Manufacturing process disturbance knowledge graph generation method and system and storage medium

Abstract

The invention discloses a method and a system for generating a manufacturing process disturbance knowledge graph, and a storage medium. The method comprises the following steps: constructing a mode layer of the knowledge graph; inputting the SMT manufacturing process disturbance data set into the trained entity recognition model for processing to obtain an entity in the SMT manufacturing process disturbance data set; extracting the context of the entity to obtain an attribute corresponding to the entity and a relationship between the entities; fusing entities in the SMT manufacturing process disturbance data set, attributes corresponding to the entities and relationships between the entities based on the mode layer, and constructing a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; performing relation completion and attribute completion on the SMT manufacturing process disturbance basic knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph; according to the method for generating the disturbance knowledge graph in the manufacturing process, system modeling and efficient utilization of bad process knowledge are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent manufacturing and industrial knowledge graph, and particularly relates to a manufacturing process disturbance knowledge graph generation method and system and a storage medium. BACKGROUND

[0002] With the continuous acceleration and deepening of enterprise digital transformation, the manufacturing industry is undergoing unprecedented changes and upgrades. Against this background, the SMT production line (Surface Mount Technology Line, i.e. Surface Mount Technology production line) as the core link of the electronic manufacturing industry, its efficient and stable operation is of great significance to ensure product quality, improve production efficiency and reduce manufacturing cost. However, in the long-term operation of the SMT production line, a large amount of detailed records and in-depth reports on process defect causes (disturbances) will inevitably accumulate. These textual materials contain rich semantic information, which are directly related to the root cause of product defects, and also reflect the potential problems and improvement space in the production process.

[0003] However, the reality is that although these defect cause records and reports contain great value, existing modeling techniques often fail to fully express the process knowledge implied therein, and are too dependent on human experience and subjective judgment, resulting in time-consuming and labor-intensive knowledge acquisition process, high cost, difficulty in effective integration and reuse, and thus limiting its widespread dissemination and in-depth application within the enterprise. In addition, with the continuous progress of SMT production line technology and the increasing complexity of production process, the types and quantities of defect cause information are also increasing rapidly, which undoubtedly further exacerbates the difficulty of knowledge acquisition and management.

[0004] In the face of the above challenges, the present application proposes a manufacturing process disturbance knowledge graph generation method, which realizes the systematic modeling and efficient utilization of defective process knowledge. SUMMARY

[0005] The purpose of the present application is to provide a manufacturing process disturbance knowledge graph generation method, system and storage medium, which realizes end-to-end intelligent construction from unstructured text to structured knowledge graph, significantly improving the automation level and decision support capability of SMT production line defect knowledge management.

[0006] In a first aspect, the present application provides a manufacturing process disturbance knowledge graph generation method, comprising: obtaining an SMT manufacturing process disturbance data set; constructing a schema layer of the knowledge graph; the schema layer defines entity types, attribute types and relationship types; inputting the SMT manufacturing process disturbance dataset into the trained entity recognition model to obtain entities in the SMT manufacturing process disturbance dataset; extracting the context of the entities to obtain attributes corresponding to the entities and relationships between the entities; fusing the entities in the SMT manufacturing process disturbance dataset, the attributes corresponding to the entities, and the relationships between the entities based on the schema layer, and constructing a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; completing relationships of the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph; completing attributes of the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0007] Optionally, the entity types include engineers, errors, problems, and solutions; the relationship types include errors found by engineers, problems directly caused by errors, chain reactions between problems, problems or errors addressed by solutions, proposers of solutions, and detectors of quality problems; and the attribute types include positions and codes of engineers, machines and employees related to problems, machines, employees, and supply chains related to solutions.

[0008] Optionally, the SMT manufacturing process disturbance dataset needs to be preprocessed before being inputted into the trained entity recognition model, including: performing word segmentation and special character removal on sentences in the SMT manufacturing process disturbance dataset in sequence to obtain a plurality of wordpieces; adding special marks at the beginning and end of the sentences to identify the beginning and end of the sentence sequence.

[0009] Optionally, the entity recognition model includes sequentially connected RoBERTa-WWM-large layers, BiLSTM layers, linear layers, and CRF layers.

