A diesel engine manufacturing knowledge graph dynamic updating method based on deep learning

Abstract

The application provides a diesel engine manufacturing knowledge graph dynamic updating method based on deep learning, comprising: converting fragmented manufacturing data generated in the manufacturing process into high-quality input data after preprocessing; obtaining data characteristics and correlation data from the high-quality input data; fusing the high-quality data corresponding to the manufacturing process of the diesel engine and converting it into new manufacturing knowledge fusion data based on the structure of the manufacturing knowledge graph; and obtaining new manufacturing knowledge data to be updated on the basis of the existing manufacturing knowledge graph and updating or expanding the existing manufacturing knowledge graph. The application obtains new manufacturing knowledge graph data through a bidirectional encoder model and a semantic alignment model, can effectively capture different manufacturing process data and their correlation in the semantic space, so as to effectively construct new manufacturing knowledge graph data and expand and update the existing manufacturing knowledge graph.

Description

Technical Field

[0001] This invention belongs to the field of mechanical manufacturing technology, specifically relating to a method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning. Background Technology

[0002] In the field of modern mechanical manufacturing, as the research and development process continues, large-scale manufacturing data is constantly accumulated, and new manufacturing process data is constantly generated, including design change records, analysis reports, test verification results, and problem solutions. Complex product development faces the following challenges:

[0003] 1. The ineffective utilization of manufacturing process data leads to incomplete coverage of manufacturing information in the original manufacturing knowledge graph. In severe cases, this affects the evolution quality and application effectiveness of the manufacturing knowledge graph. This manifests as follows: when the original manufacturing knowledge graph attempts to incorporate new manufacturing knowledge that exceeds its original scope, the lack of corresponding concepts, entities, or relational patterns as anchors prevents accurate semantic alignment of the manufacturing knowledge data; if logical contradictions arise between the old and new manufacturing knowledge, or if there is redundant substructure data that cannot be connected or verified, the management and maintenance costs of the manufacturing knowledge graph will increase significantly.

[0004] 2. The relationships between manufacturing quality information are often implicit in fragmented quality data. Their semantics have significant contextual dependencies and ambiguity, making it difficult for rule-based extraction methods to effectively capture and express these contextual and implicit complex relationships. This may disrupt the stability of the updating and evolution of the manufacturing knowledge graph. Therefore, it is extremely important to effectively establish a robust knowledge fusion and conflict resolution mechanism that is oriented towards dynamic evolution.

[0005] Therefore, traditional manufacturing knowledge data update methods lack the ability to integrate scattered process information and semantic information such as quality characteristics, and a method for dynamically updating manufacturing knowledge graphs is needed. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of traditional manufacturing knowledge data updates, such as the lack of deep semantic understanding and integration. Instead, it provides a deep learning-based method for dynamically updating diesel engine manufacturing knowledge graphs. By utilizing a bidirectional encoder model and a semantic alignment model to obtain manufacturing knowledge graph data, it effectively captures data from different manufacturing processes and their relationships in the semantic space, thereby enabling the effective construction of new manufacturing knowledge graph data.

[0007] To achieve the above objectives, the technical solution provided by this invention is:

[0008] A deep learning-based method for dynamically updating a knowledge graph of diesel engine manufacturing includes:

[0009] Step 1: Preprocess the fragmented manufacturing data generated during the manufacturing process, and transform the preprocessed fragmented manufacturing data into high-quality input data;

[0010] Step 2: Obtain data characteristics and correlation data from the high-quality input data to characterize the manufacturing information characteristics and the correlation between manufacturing information;

[0011] Step 3: Align the data characteristics and related relationship data one by one in the semantic space to obtain structured manufacturing knowledge graph data;

[0012] Step 4: Update or expand the existing manufacturing knowledge graph based on the manufacturing knowledge graph data obtained in Step 3.

[0013] As a further limitation of the present invention, step one includes:

[0014] Fragmented manufacturing data generated during the cross-stage and cross-system manufacturing process of diesel engines is preprocessed by sequentially performing data cleaning, noise filtering, time-series alignment, and multimodal unified representation. This process extracts structured data from equipment sensors, process logs, quality inspection reports, and other sources in the diesel engine manufacturing process. The structured data is then transformed into high-quality input data through pre-constructed entity data standardization annotation. Specifically, the pre-constructed entity data standardization annotation involves standardizing the entity data of process parameters, equipment status, and material batches.

