samples, thereby strengthening the model's ability to identify effective triples; in the reasoning phase, for a query to be completed, the model enumerates all entities in the knowledge graph and calculates the loss value of their corresponding triples, and the entity with the smallest loss is selected as the candidate completion result; further, a graph structure-based reordering strategy is introduced to constrain the candidate entity set within the k-hop neighbor range of the source entity in the training graph.

[0014] Preferably, step S5 specifically includes: Step S5-1: In the input analysis stage, the pre-trained domain semantic model is used to perform intent recognition and entity analysis on the user's input question, and extract key process elements; Step S5-2: In the graph reasoning stage, first map the keywords to the node and relationship space in the knowledge graph cognitive dimension, and give parameter optimization suggestions; then map the numerical data to the nodes of the domain knowledge graph in the knowledge graph perception dimension, match the recommended domain nodes according to the parameter optimization suggestions, match the corresponding core knowledge graph node intersection, and obtain the specific values of the recommended processing parameter scheme; finally, match the specific product in the instance dimension according to the processing parameter scheme as the product recommendation scheme; Step S5-3: In the output generation stage, through the language ability of the large language model, combined with the parameter optimization suggestions, the processing parameter scheme and the product recommendation scheme, combined with the reordering strategy of context perception, the most relevant knowledge fragments are selected as external evidence, and are injected into the generative language model to assist in generating accurate, interpretable and process logic answers. The framework provides an efficient and intuitive knowledge interaction interface for technicians in the field of grinding and polishing process, realizes the paradigm shift from "passive consultation" to "active questioning", and promotes the service and intelligent transformation of manufacturing knowledge.

[0015] Preferably, the step S6 specifically comprises: After processing, the surface quality of the workpiece is quantitatively evaluated by using online or offline detection means, and key quality parameters are obtained, including surface roughness, material removal rate, surface integrity and geometric precision retention; For edge grinding and polishing processing, the weighted sum of the arithmetic mean deviation Ra and the skewness Rsk of the profile can be taken as the evaluation index of the surface roughness, and then the quality score is calculated; the quality score is compared to evaluate the effect of process parameter recommendation, and for the process parameter scheme with a quality score lower than 60%, the prompt word is returned to the large model to regenerate the recommended scheme; the quality feedback data obtained by the experiment is fed back to the grinding and polishing process knowledge graph to realize the dynamic updating mechanism of the graph perception dimension.

[0016] In view of the above technical scheme, the present application has the following beneficial effects: The present application proposes a mode layer general construction method SDPKM for processing technology in the industrial field, and gives a specific processing technology characteristic design taking grinding and polishing process as an example, which facilitates the construction of field knowledge graph.

[0017] The present application proposes a multi-dimensional expression method MDPKA for the data layer of the field knowledge graph, which improves the coupling and query efficiency of different types of data. For numerical data, a double-domain knowledge graph construction method is proposed to construct a vertical graph, which improves the query efficiency from the bottom storage structure.

[0018] The present application respectively uses a large language model and a pre-trained language model for knowledge extraction and graph completion, which reduces the training cost and improves the usability of the method.

[0019] The application proposes a semi-automatic construction method of a polishing process parameter question and answer system based on a knowledge graph, which greatly reduces the technical threshold of users and is conducive to the landing of actual production scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 is a flow block diagram of a polishing process parameter question and answer system design method using a knowledge graph in the application; Figure 2 is a mode layer design flow chart based on a six-dimensional process knowledge modeling method in the application; Figure 3 is a polishing process mode layer constructed based on a six-dimensional process knowledge modeling method in the application; Figure 4 is a data layer design flow chart based on a multi-dimensional process knowledge expression architecture in the application; Figure 5 is a knowledge extraction flow chart based on a large language model in the application; Figure 6 is a process parameter question algorithm flow chart based on a knowledge graph in the application. DETAILED DESCRIPTION

[0021] The preferred embodiments of the application are described below in conjunction with the accompanying drawings, and it should be understood that the preferred embodiments described herein are only used to illustrate and explain the application, and are not used to limit the application.

[0022] The application proposes a design method of a knowledge graph process parameter question and answer system for polishing process, which includes: Step S1, knowledge graph mode layer design. The mode layer is designed in combination with the characteristics of the process itself and the commonality of the processing process, and is required to have the versatility of text data and numerical data, the rapidity of data type data query, and the flexibility of text data update. The application designs a six-dimensional process knowledge modeling method (Six-Dimensional Process Knowledge Modeling, SDPKM), which is suitable for most technical cases in the field of industrial processing.

