Equipment maintenance corpus intelligent processing and knowledge extraction method

Abstract

The invention discloses an equipment maintenance corpus intelligent processing and knowledge extraction method, which comprises the following steps: firstly, preprocessing an original corpus, and constructing a domain corpus; then based on a comparative learning framework, fusing field semantic knowledge to perform semantic enhancement representation learning on the corpora; secondly, an innovative sequence-to-sequence generation neural network model is provided, and parallel high-quality extraction of multi-granularity knowledge of a text level, a sentence segmentation level, an entity level and the like is achieved through special mark guidance; then, based on a cognitive map technology, the extracted multi-source heterogeneous knowledge is fused to construct an interpretable map embedded knowledge base, and flexible knowledge reasoning service is supported; and finally, based on the constructed knowledge base, developing intelligent applications such as an equipment fault diagnosis system, personalized maintenance scheme generation and preventive maintenance decision. Compared with the prior art, breakthrough progress is achieved, multiple artificial intelligence technology innovations are fused, and the method has great contribution in theoretical significance and application value.

Description

Technical Field

[0001] The present invention relates to the fields of natural language processing and knowledge engineering, and in particular to a method for intelligent processing and knowledge extraction of equipment maintenance corpus. Background Art

[0002] Equipment maintenance is a common and critical task in the manufacturing industry and operation and maintenance services. Equipment maintenance personnel need to accurately diagnose the cause of the failure based on equipment failure information and maintenance knowledge, and take correct maintenance steps and replace parts to ensure the normal operation of the equipment. Traditional equipment maintenance often relies on manual experience accumulation, which is inefficient and difficult to ensure maintenance quality. With the increasing complexity of equipment, it is urgent to develop intelligent corpus processing and knowledge management technologies to assist maintenance personnel in efficiently completing diagnosis and maintenance tasks.

[0003] In recent years, some scholars and enterprises have tried to apply artificial intelligence technologies such as natural language processing and machine learning to the field of equipment maintenance. Existing technologies mainly include rule-based information extraction, supervised learning sequence labeling models, etc. However, most of these methods are limited to the extraction of knowledge elements of a single granularity (such as entity or relationship extraction), and it is difficult to obtain multi-granularity knowledge information at the same time. In addition, the existing knowledge base construction methods lack a deep understanding of the semantics of the domain, and are mainly based on statistical rules or symbolic methods, which cannot support flexible knowledge reasoning and interpretation.

[0004] It can be seen that the existing technology still has some significant deficiencies when processing equipment maintenance corpus and building knowledge base, mainly including: limited domain semantic understanding ability, single knowledge extraction granularity, lack of interpretable knowledge base, and relatively single intelligent application scenario. In view of these problems, it is urgent to propose an innovative equipment maintenance corpus intelligent processing and knowledge extraction method to better support decision analysis and intelligent application of equipment maintenance. Summary of the invention

[0005] In order to solve the problems of the prior art, an embodiment of the present invention provides a method for intelligent processing and knowledge extraction of equipment maintenance corpus. The technical solution is as follows:

[0006] On the one hand, a method for intelligent processing and knowledge extraction of equipment maintenance corpus is provided, comprising the following steps:

[0007] (1) Preprocessing the original equipment maintenance corpus, performing standardization processing such as word segmentation and part-of-speech tagging on the corpus, and constructing an equipment maintenance domain corpus, which contains terminology and semantic knowledge in the equipment maintenance domain;

[0008] (2) learning semantic representation from the preprocessed corpus, and fusing domain terminology vocabulary and semantic knowledge in the domain corpus based on a contrastive learning framework to generate semantic representation with domain semantic enhancement;

[0009] (3) inputting the semantic representation into a neural network model based on sequence-to-sequence generation, and simultaneously generating and extracting equipment maintenance knowledge information of multiple granularities from the semantic representation through predefined special tags, wherein the multiple granularities include text-level knowledge, sentence-level knowledge, and entity-level knowledge;

[0010] (4) Based on the extracted multi-granularity knowledge information, a cognitive graph knowledge base is constructed, and the cognitive graph knowledge base supports knowledge reasoning and decision support services based on graph embedding.

[0011] Furthermore, the contrastive learning framework in step (2) includes: constructing positive and negative sample pairs, wherein the positive sample pair is a semantically similar corpus pair, and the negative sample pair is a semantically dissimilar corpus pair; mapping the terms in the domain corpus into predefined semantic vectors, and mapping the concepts into adjustable semantic vectors; and minimizing the loss function of the sum of the similarities of the semantic representations of the positive and negative sample pairs.