[0010] Optionally, the inputting the SMT manufacturing process disturbance dataset into the trained entity recognition model to obtain entities in the SMT manufacturing process disturbance dataset includes: inputting sentences in the SMT manufacturing process disturbance dataset into the RoBERTa-WWM-large layers to perform language modeling and obtain high-dimensional vector representations of all wordpieces in the sentences; inputting the high-dimensional vector representations of all wordpieces in the sentences into the BiLSTM layers to process and capture dependency relationships between each wordpiece and its preceding and subsequent context, and generate global semantic information vector representations of all wordpieces; The global semantic information vector representation of all word elements is input into a linear layer for linear processing to obtain the linearly transformed vector representation of all word elements; The global semantic information vector representation of all word elements is input into a linear layer for linear processing to calculate the score of each word element for each entity label based on the entity type set defined by the pattern layer, and obtain the Emission Score vector of all word elements; The Emission Score vector of all word elements is input into a CRF layer to predict the optimal entity label for all word elements and obtain the global optimal label sequence; According to the global optimal label sequence, the entities in the SMT manufacturing process disturbance data set are obtained.

[0011] Optionally, the relationship completion of the SMT manufacturing process disturbance basic knowledge graph is performed to obtain a second SMT manufacturing process disturbance knowledge graph, including: Based on the Cypher query language of the Neo4j database, the missing relationship is found; Based on the missing relationship, the context information of the node corresponding to the missing relationship is extracted: the node is an entity; the context information of the node corresponding to the missing relationship includes a node attribute set, a node existing relationship type set, and a relationship and attribute set of associated nodes.

[0012] The context information of the node corresponding to the missing relationship is input into the trained first LLM model for judgment to obtain a judgment result of the missing relationship; the judgment result of the missing relationship includes a relationship existence judgment result, a first confidence score, and a technical basis explanation; A first confidence threshold is set, and the missing relationship with a first confidence score not less than the first confidence threshold is supplemented into the SMT manufacturing process disturbance basic knowledge graph to obtain the second SMT manufacturing process disturbance knowledge graph.

[0013] Optionally, the attribute completion of the second SMT manufacturing process disturbance knowledge graph is performed to obtain a final SMT manufacturing process disturbance knowledge graph, including: Based on the attribute existence detection query of the Neo4j database, the missing attribute is found, and a queue for completion is generated; The context information of the node corresponding to each missing attribute in the queue for completion is extracted in multiple dimensions to obtain the context information of the node corresponding to the missing attribute; The context information of the node corresponding to the missing attribute is input into the trained second LLM model for judgment to obtain a judgment result of the missing attribute; the judgment result of the missing attribute includes a generated attribute value, a second confidence score, and a technical rationality explanation; A second confidence threshold is set, and the missing attributes with the second confidence score not less than the second confidence threshold are supplemented into the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0014] In a second aspect, the present application provides a manufacturing process disturbance knowledge graph generation system, comprising: A data acquisition module is configured to acquire an SMT manufacturing process disturbance dataset. A schema definition module is configured to construct a schema layer of the knowledge graph, wherein the schema layer defines entity types, attribute types and relationship types. An entity recognition module is configured to input the SMT manufacturing process disturbance dataset into a trained entity recognition model for processing to obtain entities in the SMT manufacturing process disturbance dataset. A relationship attribute extraction module is configured to extract the context of the entities to obtain attributes corresponding to the entities and relationships between the entities. A preliminary construction module is configured to fuse the entities, the attributes corresponding to the entities and the relationships between the entities in the SMT manufacturing process disturbance dataset based on the schema layer, and construct a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database. A first completion module is configured to perform relationship completion on the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph. A second completion module is configured to perform attribute completion on the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0015] In a third aspect, the present application provides a computer readable storage medium storing a computer program, wherein the computer program is executed to implement the manufacturing process disturbance knowledge graph generation method of the first aspect.

[0016] Compared with the prior art, the present application has the following beneficial effects: The application provides a manufacturing process disturbance knowledge graph generation method, system and storage medium. The method first defines a mode layer of the knowledge graph, defines entity types, attribute types and relationship types; secondly, based on an obtained SMT manufacturing process disturbance data set, entity recognition is performed in combination with an entity recognition model, and knowledge fusion is performed according to extracted relationships and attributes to obtain a first SMT manufacturing process disturbance knowledge graph; finally, relationship and attribute completion is performed based on the first SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph. The application realizes automatic extraction of key entities such as "equipment-error-problem-solution" and semantic relationships thereof from unstructured text, establishes a queryable and inferable structured knowledge network based on knowledge graph technology, and simultaneously realizes dynamic completion and optimization of the knowledge graph by using a large language model, thereby significantly improving the automation level and decision-making efficiency of adverse reason analysis, and providing intelligent support for SMT manufacturing process optimization and quality control. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 Fig. 1 shows a manufacturing process disturbance knowledge graph generation method flowchart in an embodiment of the application; Figure 2 Fig. 2 shows an SMT manufacturing process disturbance knowledge graph mode layer in an embodiment of the application; Figure 3 Fig. 3 shows a framework of an entity recognition model in an embodiment of the application; Figure 4 Fig. 4 shows a first SMT manufacturing process disturbance knowledge graph in an embodiment of the application.

