An Automatic Generation Method for OPC UA Information Models Based on LLM Knowledge Completion
Patent Information
- Application Number
- CN202610737478.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-05-27
- Publication Date
- 2026-07-03
Smart Images
Figure CN122332480A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of LLM research and application technology, and in particular relates to an automatic generation method of OPC UA information model based on LLM knowledge completion. Background Technology
[0002] Currently, enterprises require more granular data representation models for field equipment in scenarios such as condition monitoring, predictive maintenance, and manufacturing execution optimization. OPC Unified Architecture (OPC UA), due to its cross-platform compatibility, service-oriented architecture, and support for hierarchical modeling, is widely regarded as the most universal industrial semantic modeling framework. This framework provides a unified description system for equipment, encompassing "type—object—attribute," enabling the import of equipment operating parameters, topological relationships, and service interfaces into the server's address space using a unified model. However, for most existing equipment lacking standardized design documents, its data primarily exists in structured forms such as relational database tables, CSV reports, historical curve tables, or variable lists. These tables contain rich "semi-structured semantics," such as naming conventions, value ranges, and implicit hierarchies between fields. The construction of existing OPC UA information models largely relies on existing engineering description models or semantic ontology systems, such as model mapping based on AML or UML, or utilizing Semantic Web and OWL ontology technologies for reasoning and consistency verification. These methods often assume that the equipment already has standardized engineering documents or structured semantic descriptions, which makes them difficult to apply to scenarios that are common in industrial settings, where only the underlying structured data tables are retained.
[0003] The existence of these problems means that in the actual implementation process in enterprises, a large number of devices still need to be mapped manually node by node, which consumes a lot of engineering resources. Summary of the Invention
[0004] The purpose of this invention is to provide an automatic generation method for OPC UA information models based on LLM knowledge completion, so as to solve the above-mentioned technical problems.
[0005] To address the aforementioned technical problems, the specific technical solution of this invention, a method for automatically generating OPC UA information models based on LLM knowledge completion, is as follows:
[0006] An automatic generation method for OPC UA information models based on LLM knowledge completion includes the following steps:
[0007] Step 1: Construct an OPC UA information model knowledge graph containing multiple heterogeneous industrial devices;
[0008] Step 2: Construct a dual-view inference LLM module, design a context-adaptive inference example selector and a Bayesian disambiguation module; the context-adaptive inference example selector is used to dynamically match reference examples and construct prompt words according to the semantic category of unknown entities, drive LLM to perform chain inference, and predict alignment nodes and link relationships; the Bayesian disambiguation module is used to fuse the confidence of LLM output with the prior knowledge graph topology structure, and disambiguate candidate nodes through sequential updates, outputting the most likely alignment nodes;
[0009] Step 3: Construct a knowledge completion LLM module based on structural prefixes. The module includes a relation-guided multi-layer graph attention network and a projection layer.
[0010] Step 4: Train the knowledge completion LLM by combining "continuous vector prefix" and "text vector" for fine-tuning. Deploy the trained OPC UA information model automatic generation system to the industrial site and run it to generate standardized XML text of the information model.
[0011] Furthermore, step 1 specifically includes:
[0012] Step 1.1: Use the Nodeset XML text standard for industrial equipment released by the OPC Foundation as the reference model, and collect the user manuals and operation manuals provided by the equipment manufacturers to supplement the missing information nodes or reference relationships in the reference model;
[0013] Step 1.2: For the XML reference model text, open-source tools are used to parse and extract the core components, attribute nodes and their reference relationships of the equipment; for the user manual and operation manual text, NLP models are used to extract entities and relationships from the text; the extracted entity nodes and reference relationships are converted into triples after manual review and stored in the Neo4j database to form a unified information model knowledge graph containing multiple heterogeneous industrial devices.
[0014] Furthermore, in step 2, the context-adaptive reasoning example selector dynamically identifies the semantic category of the unknown entity through prompt words during the context reasoning stage, and adaptively matches the corresponding category of reference examples from the reference example repository; subsequently, the matched reference examples are sequentially concatenated with the task description, context reasoning instructions, unknown entities, candidate nodes, and relationship sets to dynamically construct context reasoning prompt words and input them into the LLM, predicting the alignment nodes and link relationships of the unknown entity through chain reasoning driven by matching examples.
