Knowledge distillation and semantic analysis combined equipment maintenance knowledge representation method

Abstract

The invention discloses an equipment maintenance knowledge representation method combining knowledge distillation and semantic analysis. The method comprises the steps of collecting structured data, unstructured text data and multi-modal data, constructing a semantic knowledge graph by using a natural language processing technology, extracting visual knowledge by using a computer vision technology, and fusing the visual knowledge with semantic knowledge to form a multi-modal knowledge base. And performing interactive knowledge distillation on the text data by using a large language model, generating optimized knowledge key points, and fusing the optimized knowledge key points with the multi-modal knowledge base to generate a comprehensive knowledge base. Knowledge is hierarchically organized according to the domain ontology and stored in a graph database, a natural language query interface is constructed, intelligent answering is performed by integrating technologies, question and answer corpus training is automatically generated, and automatic question and answer services are provided for users. According to the method, heterogeneous data integration, multi-modal knowledge fusion, hierarchical knowledge organization, natural language interaction and man-machine collaborative optimization are realized, and an intelligent equipment maintenance knowledge base can be efficiently constructed.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an equipment maintenance knowledge representation method combining knowledge distillation with semantic analysis. Background Art

[0002] Equipment maintenance is an important part of ensuring the normal and safe operation of industrial equipment and related systems. The equipment maintenance knowledge base refers to a knowledge system that collects and organizes relevant knowledge such as equipment types, failure modes, and maintenance steps. This knowledge base provides maintenance personnel with timely and accurate maintenance knowledge support, which is of great significance for improving maintenance efficiency, reducing operation and maintenance costs, and ensuring the safe operation of equipment.

[0003] The existing equipment maintenance knowledge base mainly obtains knowledge through manual compilation, such as maintenance manuals, operating guides, etc. This method has the disadvantages of a single source of knowledge and low maintenance efficiency. Other methods attempt to automatically extract knowledge from unstructured texts such as maintenance reports and forums, but they are limited to text data and lack support for structured data and multimodal data. In addition, the existing knowledge bases mostly adopt a flat knowledge organization form, which cannot reflect the complex domain concept relationship, and the knowledge query and utilization efficiency is low. In general, it is difficult for traditional methods to efficiently build a comprehensive and intelligent equipment maintenance knowledge base.

[0004] In view of the problems in the existing technology such as single source of knowledge acquisition, single form of knowledge representation, inefficient knowledge base construction, lack of interactivity and explainability in knowledge utilization, there is an urgent need for an intelligent knowledge representation method that can integrate heterogeneous data sources, comprehensive semantic analysis and knowledge distillation technology, support multimodal knowledge fusion, realize hierarchical knowledge organization and natural language interaction, so as to efficiently build an explainable and easy-to-use equipment maintenance knowledge base. Summary of the invention

[0005] In order to solve the problems of the prior art, an embodiment of the present invention provides an equipment maintenance knowledge representation method that combines knowledge distillation with semantic parsing. The technical solution is as follows:

[0006] On the one hand, a method for representing equipment maintenance knowledge by combining knowledge distillation with semantic parsing is provided, comprising the following steps:

[0007] Collect structured data and unstructured data from multiple channels, the structured data includes tables and lists, the unstructured data includes natural language text, and multimodal data such as device images and videos;

[0008] Preprocess the collected data;

[0009] Use natural language processing technology to perform semantic analysis on text data, extract entities and relationships, and build an ontology for the field of equipment maintenance;

[0010] Store semantic knowledge in the graph database according to the ontology model to form a semantic knowledge graph;

[0011] Use computer vision technology to analyze image and video data, identify visual information such as equipment components and fault manifestations, and integrate it with semantic knowledge graphs to build a multimodal knowledge base;

[0012] Use large language models to perform knowledge distillation on text data and generate a set of key knowledge points;

[0013] Domain experts review key knowledge points and provide feedback to the large language model to guide the next round of distillation, forming an iterative optimization closed loop;

[0014] Integrate optimized knowledge points with semantic knowledge graph to generate a comprehensive knowledge base;

[0015] According to the domain ontology, the comprehensive knowledge is organized by dimension, stored in the graph database in the form of attribute graph, and a multi-level index is constructed;

[0016] Build a natural language query interface to respond to user queries, integrate multiple technologies to output relevant knowledge points, generate natural language to present conversational answers, and automatically generate question-answer pairs to expand the corpus;

[0017] The retrieval-based question-answering system is trained using question-answering corpus to provide automatic question-answering services, which are combined with interactive query results to give comprehensive answers.

