OpCUA information model construction method based on unstructured text data

Abstract

The application discloses an OPC UA information model construction method based on unstructured text data, mainly solves the problems that existing OPC UA information model modeling often depends on structured data and lacks description information from text data, and realizes the steps, including: designing a marking label based on corpus text features and meeting the modeling requirements of an industrial field, and designing an organization relationship of the label; performing BIO serialization marking on extracted sample text materials to obtain training samples; constructing a Chinese pre-training model based on a BERT model to extract word vectors; processing, classifying and marking the word vectors based on CRF technology; extracting effective entities in the text by using the completed model, and organizing the model according to pre-defined organization rules. The application constructs industrial unstructured text sample data, designs an entity label extraction method using text corpus and applies the method to information model construction, widens the application scene of the information model construction method, and has simple and clear operation method and practical value.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of industrial automation, and particularly relates to a construction method of an OPC UA information model, and specifically relates to an OPC UA information model construction method based on unstructured text data, which can be used for intelligent model abstraction construction in an industrial field environment and realizes industrial intelligence. BACKGROUND

[0002] Digital transformation of production line equipment is becoming more and more necessary. OPC is a data security exchange interoperability standard applied in the automation industry and other industries, which is formulated by industry suppliers, end users and software developers. These specifications define the interfaces between clients and servers and between servers, such as accessing real-time data, monitoring alarms and events, accessing historical data and other applications. The OPC standard was first published in 1996, and its purpose is to abstract the specific protocols of programmable logic controllers (PLC) such as Modbus, Profibus, etc. into a standardized interface, as an "intermediary" role to convert general OPC "read-write" requests into specific device protocols to directly interface with HMI / SCADA systems, and vice versa. As a result, a complete product industry has emerged, and end users can optimize products by using it to achieve seamless interaction between systems through the OPC protocol. Currently, the OPC Unified Architecture (OPC UA) information model has been used as an effective method to build digital images of physical devices in various industrial fields, so that sensing and control can be achieved through virtual copies.

[0003] Since the OPC UA information model is object-oriented, the common method to construct the information model is to manually define the nodes by experienced engineers. Because of their understanding of the entire product line, they know the properties of the equipment on the production line. In addition, using model construction tools is also an option. Some researches also try to construct the model by mapping other format models to the OPC UA information model. Some other researches construct the information model by using other mode industrial information industrial databases.

[0004] However, all the above traditional methods ignore the use of rich unstructured text information in the industrial field. In the design of feasible solutions for information model construction, existing research always focuses on structured data such as knowledge graph, other models, etc. There is some redundant information in valuable unstructured data (such as text) that may interfere with the extraction of useful information. Therefore, it is important to find a novel method to obtain the information we want from complex text. In this process, it is also essential to design entity structure rules with prior knowledge. At present, domestic enterprises and research institutions still lack sufficient research on OPC UA information model modeling technology, which is still in a relatively primary stage. There is no mature OPC UA information model modeling technology application solution for industrial Internet of Things transformation, especially the lack of research on OPC UA information model construction corpus source.

[0005] In addition, in practical applications, engineers establishing information models are difficult to master all the mechanisms of industrial equipment, and when there are unknown devices, it is often difficult to obtain accurate model information. A widely applicable and easy-to-practice information model establishment method is needed.

[0006] In view of this situation, the application research of the OPC UA information model construction method based on unstructured text has great research significance, and can well promote the intelligent transformation progress of the industrial field and reduce the difficulty of intelligent transformation of the industrial field. SUMMARY

[0007] The present application focuses on the text of the field device and proposes an information model construction process that can make full use of the text data not used in the previous process. The present application uses the BERT model combined with the conditional random field CRF to extract labeled entities from the original text, and uses the predefined label relationship to construct the information model through the extracted entities. The present application first realizes the text entity recognition method in the industrial field, establishes the sample set of industrial sequence labeling, fine-tunes the BERT model, and thus obtains good extraction results. The present application designs a new information model construction method: using the language processing model BERT to analyze the text and extracting entities from the text through the running sequence labeling. Through the predefined label relationship, the entities are organized in a predetermined manner to meet the standards of OPC UA.