[0015] As a further limitation of the present invention, step two includes:

[0016] A bidirectional encoder model pre-trained based on the Transformer architecture is used to identify manufacturing characteristics from the high-quality input data. Explicit correlations of data characteristics are extracted based on these manufacturing characteristics, and overall consistency optimization is performed to obtain manufacturing characteristic correlation triples corresponding to existing manufacturing knowledge graphs. During the pre-training phase of the bidirectional encoder model, a self-supervised task for logical ordering prediction and a self-supervised task for quality causal inference, oriented towards domain manufacturing processes, are added to inject causal chain knowledge related to manufacturing quality into the bidirectional encoder model.

[0017] A finely tuned bidirectional encoder model is jointly trained using data sequence standardization labeling and association classification. Based on the bidirectional encoder model, process entities in the diesel engine manufacturing process are identified from the manufacturing characteristic association triplet. At the same time, an attention mechanism guides the bidirectional encoder model to extract implicit associations related to process entities from the high-quality input data to initially obtain the manufacturing characteristic association triplet in the manufacturing entity.

[0018] As a further limitation of the present invention, step three includes:

[0019] The manufacturing characteristic data output by the bidirectional encoder model and the vector data in the association data are accurately extracted from the manufacturing characteristic association triplet. Based on the semantic similarity calculation of the Jaccard coefficient, the similar nodes and associations of manufacturing characteristics are merged after semantic similarity calculation of the existing manufacturing knowledge graph and the newly generated manufacturing knowledge graph in the diesel engine manufacturing process, thereby realizing the alignment of related concepts of manufacturing characteristics and associations. Specifically, the semantic alignment method corresponding to the semantic alignment framework based on the Siamese-BERT network model is used to align the manufacturing characteristics and their associations in the semantic space one by one to generate the new manufacturing knowledge fusion data. According to the set similarity threshold, it is determined whether the manufacturing knowledge fusion data is the same manufacturing characteristic or the same association in the semantic space in the existing manufacturing knowledge graph. The judgment result guides the merging of manufacturing characteristics or associations and provides reliable new manufacturing knowledge fusion data for feature node similarity calculation in the subsequent dynamic update of the three-dimensional manufacturing knowledge graph, thereby obtaining the manufacturing knowledge graph data to be updated or expanded in the manufacturing process data.

[0020] As a further limitation of the present invention, step four includes:

[0021] Based on the calculation results of semantic similarity or a combination of semantic similarity and attribute similarity, the existing manufacturing knowledge graph and the newly generated manufacturing knowledge graph data in step three are used to merge similar nodes and similar associations, thereby forming a complete manufacturing knowledge graph data to be updated or expanded. Then, the existing manufacturing knowledge graph is updated or expanded based on its basic structure and its extended structure.

[0022] Through steps one through four, the BERT model and semantic similarity calculation method are used to repeat the above steps for updating or expanding the knowledge graph, thereby forming an updated and expanded knowledge graph representing the manufacturing process of complex product development. The specific expression for the similarity calculation is as follows:

[0023]

[0024] In the formula, This represents two characteristic concepts in the creation of a knowledge graph. This indicates that two characteristic concepts in a knowledge graph share the same character. This represents all the characters contained in two characteristic concepts within a knowledge graph.

[0025]

[0026] In the formula, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text, This indicates that two attribute texts in the knowledge graph contain the same common character. This represents all the characters contained in the two attribute texts in the knowledge graph.

[0027]

[0028] In the formula, This represents a given specific weight.

[0029] The advantages of this invention are:

[0030] This invention obtains new manufacturing knowledge graph data through a bidirectional encoder model and a semantic alignment model, which can effectively capture data on different manufacturing processes and their relationships in the semantic space, so as to effectively construct new manufacturing knowledge graph data and expand and update the existing manufacturing knowledge graph.

[0031] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0032] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which:

[0033] Figure 1 This invention provides a flowchart of a method for dynamically updating a knowledge graph for diesel engine manufacturing based on deep learning;

[0034] Figure 2 This invention provides a schematic diagram of the diesel engine manufacturing knowledge graph expansion and update process.

[0035] Figure 3: A schematic diagram of the three-layer network structure for accurate extraction of associations in the BERT model provided by this invention;

[0036] Figure 4 : A schematic diagram of the model framework of the semantic alignment module based on the Siamese-BERT network provided by this invention;

[0037] Figure 5 This illustration shows the output results of data processing in the knowledge graph update case of diesel engine connecting rod manufacturing based on quality feature data provided by the present invention.

[0038] Figure 6 The illustration shows the result of obtaining the knowledge graph of manufacturing process quality characteristics in the case study of updating the knowledge graph of diesel engine connecting rod manufacturing based on quality feature data provided by the present invention.

[0039] Figure 7 This illustration shows how the manufacturing knowledge graph of a manufacturing entity is updated based on data in the diesel engine connecting rod manufacturing knowledge graph update case based on quality feature data provided by the present invention.