[0018] Furthermore, after preprocessing the collected data, the unstructured text data is subjected to word segmentation, part-of-speech tagging, named entity recognition and relationship extraction.

[0019] Furthermore, the visual information includes equipment component identification, fault type identification and fault location identification.

[0020] Furthermore, the semantic parsing utilizes a deep learning model to implement named entity recognition and relationship extraction.

[0021] Furthermore, the domain ontology defines concepts such as equipment type, component structure, failure mode, maintenance steps, and their relationships.

[0022] Furthermore, the knowledge in the comprehensive knowledge base is organized according to dimensions such as equipment type, failure mode, and maintenance parts to form a hierarchical tree structure.

[0023] Furthermore, the natural language query interface supports semantic and context-based queries, and returns multimodal content such as query-related knowledge points, images, and videos.

[0024] Furthermore, when using question-answering corpus to train the question-answering system, transfer learning is used to fine-tune the question-answering model to improve the model quality.

[0025] Furthermore, the natural language generation is based on a generative language model, combined with templates and rules.

[0026] Furthermore, computer vision techniques such as object detection and instance segmentation are used to construct visual knowledge.

[0027] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0028] The present invention realizes intelligent integration of heterogeneous data sources. It can integrate multiple data sources such as structured data, unstructured text data, images and videos, and automatically extract knowledge through technologies such as semantic analysis, visual understanding and knowledge distillation, which greatly improves the efficiency of knowledge acquisition and expands the scope of knowledge sources.

[0029] Constructing a multimodal and multidimensional knowledge base The present invention organically integrates heterogeneous knowledge such as semantic knowledge, visual knowledge, and distilled knowledge to form a multimodal knowledge base. At the same time, knowledge is hierarchically organized based on domain ontology, presenting a rich multidimensional knowledge structure, which is conducive to the comprehensive description of maintenance domain knowledge.

[0030] Improved the intelligence of knowledge representation and query The present invention uses a graph database to store knowledge and supports semantic-based relationship query and reasoning. Through natural language interaction, users can naturally and efficiently obtain the required knowledge, which improves the convenience of knowledge utilization.

[0031] Enhanced knowledge interpretability The present invention combines semantic interpretation with knowledge key point interpretation, provides rich interpretative support for query results, makes knowledge more transparent and interpretable, and improves users' understanding and trust in maintenance knowledge.

[0032] The present invention realizes closed-loop optimization of human-machine collaboration. The interactive knowledge distillation and active learning mechanism are introduced. Human feedback will continuously guide the iterative learning of the model and continuously optimize the quality of knowledge, thus truly realizing the closed-loop of human-machine collaboration and knowledge evolution.

[0033] The technical solution proposed in the present invention has good versatility and can be quickly migrated to the construction of knowledge bases in other fields, such as medical health, legal precedents, etc., by adjusting the ontology model and data source, showing broad application prospects.

[0034] In summary, the present invention combines a variety of artificial intelligence technologies, breaks through the bottleneck of traditional knowledge base construction methods, effectively improves the intelligence level, knowledge expression ability and practicality of the knowledge base, and can provide high-quality knowledge service support for equipment maintenance and more fields. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0036] Figure 1 It is a schematic diagram of an equipment maintenance knowledge representation method combining knowledge distillation and semantic analysis in an embodiment of the present invention. DETAILED DESCRIPTION

[0037] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0038] The present invention relates to an equipment maintenance knowledge representation method that combines knowledge distillation with semantic analysis, aiming to solve the problems in the prior art such as single source of knowledge acquisition, weak knowledge representation form, low efficiency of knowledge base construction, and lack of interactivity and interpretability in knowledge utilization.