[0018] Figure 5 Fig. 5 shows an entity type evaluation index result in an embodiment of the application; Figure 6 Fig. 6 shows a loss value change curve in a model training process in an embodiment of the application; Figure 7 Fig. 7 shows a comparison of multiple model evaluation index results in an embodiment of the application. DETAILED DESCRIPTION

[0019] The application will be further described below with reference to the drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the application, and cannot be used to limit the protection scope of the application.

[0020] Embodiment 1

[0021] As shown in Fig. 1, an embodiment of the application introduces a manufacturing process disturbance knowledge graph generation method, which includes the following steps: Figure 1 ​S01: Obtain an SMT manufacturing process disturbance dataset; S02: Construct a schema layer of the knowledge graph; the schema layer defines entity types, attribute types, and relationship types; S03: Input the SMT manufacturing process disturbance dataset into a trained entity recognition model for processing to obtain entities in the SMT manufacturing process disturbance dataset; S04: Extract the context of the entities to obtain attributes corresponding to the entities and relationships between the entities; S05: Fuse the entities in the SMT manufacturing process disturbance dataset, the attributes corresponding to the entities, and the relationships between the entities based on the schema layer, and construct a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; S06: Complete the relationships of the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph; S07: Complete the attributes of the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0022] The embodiment proposes a method for generating a manufacturing process disturbance knowledge graph, which first constructs a schema layer, then inputs an SMT manufacturing process disturbance dataset into a trained entity recognition model for recognition to obtain entities, then constructs a first SMT manufacturing process disturbance knowledge graph in a Neo4j database based on the schema layer and the recognized entities and extracted relationships and attributes, and finally completes the relationships and attributes of the first SMT manufacturing process disturbance knowledge graph through an LLM large language model to obtain a final SMT manufacturing process disturbance knowledge graph. The end-to-end intelligent construction from unstructured text (SMT manufacturing process disturbance dataset) to structured knowledge graph (SMT manufacturing process disturbance knowledge graph) significantly improves the automation level and decision support capability of SMT production line defect knowledge management.

[0023] In the embodiment, step S02 constructs a schema layer of the knowledge graph, as shown in Figure 2 The schema layer of the knowledge graph is equivalent to ontology modeling. Ontology is a formalized description of concepts, attributes, and their mutual relationships in a specific field, aiming to provide a unified and clear semantic understanding framework for machines and humans. In a knowledge graph, ontology as the core content of the concept layer specifies "what concepts are", "what relationships exist between these concepts", and "what attributes each concept has", thereby supporting knowledge organization, representation, and reasoning, and is a key foundation for achieving semantic consistency and intelligent processing.

[0024] The schema layer defines entity types, attribute types, and relationship types; 1) Entity types include: engineers, errors, problems, and solutions; Specifically, entities typically represent objective things or abstract concepts in the real world, such as people, places, diseases, or events. Regarding the collection of entities, this embodiment, through analysis of a company's database, summarizes four types of entities: Engineer, Error, Problem, and Solution. Table 1 provides specific definition examples: Table 1 Example of Entity Definition

[0025] Relationship types include errors discovered by engineers, problems directly caused by errors, chain reactions between problems, problems or errors addressed by solutions, and the proposer of solutions and the detector of quality problems; Specifically, in a triple, the entity relation acts as a bridge connecting the subject entity and the object entity, and is the core part of the triple representing semantics. A triple represents knowledge in the form of "head entity, relation, tail entity," where the relation defines the semantic link between two entities. Based on the defined entity types, this embodiment identifies and extracts six types of entity relations: HAS_ERROR, LEADS_TO, CAUSES, SOLVES, PROPOSED_BY, and DETECTED_BY; Table 2 shows examples of entity relation definitions. Table 2 Example of Entity Relationship Definition

[0026] Attribute types include engineer job title and code name, machines and employees related to the problem, machines, employees, and supply chain related to the solution.

[0027] In this embodiment, the SMT manufacturing process disturbance dataset needs to be preprocessed before step S03, specifically including: The statements in the SMT manufacturing process disturbance dataset are sequentially segmented and special characters are removed to obtain a number of tokens. Add special markers at the beginning and end of a statement, such as "[CLS]" to indicate the beginning of a statement and "[SEP]" to indicate the end of a statement.