[0040] Figure 8 This illustration shows the updated manufacturing knowledge graph after fusion in the diesel engine connecting rod manufacturing knowledge graph update case based on quality feature data provided by the present invention. Detailed Implementation

[0041] The embodiments of the present invention are described in detail below. These embodiments are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0042] See Figure 1 and Figure 2 This invention provides a method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning, comprising the following steps:

[0043] Step 1: Preprocess the fragmented manufacturing data generated during the manufacturing process, and transform the preprocessed manufacturing data into high-quality input data.

[0044] In practical applications, the first step is to prepare data on the diesel engine manufacturing process, including existing manufacturing knowledge graph data, industry standard documents, process data, quality data, and analysis reports.

[0045] The preferred approach involves data cleaning, noise filtering, temporal alignment, and multimodal data unification. Structured or unstructured data from different sources, such as equipment sensors, process logs, and quality inspection reports, are processed using the mainstream Jieba algorithm. This process sequentially extracts, standardizes, cleans, and semantically annotates fragmented manufacturing process data, including cross-domain and cross-system discrete and fragmented diesel engine manufacturing process data. This transforms the data into manufacturing knowledge graph data that meets the needs of updating or expanding existing manufacturing knowledge graphs. The result is high-quality input data that eliminates the impact of data noise, missing data, and semantic inconsistencies. This high-quality input data is then transformed into input data to be fed into the BERT model.

[0046] To address the unique characteristics of the diesel engine manufacturing field, this invention employs a pre-trained BERT model for automatic identification and extraction of manufacturing knowledge during the dynamic updating of the diesel engine manufacturing knowledge graph. Furthermore, it constructs manufacturing knowledge based on a holistic optimization of manufacturing process data and related relationships within the BERT model's triplet. Specifically, this includes:

[0047] (1) Preprocessing fragmented manufacturing data to obtain high-quality input data; specifically, preprocessing fragmented manufacturing data by data cleaning, noise filtering, time-series alignment, and unified multimodal data representation to transform structured and unstructured data from different sources such as equipment sensors, process logs, and quality inspection reports into high-quality input data. Considering the specific characteristics of the manufacturing field, this embodiment of the invention also standardizes and labels entities such as process parameters, equipment status, and material batches.

[0048] (2) The BERT pre-trained model is used to identify and extract the characteristics and relationships in the manufacturing data (a bidirectional encoder model based on the Transformer architecture is used to perform hierarchical semantic capture of explicit characteristics and implicit relationships in the manufacturing data), so as to realize the automatic extraction of data characteristics and potential relationships. Specifically, the pre-processed multi-source data is constructed into a unified token sequence: for unstructured data such as process specification text and equipment log description, it is directly converted into tokens through WordPiece word segmentation; for structured data such as sensor time series data and quality inspection parameter tables, it is first mapped into discrete semantic labels or encoding vectors (such as "spindle speed - 3500 rpm", "roughness - Ra0.8"), and then embedded into the same semantic space. Through the self-attention mechanism of the bidirectional encoder, the BERT model can simultaneously capture the forward and backward contextual dependencies of the manufacturing data and identify long-distance cross-text process relationship patterns such as "cutting speed - tool wear - surface quality".

[0049] (3) By using semantic cosine similarity calculation and semantic space alignment mechanism, the integration of manufacturing information characteristics and relationships across stages and systems is realized, and the transformation of fragmented manufacturing data into manufacturing knowledge is achieved.

[0050] Step 2: Based on the bidirectional encoder model, capture data characteristics representing manufacturing information and the correlations between manufacturing information from high-quality input data. Specifically:

[0051] (1) Before capturing high-quality input data, this embodiment of the invention first uses existing manufacturing knowledge graph data to fine-tune the BERT model and its extended models. Specifically, when identifying manufacturing characteristics, the BERT model uses the BERT-BiLSTM-CRF algorithm as the core algorithm, and when extracting the association relationship based on the manufacturing characteristics, the BERT model uses the SpanBERT algorithm to extract the manufacturing characteristic association relationship. This embodiment of the invention further trains the basic model by using a small amount of incremental data from specific use cases to change the parameter weights of the neural network in the BERT model.

