Equipment operation and maintenance knowledge medium table construction method based on large model

Abstract

The invention discloses an equipment operation and maintenance knowledge middle table construction method based on a large model, and the method comprises the following steps: preparing a text database, reading text data from the text database, and carrying out the preprocessing of the text data; carrying out segmentation, vectorization and dimension standardization processing on the preprocessed text data; creating and storing a vector index; the form of the text data after vectorization and dimension normalization and the constructed vector index jointly form an equipment operation and maintenance large model plug-in knowledge base, and on the basis, knowledge application is carried out. According to the construction of the knowledge medium table based on the equipment operation and maintenance large model, accurate and comprehensive equipment operation and maintenance information can be provided, the response speed and efficiency of the system are improved, and the expandability and flexibility of the system are improved. Therefore, the requirements of the user can be met, and better user experience is provided.

Description

Technical Field

[0001] The present invention relates to a method for constructing a knowledge platform, and in particular to a method for constructing a equipment operation and maintenance knowledge platform based on a large model. Background Art

[0002] At present, big models are developing rapidly in the field of equipment operation and maintenance and are beginning to play an important role. Based on a large amount of professional knowledge and experience in the field of operation and maintenance, they can intelligently generate decision support, help operation and maintenance personnel formulate operation and maintenance strategies more scientifically, and can also identify failure modes in equipment operation and maintenance, and provide fast and accurate fault diagnosis and solutions, reducing the time and cost of equipment failure handling.

[0003] However, currently, large models also have some shortcomings in the field of equipment operation and maintenance. For example, they lack the support of proprietary knowledge in the field of industry-specific knowledge question and answer, and may produce some illusions in reasoning ability, making the large model question and answer results inaccurate and in time. Knowledge recall is a method of quickly and accurately retrieving knowledge related to the query from the industry-specific knowledge base, providing efficient information retrieval and answer generation. Therefore, building a knowledge middle platform to integrate large models and knowledge recall technology can effectively make up for these shortcomings. Summary of the invention

[0004] In order to solve the shortcomings of the above-mentioned technology, the present invention provides a method for constructing an equipment operation and maintenance knowledge platform based on a large model.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for constructing a device operation and maintenance knowledge platform based on a large model, comprising the following steps:

[0006] Step 1: Prepare a text database to collect knowledge documents in the field of equipment operation and maintenance;

[0007] Step 2: Read text data from the text database and preprocess the read text data;

[0008] Step 3: Text segmentation: Split the preprocessed text data into text paragraphs;

[0009] Step 4: Text vectorization: embed the segmented text paragraphs and vectorize them so that the semantic information of the text is encoded into a numerical vector; then the text vector is dimensionally normalized;

[0010] Step 5: Create and save vector indexes, including:

[0011] Step 5.1: Build vector index: Use the vector index structure to build the index of the calculated text vector;

[0012] Step 5.2: Store vector index: Store the constructed vector index structure in the form of a file in the vector database, and save the vector index in the form of a text file or a binary file;

[0013] Step 6: The form of the text data after vectorization and dimension normalization, and the constructed vector index together form the equipment operation and maintenance large model plug-in knowledge base;

[0014] Apply knowledge middle platform based on the equipment operation and maintenance big model.

[0015] Preferably, in step 1, the text database is composed of professional text materials in the field of equipment operation and maintenance for uploading to a specified path, including but not limited to equipment technical data, equipment failure analysis reports, equipment maintenance records, and the formats of the text materials include but are not limited to pdf, txt, md, and docx.

[0016] Preferably, in step 1, the preprocessing method includes text cleaning and screening, and ultimately only valid text information is retained.

[0017] Preferably, in step 3, the text segmentation method includes: segmentation by rules, segmentation by length, and segmentation by semantics.

[0018] Preferably, in step 4, a pre-trained text2vec model is used to perform text vectorization processing, and the specific process is as follows:

[0019] Step 4.1: Text preprocessing: Preprocess each text to convert it into a word sequence;

[0020] Step 4.2: Build a vocabulary: Based on the text data preprocessed in 4.1, build a vocabulary to record all the words that have appeared. Each word will be associated with a unique index value.

[0021] Step 4.3: Train the embedding model: Use the embedding algorithm provided by text2vec to train the vocabulary and obtain the embedding vector of the word. During the training process, the model will learn the semantic information of the word based on the context.

[0022] Step 4.4: Text segment embedding: Select a text vectorization method based on the specific task to perform text segment embedding processing; text vectorization methods include word frequency-based representation and word embedding-based representation;

[0023] Step 4.5: Dimension normalization: Dimension normalization is performed on the text segment vector to make the vector consistent in numerical range;

[0024] Step 4.6: Vector storage: Store the normalized vectors in a vector database in text or binary form, taking each vector as a record.