[0015] Furthermore, in step 2, the Bayesian disambiguation module deduplicates the node set output by the dual-view LLM inference to construct a conflict hypothesis space, and calculates the structural prior probability of each candidate node based on the topological attributes of the knowledge graph; it extracts the log probability of the candidate node when decoding the dual-view LLM as the generation likelihood, and uses Bayes' theorem to fuse it with the structural prior probability to obtain the initial posterior confidence distribution; when the maximum confidence is lower than a preset threshold, it triggers external corpus retrieval to obtain supplementary evidence, and performs Bayesian sequential updates on the confidence distribution until convergence, and selects the top three nodes with the highest probabilities as the final disambiguation result.
[0016] Furthermore, the graph attention network described in step 3 consists of a message passing module and multi-head attention. After receiving the alignment node and the link relationship, the graph attention network constructs a set of adjacent parent nodes connected to the alignment node by the link relationship. The message function generates the adjacency message of each parent node through linear transformation. The multi-head attention mechanism calculates the attention weights of different adjacency messages and performs weighted aggregation. The above process is then encoded in multiple layers to obtain the local embedding of the alignment node. Finally, a learnable global vector is introduced, and the comprehensive embedding representation of the alignment node is obtained through a linear fusion strategy.
[0017] The projection layer transforms the low-dimensional embedding vector of the aligned node into a high-dimensional vector, matching its dimension with that of the hidden layer in the LLM. This vector is then used as a virtual prefix and appended to the LLM text prompt vector after word segmentation to form a structure-guided input sequence.
[0018] Furthermore, in step 4, during training, the graph attention network is used to extract the structured vector representation of adjacent triples based on the alignment nodes and link relationships. These vectors are then projected and injected as prefixes into the beginning of the LLM input sequence, so that the LLM can simultaneously receive uncompressed graph structure information and text prompts during training.
[0019] Furthermore, step 4 specifically includes:
[0020] Step 4.1: Deploy the information model automatic generation system on an OPC UA server at the edge layer;
[0021] Step 4.2: Collect structured equipment operation data from the industrial database and input it into the information model automatic generation system;
[0022] Step 4.3: The system links unknown entities in the structured text to the corresponding information model knowledge graph based on semantic reasoning and knowledge completion, and stores the complete model graph in the Neo4j database;
[0023] Step 4.4: Parse node tags and relationship attributes using the Neo4j database plugin to automatically generate an ontology file. Call the framework to parse the ontology file and use the DOM mechanism to organize the hierarchical structure. Finally, use the UA-ModelCompiler compiler to complete file verification and conversion, generating a model description file that conforms to the OPC UA specification.
[0024] The automatic generation method for OPC UA information models based on LLM knowledge completion proposed in this invention has the following advantages:
[0025] 1. Significantly reduces manual intervention and improves modeling efficiency.
[0026] Traditional OPC UA information model construction relies on manual mapping of device nodes one by one, which is time-consuming and labor-intensive. This invention uses LLM to automatically complete semantic reasoning, entity alignment, and relation completion, and can automatically generate standardized models from structured device data (such as database tables, CSV, variable lists, etc.), significantly reducing engineering resource input.
[0027] 2. Adapt to the "weak semantic" data environment commonly found in industrial settings.
[0028] Existing methods often rely on standardized engineering documents (such as AML and OWL ontology), while this invention is specifically designed for scenarios common in industrial settings where only the underlying structured data tables are retained. It can extract semantic information from semi-structured field names, value ranges, and implicit hierarchies, and has stronger practicality and generalization capabilities.
[0029] 3. Dual-perspective reasoning mechanism improves the accuracy of entity alignment.
[0030] By combining the dual perspectives of "context-example-driven chained reasoning" and "LLM intrinsic prior knowledge reasoning," and introducing a context-adaptive example selector, the bias of a single reasoning approach is avoided, and the prediction accuracy of alignment nodes and link relationships for unknown entities is significantly improved.
[0031] 4. The Bayesian disambiguation module effectively resolves candidate node conflicts.