[0023] Step S2, knowledge graph data layer design. Unlike the method of simple knowledge graph storing knowledge nodes in data layer according to mode layer classification, the application designs a multi-dimensional process knowledge architecture (MDPKA). In this method, the mode layer constructed is stored in a top-down order, and its data layer is divided into different dimensions, including cognitive dimension, perception dimension and instance dimension. At the same time, since the perception dimension and the instance dimension mainly represent numerical data, a double-domain graph structure is used for storage, and labels are added for knowledge of different sub-definition types, further improving the knowledge integration degree and index efficiency.

[0024] Step S3, knowledge extraction based on large language model. Different knowledge extraction methods are designed based on MDPKA. For cognitive dimension, agent is designed to realize knowledge extraction and node classification effect. First, large model prompt words are designed for preliminary extraction, in which the prompt word design strictly follows the output format requirement to ensure that each element contains "subject", "relation" and "object" fields. Then, knowledge cleaning is performed multiple times, and Chinese language syntax is designed for expression standardization. Finally, the knowledge returned in JSON array form is input into the classification model based on large language model, which conforms to the mode layer structure constructed by SDPKM. For perception dimension, parameter unit unification algorithm and structured data mapping rule are designed to import knowledge nodes. For instance dimension, matching algorithm is used to construct nodes.

[0025] Step S4, knowledge graph completion based on pre-trained model. For the cognitive dimension of the graph, a method based on pre-trained language model (such as BERT, RoBERTa, BioBERT, SciBERT, etc.) is used to preliminarily complete the graph, which uses its deep semantic understanding ability to extract implicit relationships from unstructured text, and scores the confidence of candidate triples, fine-tunes and integrates existing entity relationships, supplements missing relationships and entities, and improves the quality and practicality of the graph, providing a more solid knowledge foundation for subsequent intelligent knowledge question answering.

[0026] Step S5, process parameter question answering based on knowledge graph. Based on the improved grinding and polishing process knowledge graph, a knowledge graph question answering system is constructed using retrieval enhancement algorithm combined with semantic analysis.

[0027] Step S6, closed-loop verification based on grinding and polishing platform. Through the constructed grinding and polishing processing verification platform, the quality parameters are returned to the graph for updating according to the processing parameters recommended by the knowledge question answering system, and if the quality of the workpiece after processing does not meet the standard, the parameters are re-recommended for repeated experiments.

[0028] Step S1, knowledge graph schema layer design: the knowledge graph has a flexible storage paradigm, which can be stored in different forms according to different application scenarios and requirements.

[0029] Step S1-1, basic principles of schema layer design: the present application proposes the basic principles of knowledge graph schema layer design. The schema layer design of the domain knowledge graph should conform to the natural language expression habits and the habitual way of thinking of human cognitive learning, in order to meet the explainability and reliability of the knowledge graph. Therefore, the schema layer is designed for different downstream tasks from the perspectives of cognition and practice.

[0030] Step S1-2, taking grinding and polishing processing technology as an example: data types can be divided into text data (such as "wide grinding wheel grinding relies on increasing grinding width to improve grinding efficiency", "3M product number 334U purple dry grinding square sand roll, P240, 70mmx12m") and numerical data ("the width of the cylindrical grinding wheel can reach 300mm"). The present application adopts the open world assumption, fixes the numerical data schema layer to ensure query speed and data quality, and gives the update space of the text data schema layer. Specifically, knowledge can be divided into grinding and polishing case data and grinding and polishing theoretical knowledge, which correspond to numerical data in actual processing and processing rules obtained through data analysis and experience summary. Grinding and polishing knowledge mainly includes process principle, workpiece, equipment, parameter, process and post-processing. Therefore, the six-dimensional process knowledge modeling method SDPKM is proposed.

[0031] The SDPKM schema design method includes the following six core dimensions: process principle (Principle), workpiece (Part), equipment (Plant), parameter (Parameter), process (Process), and post-processing (Post-processing).

[0032] The six core dimensions conform to the core steps of the process and the process planning flow in industrial processing.