[0008] The purpose of the present application is achieved by the following technical solution: an OPC UA information model construction method based on unstructured text data, comprising the following steps:

[0009] (1) Design an annotation label based on the text features of the corpus, which meets the modeling needs of the industrial field, and design the organization relationship of the label, that is, use different levels of description labels, the highest level is the model description object, and the attributes and attribute values of the object are below it;

[0010] (2) The sample text materials extracted from the device description document are BIO serialization annotation based on description tags at different levels to obtain training samples;

[0011] (3) Construct a Chinese pre-trained model based on the BERT model to extract word vectors from the sample text annotated in step (2);

[0012] (4) Based on CRF technology, process, classify and label the word vectors extracted in step (3);

[0013] (5) Use the training samples obtained in step (2) to train the text processing model that combines BERT and CRF;

[0014] (6) Use the trained text processing model to extract valid entities from the text and organize the model according to the predefined organization rules. That is, use the text processing model to extract entity text from the labels and organize the entities according to the predefined label relationships. Finally, construct an object-oriented OPC UA information model for the relevant entity labels.

[0015] Furthermore, in step (1), the information model node of the object-oriented OPC UA is abstracted into entity labels, and the entity label relationship is defined according to the original organizational structure of the information model node. Subsequently, the model can be constructed according to the predefined relationship.

[0016] Furthermore, in step (1), for the information model of OPC UA, labels are designed that include three levels: objects, object attributes and characteristics, and attribute values. Specifically, these include four types of labels: model description object OBJ, component COM, attribute ATT, and attribute value VAL.

[0017] Furthermore, the annotation of industrial data for the BIO-annotated sample data described in step (2) is as follows:

[0018] ① Use tags that meet the modeling requirements of the industrial field to perform BIO annotation on unstructured text data in the industrial field;

[0019] ② The labeling itself takes into account the label organization relationship of the information model and defines the label relationship that conforms to the structural requirements of the OPC UA information model. The model labels are designed according to three layers: object, attribute features and related attribute values, which can directly correspond to the objects, nodes and variables of the OPC UA information model.

[0020] Furthermore, in step (2), the text material is labeled with four types of tags: the designed model description object OBJ, the component COM, the attribute ATT, and the attribute value VAL. Redundant words that are not part of the target information are classified as O tags, the first character of the target information is labeled as B, and the subsequent characters are labeled as I.

[0021] Furthermore, in step (2), after the annotation is completed, the original text is further enhanced: the annotated text is translated by back-translating the text containing the target information. According to the distribution of O tags, the back-translated text is inserted into the appropriate position in the original text. The insertion method is as follows: by calculating the frequency of O tags in a certain paragraph, if the frequency value exceeds the set threshold, the back-translated text to be inserted is inserted into the current paragraph position; after all paragraphs are traversed, if there is still back-translated text remaining, the remaining part is inserted into the end of the entire sample.

[0022] Furthermore, in steps (3) and (4), the labeling and entity extraction methods are applied to the training process with a small number of samples:

[0023] ①The BERT model is a pre-trained model. During the pre-training process, it utilizes large-scale computing resources to extract implicit information from Chinese text. In this study, a fine-tuning method is used during the training process, and a small number of training rounds (3 to 6 rounds) are conducted to achieve model transfer, which meets the needs of industrial sites and actual operation.

[0024] ②The BERT model can make good use of the contextual information of the text. By using the masking method, it can mask some words and predict the masked words, thus obtaining a model containing rich contextual information.

[0025] ③ Use CRF to process the word vector information output by the BERT model, output serialized annotation results, and optimize the entire sequence, which meets the overall performance requirements of industrial scenarios.