[0032] It integrates LLM output confidence with prior knowledge graph topology and supports sequential updates from external corpus retrieval, enabling the selection of the most reasonable alignment result from multiple candidate nodes and reducing the risk of error propagation caused by ambiguity.
[0033] 5. Structural prefix injection enhances LLM's ability to perceive graph structures.
[0034] By extracting adjacency structure embeddings through a relation-guided multi-layer graph attention network and injecting them as continuous vector prefixes into the LLM input sequence, the LLM can simultaneously perceive graph topology and text semantics during training and inference, thereby improving the contextual consistency of knowledge completion.
[0035] 6. End-to-end automated deployment and standardized output
[0036] The system can be deployed on an edge layer OPC UA server, automatically receive structured data from industrial databases, complete semantic reasoning, knowledge completion, and graph storage, and finally generate an XML model description file that conforms to the OPC UA specification. This file can be directly imported into the server address space to achieve semantic interoperability and standardized access for devices.
[0037] 7. It has good scalability and maintainability.
[0038] The system uses a knowledge graph (Neo4j) as a unified information model store, supporting the fusion modeling of multiple heterogeneous devices. New device types or relationships can be added by supplementing the reference model and operation manual text, without requiring a redesign of the system architecture.
[0039] In summary, this invention has significant beneficial effects in reducing labor costs, improving the level of modeling automation, adapting to complex industrial data environments, and ensuring model accuracy and standardization. Attached Figure Description
[0040] Figure 1 This is a flowchart of the method in an embodiment of the present invention.
[0041] Figure 2 This is a flowchart of the process for constructing an information model knowledge graph in an embodiment of the present invention.
[0042] Figure 3 This is the entity semantic dual-perspective reasoning LLM structure diagram in this embodiment of the invention.
[0043] Figure 4 This embodiment of the invention uses knowledge-based structural prefix completion to complete the LLM structure diagram.
[0044] Figure 5 This is a flowchart of generating the information model XML text in an embodiment of the present invention. Detailed Implementation
[0045] To better understand the purpose, structure, and function of this invention, the following description, in conjunction with the accompanying drawings, provides a more detailed account of an automatic generation method for OPC UA information models based on LLM knowledge completion.
[0046] like Figure 1As shown, the present invention provides an automatic generation method for OPC UA information models based on LLM knowledge completion, comprising the following steps:
[0047] Step 1: Construct an OPC UA information model knowledge graph containing multiple heterogeneous industrial devices;
[0048] like Figure 2 As shown, the specific process for constructing the knowledge graph of the industrial heterogeneous equipment information model in step 1 is as follows:
[0049] Step 1.1: Information Model Source: The OPC Foundation's industrial equipment standard Nodeset XML text is used as the reference model. User manuals, operation manuals, and other texts provided by equipment manufacturers are collected to supplement the missing information nodes or reference relationships in the reference model.
[0050] Step 1.2: Knowledge Graph Construction: For XML reference model text, open-source tools such as python-opcua and EclipseMilo are used to parse and extract the core components, attributes, and other nodes of the equipment, as well as the reference relationships between them; for user manuals and operation manuals, NLP models such as SpaCy are used to extract entities and relationships from the text; after manual review, the extracted entity nodes and reference relationships are converted into triples and stored in the Neo4j database to form a unified information model knowledge graph containing multiple heterogeneous industrial devices.
[0051] Step 2: Construct a dual-view inference LLM module. The dual-view inference LLM module is mainly used to design a context-adaptive inference example selector and a Bayesian disambiguation module. The context-adaptive inference example selector is used to dynamically match reference examples and construct prompt words based on the semantic category of unknown entities, driving the LLM to perform chained inference and predicting alignment nodes and link relationships. The Bayesian disambiguation module is used to fuse the confidence of the LLM output with the prior knowledge graph topology structure, and disambiguate candidate nodes through sequential updates to output the most likely alignment nodes.