[0028] In this embodiment, the entity recognition model in step S03 includes: a RoBERTa-WWM-large layer, a BiLSTM layer, a linear layer, and a CRF layer connected in sequence, such as... Figure 3 The diagram shows the framework of the entity recognition model. In this embodiment, step S03 inputs the SMT manufacturing process disturbance dataset into the trained entity recognition model for processing to obtain the entities in the SMT manufacturing process disturbance dataset, including: The statements in the SMT manufacturing process disturbance dataset are input into the RoBERTa-WWM-large layer for language modeling to obtain high-dimensional vector representations of all word units in the statement; The high-dimensional vector representations of all word units in the statement are input into the BiLSTM layer for processing to capture the dependency relationship between each word unit and its preceding and following context, and generate a global semantic information vector representation of all word units; The global semantic information vector representation of all word units is input into the linear layer for linear processing to obtain the linearly transformed vector representation of all word units; The global semantic information vector representation of all word units is input into the linear layer for linear processing to calculate the score of each word unit's global semantic information vector representation for each entity label, obtaining an Emission Score vector for all word units; the entity label is based on the entity type set defined by the pattern layer; The Emission Score vector of all word units is input into the CRF layer to predict the optimal entity label for all word units, obtaining a global optimal label sequence; According to the global optimal label sequence, the entities in the SMT manufacturing process disturbance dataset are obtained.

[0029] Specifically, the RoBERTa-WWM-large layer in the entity recognition model converts input text (statements) into vector representations with contextual semantics. It models the entire sequence through multiple Transformer encoder layers, and finally outputs a fixed-dimensional contextual semantic vector for each token, providing semantic-rich feature input for downstream models.

[0030] Specifically, the BiLSTM layer effectively captures the dependency relationship between each word and its preceding and following context by simultaneously introducing forward and backward LSTM networks, thereby generating a representation containing global semantic information and utilizing the gating mechanism to retain key information and suppress useless interference. It introduces temporal order and directionality based on the token semantic representation provided by RoBERTa-WWM-large, enhancing the ability to capture syntactic structure and dynamic context dependencies. For each token in the input sequence, BiLSTM outputs a vector that combines the token's left and right context long-term dependency information.

[0031] Specifically, the main role of the CRF layer is to optimize the results of sequence labeling. The BiLSTM layer provides context information for each token, but it independently predicts the label for each token, ignoring the dependencies between labels. The CRF layer models these dependencies by introducing transition probabilities between labels, ensuring the rationality of the entire label sequence. In addition, CRF trains the model by maximizing the conditional probability of the label sequence, ensuring that the output label sequence best matches the true annotation. This enables the model to make more accurate label predictions in the global context.

[0032] In this embodiment, the RoBERTa-WWM-large layer is improved based on the Bert model, and the specific improvements are as follows: Compared with the Bert model, the RoBERTa-WWM-large makes the following improvements: 1) Remove the next sentence prediction task (NSP): Unlike BERT's two pre-training tasks (Masked Language Model and Next Sentence Prediction), RoBERTa-WWM-large removes the Next Sentence Prediction task, considering that this task has limited improvement in language understanding. After removing NSP, RoBERTa-WWM-large can focus more on the language modeling task, thereby improving the effect.

[0033] 2) Dynamic masking (Dynamic Masking): BERT uses a fixed mask during pre-training for the MLM task, which means that the mask selection is fixed during training. RoBERTa-WWM-large uses a dynamic masking strategy, that is, the mask is different in each training step, which enhances the diversity of training and helps to improve the generalization ability of the model.

[0034] 3) More training data: RoBERTa-WWM-large uses more training data than BERT. In the pre-training of BERT, BooksCorpus and English Wikipedia are used as two large corpora, while RoBERTa-WWM-large adds CC-News, OpenWebText, Stories, etc. Datasets greatly expand the range and diversity of training corpus, enabling the model to better capture a wider range of language features.

[0035] 4) Longer training time and larger batch size: The original BERT model uses about 16GB of corpus, 1M total training steps, and a batch size of 256; while RoBERTa-WWM-large uses a larger scale of 160GB training data, the number of training steps increases to between 500K and 1.5M, and the batch size increases to 2,000 or even 8,000. Under such a configuration, RoBERTa-WWM-large can more fully learn the semantic relationship in the massive corpus, and obtain more stable and optimal parameters.