[0052] (2) Based on the pre-trained BERT model, this embodiment of the invention first uses the BERT-BiLSTM-CRF algorithm, which has context awareness and entity boundary discrimination capabilities, as the core algorithm from the high-quality input data obtained in step one to automatically identify manufacturing characteristic-related semantics, and then extracts the correlation relationships based on the identified manufacturing characteristics. Specifically:

[0053] (21) This embodiment of the invention employs a pre-trained BERT model with context awareness and entity boundary discrimination capabilities to automatically identify manufacturing characteristics from high-quality input data. The pre-trained BERT model uses a three-level progressive architecture to abstract and refine the sequence labeling of the pre-processed fragmented manufacturing data layer by layer, thereby completely identifying the manufacturing characteristics of entities in the manufacturing process. In the pre-training stage, in addition to the conventional Masked Language Model (MLM) task, two self-supervised tasks for the manufacturing domain are added: process logic sequencing prediction (such as reordering discrete operation steps to learn the process sequence logic) and quality causal inference (predicting the impact of key process parameters on downstream quality inspection indicators by masking key process parameters), forcing the model to inject causal chain knowledge of the manufacturing process into the bidirectional encoding. The specific process of the BERT model to identify manufacturing characteristics in the diesel engine manufacturing process is as follows: (211) In this embodiment of the invention, the manufacturing characteristics, manufacturing processes, process parameters and their attributes in the preprocessed manufacturing process data are analyzed and context-aware encoded by the BERT model. The concepts of manufacturing characteristics, processes, process parameters and their attributes in the data are encoded by the BERT model to achieve the understanding and preliminary identification of the global semantic relationship between manufacturing knowledge concepts; (212) The context vector sequence output by the BERT model in this embodiment of the invention is fed into the BiLSTM network. Combined with the forward and backward information propagation of the BiLSTM network, the logical order and correlation coupling of manufacturing characteristics in the development process are captured, and the model's recognition accuracy for manufacturing-related concepts with strong semantic boundaries and context dependence is improved, thereby ensuring that manufacturing characteristics and their related attributes can be effectively captured by the BERT model; (213) In this embodiment of the invention, the Conditional Random Fields (CRF) layer is introduced into the output layer of the BERT model to optimize the overall consistency of the identified characteristics and correlation relationships, thereby realizing the identification and extraction of manufacturing characteristics and attributes from process data.

[0054] (22) This embodiment of the invention is based on a pre-trained BERT model and uses the SpanBERT algorithm to extract the manufacturing characteristic association relationship in the diesel engine manufacturing process data. The overall implementation process includes: (221) Based on the manufacturing characteristic identification results, a candidate characteristic span set is constructed, and each entity span is input into the SpanBERT model. Through random masking and reconstruction training, a high-dimensional semantic representation is generated for each characteristic span; (222) For each candidate characteristic pair, their respective semantic vectors are concatenated and input into the classifier to calculate the probability distribution of the relationship category. The most likely relationship category between the characteristic pairs is output, such as the relationship of causing or influencing; (223) The decision network in the semantic alignment model outputs the decision result in the form of manufacturing characteristic-association relationship-manufacturing characteristic triples. These triples together constitute the semantic association network in the manufacturing knowledge graph.

[0055] (23) This embodiment of the invention optimizes the overall consistency of the manufacturing characteristics identified by the BERT model and their associations in the semantic space, realizing the identification and extraction of manufacturing characteristics and their attributes from manufacturing process data. This embodiment preferably uses the SpanBERT algorithm to construct an end-to-end association extraction framework, accurately extracting associations related to manufacturing characteristics and accurately calculating the accuracy of node similarity calculations in the subsequent manufacturing atlas. Specifically, see [link to relevant documentation]. Figure 3 This invention presents a three-layer network structure for accurate extraction of association relationships using the BERT model, which integrates the SpanBERT model, an association classifier, and a decision-making module. Compared with traditional word vector models, this invention utilizes the capabilities of the BERT model in semantic extraction, context dependency modeling, and cross-domain semantic transfer to capture data on different manufacturing processes and their association relationships in the semantic space, effectively constructing new manufacturing knowledge graph data. More specifically, the three-layer network structure for accurate extraction of association relationships in the BERT model designed in this invention includes: (231) a SpanBERT encoding layer, (232) a candidate span generation layer, (233) a relationship classifier layer, and (234) a decision output layer, wherein: (231) the SpanBERT encoding layer first receives the manufacturing characteristics and attribute data extracted by the BERT model and formats them as [token sequence, [Entity boundary label], and then generate context-aware semantic vectors through a deep bidirectional Transformer encoder; (232) The candidate span generation layer constructs a candidate span set based on the characteristic recognition results of the corresponding context-aware semantic vectors, and uses a span filtering mechanism in the lightweight candidate span set to remove invalid combinations and retain candidate manufacturing characteristic pairs with high confidence in the candidate span set; (233) The relation classifier layer concatenates the manufacturing characteristic pairs of the candidate spans on the semantic vectors and inputs them into a fully connected network, and calculates the probability distribution of the association relationship category through the relation classifier layer; (234) After obtaining the probability distribution of the association relationship category, the judgment output layer outputs the entity characteristic association relationship triplet that conforms to the manufacturing domain constraints through threshold judgment and logical verification.