[0012] Furthermore, the neural network model based on sequence-to-sequence generation adopts a Transformer encoder-decoder architecture and fuses contextual information through an attention mechanism during the decoding process.

[0013] Furthermore, the decoder is a pointer generation network that can copy words from the source corpus or generate new words.

[0014] Furthermore, the multi-granularity knowledge extraction is guided by special marking prefixes in the decoding process, and different prefixes correspond to knowledge elements of different granularities.

[0015] Furthermore, a gradient decay strategy is adopted to control the generation quality and order of knowledge elements of different granularities by weighting the prediction probabilities of corresponding tags.

[0016] Furthermore, the method for constructing the cognitive graph knowledge base in step (4) includes: extracting triples from a semi-structured data source to construct an initial cognitive graph; training a graph embedding model of nodes and relationships based on node neighborhood structure and node content information; designing a latent variable model to integrate entity features, context information and graph structure information to generate a graph embedding representation that integrates multi-source heterogeneous knowledge.

[0017] Furthermore, in the cognitive graph knowledge base, similar nodes or paths are calculated through similarity measurement, and an explanation model based on the attention mechanism is developed to support explainable knowledge reasoning services.

[0018] Furthermore, the decision support service includes: interactive fault diagnosis and cause inference based on the question-answering system and the constructed knowledge base; generating personalized maintenance plans for equipment in combination with the extracted maintenance steps and experience; evaluating possible future failure risks and recommending preventive maintenance measures.

[0019] Furthermore, the method is applied to intelligent corpus processing, knowledge base construction and related maintenance decision support services in the field of equipment maintenance.

[0020] The technical solution provided by the embodiment of the present invention has the following beneficial effects:

[0021] 1. Improved the ability to understand the semantics in the field of equipment maintenance. Through the continued pre-training of the pre-trained language model and the semantic enhancement strategy based on contrastive learning, the domain-specific terminology and semantic knowledge are fully utilized, which significantly improves the ability to understand the semantics in the field of equipment maintenance and lays the foundation for subsequent knowledge extraction and application.

[0022] 2. High-quality multi-granularity knowledge parallel extraction is achieved. The neural network model based on sequence-to-sequence generation proposed in this invention can efficiently extract knowledge information of different granularities such as text level, sentence level and entity level at the same time through labeling guidance and gradient decay mechanism, meeting the diverse needs of equipment maintenance decision-making.

[0023] 3. Constructed an interpretable cognitive graph knowledge base. Based on cognitive graph construction and graph embedding technology, combined with multi-source heterogeneous knowledge association, constructed a cognitive graph knowledge base that can support interpretable knowledge reasoning and decision-making services, which significantly improved the intelligence level of the knowledge base.

[0024] 4. We have expanded the cognitive graph knowledge base based on which intelligent application scenarios for equipment maintenance are built. By combining the results of multi-granular knowledge extraction, we have developed intelligent applications such as interactive fault diagnosis systems, personalized maintenance plan generation, and preventive maintenance decision-making, which have greatly expanded the intelligent service capabilities in the field of equipment maintenance.

[0025] 5. Both theoretical innovation and application value This invention integrates a number of cutting-edge artificial intelligence technologies such as natural language processing, machine learning, and knowledge graphs, and has innovations in many aspects such as semantic enhancement, knowledge extraction, and knowledge base construction, and has important theoretical significance. At the same time, it is closely oriented to the actual needs of the equipment maintenance field, can significantly improve the level of intelligent equipment maintenance, and has broad application prospects.

[0026] In general, the present invention has made a significant breakthrough compared with the prior art, and has outstanding performance in both theoretical innovation and practical application value. It is an important technological advancement in the field of equipment repair and intelligent maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the accompanying drawings required for the description of the embodiments. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0028] Figure 1 It is a schematic flowchart of a method for intelligent processing and knowledge extraction of equipment maintenance corpus in an embodiment of the present invention. Specific embodiments

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the following will further describe the embodiments of the present invention in detail with reference to the accompanying drawings.

[0030] The invention discloses a method for intelligent processing and knowledge extraction of equipment maintenance corpus, aiming to solve problems such as insufficient domain semantic understanding ability, single knowledge extraction granularity, lack of interpretable knowledge base, and limited intelligent application scenarios in the prior art. The following will elaborate on the specific embodiments of this method. See Figure 1 :

[0031] 1. Corpus preprocessing and construction of domain corpus

[0032] First, perform standardized preprocessing on the original equipment maintenance corpus, including steps such as word segmentation, part-of-speech tagging, and stop-word removal, to prepare high-quality corpus input.