[0025] Preferably, in step 4.5, the dimensionality normalization method includes L2 norm normalization, by calculating the L2 norm of the vector, that is, the Euclidean norm, and then dividing each dimension of the vector by the L2 norm, thereby scaling the length of the vector to 1.

[0026] Preferably, in step 5.1, when constructing the vector index, a suitable index structure needs to be selected according to the task requirements, and the index structure includes KD-Tree, Ball Tree, and inverted index;

[0027] In the process of building the index, the calculated text vectors are inserted into the corresponding index structure, and similar vectors are clustered together according to the similarity between the vectors to form the nodes or inverted lists of the index.

[0028] Preferably, in step 5.2, when the text file format is selected for storage: the index structure is saved in an easily readable text format, with each line representing a node or inverted list of a vector index, and the vector and related information corresponding to each node or list;

[0029] When choosing to store in binary file format: use the serialization method to convert the index structure into binary data and save it to the specified path.

[0030] Preferably, in step 6, the knowledge platform application process includes the following:

[0031] Step 6.1: Pre-process the user's questions to make the format of the questions meet the processing requirements;

[0032] Step 6.2: Text vectorization: Use the text vectorization method in step 4.4 and the dimension normalization method in step 4.5 to convert the user question into a vector representation;

[0033] Step 6.3: Vector similarity matching: Use the vector similarity calculation method to perform similarity matching between the vector of the user's question and the vector in the knowledge base, and find the knowledge base vector that is most similar to the user's question;

[0034] Step 6.4: Question response: Based on the user's question, relevant information is searched from the knowledge base, and a large language model is used for further parsing, generation and processing to provide an answer or response.

[0035] Preferably, in step 6.4, the question reply includes the following processing steps:

[0036] Step 6.4.1: Extract relevant information: After retrieving the knowledge base vector that is most similar to the user's question from the knowledge base, extract information related to the user's question from the knowledge base with the help of the vector index. The relevant information includes the answer to the question, the solution, and the relevant text paragraphs;

[0037] Step 6.4.2: Parse, construct and generate a new prompt: Based on the results matched from the knowledge base in step 6.3, the system parses and constructs to generate a new prompt; the prompt is used to indicate the context of the input information and the parameter information of the input model to further interact with the large language model;

[0038] Step 6.4.3: prompt calls the equipment operation and maintenance model: the equipment operation and maintenance model generates a reply with relevant content based on the new prompt received.

[0039] The present invention discloses a method for constructing a device operation and maintenance knowledge platform based on a large model. First, by combining the professional knowledge in the field of device operation and maintenance with a large language model, more accurate and comprehensive information related to device operation and maintenance can be provided. This allows users to obtain more specific and practical answers by asking questions, solving problems or obtaining relevant suggestions.

[0040] Secondly, the construction of the knowledge middle platform can improve the response speed and efficiency of the system. Since the information in the knowledge base has been pre-processed and stored as vector representation, and retrieved using vector indexes, it is possible to quickly match the vector of the user's question and extract relevant information from the knowledge base. In this way, the system can give accurate responses in a short time and improve the user experience.

[0041] In addition, the construction of the knowledge center improves the scalability and flexibility of the system. By separating the knowledge base from the large model of equipment operation and maintenance, the content of both can be updated and expanded independently. This means that the content of the knowledge base can be flexibly updated according to new equipment operation and maintenance knowledge or user needs without retraining the entire large language model. This can respond to changing needs more quickly and provide more accurate and practical information.

[0042] In summary, the construction of the knowledge middle platform based on the equipment operation and maintenance model of the present invention brings many beneficial effects, including providing accurate and comprehensive equipment operation and maintenance information, improving the response speed and efficiency of the system, and improving the scalability and flexibility of the system. This can meet the needs of users and provide a better user experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a logic diagram of the technical solution of the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0045] The present invention discloses a method for constructing a device operation and maintenance knowledge middle platform based on a large model. By constructing a knowledge middle platform in the field of equipment operation and maintenance and applying a knowledge recall method, exclusive support is provided for knowledge question and answer tasks of the large model in the field of equipment operation and maintenance.

[0046] The basic idea of ​​building a knowledge platform in the field of equipment operation and maintenance is to collect industry-specific knowledge texts, segment and vectorize these texts, and then perform vector retrieval based on user input, and input the retrieval results as context into the big model to achieve the induction and summary of relevant knowledge. This combination can provide more accurate information retrieval, enabling the big model to perform more targeted reasoning and question-answering based on specific knowledge in the exclusive field.