[0056] To address the limitations of the general BERT model in semantic understanding within specific contexts, this invention employs a pre-trained BERT model and its extended versions to fine-tune an existing manufacturing knowledge graph within the manufacturing domain. This allows the BERT model to maximize its understanding of manufacturing characteristics and their explicit and implicit relationships in the semantic space, achieving domain-specificity and adaptability of the manufacturing knowledge graph during continuous updates. In the fine-tuning stage of the BERT model, this invention uses joint training of sequence labeling and relation classification: a CRF layer is used to identify process entities (such as equipment, parameters, and defect types), while an attention-guided relation extraction head performs semantic relation classification on entity pairs, explicitly extracting relations such as "influence" and "belong to," thereby constructing preliminary relation triples. Next:

[0057] In the second step of this invention, the pre-trained BERT model, after fine-tuning, extracts information such as high-quality input data from fragmented process document data during the manufacturing process in a timely manner, and can effectively capture the explicit and implicit relationships of different quality characteristics and their correlations in the semantic space.

[0058] Step 3: Align the data characteristics and relationships across stages and systems in the semantic space to obtain structured manufacturing knowledge graph data. Specifically, this embodiment of the invention uses a semantic alignment module based on the Siamese-BERT algorithm to merge similar nodes and relationships of manufacturing characteristics between existing manufacturing knowledge graphs and newly generated manufacturing knowledge graphs during diesel engine manufacturing, achieving alignment of related concepts of manufacturing characteristics and relationships; after performing semantic similarity calculations, it merges similar nodes and relationships of manufacturing characteristics between existing manufacturing knowledge graphs and newly generated manufacturing knowledge graphs during diesel engine manufacturing, obtaining manufacturing knowledge graph data to be updated or expanded in the manufacturing process data. Specifically:

[0059] This invention employs a semantic alignment method based on the Siamese-BERT algorithm to align manufacturing characteristics and their relationships in the semantic space one by one. More specifically: This invention's semantic alignment framework based on the Siamese-BERT network model aligns the relationships in the manufacturing knowledge graph data to be updated one by one, see [link to relevant documentation]. Figure 4The semantic alignment model framework in this embodiment includes a shared encoder, a pooling and projection layer, a similarity calculation layer, and a contrastive loss optimization layer. In the shared encoder, the encoders of two identical BERT models share weights, ensuring that manufacturing characteristics and relationships from different inputs are mapped to a unified semantic space. The pooling and projection layer projects the output manufacturing knowledge graph data to be updated into a 128-dimensional semantic space after mean pooling. The similarity calculation layer calculates the cosine similarity between manufacturing characteristics and their relationships in the semantic space and introduces learnable similarity calculation parameters and manufacturing knowledge graph update expansion parameters to scale the manufacturing knowledge graph in the semantic space. The contrastive loss optimization module uses Triplet Loss or Contrastive Loss for end-to-end training to optimize the updated and expanded data of the manufacturing knowledge graph. Specifically, it includes:

[0060] (31) The structured triples composed of features and relationships extracted from different manufacturing data are vectorized and input into the Siamese-BERT model. Specifically, this includes: (311) Reading the relationships extracted by the SpanBERT encoding layer and generating standard JSON format triples; (312) Comparing the intersection of the manufacturing features of newly extracted entities in the BERT model with the set of node IDs in the existing manufacturing knowledge graph; for example: if the new entity ID exists in the existing graph (such as "bolt"), directly extract all the relationship triples of the node in the graph (such as "bolt-connection-link") as channel A input, and the new entity relationship as channel B input to form a positive sample pair; if the entity ID does not exist, the new node relationship is used as channel B input, and the heterogeneous node relationship is randomly selected in the current graph according to the entity type as channel A input to form a negative sample pair; (313) Serializing the text pairs and converting the serialized text pairs into the standard BERT output format.