[0036] In this embodiment, the training steps of the entity recognition model include: Obtaining a set of SMT manufacturing process failure cause data; the set of failure cause data includes labeled entity tags; Taking the set of SMT manufacturing process failure cause data as a sample set, and inputting the sample set into the entity recognition model for training to obtain optimal parameters of the entity recognition model; Based on the optimal parameters of the entity recognition model, a trained entity recognition model is obtained.

[0037] In this embodiment, the context of the entity is extracted in S04 to obtain the attributes corresponding to the entity and the relationship between the entities; this step is based on artificial preliminary extraction.

[0038] In this embodiment, S05 fuses the entities, attributes corresponding to the entities, and relationships between the entities in the SMT manufacturing process disturbance data set based on the mode layer, and constructs a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; as shown in the first SMT manufacturing process disturbance knowledge graph partial example diagram. Figure 4

[0039] Specifically, after the above entity recognition, relationship attribute extraction, and knowledge fusion, a visualization example of the knowledge graph based on the Neo4j database is adopted. SMT process failure analysis (analysis of SMT manufacturing process disturbance) is a complex quality management process, which can be divided into printing, soldering, reflow soldering, detection and other key processes. Based on the graph mode layer, the data layer of the SMT process disturbance knowledge table is constructed, the extracted entities such as failure types, causes, processes, equipment, and parameters and their associated relationships are stored in the Neo4j database, and further connected to the front-end visualization module to realize the graphical display of the failure cause knowledge graph.

[0040] Figure 4 ​An example of the first SMT manufacturing process disturbance knowledge graph, which is a structured data model composed of nodes and relationships. Each node represents a specific entity and can have one or more labels to identify the type of entity and contain several attributes. Relationships represent the connection between two nodes and have directionality, as well as types and attributes to express the specific meaning of the relationship. The Cypher query language can be used to flexibly retrieve and explore the graph structure.

[0041] In this embodiment, S06 completes the relationships of the SMT manufacturing process disturbance base knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph, which includes: S61: Find missing relationships based on the Cypher query language of the Neo4j database; Specifically, the following two query modes are used: Table 3 Query Mode

[0042] The query results are returned in the form of triples (defect node name, candidate cause / solution node name, relationship type to be verified), providing a clear completion target for subsequent processing.

[0043] S62: Based on the missing relationship, extract the context information of the missing relationship corresponding node: the node is an entity; the context information of the missing relationship corresponding node includes the node's own attribute set, the node's existing relationship type set, and the relationship and attribute set of the associated node.

[0044] Specifically, to support subsequent semantic reasoning, the structured feature information of the relevant nodes needs to be extracted. The complete context of the target node and its associated network is obtained through an extended Cypher query, which returns context information containing three dimensions: 1) the node's own attribute set; 2) the node's existing relationship type set; 3) the label and attribute set of the associated node. This multi-level context representation provides sufficient semantic basis for relationship reasoning.

[0045] S63: Input the context information of the missing relationship corresponding node into the trained first LLM model to obtain the judgment result of the missing relationship; the judgment result of the missing relationship includes: relationship existence judgment result, first confidence score, and technical basis explanation; Specifically, a large language model (LLM) is used as a probability reasoning engine for the existence of relationships, outputting three key fields: 1) relationship existence judgment (Boolean value); 2) confidence score (0-1 continuous value); 3) technical basis explanation (string). This structured output ensures machine processability while retaining human interpretable decision-making processes.

[0046] Wherein, the training process of large language model (LLM) is well known to those skilled in the art, and will not be described in detail here. According to different input training data, a trained first LLM model and a trained second LLM model are obtained respectively.

[0047] S64: Set a first confidence threshold, and supplement the missing relationship whose first confidence score is not less than the first confidence threshold into the SMT manufacturing process disturbance base knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph.

[0048] Specifically, the first confidence threshold θ = 0.7 is set as the quality control standard for relationship storage. For candidate relationships that satisfy the first confidence confidence1≥θ, the following operations are performed: 1) create a corresponding relationship edge in Neo4j. 2) Add metadata properties (including confidence, creation method mark, etc.) to the relationship.