[0033] Step S1-3, visualization of grinding and polishing knowledge graph: ontology language modeling is adopted, and the schema layer skeleton is visualized through Protégé software.

[0034] Step S2, knowledge graph data layer design: for the constructed schema layer, the storage details of related knowledge need to be further specified, such as for the "grinding and polishing process" mode under the "wide grinding wheel grinding" node, "grinding wheel width 300mm" as node attribute, or separately construct "grinding wheel" node, "width 300mm" as node attribute or relationship attribute.

[0035] Step S2-1, multi-dimensional process knowledge expression architecture: a multi-dimensional process knowledge expression architecture MDPKA is proposed, which divides the data layer into cognitive dimension, perception dimension and instance dimension. The cognitive dimension stores theoretical knowledge, which can further improve the complexity of the mode layer sub-level according to the subdivision of the polishing process. The perception dimension stores the pure numerical data in the case data, which mainly takes the four core dimensions of processing workpiece, processing equipment, processing parameter and post-processing as the attributes Property of the case data node. The instance dimension has timeliness and is used to store product information conforming to the node attributes of the perception dimension, such as specific polishing robot model, polishing machine tool model, clamp model, tool model, etc.

[0036] Step S2-2, cognitive dimension design: the cognitive dimension stores theoretical knowledge, and the polishing process is classified according to different underlying physical and chemical principles. For different underlying principles of polishing processing method, it is respectively disassembled, and the related parameters are taken as a set to ensure the completeness of the mode layer. The disassembly rule is that the lower layer classification is opposite and does not coincide with each other, and the mapping from the upper layer to the lower layer satisfies the injection condition, so the inclusion relationship in Protégé is used, and the reasoning rule is that the entity belongs to the tail category of the inclusion relationship, which must belong to the head category of the inclusion relationship, and the mapping between the parameters in the same layer is bijective. The cognitive dimension node can further give sub-labels. Taking the cognitive dimension node of the polishing process as an example: Process principle: three-stage theory of abrasive particles, brittle material fracture, plastic material shear slip, chemical corrosion, elastic impact, magnetorheological polishing, Preston equation, Hertz contact theory Processing workpiece: workpiece material, workpiece structure, workpiece type, workpiece roughness, material melting point, material hardness, material ductility, chemical activity, subsurface damage Processing equipment: robot type, end type, clamp type, power type, control system Processing parameter: spindle speed, feed rate, floating force, processing angle, abrasive attribute, polishing liquid Sub-process: mirror polishing, electrochemical polishing, mechanical grinding, chemical mechanical polishing Post-processing: heat treatment, magnetic flaw detection, cleaning, rust prevention, surface coating, surface passivation Step S2-3, perception dimension design: the basic framework of the perception dimension can store the data of the smallest unit of process, and the design practice layer couples different unit actions to give the evaluation parameters of the overall scheme. At the same time, the perception dimension proposes the polishing data category, which has a correlation with all the process parameters to realize the connection and integration of multiple influencing factors.

[0037] The dual-domain knowledge graph is constructed for knowledge storage; the core graph is sorted in a process ordering mode layer, which is used as a node attribute to construct a perception data node; and the domain graph stores the output parameters of sensors in actual industrial production processes as node attributes according to the node label Label named by the SDPKM method.

[0038] Step S3, knowledge extraction: The unstructured data is extracted according to the large model prompt word of the above mode layer. In the process of extracting most domain knowledge graphs, the main difficulty and workload lies in data format conversion and data cleaning.

[0039] Step S3-1, data preprocessing: The data format conversion selects Markdown or JSON format with clear semantic specifications, and gives the corresponding prompt to improve the accuracy of the large model in understanding this type of format document. In order to enhance the accuracy of the large model in understanding the target format, a prompt template is designed and introduced to clearly indicate the model to output the results in the specified format. For example, add the following prompt in the prompt word: “Please organize the following content in Markdown format, use # for main title, ## for sub-title, and - for item list” or “Please return in JSON object form, containing 'entity','relation', 'event' three keys”. Through such guidance, the standardization and consistency of the model output are significantly improved. For table data, OCR technology combined with table structure recognition algorithm is used for automatic extraction; for picture data, screenshot storage is performed.