[0061] (32) Semantic consistency is measured by calculating the cosine similarity of different manufacturing characteristics or relationships (since the range of cosine similarity is [-1, 1], the closer the cosine similarity is to 1, the stronger the relationship between different concepts). In this embodiment of the invention, a hierarchical refined calculation method is used to calculate the cosine similarity of different manufacturing characteristics or relationships to measure semantic consistency; the specific calculation process is as follows: (321) Vectorization of relationships in semantic space: Specifically, the manufacturing characteristics or relationships extracted from different sources (such as MES system, quality inspection report, equipment log) are vectorized. For characteristic node text Input the Siamese-BERT shared encoder to extract the text of the feature nodes. Corresponding hidden state ,definition The implicit relation vector in the semantic space is obtained by dimensionality reduction through the projection layer. , , implicit relations in semantic space Perform normalization. Specifically, concatenate the relationship description text and the context of the beginning and end entities in the structured triples into the format "[entity1]relation[entity2]" (e.g., "spindle speed causes vibration defects"), and generate a relation semantic vector. (322) Cosine similarity calculation: For any two manufacturing characteristic or correlation vectors and The cosine similarity is calculated and expressed as: To enhance differentiation, a temperature scaling parameter is introduced. Sharpening is represented as: In actual judgment, the original cosine value is used, and scaling is only used for gradient optimization during training. (323) Hierarchical threshold judgment and semantic consistency evaluation: Dynamic judgment thresholds are set according to the manufacturing ontology level, specifically including: core concept layer (such as "spindle speed", "cutting force"), threshold High-precision matching is required; auxiliary attribute layer (such as "operator"), threshold Allow for moderate fuzzy matching; for cross-domain relationship layers (such as "process parameters" and "quality indicators"), a threshold is required. To ensure the semantic accuracy of the relations. (324) Similarity matrix construction and optimal matching solution: For the node sets of two knowledge graphs. and , build similarity matrix ,in The Hungarian algorithm or greedy matching is used to find the optimal node correspondence, avoiding many-to-one erroneous mappings. Mutual exclusion constraints are introduced during the matching process: if nodes in the knowledge graph are generated... Nodes in creating knowledge graphs If a defined relation boundary already exists, then its matching object... At the node The relationships between the graphs must satisfy a similar relational structure (verified through subgraph similarity).

[0062] (33) Based on the set similarity threshold, determine whether the new manufacturing data is the same manufacturing characteristic or the relationship in the semantic space, guide the merging of manufacturing characteristics or relationships, and obtain the manufacturing knowledge data (such as JSON files) that need to be updated or expanded in the manufacturing process data.

[0063] Step 4: Update or expand the existing manufacturing knowledge graph based on the manufacturing knowledge graph data obtained in Step 3.

[0064] Specifically, in order to merge similar nodes and relationships among manufacturing characteristics in the manufacturing knowledge graph after semantic similarity calculation, this embodiment of the invention, based on the semantic similarity calculation results of the Jaccard coefficient, merges the existing manufacturing knowledge graph and the newly generated manufacturing knowledge graph data using manufacturing process data into similar nodes and similar relationships after performing semantic similarity calculation. The specific calculation formula is as follows:

[0065]

[0066] In the formula, and This represents two characteristic concepts in the creation of a knowledge graph. This indicates that both contain the same characters. It represents all the characters in both.

[0067] This invention utilizes semantic similarity calculation and semantic space alignment mechanisms to achieve the fusion of manufacturing characteristics and their relationships in cross-stage and cross-system manufacturing information. Specifically, it fuses manufacturing characteristics and their relationships in semantic space through semantic cosine similarity calculation and semantic space alignment mechanisms. More specifically, this invention maps the entity manufacturing features and relationships output by the bidirectional encoder model to a unified low-dimensional semantic space (typically 128-256 dimensions), and performs semantic fusion using domain-adapted cosine similarity metrics. Preferably, this invention employs hierarchical cosine distance calculation based on manufacturing ontology, applying adjustable weights to vectors at different ontology levels of manufacturing knowledge in the semantic space (e.g., "equipment level - process level - quality level"), making cross-level similarity calculations more aligned with manufacturing logic. To achieve alignment of manufacturing characteristics and relationships in the semantic space, this invention also introduces adversarial domain adaptation training, adding a domain discriminator to classify encoded vectors from different systems (e.g., ERP, MES, SCADA). The bidirectional encoder learns domain-independent general representations through a gradient inversion layer (GRL), ensuring high semantic consistency between concepts such as "tool life" (which are domain-independent) and the semantic vectors in the tool management system and quality traceability system. Furthermore, to ensure semantic alignment, this embodiment of the invention also incorporates a dynamic semantic routing mechanism during the BERT model design phase. This mechanism includes: constructing a learnable memory network to store high-frequency process-quality association patterns (such as expert rules in a knowledge base); when the cosine similarity between a new association extracted by the BERT model's encoder and the implicit association pattern stored in the memory network exceeds a threshold, a memory-enhanced attention mechanism is triggered. During the BERT model training phase, historical knowledge is injected into the current relation extraction process, achieving iterative condensation from fragmented manufacturing data to manufacturing knowledge. The final BERT model, by setting a semantic similarity threshold (e.g., cosine similarity > 0.85) and a confidence score, maps the extracted entity relation triples (including explicit and implicit) to the ontology architecture of the manufacturing knowledge graph.