[0026] Furthermore, in step (6), after obtaining the trained model, the actual text appearing in the industrial production environment to be processed is input, and the corresponding processing is performed to obtain the keywords in the text, i.e. entities; on this basis, the model structure is organized using predefined relationships between entities, and finally the construction of the OPC UA information model is completed.

[0027] Further, in step (6), after obtaining the tags, they are mapped to an abstract model according to their organizational relationship; the abstract model itself does not contain any specification information, and its structure only contains an object-oriented hierarchical structure. Specifically: words labeled OBJ are considered to be the objects described by the model. They are located at the top level of the abstract model and are the main body of the model. When there are multiple words labeled OBJ in the segment describing the same device, only one of them is used as the model description object; words labeled COM are the components of the model. In the abstract model, these words are located around the model body - the model description object and are connected to it, indicating that there is a relationship of inclusion and belonging between the two; words labeled ATT are the description of the model, reflecting the performance of the model. These words also exist around the description object and are connected to it, but the connection relationship is an attribute relationship; finally, there are words labeled VAL. These words are the specific values ​​when describing the model. In the abstract model, their position is around the words labeled ATT and connected to them.

[0028] Furthermore, in step (6), the nodes in the OPC UA information model are actually described in XML format. The way the file is written is subject to strict definition requirements. It is necessary to transform the abstract model with a defined structural framework into a model file that meets the definition requirements through a certain corresponding method, so as to realize the transformation from object model to node model. The corresponding method between the abstract model and the OPC UA information model is as follows: the UA ObjectType element represents the object type node, which corresponds to the main body and the components of the abstract model. The OPC UA node reference information of the latter also contains its connection relationship with the model body. The UA Variable represents the variable node, which corresponds to the model description in the abstract model. The specific value of the UA Variable node is determined by the specific value when describing the model. According to the above corresponding method, the OPC UA information model in XML format is constructed.

[0029] Compared with the prior art, the present invention has the following advantages:

[0030] 1. This invention fully utilizes unstructured text data in industrial environments, broadening the range of materials that can be used for information modeling. Existing research always focuses on structured data, such as knowledge graphs and other models, while valuable unstructured data (such as text) is often discarded due to the presence of redundant information. This invention finds a novel method—entity extraction—to extract desired information from complex text and reorganize the information extracted from the text to form a structured model.

[0031] 2. This invention designs and constructs industrial data entity extraction samples, combines the labels required by the information model, and uses the BIO annotation method to annotate the collected raw data, thus creating a replicable path for the utilization of text in industrial scenarios.

[0032] 3. This invention specifically addresses the scenario of information model construction in industry, designing a method for building models based on BERT and CRF. It fully considers the characteristic of limited sample sizes in industrial environments, achieving word segmentation and accurate word vectorization of the input text through a Chinese pre-trained model. This avoids problems such as missing word vector information and inaccurate sequence annotation caused by insufficient samples. Furthermore, the design features based on BERT and CRF models significantly reduce training costs in practical applications, eliminating reliance on large-scale computing devices and significantly improving the method's versatility. Attached Figure Description

[0033] Figure 1 This is the overall flowchart of the present invention.

[0034] Figure 2 This is a schematic diagram of BIO annotation for sample text data in this invention.

[0035] Figure 3 This is a schematic diagram of the BERT+CRF model for information extraction and processing in this invention.

[0036] Figure 4 This is a schematic diagram of the OPC UA information model of a simulation example of the present invention. Detailed Implementation

[0037] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0038] Before introducing the OPC UA information model construction method based on unstructured text data provided by this invention, we will briefly introduce the concepts of the entity extraction neural network model BERT+CRF and the OPC UA information model.