[0033] At the same time, construct an equipment maintenance domain corpus, collect and organize the proprietary term vocabulary, concept phrases, and semantic knowledge in this field. This domain corpus provides domain knowledge support for subsequent semantic enhancement and knowledge extraction.

[0034] Specifically, the goal of this step of corpus preprocessing and construction of domain corpus is to obtain high-quality corpus input and construct a domain knowledge base to provide support for subsequent enhanced semantic representation and knowledge extraction.

[0035] 1.1 Corpus preprocessing

[0036] Perform the following standard preprocessing steps on the original equipment maintenance corpus:

[0037] Word segmentation: Split the text sequence into a sequence of word tokens, and rule-based or statistical word segmentation tools can be used;

[0038] Part-of-speech tagging: Assign part-of-speech tags (such as nouns, verbs, etc.) to each word token, and hidden Markov models or neural network sequence tagging models can be used;

[0039] Stop-word removal: Remove high-frequency function words in the corpus, such as "de", "le", etc., using a predefined stop-word list;

[0040] Case normalization: convert all text to the same case;

[0041] Special character filtering: remove special symbols, numbers, URLs and other meaningless characters in the corpus.

[0042] 1.2 Domain Corpus Construction

[0043] Terminology collection: Automatically or manually collect relevant terms, phrases, abbreviations and other proprietary terms from equipment maintenance manuals, terminology dictionaries and other field resources;

[0044] Semantic knowledge annotation: annotate the collected terminology with semantic types (such as fault causes, repair parts, etc.), and use remote supervision methods such as seed entity expansion;

[0045] Knowledge base construction: Build a structured ontology knowledge base based on annotated terminology vocabulary, or unstructured vector storage.

[0046] 2. Semantic Representation Enhancement Based on Contrastive Learning

[0047] Based on pre-trained language models (such as BERT), the model is further pre-trained (Fine-tuning) on ​​a large-scale corpus to further improve the model's ability to understand the semantics in the field of equipment maintenance.

[0048] A contrastive learning framework is designed to learn domain semantic representation in a self-supervised manner by constructing semantically similar / dissimilar positive and negative sample pairs and optimizing the loss function that minimizes the sum of the similarities of the semantic representations of the positive and negative sample pairs.

[0049] The terms in the domain corpus are mapped to fixed semantic vectors, and the domain concepts are mapped to adjustable semantic vectors. The domain semantic vectors are fused by attention weighting to enhance the semantic representation of the text.

[0050] Specifically, the goal of this step of semantic representation enhancement based on contrastive learning is to learn contextual representations with domain semantic enhancement, providing high-quality semantic representation input for subsequent multi-granularity knowledge extraction.

[0051] 2.1 Contrastive Estimation Based on Contrastive Learning

[0052] Positive sample construction: traverse the corpus and construct two sentence pairs whose sequence similarity scores exceed the preset threshold as positive sample pairs;

[0053] Negative sample construction: For each positive sample pair, randomly sample other sentences to construct negative sample pairs;

[0054] Based on the contrastive learning framework, the semantic representation similarity of positive samples is maximized, while the semantic representation similarity of negative samples is minimized.

[0055] 2.2 Domain Semantic Knowledge Fusion

[0056] Term mapping: Mapping the terms in the domain corpus into semantic vector representations of fixed dimensions;

[0057] Concept mapping: Mapping more abstract domain concepts into adjustable semantic vector representations;

[0058] Attention Weighting: The term / concept semantic vector is fused with the contextual semantic representation through a multi-head attention mechanism.

[0059] 3. Multi-granularity knowledge extraction based on sequence-to-sequence generation

[0060] An innovative sequence-to-sequence (Seq2Seq) neural network model is proposed, which models knowledge extraction as a sequence-to-sequence generation task, effectively reducing the cost of manual feature engineering.

[0061] The encoder-decoder architecture of Transformer is adopted. The encoder uses a multi-head attention mechanism to obtain contextual semantic information, and the decoder uses a pointer generation network PointerGenerator, which can copy words from the source text or generate new words based on the context.

[0062] During the decoding process, the model is guided by pre-designed special tag prefixes (such as [Text], [Sentence], and [Entity]) to generate knowledge elements of different granularities. For example, the [Text] prefix guides the generation of overall text-level knowledge, while the [Sentence] prefix guides the generation of sentence-level knowledge.