[0049] In this embodiment, step S07 performs attribute completion on the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph, including: S71: Based on the attribute existence detection query of the Neo4j database, find the missing attributes and generate a queue for completion; S72: For each missing attribute corresponding to the node in the queue for completion, perform multi-dimensional context extraction to obtain context information of the node corresponding to the missing attribute; S73: Input the context information of the node corresponding to the missing attribute into the trained second LLM model for judgment to obtain a judgment result of the missing attribute; the judgment result of the missing attribute includes: generated attribute value, second confidence score and technical rationality explanation; S74: Set a second confidence threshold, and supplement the missing attribute whose second confidence score is not less than the second confidence threshold into the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0050] Specifically, the attribute completion process is committed to intelligently filling in missing professional attributes for various nodes in the knowledge graph. By comprehensively scanning all nodes in the graph, it accurately identifies defect, cause or solution nodes with empty key attribute fields. For each node with a missing attribute, the system extracts the node's label type, existing attribute set and all associated node information to build a complete context. These structured data are sent to a LLM model (trained second LLM model) optimized by professional domain knowledge for in-depth analysis. The model will intelligently generate attribute values that meet industry standards based on node type, such as generating severity scores of 1-5 for defect nodes, determining the occurrence frequency of "high / medium / low" for cause nodes, or evaluating the implementation cost level for solutions. Each attribute completion is accompanied by a confidence score and detailed technical basis explanation. Finally, a second confidence threshold of 0.7 is set for automatic filtering, and only attribute values that meet the standard are officially written into the graph, while the reasoning process and judgment basis are retained as metadata.

[0051] In summary, the core design features of the manufacturing process disturbance knowledge graph generation method proposed in this embodiment are reflected in the following four key dimensions: 1) A hybrid expert system design is adopted at the architecture level, organically integrating the structured knowledge representation of Neo4j and the semantic reasoning capabilities of large language models, forming a complementary automatic completion mechanism; 2) By introducing a strict confidence threshold (≥ 0.7) filtering mechanism, the system output meets the industrial quality requirements, effectively balancing completion efficiency and accuracy; 3) Emphasis is placed on the explainability of the decision-making process, requiring LLM to provide detailed technical basis (reason field) for human review to provide transparent judgment basis; In terms of scalability, a modular design paradigm is adopted to support seamless access to new relationship types and attribute fields, ensuring that the system can continuously adapt to the evolving needs of domain knowledge.

[0052] Embodiment 2

[0053] Based on the manufacturing process disturbance knowledge graph generation method provided in Embodiment 1, this embodiment further illustrates the present application in combination with specific embodiments.

[0054] I. Verify the overall performance of the entity recognition model on all entity categories: 1) Data set explanation: The SMT production line has accumulated a large number of records and reports about process defect causes during long-term operation. However, there is no related question data set in this field. Therefore, this embodiment attempts to construct a data set suitable for SMT process defect cause analysis.

[0055] The main data corpus used in this embodiment refers to the long-term accumulated process defect records, repair reports, quality inspection analysis documents and other typical technical materials of a company's SMT production line, as well as the SMT process related standards published by standard organizations. These materials cover the professional knowledge of the SMT production line, including the occurrence of defects, cause classification, detection methods and treatment methods. At the same time, this paper also analyzes the content of the related industry forum and technical support platform, and finds that the expression structure of SMT process defect related problems has certain universality, which can be summarized into several high-coverage templates. Therefore, this paper uses the template construction method to build the data set, and obtains an original text data set of 2998 sentences, totaling 31487 characters. The number of different entity categories is shown in Table 4.

[0056] Table 4 Number of different entity categories

[0057] 2) Experimental environment This embodiment is based on the PyTorch open source machine learning library to build the RoBERTa-WWM-large-BiLSTM-CRF model. The experimental environment is: Windows 11 operating system, AMD Ryzen 9 7945HX with Radeon Graphics, NVIDIA GeForce RTX 4060 Laptop GPU graphics card, 16GB memory. This embodiment experiment is based on Python3.8, deep learning framework Pytorch2.2.2 and transformers4.46.3 to build the model network.

[0058] In this experiment, we set the following key parameters to ensure the effectiveness and accuracy of the experiment, as shown in Table 5.

[0059] Table 5 Parameter setting table

[0060] 3) Experimental results This embodiment uses precision P (Precision), recall R (Recall) and F1 value (F1-score) to evaluate the prediction results of the four entities, which are "Engineer", "Error", "Problem" and "Solution", as shown in Table 6. Figure 5 From Figure 5 ​It can be seen that the entity recognition model has excellent entity extraction capability, especially in the Engineer category, which achieves a near-full score (F1 value of 0.97). For categories such as Error, Problem, and Solution, which have similar semantics, the model still maintains high recall and F1 values, demonstrating strong generalization ability and stability.

[0061] II. Verify the performance of the entity recognition model compared with existing models To verify the performance advantage of the entity recognition model in the SMT process disturbance entity recognition task, three typical models, BERT-BiLSTM-CR, BiLSTM-CRF, and LSTM-CRF, are selected for comparison, and are trained on the same self-made SMT process defect reason dataset.