[0040] Step S3-2, knowledge extraction: The triple extraction can be divided into three sub-tasks, which respectively process the relationship between entities, the relationship between entities and events, and the relationship between events. The step-by-step construction method supports batch processing and incremental updating, and is suitable for efficient processing of large-scale corpus.

[0041] Step S3-3, label completion: After extraction, the entity, relationship and event type are abstracted, clustered and generalized, and the synonyms are merged. The mode layer skeleton is associated and stored as a label. The label is beneficial to establish an index in the Neo4j storage form, and further improve the query efficiency.

[0042] Step S3-4, graph visualization: The extracted triple and label are imported into the Neo4j software for interactive visualization display. The visualization parameters are set through the Cypher language, such as limiting the displayed entity type, controlling the number of nodes displayed, and adjusting the layout algorithm to meet the clear display needs of large-scale graphs. Graph visualization is beneficial to analyze the extraction quality, data characteristics and perform error auxiliary checking.

[0043] Step S4, graph completion: a graph completion algorithm that fuses semantic representation and graph structure information is used to complete the missing triples (h, r, t) in the knowledge graph. The task form mainly includes link prediction and triple classification. Link prediction is used to infer the missing entity in the query (h, r,?) or (?, r, t), and the evaluation indexes are mean reciprocal rank (MRR), hit rate (H@1, H@3, H@10), average rank (MR) and conditional rank (CR@1). Triple classification is used to judge the correctness of the given triple (h, r, t), and the evaluation indexes are accuracy (Acc) and F1 value.

[0044] The application realizes semantic embedding representation based on a pre-trained language model BERT, and improves the ability of the model to distinguish positive and negative samples through a structured training strategy. Specifically, in the training phase, a training set including positive samples and negative samples is constructed, and the loss of the positive samples is minimized and the loss of the negative samples is maximized by optimizing the loss function, thereby strengthening the recognition ability of the model to effective triples. In the reasoning phase, for the query to be completed, the model enumerates all entities in the knowledge graph and calculates the loss value of the corresponding triples, and the entity with the minimum loss is taken as the candidate completion result. To further improve the prediction accuracy and reliability, a reordering strategy based on the graph structure is introduced, which constrains the candidate entity set within the k-hop neighbor range of the source entity in the training graph, thereby reducing the size of the candidate pool, improving the calculation efficiency, and enhancing the structural consistency of the completion result. The combination of semantic representation and graph topology information significantly improves the efficiency and accuracy of knowledge graph completion, and is suitable for the automatic construction and completion of large-scale knowledge graphs.

[0045] Step S5, knowledge graph question answering: based on the improved grinding process knowledge graph, the Retrieval-Augmented Generation (RAG) algorithm and deep semantic analysis technology are combined to break through the limitations of traditional keyword matching and realize accurate understanding and efficient response to natural language questions. Specifically, in the question answering process, there are three stages: input analysis, graph reasoning and output generation.

[0046] Step S5-1: In the input analysis stage, the pre-trained domain semantic model is used to identify the intent and analyze the entities of the user input question, and extract key process elements (such as "workpiece material", "surface roughness requirement", "equipment type", etc.).

[0047] Step S5-2: In the graph reasoning stage, first map the keywords to the node and relationship space in the knowledge graph cognitive dimension, and give parameter optimization suggestions. Then, map the numerical data to the nodes of the domain knowledge graph in the knowledge graph perception dimension, match the recommended domain nodes according to the parameter optimization suggestions, and then match the corresponding core knowledge graph node intersection to obtain the specific values of the recommended processing parameter scheme. Finally, match the specific product in the instance dimension according to the processing parameter scheme as the product recommendation scheme.

[0048] Step S5-3: In the output generation stage, through the language ability of the large language model, combined with the parameter optimization suggestions, the processing parameter scheme and the product recommendation scheme, combined with the reordering strategy of context perception, the most relevant knowledge fragments are selected as external evidence, and injected into the generative language model to assist in generating accurate, interpretable and process logic consistent answers. This framework provides an efficient and intuitive knowledge interaction interface for technicians in the field of grinding and polishing processes, realizes the paradigm shift from "passive consultation" to "active questioning", and promotes the service and intelligent transformation of manufacturing knowledge.

[0049] Step S6, closed loop verification: build a grinding and polishing processing verification experiment platform, integrate high-precision grinding and polishing equipment, multi-modal sensing system (monocular camera, tachometer), and reproduce the key process parameter combinations suggested by the knowledge question and answer system, such as abrasive particle size, grinding wheel speed, feed speed, pressure control, cooling method and processing time, etc.