[0063] By introducing the gradient decay strategy, the prediction probabilities corresponding to different prefix tags will weightedly affect the generation quality and order of knowledge elements of this granularity, thereby achieving high-quality multi-granularity parallel knowledge extraction.

[0064] Specifically, the goal of this step of multi-granularity knowledge extraction based on sequence-to-sequence generation is to design an innovative neural network model, model knowledge extraction as a sequence-to-sequence generation task, and at the same time achieve parallel high-quality extraction of multi-granularity knowledge through special labeling guidance.

[0065] 3.1 Transformer Encoder-Decoder

[0066] Encoder: Adopts the standard Transformer encoder architecture, including multiple encoder layers; each layer consists of multi-head self-attention and feed-forward networks, which can effectively model long-range contextual dependencies;

[0067] Decoder: A Pointer-GeneratorNet is used to copy words from the source text or generate new words based on the context; the decoded information is well aligned with the source input by sharing the encoder output and the attention mechanism.

[0068] 3.2 Multi-granularity knowledge extraction strategy

[0069] Tag definition: Design special tags as prefixes, such as [Text] to indicate overall text knowledge, [Sentence] to indicate sentence knowledge, [Entity] to indicate entity knowledge, etc.;

[0070] Tag guidance: During the training process, the input data is segmented into different granularities and prefixed accordingly. The model learns to generate knowledge elements that start with tags and correspond to corresponding granularities.

[0071] Gradient decay: The prediction probabilities of different labels are used as weights, and the weighted average gradient is used to control the quality and order of knowledge generation at different granularities.

[0072] 4. Knowledge base construction and reasoning services based on cognitive graphs

[0073] First, triples (such as <fault cause, result, equipment damage>) are extracted from semi-structured data sources such as equipment maintenance manuals and fault cases to construct an initial cognitive map.

[0074] Based on the coupled coding model of node neighborhood structure and node content information, the embedding representation of nodes (entities) and relationships is trained to obtain the initial embedding of the cognitive graph.

[0075] Design a latent variable model to integrate multi-source heterogeneous knowledge such as entity features, local / global context, and graph topology to generate a graph embedding representation that incorporates rich information.

[0076] In the embedding space, similar entities / relations are calculated through similarity measurement, and an attention-based multi-hop knowledge path visualization model is developed to support interactive and explainable knowledge reasoning.

[0077] Specifically, the goal of this step of knowledge base construction and reasoning service based on cognitive graph is to build a cognitive graph knowledge base with rich knowledge and interpretability, and support flexible intelligent services based on graph embedding.

[0078] 4.1 Cognitive Map Construction

[0079] Semi-structured data extraction: Extract structured triples of knowledge from semi-structured data such as equipment maintenance manuals and fault cases, such as <engine damage, cause, oil tank leakage>;

[0080] Initial graph construction: Organize the extracted triples by entity-relationship to construct the initial cognitive graph;

[0081] Node / edge encoding: Encode entity nodes and relationship edges, and convert node text into vector representation.

[0082] 4.2 Graph Embedding and Heterogeneous Knowledge Fusion

[0083] Node neighborhood encoding: Design a graph neural network model that integrates node content and node neighborhood structure information, and calculates node embedding representation for each node and its neighbors;

[0084] Relation encoding: Based on the node pair embedding representation and relation edge features, the relation embedding is calculated through a feedforward network;

[0085] Latent variable model: Set latent variables for each node, integrate heterogeneous information such as node features, context, and graph topology knowledge, and generate node and relationship embeddings with rich information.

[0086] 4.3 Knowledge Reasoning and Visual Explanation

[0087] Similarity calculation: Based on the node or relationship embedding vector, the similarity can be calculated to directly find similar faults, replace parts, etc.

[0088] Path reasoning: Through a supervised path ranking model, the optimal explanation path connecting two nodes can be found;

[0089] Attention Visualization: Develop attention-based interactive tools to intuitively display the reasoning process and path explanation.

[0090] 5. Intelligent applications and decision support services

[0091] Based on the constructed knowledge base, a human-computer interactive question-answering system is developed to support interactive fault diagnosis and cause inference.

[0092] Combine the maintenance step knowledge and reusable experience extracted from the corpus to generate high-quality personalized maintenance plans and standardize the maintenance process.

[0093] Assess possible future failure risks and provide preventive maintenance recommendations to equipment maintenance personnel based on knowledge base reasoning to improve equipment operation reliability.

[0094] Specifically, the step of intelligent application and decision support service is based on the constructed cognitive graph knowledge base and combines the multi-granularity knowledge extraction results to develop an intelligent application system in the field of equipment maintenance.