[0039] BERT (Bidirectional Encoder Representations from Transformers) is a pre-trained model framework proposed by a Google research team. It has been pre-trained on a massive corpus by the team to obtain good parameters for the encoder-decoder neural network in the model. In addition, the BERT model employs a masking training method, pre-masking a portion of the words in the training text before inputting them into the model for lexical vector representation, creating an effect similar to a cloze test, achieving better results than previous Bidirectional Long-Short Term Memory (BiLSTM) neural networks. CRF, on the other hand, is a sequence labeling algorithm that takes an input sequence and outputs target sequence labels; it is also considered a seq2seq model. For example, in this invention, the input sequence for named entity recognition is a string of words, and the output sequence is the corresponding entity labels—object (OBJ), component (COM), attribute (ATT), and attribute value (VAL).

[0040] As a key technology in OPC-UA applications, information modeling has always been a focus of academic research. The OPC-UA information model uses object-oriented techniques to describe device objects, defining device attribute variables, operating methods, and relationships with other objects. This more effectively represents the semantics of data and promotes interoperability. The OPC-UA information model consists of multiple nodes, including objects, object types, variables, and variable types. These can be categorized hierarchically into three levels: model description objects, object attribute characteristics, and attribute values.

[0041] In the application of OPC UA information modeling technology, especially in the construction of information models for industrial production scenarios, structured data is often used as the source of modeling materials, which faces a series of challenges such as limited modeling methods, complex modeling approaches, and low utilization of industrial information. To address these challenges, this invention proposes a method for constructing OPC UA information models based on unstructured text data. This method involves establishing small-scale labeled industrial data samples and using a pre-trained model for semantic analysis. Figure 1 The implementation steps of this invention are as follows:

[0042] Step 1: Design appropriate and relevant annotation tags based on corpus text features, meeting the needs of industrial site modeling, and design the tag organization relationship, using different levels of descriptive tags, with the highest level being the model description object, followed by object attributes and attribute values. For the OPC UA information model, a tag system containing three levels is needed: object, object attribute characteristics, and attribute values. Taking robotic arm equipment as an example, there are four types of tags: Model Description Object (OBJ), Component (COM), Attribute (ATT), and Attribute Value (VAL). The Model Description Object (OBJ) tag is established based on the following consideration: a specific document only targets a specific model, and all OBJ class terms can correspond to the same model description object. Based on this, the model description object can be abstracted into a general equipment category, making the category information completely correspond to the object node type in the OPC UA information model. Component (COM) is essentially also an object, but because its actual level is lower, this type of entity is combined with composition relationships. Attribute (ATT) and Attribute Value (VAL) are used to describe the data-related characteristics exhibited by the mechanical equipment as a physical object. This labeling system includes information at multiple levels, including objects, object attributes and characteristics, and attribute values. The highest level is the model description object (OBJ), the next level is the components (COM) and attributes (ATT) that describe the object's attribute characteristics, and the last level is the attribute value (VAL).

[0043] Step Two: Perform BIO (Begin Inside Outside) serialization annotation on the sample text materials extracted from the equipment description documents (which can be in HTML or PDF format) based on different levels of description tags to obtain training samples. Industrial sites and mechanical equipment themselves contain a large amount of unstructured text data. For example, robotic arm equipment often provides equipment manuals and parameter performance descriptions, while the operating software also contains related descriptions and introductions. Sample text materials can be easily extracted from these documents. The text materials are labeled using the four types of tags designed in Step One. Redundant words that are not target information are classified as O tags, the first character of the target information is labeled as B, and subsequent characters are labeled as I. For example, "maximum load" is labeled as B-ATT (used to label the beginning "maximum") and I-ATT (used to label the subsequent "large load"), respectively. After annotation, the original text is further enhanced: the annotated text is translated by back-translating the text containing the target information. According to the distribution of O tags, the back-translated text is inserted into the appropriate position in the original text. The insertion method is as follows: by calculating the frequency of O tags in a certain paragraph, if the frequency value exceeds the set threshold, the back-translated text to be inserted is inserted into the current paragraph position; after all paragraphs are traversed, if there is still back-translated text remaining, the remaining part is inserted into the end of the entire sample.