[0062] The loss value change curve in the training process of each model is shown in Figure 6 By comparison, the superior performance of the entity recognition model of the present embodiment in terms of training convergence speed and recognition accuracy can be intuitively reflected.

[0063] Among the four models, the entity recognition model in the present embodiment shows the best loss reduction trend and final performance, and its advantage mainly lies in the dual improvement of semantic modeling and sequence modeling ability. The model uses the powerful context semantic representation ability of RoBERTa-WWM-large, which is more accurate and detailed in semantic understanding than traditional word vectors or BERT. At the same time, it combines the modeling ability of BiLSTM for bidirectional information and the modeling of the dependency relationship between labels by CRF, so that the entire structure shows strong learning ability and generalization ability in the sequence labeling task. From the loss value change, RoBERTa-WWM-large-BiLSTM-CRF not only has the lowest loss at the beginning of training, but also has the fastest convergence speed and the lowest final loss value (0.31), which is much better than other models, reflecting its significant advantage in feature expression and sequence prediction.

[0064] The experimental results are shown in Table 6, Figure 7 Table 6 Evaluation Results Table 6 Evaluation Results

[0065] In the entity recognition task, the entity recognition model shows significant advantages compared with other models, with Macro F1 reaching 0.83, surpassing traditional LSTM-CRF (0.41), BiLSTM-CRF (0.53) and BERT-BiLSTM-CRF (0.81) models. The model is more excellent in capturing context information and semantic understanding, especially in the recall rate (reaching 0.87), effectively reducing the missing label phenomenon, reflecting the strong pre-training characteristics and better language expression ability of RoBERTa-WWM-large. Overall, RoBERTa-WWM-large performs best in this task, significantly improving the entity recognition effect.

[0066] Embodiment 3

[0067] The embodiment of the application introduces a manufacturing process disturbance knowledge graph generation system, comprising: A data acquisition module is configured to acquire an SMT manufacturing process disturbance dataset. A schema definition module is configured to construct a schema layer of the knowledge graph, wherein the schema layer defines entity types, attribute types and relationship types. An entity recognition module is configured to input the SMT manufacturing process disturbance dataset into a trained entity recognition model for processing to obtain entities in the SMT manufacturing process disturbance dataset. A relationship attribute extraction module is configured to extract the context of the entities to obtain corresponding attributes of the entities and relationships between the entities. A preliminary construction module is configured to fuse the entities, corresponding attributes of the entities and relationships between the entities in the SMT manufacturing process disturbance dataset based on the schema layer, and construct a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database. A first completion module is configured to complete relationships of the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph. A second completion module is configured to complete attributes of the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

[0068] Embodiment 4

[0069] The embodiment provides a computer readable storage medium storing a computer program, wherein the computer program is executed to realize the manufacturing process disturbance knowledge graph generation method of the embodiment 1.

[0070] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code.

[0071] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart and / or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable Figure 1 The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart and / or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable

[0072] The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. Figure 1 The flowchart and / or block diagram in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present application. In this regard, each block in the flowchart and / or block diagrams can represent a module, segment, or portion of code, which comprises one or more executable Figure 1 The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks. The computer program instructions can also be loaded onto a computer, other programmable data processing apparatus, or other processing device to cause a series of operational steps to be performed on the computer, other programmable apparatus or other processing device to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions specified in the flowchart block or blocks.

[0073] The embodiments of the application in the foregoing description are only illustrative of the present application. Without departing from the principles and scope of the present application, those skilled in the art can make many forms, which should be covered by the protection of the present application.

Claims

1. A method for generating a knowledge graph of manufacturing process perturbations, characterized in that, The application relates to a method for constructing an SMT manufacturing process disturbance knowledge graph. The method comprises the following steps: acquiring an SMT manufacturing process disturbance data set; constructing a schema layer of a knowledge graph; the schema layer defines entity types, attribute types and relationship types; inputting the SMT manufacturing process disturbance data set into a trained entity recognition model for processing to obtain entities in the SMT manufacturing process disturbance data set; extracting the context of the entities to obtain attributes corresponding to the entities and relationships between the entities; fusing the entities, the attributes corresponding to the entities and the relationships between the entities in the SMT manufacturing process disturbance data set based on the schema layer, and constructing a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; completing relationships of the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph; 2. The method of claim 1, wherein, completing attributes of the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

3. The method of claim 2, wherein, The entity types comprise engineers, errors, problems and solutions; the relationship types comprise errors found by engineers, problems directly caused by errors, chain reactions between problems, problems or errors solved by solutions, proposers of solutions and detectors of quality problems; and the attribute types comprise positions and codes of engineers, machines and employees related to problems, machines, employees and supply chains related to solutions. The SMT manufacturing process disturbance data set needs to be preprocessed before being inputted into the trained entity recognition model for processing, and the preprocessing comprises the following steps: performing word segmentation and special character removal on sentences in the SMT manufacturing process disturbance data set in sequence to obtain a plurality of word units; 4. The method of claim 3, wherein, adding special marks at the beginning and end of the sentences to identify the beginning and end of the sentence sequence.