[0095] 5.1 Equipment Fault Diagnosis and Cause Analysis

[0096] Interactive question and answer: Build a human-computer dialogue system based on the knowledge base to support user natural language interaction;

[0097] Reasoning path search: Retrieve entities and relationships related to the fault reporting problem in the knowledge graph to find the optimal explanation path;

[0098] Result generation: Automatically generate a fault cause analysis report based on the optimal explanation path.

[0099] 5.2 Maintenance plan generation

[0100] Step knowledge organization: Extract maintenance-related step knowledge from the extracted multi-granularity knowledge;

[0101] Similar case search: Search the solution steps of historical similar maintenance cases in the knowledge base;

[0102] Solution integration and planning: Integrate step knowledge with reference cases and combine with physical equipment information to generate targeted maintenance solutions.

[0103] 5.3 Preventive maintenance recommendations

[0104] Risk assessment: Estimate future failure risks based on equipment operation cycle, environment and other factors;

[0105] Knowledge reasoning: Based on the graph knowledge base, the equipment status and risk combination reasoning;

[0106] Maintenance recommendations: Provides a list of recommended preventive operations such as equipment inspection and replacement.

[0107] 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.

Claims

1. A method for intelligent processing and knowledge extraction of equipment maintenance corpus, characterized in that: The following steps are involved: (1) Preprocessing the original equipment maintenance corpus, performing standardized processing of word segmentation and part-of-speech tagging on the corpus, and constructing an equipment maintenance domain corpus, wherein the domain corpus contains terminology and semantic knowledge in the equipment maintenance domain; (2) learning semantic representation from the preprocessed corpus, and fusing domain terminology vocabulary and semantic knowledge in the domain corpus based on a contrastive learning framework to generate semantic representation with domain semantic enhancement; (3) inputting the semantic representation into a neural network model based on sequence-to-sequence generation, and simultaneously generating and extracting equipment maintenance knowledge information of multiple granularities from the semantic representation through predefined special tags, wherein the multiple granularities include text-level knowledge, sentence-level knowledge, and entity-level knowledge; (4) Based on the extracted multi-granularity knowledge information, a cognitive graph knowledge base is constructed, and the cognitive graph knowledge base supports knowledge reasoning and decision support services based on graph embedding.

2. The method according to claim 1, characterized in that The contrastive learning framework in step (2) includes: constructing positive and negative sample pairs, wherein the positive sample pair is a semantically similar corpus pair, and the negative sample pair is a semantically dissimilar corpus pair; mapping the terms in the domain corpus into predefined semantic vectors, and mapping the concepts into adjustable semantic vectors; and minimizing the loss function of the sum of the similarities of the semantic representations of the positive and negative sample pairs.

3. The method according to claim 2, characterized in that The neural network model based on sequence-to-sequence generation adopts a Transformer encoder-decoder architecture and fuses contextual information through an attention mechanism during the decoding process.

4. The method according to claim 3, characterized in that The decoder is a pointer generation network that can copy words from the source corpus or generate new words.

5. The method according to claim 3, characterized in that: The multi-granularity knowledge extraction is guided by special tag prefixes in the decoding process, and different prefixes correspond to knowledge elements of different granularities.

6. The method according to claim 5, characterized in that A gradient decay strategy is adopted to control the generation quality and order of knowledge elements of different granularities by weighting the prediction probabilities of corresponding tags.

7. The method according to claim 1, characterized in that The method for constructing the cognitive graph knowledge base described in step (4) includes: extracting triples from a semi-structured data source to construct an initial cognitive graph; training a graph embedding model of nodes and relationships based on node neighborhood structure and node content information; designing a latent variable model to integrate entity features, context information and graph structure information to generate a graph embedding representation that integrates multi-source heterogeneous knowledge.

8. The method according to claim 7, characterized in that In the cognitive graph knowledge base, similar nodes or paths are calculated through similarity measurement, and an explanation model based on the attention mechanism is developed to support explainable knowledge reasoning services.

9. The method according to claim 1, characterized in that: The decision support service includes: interactive fault diagnosis and cause inference based on the question-answering system and the constructed knowledge base; generating personalized maintenance plans for equipment in combination with the extracted maintenance steps and experience; evaluating possible future failure risks and recommending preventive maintenance measures.

10. The method according to any one of claims 1 to 9, characterized in that: The method is applied to intelligent corpus processing, knowledge base construction and related maintenance decision support services in the field of equipment maintenance.

Images