[0050] After processing, use online or offline detection means to quantitatively evaluate the surface quality of the workpiece, obtain key quality parameters, including but not limited to surface roughness, material removal rate, surface integrity (presence or absence of scratches, burns, micro-cracks) and geometric accuracy retention, etc.

[0051] Based on the roughness and its reduction, the material removal efficiency is calculated to calculate the quality score of the grinding and polishing process, and the specific formula is: quality score = roughness reduction / workpiece roughness x material removal efficiency x 100%; In the specific evaluation process, different surface roughness formulas are used for evaluation considering different grinding and polishing processing scenarios. For edge grinding, the calculation formula of the arithmetic mean deviation Ra of the profile is as follows:

[0052] Where Z(x) is the profile ordinate, and l is the total length of the profile. For plane grinding, the calculation formula of the arithmetic mean deviation Sa of the plane is as follows:

[0053] Wherein, Z(x, y) is the plane longitudinal coordinate, A is the total area of the sampling surface. In order to further express the surface profile details, the root mean square deviation Rq of the profile can be used, which is more sensitive to peaks and valleys:

[0054] Generally, the industrial field can use the following formula for approximate conversion:

[0055] In addition, the skewness Rsk of the profile represents the asymmetry of the longitudinal coordinate value probability density function, which is greatly affected by discrete peaks or discrete valleys, and the calculation formula is as follows:

[0056] For edge grinding and polishing processing, the weighted sum of the arithmetic mean deviation Ra and the skewness Rsk of the profile can be taken as the evaluation index of the surface roughness, and then the quality score is calculated. The recommended effect of the process parameters is evaluated by comparing the quality score. For the process parameter scheme with a quality score lower than 60%, the prompt word is returned to the large model to generate a recommended scheme again. The quality feedback data obtained by the experiment is fed back to the grinding and polishing process knowledge graph, and a dynamic updating mechanism of the graph perception dimension is realized.

[0057] To sum up, the above is only a preferred embodiment of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A knowledge graph-based polishing process parameter question and answer system design method, characterized in that, Comprise: Step S1, design a knowledge graph mode layer with six dimensions, including: process principle, workpiece, processing equipment, processing parameters, sub-process and post-processing six dimensions; Step S2, for the constructed knowledge graph mode layer, design a knowledge graph data layer with cognitive dimension, perception dimension and instance dimension; wherein, the cognitive dimension stores theoretical knowledge, the perception dimension stores pure numerical data in the case data, and the instance dimension has timeliness and is used for storing product information conforming to the node attribute of the perception dimension; Step S3, knowledge extraction based on large language model for cognitive dimension, perception dimension and instance dimension; Step S4, complete the knowledge graph based on pre-training model; Step S5, based on the improved grinding and polishing process knowledge graph, use retrieval enhancement algorithm, combined with semantic analysis, build knowledge graph question answering system; Step S6, through the built grinding and polishing processing verification platform, according to the processing parameters recommended by the knowledge question answering system, carry out experiment, return the quality parameters to the graph for updating, if the quality of the workpiece after processing does not meet the standard, re recommend parameters for repeated experiment.

2. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, The basic principles of mode layer design include conforming to the natural language expression habit and the habitual way of thinking of human cognitive learning to meet the explainability and credibility of knowledge graph.

3. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, Ontology language modeling is adopted, and the Protégé software is used for visualization of the mode layer skeleton.

4. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, In step 2, the cognitive dimension stores theoretical knowledge, and the grinding and polishing process is classified according to the underlying physical and chemical principles of different processes. The grinding and polishing process of different underlying principles is disassembled step by step, and the related parameters are taken and set to ensure the completeness of the mode layer; The disassembly rule is that the lower classification is opposite and does not overlap, and the mapping from the upper layer to the lower layer meets the injection condition, the inclusion relationship in Protégé is used, and the reasoning rule is that the entity belongs to the tail category of the inclusion relationship, which must belong to the head category of the inclusion relationship, and the mapping between parameters in the same layer is bijective; The cognitive dimension node further gives sub-labels; The basic framework of perception dimension can store the data of the smallest unit of process, design practice layer to couple different unit actions, and give the evaluation parameters of the whole scheme; The perception dimension proposes the grinding and polishing data category, which has a correlation with all process parameters, so as to realize the connection and integration of multiple influence factors.

5. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, Build a double-domain knowledge graph for knowledge storage; The core graph is sorted according to the mode layer of the process, which is used as the node attribute to build the perception data node; The domain graph names the node label Label according to the SDPKM method, and stores the sensor output parameters in actual industrial production process as node attributes.

6. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, The step S3 specifically includes: Step S3-1, data preprocessing: the conversion of data format selects Markdown or JSON format with clear semantic specification, and gives the prompt to improve the understanding accuracy of large model for this kind of format document; Design and introduce prompt template to clearly indicate the model to output results in specified format; For table data, use OCR technology combined with table structure recognition algorithm to automatically extract; For picture data, take a screenshot and store it; Step S3-2, knowledge extraction: the triple extraction is subdivided into three subtasks, which respectively process the relationship between entities, the relationship between entities and events, and the relationship between events; Step S3-3, label completion: after extraction, the distribution abstracts, clusters and generalizes the entities, relationship and event types, merges the synonymous expressions, associates with the mode layer skeleton, and stores as labels; Step S3-4, atlas visualization: the extracted triples and labels are imported into Neo4j software for interactive visualization. The visualization parameters are set through Cypher language.

7. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, In step S4, a graph completion algorithm that fuses semantic representation and graph structure information is used to complete the missing triples (h, r, t) in the knowledge graph, and the task form includes link prediction and triple classification.

8. The design method of a knowledge graph-based polishing process parameter question and answer system according to claim 7, characterized in that, In step S4, semantic embedding representation is realized based on pre-trained language model BERT, and the ability of the model to distinguish positive and negative samples is improved through a structured training strategy: in the training phase, a training set including positive and negative samples is constructed, and the loss of positive samples is minimized and the loss of negative samples is maximized through optimization of the loss function, thereby strengthening the model's ability to identify effective triples; in the reasoning phase, for the query to be completed, the model enumerates all entities in the knowledge graph and calculates the loss value of the corresponding triples, and the entity with the smallest loss is taken as the candidate completion result; further, a graph structure-based reordering strategy is introduced to constrain the candidate entity set within the k-hop neighbor range of the source entity in the training graph.

9. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, The step S5 specifically includes: Step S5-1: in the input analysis stage, the pre-trained domain semantic model is used to perform intent recognition and entity analysis on the user input question, and extract key process elements; Step S5-2: in the graph reasoning stage, first, map the keywords to the node and relationship space in the knowledge graph cognitive dimension, and give parameter optimization suggestions; then, map the numerical data to the nodes of the domain knowledge graph in the knowledge graph perception dimension, match the recommended domain nodes according to the parameter optimization suggestions, and then match the corresponding core knowledge graph node intersection to obtain the specific values of the recommended processing parameter scheme; finally, match the product in the instance dimension according to the processing parameter scheme as the product recommendation scheme; Step S5-3: in the output generation stage, the language ability of the large language model is combined with the parameter optimization suggestion, the processing parameter scheme and the product recommendation scheme, and the reordering strategy with context perception is used to select the most relevant knowledge fragments as external evidence, which are injected into the generative language model to assist in generating accurate, interpretable and process-logic-compliant answers. This framework provides an efficient and intuitive knowledge interaction interface for technicians in the grinding and polishing process field, realizes the paradigm shift from "passive consultation" to "active questioning", and promotes the service and intelligent transformation of manufacturing knowledge.

10. The knowledge graph-based polishing process parameter question and answer system design method of claim 1, wherein, The step S6 specifically includes: After processing, the workpiece surface quality is quantitatively evaluated by online or offline detection means to obtain key quality parameters, including surface roughness, material removal rate, surface integrity and geometric precision retention; For edge polishing processing, the weighted sum of the arithmetic mean deviation Ra and the skewness Rsk of the profile can be taken as the evaluation index of surface roughness, and then the quality score is calculated; the quality score is compared to evaluate the effect of process parameter recommendation, and for the process parameter scheme with a quality score lower than 60%, the recommendation scheme is regenerated by returning to the large model through a prompt word; the quality feedback data obtained through experiments is fed back to the polishing process knowledge graph to realize the dynamic updating mechanism of the graph perception dimension.

Images