5. The method of claim 4, wherein, The entity recognition model comprises sequentially connected RoBERTa-WWM-large layers, BiLSTM layers, linear layers and CRF layers. The entity recognition model comprises sequentially connected RoBERTa-WWM-large layers, BiLSTM layers, linear layers and CRF layers. The method comprises the following steps: inputting the sentences in the SMT manufacturing process disturbance data set into the RoBERTa-WWM-large layers for language modeling to obtain high-dimensional vector representations of all word units in the sentences; inputting the high-dimensional vector representations of all word units in the sentences into the BiLSTM layers for processing to capture the dependency relationship between each word unit and its context, and generate global semantic information vector representations of all word units; inputting the global semantic information vector representations of all word units into the linear layers for linear processing to calculate scores of global semantic information vector representations of each word unit with respect to entity labels, and obtain Emission Score vectors of all word units; the entity labels are set based on the entity types defined by the schema layer; inputting the Emission Score vectors of all word units into the CRF layers to predict optimal entity labels of all word units, and obtain a global optimal label sequence; obtaining the entities in the SMT manufacturing process disturbance data set according to the global optimal label sequence.

6. The method of claim 5, wherein, The relationship completion is performed on the SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph, including: Finding a missing relationship based on a Cypher query language of a Neo4j database; Based on the missing relationship, context information of a node corresponding to the missing relationship is extracted, the node being an entity, and the context information of the node corresponding to the missing relationship including a node attribute set, a node existing relationship type set, and a relationship and attribute set of associated nodes; The context information of the node corresponding to the missing relationship is input into a trained first LLM model to obtain a judgment result of the missing relationship, the judgment result of the missing relationship including a relationship existence judgment result, a first confidence score, and a technical basis description; A first confidence threshold is set, and a missing relationship with a first confidence score not less than the first confidence threshold is supplemented into the SMT manufacturing process disturbance basic knowledge graph to obtain the second SMT manufacturing process disturbance knowledge graph.

7. The method of claim 6, wherein, The attribute completion is performed on the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph, including: Based on attribute existence detection queries of a Neo4j database, a missing attribute is found, and a queue for completing is generated; For each node corresponding to a missing attribute in the queue for completing, multi-dimensional context extraction is performed to obtain context information of the node corresponding to the missing attribute; The context information of the node corresponding to the missing attribute is input into a trained second LLM model to obtain a judgment result of the missing attribute, the judgment result of the missing attribute including a generated attribute value, a second confidence score, and a technical rationality description; A second confidence threshold is set, and a missing attribute with a second confidence score not less than the second confidence threshold is supplemented into the second SMT manufacturing process disturbance knowledge graph to obtain the final SMT manufacturing process disturbance knowledge graph.

8. A manufacturing process disturbance knowledge graph generation system, comprising: It includes: A data acquisition module is configured to acquire an SMT manufacturing process disturbance dataset; A schema definition module is configured to construct a schema layer of a knowledge graph; The schema layer defines an entity type, an attribute type, and a relationship type; An entity recognition module is configured to input the SMT manufacturing process disturbance dataset into a trained entity recognition model to obtain entities in the SMT manufacturing process disturbance dataset; A relationship attribute extraction module is configured to extract a context of the entities to obtain attributes corresponding to the entities and relationships between the entities; A preliminary construction module is configured to fuse the entities, the attributes corresponding to the entities, and the relationships between the entities in the SMT manufacturing process disturbance dataset based on the schema layer, and construct a first SMT manufacturing process disturbance knowledge graph based on a Neo4j database; A first completion module is configured to perform relationship completion on an SMT manufacturing process disturbance basic knowledge graph to obtain a second SMT manufacturing process disturbance knowledge graph; A second completion module is configured to perform attribute completion on the second SMT manufacturing process disturbance knowledge graph to obtain a final SMT manufacturing process disturbance knowledge graph.

9. A computer-readable storage medium, characterized in that, The computer program stored in the computer readable storage medium, when executed, implements the method for generating the manufacturing process disturbance knowledge graph according to any one of claims 1-7.

Images