[0039] To achieve the above objectives, the present invention proposes a method that comprehensively utilizes technologies such as knowledge distillation, semantic parsing, and visual understanding to achieve intelligent fusion of heterogeneous data, multi-level knowledge organization, and natural language interaction.

[0040] The method flow is as follows, see Figure 1 :

[0041] 1. Multi-source heterogeneous data collection

[0042] Collect structured data (tables, lists) and unstructured data (natural language text) from channels such as equipment maintenance manuals, service records, and forum discussions, as well as multimodal data such as equipment images and videos. Perform pre-processing such as cleaning, deduplication, and standardization on the collected data to prepare for subsequent processing.

[0043] 2. Semantic Knowledge Graph Construction

[0044] Use natural language processing technology to perform semantic analysis on text data, including word segmentation, part-of-speech tagging, named entity recognition, and relationship extraction. Based on the extracted entities and relationships, build an ontology for the field of equipment maintenance. Store the extracted semantic knowledge in a graph database (such as Neo4j) according to the ontology model to form a structured semantic knowledge graph.

[0045] 3. Visual Knowledge Extraction

[0046] Use computer vision technologies such as target detection and instance segmentation to analyze equipment image / video data, identify visual information such as equipment components, fault manifestations (such as cracks, wear), working environment, etc., and integrate it with semantic knowledge graphs to build a multimodal knowledge base.

[0047] 4. Interactive Knowledge Distillation

[0048] A large language model is used to perform knowledge distillation on text data to generate a preliminary set of key knowledge points. Domain experts review the key knowledge points and perform operations such as supplementation, correction, and annotation. The manual review results are fed back to the language model to guide the next round of distillation, forming an iterative optimization closed loop. Finally, high-quality key knowledge points are integrated with the semantic knowledge graph to generate a comprehensive knowledge base.

[0049] 5. Knowledge organization and storage

[0050] According to the domain ontology, the comprehensive knowledge is organized by dimensions such as equipment type, failure mode, and maintenance steps, and stored in the graph database in the form of attribute graphs. Multi-level indexes are built in the knowledge base to support efficient retrieval.

[0051] 6. Natural language query and interactive question and answer

[0052] Build a natural language query interface based on a comprehensive knowledge base, where users can ask questions through text / voice. Integrate semantic parsing, relational reasoning, knowledge distillation and other technologies during querying to output knowledge points related to the query. Generate natural language for the output results, integrate semantic interpretation and knowledge point interpretation, and provide conversational interactive answers. At the same time, automatically generate question-answer pairs based on queries and feedback, and continuously expand the question-answer corpus.

[0053] 7. Question and answer corpus training and services

[0054] Using the automatically generated question and answer corpus, the retrieval-based question and answer system is trained by transfer learning to continuously improve its quality. The question and answer system can provide users with automatic question and answer services and answer common equipment maintenance questions. For complex questions, the automatic question and answer results are combined with the interactive query results to give a comprehensive answer.

[0055] The present invention integrates a variety of artificial intelligence technologies, realizes the intelligent knowledge extraction and fusion of heterogeneous data sources, organizes knowledge through graph databases, supports flexible query and relational reasoning, and provides natural language interaction and explanatory services, which can provide efficient, intelligent and explainable knowledge support for the field of equipment maintenance.

[0056] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

[0057] Based on the above specific implementations, the present invention may also adopt the following other implementations or variations without departing from the overall technical solution and innovative gist:

[0058] 1. Heterogeneous Data Source Expansion In addition to the data sources listed above, the present invention can also integrate other heterogeneous data sources, such as:

[0059] Equipment operation log data, sensor data and other machine data

[0060] Equipment 3D model data, simulation data and other engineering data

[0061] Multimodal data such as maintenance personnel's voice conversation records can comprehensively capture information about the entire life cycle of equipment by integrating more heterogeneous data sources, thereby building a more complete and multi-dimensional knowledge base.