[0044] Step 3: Construct a pre-trained Chinese model based on BERT (Bidirectional Encoder Representations from Transformers) to extract word vectors from the labeled sample text in Step 2. Utilize the pre-trained Chinese model provided by Google to convert text to word vectors for subsequent model training, better extracting contextual information and improving the accuracy of subsequent label classification. Using the bidirectional encoder representation in BERT can effectively achieve word segmentation and word vectorization. The BERT model itself contains implicit information about the Chinese text while processing the input text, and also obtains contextual information from the input text.

[0045] Step 4: Process, classify, and label word vectors using CRF (Conditional Random Field) technology: The word vectors obtained in Step 3 are serialized and labeled using CRF. Using CRF to process them can better utilize grammatical and other rule information. After the word vectors output by the BERT model are serialized and labeled by the CRF layer, each word can be identified as a target entity with its corresponding label. The label categories include object (OBJ), component (COM), attribute (ATT), and attribute value (VAL).

[0046] Step 5: Use the training samples obtained in Step 2 to train the model (BERT+CRF) from Steps 3 and 4 to achieve a high accuracy rate. This model will then be used for actual annotation and entity recognition of unannotated text. The training process is deployed in a cloud GPU or local GPU computing environment: By writing Python programs and corresponding execution scripts, the aforementioned sample database is accessed, and the model is trained in a Linux environment. After 3 to 6 rounds of fine-tuning, the BERT+CRF model can achieve a good recognition accuracy.

[0047] Step Six: Extract valid entities from the text using the trained model and organize the model according to predefined rules. Specifically, use the model from Step Five to extract entity text from the labels and organize the entities according to predefined label relationships. Finally, construct an object-oriented OPC UA information model for the relevant entity labels.

[0048] After obtaining the tags, they can be mapped to an abstract model based on their organizational relationships. The abstract model itself does not contain any specification information; its structure only contains an object-oriented hierarchical structure. Specifically: words labeled OBJ are considered objects described by the model. They are located at the top level of the abstract model and are the main body of the model. When multiple words labeled OBJ exist in a passage describing the same device, only one is used as the model description object. Words labeled COM are components of the model. In the abstract model, these words are located around and connected to the main body of the model—the model description object—indicating a containment and belonging relationship between them. Words labeled ATT are descriptions of the model, reflecting its performance. Similar to words labeled COM, these words also exist around and are connected to the description object, but the connection relationship is an attribute relationship, not a containment and belonging relationship. Finally, there are words labeled VAL. These words are the specific numerical values ​​used to describe the model. Therefore, in the abstract model, they appear around and are connected to words labeled ATT.

[0049] In the OPC UA information model, nodes are actually described in XML format, and their file writing method is subject to strict definition requirements. Therefore, it is necessary to transform the abstract model with a defined structural framework into a model file that meets the definition requirements through a certain corresponding method, so as to realize the transformation from object model to node model.

[0050] The correspondence between the abstract model and the OPC UA information model is as follows: The UA ObjectType element represents the object type node, corresponding to the main body and components of the abstract model. The OPC UA node reference information of the latter also contains its connection relationship with the main body of the model; UA Variable represents the variable node, corresponding to the model description in the abstract model; the specific value of the UA Variable node is determined by the specific value when describing the model.

[0051] Following the above methods, an OPC UA information model in XML format can be constructed.

[0052] This invention, on the one hand, fully leverages the advantages of language processing neural network models for text processing. While making full use of unstructured industrial data, it utilizes entity extraction methods to extract desired information from complex text and reorganizes this information into a structured model. This allows the proposed method to better utilize the abundant unstructured data in industrial settings and obtain more information compared to traditional methods. On the other hand, this invention utilizes a self-built small-sample corpus database in an industrial environment. Based on the characteristics of Chinese text, it uses a pre-trained model to achieve word segmentation and accurate word vectorization of input text with small samples and low training overhead. This avoids problems such as missing word vector information and inaccurate serialization annotation caused by insufficient samples.