[0052] like Figure 3 As shown, the dual-view inference LLM module in step 2 includes a candidate recall module, a dual-view inference LLM module, and a Bayesian disambiguation module, which are specifically implemented as follows:
[0053] Step 2.1: The candidate recall module uses edit distance and 3-gram Jaccard (a text similarity calculation method based on Jaccard similarity) to calculate the character similarity between unknown entities on the device and known nodes in the knowledge graph. Simultaneously, it introduces an embedding model based on Sentence-BERT (a sentence embedding model improved based on the BERT architecture) to calculate the semantic similarity between the two. Finally, a linear fusion mechanism is used to select the top-k candidate node set with the highest similarity, as shown in the following formula:
[0054] (1)
[0055] (2)
[0056] (3)
[0057] in Represents an unknown entity With known nodes Character similarity, Indicates from string arrive The minimum number of editing operations, and The lengths of the two strings are respectively. Represents a collection of strings of length 3; Represents an unknown entity With known nodes semantic similarity, and These are the embedding vectors for the unknown entity and the known node, respectively. and These are the L2 norms of the two vectors, used to measure the length of the vectors; Represents an unknown entity With known nodes The overall similarity These are the weight parameters for the semantic layer;
[0058] Step 2.2: The dual-view inference LLM module predicts three alignment nodes and one link relationship for the unknown entity from candidate nodes and candidate relationships using example-driven CoT chain inference and the LLM's inherent pre-trained knowledge. Candidate nodes are selected by the candidate recall module, and candidate relationships are derived from common relationships in OPC UA modeling, such as HasComponent and HasProperty. In the context inference stage, a context-adaptive inference example selector is introduced. This selector dynamically identifies the semantic category of the unknown entity using prompts and adaptively matches corresponding reference examples from the reference example repository. Subsequently, the matched reference examples are sequentially concatenated with the task description, context inference instructions, unknown entity, candidate nodes, and relationship set to dynamically construct context inference prompts, which are then input into the LLM. The alignment nodes and link relationships of the unknown entity are predicted through example-driven chain inference, thus avoiding inference bias caused by uniform examples. In the prior entity inference stage, the LLM does not rely on any examples or context, directly inferring the alignment nodes and link relationships from the unknown entity. The context inference and prior knowledge inference prompts are as follows:
[0059] (4)
[0060] (5)
[0061] in Indicates contextual inference prompts, This indicates the description of the reasoning task. This refers to the task instructions and inference chain in the reference example. Indicates a contextual reasoning task instruction. These represent unknown entities, candidate node sets, and candidate relation sets, respectively. Indicates clues for prior reasoning, Indicates instructions for prior reasoning tasks;
[0062] Step 2.3: The Bayesian disambiguation module reconstructs the conflict hypothesis space from the node set output by the dual-view LLM inference and calculates the structural prior probability of each candidate node based on the topological attributes of the knowledge graph; it extracts the log probability of the candidate nodes during dual-view LLM decoding as the generation likelihood and fuses it with the structural prior probability using Bayes' theorem to obtain the initial posterior confidence distribution; when the maximum confidence is lower than a preset threshold, it triggers external corpus retrieval to obtain supplementary evidence, and performs Bayesian sequential updates on the confidence distribution until convergence, selecting the three nodes with the highest probabilities as the final disambiguation result, as shown in the following formula:
[0063] (6)
[0064] (7)
[0065] (8)
[0066] in This represents the prior structural probability of a candidate node. This indicates the topological depth of the candidate node in the graph. This represents the confidence distribution calculated after fusing structural priors and dual-view evidence. This indicates that supplementary evidence has been obtained. The system then updates the node confidence level;
[0067] Step 3: Construct a knowledge completion LLM module based on structural prefixes. The module includes a relation-guided multi-layer graph attention network and a projection layer.