[0062] 2. Rich knowledge representation forms In addition to using semantic knowledge graphs and attribute graphs to store and organize knowledge, the present invention can also use other knowledge representation methods, such as:

[0063] Store structured knowledge as a relational or NoSQL database

[0064] Representing unstructured knowledge as vector embeddings

[0065] Using knowledge base templates or ontology languages ​​to describe knowledge and combining multiple knowledge representation forms can efficiently store, retrieve and reason about knowledge according to different scenarios and needs.

[0066] 3. Multimodal knowledge fusion enhancement In the visual knowledge extraction step, in addition to analyzing device images and videos, the following technologies can also be integrated:

[0067] Speech recognition technology extracts equipment and fault information from speech data

[0068] Machine learning technology mines patterns in equipment operation logs, sensor data, etc.

[0069] Computer graphics technology can analyze the three-dimensional model of the equipment and build a richer and more comprehensive multimodal knowledge base by integrating information from more modalities.

[0070] 4. Human-computer interaction optimization In the process of knowledge distillation and query interaction, in addition to the above human-computer interaction methods, the following enhancement mechanisms can also be introduced:

[0071] Based on dialogue management technology, guide users to express their needs more efficiently

[0072] Through active learning algorithms, the system actively requests human feedback and annotations

[0073] By using explainable AI technology and enhancing the explainability of system output, knowledge quality and system usability can be improved by optimizing human-computer interaction.

[0074] The above-mentioned variations and application extensions are intended to further enrich and optimize the technical solutions of the present invention and enhance its versatility and practicality, but do not deviate from the innovative spirit and core technical scope of the present invention.

Claims

1. A method for representing equipment maintenance knowledge by combining knowledge distillation with semantic analysis, comprising the following steps: Collect structured data and unstructured data from multiple channels, the structured data includes tables and lists, the unstructured data includes natural language text, and collect multimodal data such as device images and videos; Preprocess the collected data; Use natural language processing technology to perform semantic analysis on text data, extract entities and relationships, and build an ontology for the field of equipment maintenance; Store semantic knowledge in the graph database according to the ontology model to form a semantic knowledge graph; Use computer vision technology to analyze image and video data, identify visual information of equipment components and fault manifestations, and integrate it with semantic knowledge graphs to build a multimodal knowledge base; Use large language models to perform knowledge distillation on text data and generate a set of key knowledge points; Domain experts review key knowledge points and provide feedback to the large language model to guide the next round of distillation, forming an iterative optimization closed loop; Integrate optimized knowledge points with semantic knowledge graph to generate a comprehensive knowledge base; According to the domain ontology, the comprehensive knowledge is organized by dimension, stored in the graph database in the form of attribute graph, and a multi-level index is constructed; Build a natural language query interface to respond to user queries, integrate multiple technologies to output relevant knowledge points, generate natural language to present conversational answers, and automatically generate question-answer pairs to expand the corpus; The retrieval-based question-answering system is trained using question-answering corpus to provide automatic question-answering services, which are combined with interactive query results to give comprehensive answers.

2. The method according to claim 1, characterized in that After preprocessing the collected data, the unstructured text data is subjected to word segmentation, part-of-speech tagging, named entity recognition, and relation extraction.

3. The method according to claim 1, characterized in that The visual information includes equipment component identification, fault type identification and fault location identification.

4. The method according to claim 1, characterized in that The semantic parsing utilizes a deep learning model to implement named entity recognition and relationship extraction.

5. The method according to claim 1, characterized in that The domain ontology defines the concepts of equipment type, component structure, failure mode, maintenance steps and their relationships.

6. The method according to claim 1, characterized in that The knowledge in the comprehensive knowledge base is organized according to the dimensions of equipment type, failure mode, and maintenance component to form a hierarchical tree structure.

7. The method according to claim 1, characterized in that The natural language query interface supports semantic and context-based queries, and returns multimodal content such as query-related knowledge points, images, and videos.

8. The method according to claim 1, characterized in that: When using question-answering corpus to train the question-answering system, transfer learning is used to fine-tune the question-answering model to improve model quality.

9. The method according to claim 1, characterized in that: The natural language generation is based on a generative language model and is implemented in combination with templates and rules.

10. The method according to claim 1, characterized in that Computer vision techniques such as object detection and instance segmentation are used to build visual knowledge.

Images