[0068] In the BERT model, when the similarity between two manufacturing characteristic concepts exceeds a certain threshold, they are considered to be the same or similar concepts. Then, the two identical or similar concepts are merged into one concept, and different local manufacturing knowledge graph data are gradually fused to form a complete manufacturing knowledge graph. In the actual fusion process of the manufacturing knowledge graph, there may be cases where nodes have the same name but different attribute values. In this case, the low similarity between feature nodes of two manufacturing knowledge graphs may prevent the merging of manufacturing characteristics in the manufacturing knowledge graph. Therefore, in the application of the BERT model, this embodiment of the invention, in addition to node semantic similarity, also incorporates the similarity of manufacturing node attribute content for further evaluation of manufacturing characteristics.

[0069] The expression for calculating attribute similarity in this embodiment of the invention is as follows:

[0070]

[0071] In the formula, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text.

[0072] Because the embodiments of the present invention introduce attribute similarity calculation in the above calculation, by giving specific weights The comprehensive similarity of the knowledge nodes is calculated, and the expression is:

[0073] ;

[0074] Based on the calculation results of the semantic similarity or comprehensive similarity mentioned above, it is determined whether the node information can be fused, thereby realizing the fusion update or expansion of the basic structure and extended structure of the manufacturing knowledge graph, so as to form a complete knowledge graph update and expansion for expressing the manufacturing knowledge graph data in the process of complex product development.

[0075] For further reference during the practical application phase. Figure 3 In step three of the present invention embodiment, the network structure of the BERT model, consisting of the SpanBERT encoding layer, candidate span generation layer, relation classifier layer, and decision output layer, is used to accurately extract the association relationships in the preprocessed high-quality data during the manufacturing process. The detailed processing flow includes: (1) Input data reconstruction and candidate span set construction: For each manufacturing document fragment, a candidate feature span set is constructed. The span length constraint (maximum span ≤ 8 tokens) and type compatibility filtering strategy are adopted to reduce the number of candidate pairs. (2) Span-level semantic representation generation: Each candidate span is generated. Inputting a pre-trained SpanBERT model, a high-dimensional semantic representation vector is generated through span boundary masking and an internal attention focusing mechanism. (3) Classification and probability calculation of association relationships: For each pair of candidate spans Concatenate its semantic vectors into The data is input into the relation classifier of the multilayer perceptron. The relation classifier defines the relation category set as {cause, influence, belong to, located, no association}, where the "cause" relation has the highest weight and is used to capture the quality causal chain. (4) Domain constraint verification and triple output: The manufacturing domain rule base is introduced to perform posterior verification of the classification results. For example, causal logic rules, temporal dependency rules, numerical constraint rules, etc. The relation pairs verified by the rules are output in the form of triples. All triples constitute the semantic association network data of the manufacturing process knowledge graph, which is stored in JSON-LD format and retains entity ID, relation type, confidence score and source document traceability information, providing high-quality structured input for subsequent graph fusion.

[0076] In updating the knowledge graph of diesel engine connecting rod manufacturing based on quality feature data, the BERT-BiLSTM-CRF algorithm is first used to automatically identify semantics related to quality characteristics. Then, the SpanBERT algorithm is used to extract the relationships between quality characteristics from the quality data. Finally, a semantic alignment method based on the Siamese-BERT algorithm is used to align the concepts related to quality characteristics and their relationships, forming a quality knowledge graph in JSON format. The result is as follows: Figure 6 The results are shown.

[0077] During the merging and updating of the knowledge graph for diesel engine connecting rod manufacturing, existing knowledge graphs for connecting rod manufacturing structures, such as... Figure 7 As shown, using a semantic similarity calculation method based on the Jaccard coefficient, according to Figure 6 The quality data knowledge graph shown will be used to... Figure 7 The manufacturing knowledge graph shown is updated. Through candidate node matching and identification, node name semantic similarity calculation, node attribute similarity calculation, and comprehensive similarity calculation and graph update expansion, it is agreed that when the comprehensive similarity of candidate node pairs from two graphs is greater than or equal to 0.7, the two nodes are fused. For the nodes determined to be fused, the original attribute fields in the structure graph are retained, and the node information from the quality characteristic graph is supplemented into the structure graph, realizing the fusion and update of the graphs. The update result is as follows: Figure 8 As shown.

[0078] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the scope of the technology disclosed in the present invention, and such modifications or substitutions should all be covered within the scope of protection of the present invention.