[0068] like Figure 4 As shown, the knowledge completion LLM module based on structural prefixes in step 3 is specifically implemented as follows:
[0069] Step 3.1: The graph attention network consists of a message passing module and multi-head attention. After receiving the alignment node and the link relationship, the graph attention network constructs a set of adjacent parent nodes connected to the alignment node by the link relationship. The message function generates the adjacency message of each parent node through a linear transformation. The multi-head attention mechanism calculates the attention weights of different adjacency messages and performs weighted aggregation. The above process is then encoded in multiple layers to obtain the local embedding of the alignment node. Finally, a learnable global vector is introduced, and the comprehensive embedding representation of the alignment node is obtained through a linear fusion strategy, as shown in the following formula:
[0070] (9)
[0071] (10)
[0072] (11)
[0073] (12)
[0074] (13)
[0075] in Indicates the parent node Through relationships Pass to the alignment node Adjacency messages, Represents the Hadamard product. , Representation and relation Relevant trainable parameters; Indicates the first Parent node in the head Attention weights Indicates the first Attention parameters for each head. , They represent the first The coefficient matrix of the aligned nodes and parent node messages in each header. This represents vector concatenation. Represents the set of adjacent parent nodes. This represents the LeakyReLU activation function; Indicates alignment node Adjacency aggregation information, Represents the coefficient matrix. The number of attention heads is represented by ∏, and the join operation is represented by ∏. Indicates alignment node Local embedding representation, Indicates the number of layers in the graph attention network; Indicates alignment node The comprehensive embedding representation, Indicates alignment node The global vector, The weight parameters represent the local embedding.
[0076] Step 3.2: The projection layer transforms the low-dimensional embedding vectors of the three aligned nodes into high-dimensional vectors, matching their dimensions with those of the hidden layers in the LLM. These vectors are then used as virtual prefixes and appended to the LLM text prompt vectors after word segmentation, forming a structure-guided input sequence, as shown in the following formula:
[0077] (14)
[0078] in This represents the high-dimensional vector obtained by projecting the node vector. This represents the hidden layer dimension of an LLM. Indicates the projection layer parameters. Represents n-dimensional node embedding;
[0079] Step 4: Train the knowledge completion LLM by combining "continuous vector prefixes" and "text vectors" for joint fine-tuning. Deploy the trained OPC UA information model automatic generation system to the industrial site and run it to generate standardized XML text of the information model.
[0080] In step 4, the knowledge completion LLM is trained by combining "continuous vector prefixes" and "text vectors" for joint fine-tuning. During training, the graph attention network is used to extract the structured vector representation of adjacent triples based on the alignment nodes and link relationships. These vectors are then projected and injected as prefixes at the beginning of the LLM's input sequence, so that the LLM can receive uncompressed graph structure information and text prompts simultaneously during training.
[0081] like Figure 5 As shown, the system deployment and operation process in step 4 is as follows:
[0082] Step 4.1: Model Deployment: Deploy the information model automatic generation system on the OPC UA server at the edge layer to realize information modeling of industrial equipment;
[0083] Step 4.2: Data Input: Collect structured equipment operation information from the industrial database and input it into the information model automatic generation system;
[0084] Step 4.3: Information Modeling: The system links unknown entities in the structured text to the corresponding device's OPC UA information model knowledge graph based on semantic reasoning and knowledge completion, and stores the complete model graph in the Neo4j database;
[0085] Step 4.4: Model Compilation: The Neo4j database plugin Neosemantics parses node tags and relationship attributes to automatically generate an ontology OWL file conforming to the RDF / OWL semantic specification. Then, the Jena framework is called to parse the OWL file and uses the DOM mechanism to organize the hierarchical structure according to the OPC official XML syntax requirements. Finally, the UA-ModelCompiler compiler is used to complete the XML file validation and transformation to generate a model description file conforming to the OPC UA specification. This file can be directly imported into the address space of the OPC UA server to achieve semantic interoperability and standardized access for devices.
[0086] It is understood that the present invention has been described through some embodiments, and those skilled in the art will recognize that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. Furthermore, under the teachings of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are within the protection scope of the present invention.
Claims
1. A method for automatically generating OPC UA information models based on LLM knowledge completion, characterized in that, Includes the following steps: Step 1: Construct an OPC UA information model knowledge graph containing multiple heterogeneous industrial devices; Step 2: Construct a dual-view inference LLM module, and design a context-adaptive inference example selector and a Bayesian disambiguation module; The context-adaptive reasoning example selector is used to dynamically match reference examples and construct prompt words based on the semantic category of unknown entities, driving LLM to perform chained reasoning and predict alignment nodes and link relationships. The Bayesian disambiguation module is used to fuse the confidence of the LLM output with the prior knowledge graph topology, and to disambiguate candidate nodes through sequential updates, outputting the most likely aligned nodes. Step 3: Construct a knowledge completion LLM module based on structural prefixes. The module includes a relation-guided multi-layer graph attention network and a projection layer. Step 4: Train the knowledge completion LLM by combining "continuous vector prefix" and "text vector" for fine-tuning. Deploy the trained OPC UA information model automatic generation system to the industrial site and run it to generate standardized XML text of the information model.
2. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, Step 1 specifically includes: Step 1.1: Use the Nodeset XML text standard for industrial equipment released by the OPC Foundation as the reference model, and collect the user manuals and operation manuals provided by the equipment manufacturers to supplement the missing information nodes or reference relationships in the reference model; Step 1.2: For the XML reference model text, open-source tools are used to parse and extract the core components, attribute nodes and their reference relationships of the equipment; for the user manual and operation manual text, NLP models are used to extract entities and relationships from the text; the extracted entity nodes and reference relationships are converted into triples after manual review and stored in the Neo4j database to form a unified information model knowledge graph containing multiple heterogeneous industrial devices.
3. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, In step 2, the context-adaptive reasoning example selector dynamically identifies the semantic category of unknown entities through prompt words during the context reasoning stage, and adaptively matches reference examples of the corresponding category from the reference example repository. Subsequently, the matched reference examples, task descriptions, contextual reasoning instructions, unknown entities, candidate nodes, and relationship sets are sequentially concatenated to dynamically construct contextual reasoning prompts, which are then input into the LLM. Through example-driven chain reasoning, the alignment nodes and link relationships of unknown entities are predicted.
4. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, In step 2, the Bayesian disambiguation module deduplicatively constructs a conflict hypothesis space from the node set output by the dual-view LLM inference and calculates the structural prior probability of each candidate node based on the topological attributes of the knowledge graph. The log probability of candidate nodes in dual-view LLM decoding is extracted as the generation likelihood, and it is fused with the structural prior probability using Bayes' theorem to obtain the initial posterior confidence distribution. When the maximum confidence is lower than the preset threshold, external corpus retrieval is triggered to obtain supplementary evidence, and the confidence distribution is updated sequentially by Bayes until convergence, and the top three nodes with the highest probabilities are selected as the final disambiguation result.
5. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, The multi-layer graph attention network described in step 3 consists of a message passing module and multi-head attention. After receiving the alignment node and the link relationship, the graph attention network constructs a set of adjacent parent nodes connected to the alignment node by the link relationship. The message function generates the adjacency message of each parent node through linear transformation. The multi-head attention mechanism calculates the attention weights of different adjacency messages and performs weighted aggregation. The above process is then encoded in multiple layers to obtain the local embedding of the alignment node. Finally, a learnable global vector is introduced, and the comprehensive embedding representation of the alignment node is obtained through a linear fusion strategy. The projection layer transforms the low-dimensional embedding vector of the aligned node into a high-dimensional vector, matching its dimension with that of the hidden layer in the LLM. This vector is then used as a virtual prefix and appended to the LLM text prompt vector after word segmentation to form a structure-guided input sequence.
6. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, In step 4, during training, the graph attention network is used to extract the structured vector representation of adjacent triples based on the alignment nodes and link relationships. These vectors are then projected and injected as prefixes into the beginning of the LLM input sequence, so that the LLM can simultaneously receive uncompressed graph structure information and text prompt vectors during training.
7. The method for automatically generating OPC UA information models based on LLM knowledge completion according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1: Deploy the information model automatic generation system on an OPC UA server at the edge layer; Step 4.2: Collect structured equipment operation data from the industrial database and input it into the information model automatic generation system; Step 4.3: The system links unknown entities in the structured text to the corresponding information model knowledge graph based on semantic reasoning and knowledge completion, and stores the complete model graph in the Neo4j database; Step 4.4: Parse node tags and relationship attributes using the Neo4j database plugin to automatically generate an ontology file. Call the framework to parse the ontology file and use the DOM mechanism to organize the hierarchical structure. Finally, use the UA-ModelCompiler compiler to complete file verification and conversion, generating a model description file that conforms to the OPC UA specification.