[0053] The effects of the method proposed in this invention will be further described below with reference to simulation examples.

[0054] Simulation Example 1:

[0055] This section will use the text content of a robotic arm device as an example to illustrate the annotation and subsequent processing. Figure 2 The image shows the annotation of robotic arm device description information extracted from publicly available equipment information online. Users can easily transfer the method of this invention to other industrial text data. Figure 2 It is evident that unstructured text data in industrial scenarios contains a rather complex amount of information, requiring detailed cleaning and annotation. The final annotated entities include four types of tags: objects (OBJ), components (COM), attributes (ATT), attribute values ​​(VAL), and no-information tags (O). The former is further subdivided into B and I.

[0056] Building an information model also requires establishing a text processing model, namely the BERT+CRF model. Figure 3 The diagram shows the neural network information extraction model used in this invention. The model structure, from bottom to top, is as follows:

[0057] Information input and word segmentation, where Chinese word segmentation is done directly by character;

[0058] The BERT model, which is constructed with bidirectional transducers as cells, is divided into encoder and decoder parts, both of which are composed of bidirectional transducers.

[0059] The CRF sequence labeling model processes the word vectors output by BERT, and according to the conditional probability distribution, it processes the word vectors to achieve the conversion from vectors to labels.

[0060] Model training is conducted in the cloud environment provided by Google. First, a script is written to define the training rounds and the location where the output neural network model parameters are stored. The script is also defined to call the model's Python file to perform calculations. Then, the script is executed on the cloud in a Linux environment to perform calculations. Finally, the node files with better training results are selected as the final neural network model.

[0061] Inputting a new descriptive text about a robotic arm in an industrial scenario into a neural network model, the model processes and annotates the text, resulting in annotations categorized into four types: Object (OBJ), Component (COM), Attribute (ATT), and Attribute Value (VAL). These annotations form the target vocabulary used for modeling. The overall style of the annotated input text is shown below. Figure 2 The sample text data is similar.

[0062] Based on the defined node relationships, the relevant labeled text is organized, with the object being the word labeled as an object. Other nodes are organized similarly, ultimately completing the construction of the OPC UA information model. An example of the constructed OPC UA information model is shown below. Figure 4 As shown.

[0063] The above description is merely a preferred embodiment of the present invention. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the technical solutions of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall still fall within the protection scope of the technical solutions of the present invention.

Claims

1. A method for constructing an OPC UA information model based on unstructured text data, characterized in that, Includes the following steps: (1) Design annotation labels based on corpus text features that meet the needs of industrial site modeling, and design the organization relationship of the labels, that is, use different levels of description labels, with the highest level being the model description object, and below that being the object's attributes and attribute values. (2) The sample text materials extracted from the device description document are BIO serialized and labeled based on description tags at different levels to obtain training samples; after labeling, the original text is enhanced: the labeled text is translated by back-translating the text containing target information. According to the distribution of O tags, the back-translated text is inserted into the appropriate position in the original text. The insertion method is as follows: by calculating the frequency of O tags in a certain paragraph, if the frequency value exceeds the set threshold, the back-translated text to be inserted is inserted into the current paragraph position; after all paragraphs are traversed, if there is still back-translated text remaining, the remaining part is inserted into the end of the entire sample. (3) Construct a Chinese pre-trained model based on the BERT model to extract word vectors from the sample text annotated in step (2); (4) Based on CRF technology, process, classify and label the word vectors extracted in step (3); (5) Use the training samples obtained in step (2) to train the text processing model that combines BERT and CRF; (6) Use the trained text processing model to extract valid entities from the text and organize the model according to the predefined organization rules. That is, use the text processing model to extract entity text from the labels and organize the entities according to the predefined label relationships. Finally, construct an object-oriented OPC UA information model for the relevant entity labels.

2. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, In step (1), the information model node of the object-oriented OPC UA is abstracted into entity labels, and the entity label relationship is defined according to the original organizational structure of the information model node. Subsequently, the model can be constructed according to the predefined relationship.

3. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, In step (1), for the information model of OPC UA, labels are designed to include three levels: objects, object attributes and characteristics, and attribute values. Specifically, these include four types of labels: model description object OBJ, component COM, attribute ATT, and attribute value VAL.

4. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, The BIO annotation sample data described in step (2) for industrial data annotation is as follows: ① Use tags that meet the modeling requirements of the industrial field to perform BIO annotation on unstructured text data in the industrial field; ② The annotation itself takes into account the label organization relationship of the information model and defines the label relationship that conforms to the structural requirements of the OPC UA information model. The model labels are designed according to three layers: object, attribute features and related attribute values, which can directly correspond to the objects, nodes and variables of the OPC UA information model.

5. The method for constructing an OPC UA information model based on unstructured text data according to claim 3, characterized in that, In step (2), the text material is labeled with four types of tags: the model description object OBJ, the component COM, the attribute ATT, and the attribute value VAL. Redundant words that are not part of the target information are classified as O tags, the first word of the target information is labeled as B, and the subsequent words are labeled as I.

6. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, In steps (3) and (4), the labeling and entity extraction methods are applied to the training process with a small number of samples: ①The BERT model is a pre-trained model. During the pre-training process, it utilizes large-scale computing resources to extract implicit information from Chinese text. In this study, a fine-tuning method is used during the training process to achieve model transfer through a small number of training rounds, which meets the needs of industrial sites and actual operation. ②The BERT model can make good use of the contextual information of the text. By using the masking method, it can mask some words and predict the masked words, thus obtaining a model containing rich contextual information. ③ Use CRF to process the word vector information output by the BERT model, output serialized annotation results, and optimize the entire sequence, which meets the overall performance requirements of industrial scenarios.

7. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, In step (6), after obtaining the trained model, input the actual text that appears in the industrial production environment to be processed, and then process it accordingly to obtain the keywords in the text, i.e. entities; on this basis, use the predefined relationships between entities to organize the model structure, and finally complete the construction of the OPC UA information model.

8. The method for constructing an OPC UA information model based on unstructured text data according to claim 1, characterized in that, In step (6), after obtaining the tags, they are mapped to an abstract model according to their organizational relationship. The abstract model itself does not contain any specification information. Its structure only contains an object-oriented hierarchical structure. Specifically: words labeled OBJ are considered to be the objects described by the model. They are located at the top level of the abstract model and are the main body of the model. When there are multiple words labeled OBJ in the segment describing the same device, only one of them is used as the model description object. Words labeled COM are the components of the model. In the abstract model, these words are located around the model body - the model description object and are connected to it, indicating that there is a relationship of inclusion and belonging between the two. Words labeled ATT are the description of the model, reflecting the performance of the model. These words also exist around the description object and are connected to it, but the connection relationship is an attribute relationship. Finally, there are words labeled VAL. These words are the specific values ​​when describing the model. In the abstract model, they appear around the words labeled ATT and are connected to them.

9. The method for constructing an OPC UA information model based on unstructured text data according to claim 8, characterized in that, In step (6), the nodes in the OPC UA information model are actually described in XML format. The way the file is written is subject to strict definition requirements. It is necessary to transform the abstract model with a defined structural framework into a model file that meets the definition requirements through a certain corresponding method, so as to realize the transformation from object model to node model. The corresponding method between the abstract model and the OPC UA information model is as follows: the UA ObjectType element represents the object type node, which corresponds to the main body and the components of the abstract model. The OPC UA node reference information of the latter also contains its connection relationship with the model body; UAVariable represents the variable node, which corresponds to the model description in the abstract model; the specific value of the UA Variable node is determined by the specific value when describing the model; according to the above corresponding method, the OPC UA information model in XML format is constructed.

Images