Claims

1. A method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning, characterized in that, include: Step 1: Preprocess the fragmented manufacturing data generated during the manufacturing process, and transform the preprocessed fragmented manufacturing data into high-quality input data; Step 2: Obtain data characteristics and correlation data from the high-quality input data to characterize the manufacturing information characteristics and the correlation between manufacturing information; Step 3: Align the data characteristics and related relationship data with the relevant concepts of manufacturing characteristics and relationships in the semantic space to obtain structured manufacturing knowledge graph data; Step 4: Update or expand the existing manufacturing knowledge graph based on the manufacturing knowledge graph data obtained in Step 3.

2. The method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning according to claim 1, characterized in that, Step one includes: Fragmented manufacturing data generated during the cross-stage and cross-system manufacturing process of diesel engines is preprocessed by sequentially performing data cleaning, noise filtering, time-series alignment, and multimodal unified representation. This process extracts structured data from equipment sensors, process logs, quality inspection reports, and other sources in the diesel engine manufacturing process. The structured data is then transformed into high-quality input data through pre-constructed entity data standardization annotation. Specifically, the pre-constructed entity data standardization annotation involves standardizing the entity data of process parameters, equipment status, and material batches.

3. The method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning according to claim 1, characterized in that, Step two includes: The BERT-based pre-trained bidirectional encoder model identifies manufacturing characteristics from the high-quality input data, extracts explicit correlations of data characteristics based on these characteristics, and performs overall consistency optimization to obtain manufacturing characteristic correlation triples corresponding to existing manufacturing knowledge graphs. During the pre-training phase of the bidirectional encoder model, a self-supervised task for logical ordering prediction and a self-supervised task for quality causal inference, oriented towards domain manufacturing processes, are added to inject causal chain knowledge related to manufacturing quality during the diesel engine manufacturing process into the bidirectional encoder model. A finely tuned bidirectional encoder model is jointly trained using data sequence standardization labeling and association classification. Based on the bidirectional encoder model, process entities in the diesel engine manufacturing process are identified from the manufacturing characteristic association triplet. At the same time, an attention mechanism guides the bidirectional encoder model to extract implicit associations related to process entities from the high-quality input data to initially obtain the manufacturing characteristic association triplet in the manufacturing entity.

4. The method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning according to claim 1, characterized in that, Step three includes: The manufacturing characteristic data output by the bidirectional encoder model and the vector data in the association data are accurately extracted from the manufacturing characteristic association triplet. Based on the semantic similarity calculation of the Jaccard coefficient, similar nodes and associations of manufacturing characteristics are merged after semantic similarity calculation of the existing manufacturing knowledge graph and the newly generated manufacturing knowledge graph in the diesel engine manufacturing process, thereby realizing the alignment of related concepts of manufacturing characteristics and associations. Specifically, this includes: using the semantic alignment method corresponding to the semantic alignment framework based on the Siamese-BERT network model to align the manufacturing characteristic data and the vector data one by one to achieve the alignment of manufacturing characteristics and associations; judging whether the manufacturing knowledge fusion data is the same manufacturing characteristic or the same association in the semantic space in the existing manufacturing knowledge graph according to the set similarity threshold, and guiding the merging of manufacturing characteristics or associations according to the judgment result to obtain the manufacturing knowledge graph data to be updated or expanded in the manufacturing process data.

5. The method for dynamically updating a knowledge graph of diesel engine manufacturing based on deep learning according to claim 1, characterized in that, Step four includes: Based on the calculation results of semantic similarity or a combination of semantic similarity and attribute similarity, the existing manufacturing knowledge graph and the newly generated manufacturing knowledge graph data in step three are used to merge similar nodes and similar associations, thereby forming a complete manufacturing knowledge graph data to be updated or expanded. Then, the existing manufacturing knowledge graph is updated or expanded based on its basic structure and its extended structure. Through steps one through four, the BERT model and semantic similarity calculation method are used to repeat the above steps for updating or expanding the knowledge graph, thereby forming an updated and expanded knowledge graph representing the manufacturing process of complex product development. The specific expression for the similarity calculation is as follows: In the formula, This represents two characteristic concepts in the creation of a knowledge graph. This indicates that two characteristic concepts in a knowledge graph share the same character. This represents all the characters contained in two characteristic concepts within a knowledge graph. In the formula, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text, Representing manufacturing characteristics in a manufacturing knowledge graph The attribute text, This indicates that two attribute texts in the knowledge graph contain the same common character. This represents all the characters contained in the two attribute texts in the knowledge graph. In the formula, This represents a given specific weight.

Images