[0047] like Figure 1 As shown, the detailed steps of the present invention are as follows:

[0048] Step 1: Prepare the text database

[0049] The text database is used to collect knowledge documents in the field of equipment operation and maintenance, and provide corpus support for building a knowledge base in the field of operation and maintenance in the knowledge center. The text database is composed of professional text materials in the field of equipment operation and maintenance uploaded to the specified path, including equipment technical information, equipment failure analysis reports, equipment maintenance records, etc. These text materials can be in various formats such as pdf, txt, md, docx, etc. The local text database is the basis for building a knowledge base in the field of operation and maintenance.

[0050] Step 2: Text reading and processing

[0051] Read text data from the text database and preprocess it, which includes text cleaning and screening, and finally retain only valid text information. Data cleaning mainly removes useless tags, spaces, punctuation marks, etc. in the text, and handles incorrect spelling and grammar. Data screening mainly removes sensitive information, such as corporate and personal identification information, and other unauthorized information.

[0052] Step 3: Text Segmentation

[0053] Split the article according to rules, length or semantics. Text segmentation can make the text information clearer and more organized, which is convenient for subsequent applications.

[0054] Text segmentation is to cut the preprocessed text data into appropriate text paragraphs. There are many methods for text segmentation, including segmentation by rules, segmentation by length, segmentation by semantics, etc. Segmentation by rules is segmentation based on certain rules or symbols, such as segmentation by punctuation marks, line breaks, etc. Segmentation by length is to divide the text into fixed-size segments according to the specified length, regardless of the semantics or structure. Semantic segmentation is segmentation based on the semantic information of the text, such as segmentation by paragraph, sentence boundary, topic, etc., to retain more semantic integrity. These different segmentation methods can be selected according to specific application scenarios and needs to obtain better text analysis and processing effects.

[0055] Step 4: Text Vectorization

[0056] Embed the segmented text paragraphs to vectorize them. You can use the pre-trained text2vec model or train your own model based on your needs.

[0057] Embedding the segmented text segments refers to the process of converting each text segment into a vector representation. The purpose of this is to encode the semantic information of the text into a numerical vector for subsequent semantic retrieval and similarity calculation. The detailed steps are as follows:

[0058] Step 4.1: Text preprocessing: Preprocess each text, such as removing stop words, punctuation marks and special characters, performing word segmentation, etc. This can convert the text into a sequence of words.

[0059] Step 4.2: Build a vocabulary: Based on the preprocessed text data, build a vocabulary to record all the words that have appeared. Each word will be associated with a unique index value.

[0060] Step 4.3: Train the embedding model: Use the embedding algorithm provided by text2vec, such as Word2Vec or GloVe, to train the vocabulary and get the embedding vector of each word. During the training process, the model will learn the semantic information of the word based on the context.

[0061] Step 4.4: Text segment embedding: There are many methods for text vectorization. Commonly used methods include word frequency representation and word embedding representation. The choice should be based on the specific task.

[0062] Representation method based on word frequency: The bag-of-words model represents text as a vector, where each dimension corresponds to a word and the value represents the frequency of occurrence of the word in the text. The TF-IDF method is based on the bag-of-words model and considers the importance of words in the entire corpus, weighting them by calculating the word frequency and inverse document frequency.

[0063] Representation method based on word embedding: It averages or weighted averages the embedding vectors of each word in each text segment to obtain the vector representation of the text segment. The average operation is to simply add the embedding vectors of all words and divide it by the number of words, while the weighted average operation assigns different weights to the word vectors according to the importance of the word, and then adds the weighted word vectors to obtain the vector representation of the text segment. The average operation can provide an overall semantic information, evenly distributing the contribution of each word to the entire text segment. The weighted average operation can weightedly combine the word vectors according to the importance of the word to more accurately reflect the semantics of the text segment.

[0064] Step 4.5: Dimension normalization: Normalize the text segment vector to ensure the consistency of the vector in the numerical range. The common normalization method is L2 norm normalization.

[0065] L2 norm normalization is a common normalization method that calculates the L2 norm (Euclidean norm) of a vector and then divides each dimension of the vector by the L2 norm to scale the length of the vector to 1. This eliminates the size differences of the vectors, making comparisons between different vectors more fair and accurate.

[0066] The purpose of dimension normalization is to ensure the consistency of the vector in the numerical range, make the values ​​of different dimensions comparable, and avoid the influence of certain dimensions on the overall vector representation. Through normalization operations, the vector can be made more stable and interpretable in terms of value, which helps the reliability of subsequent similarity calculation tasks.

[0067] Step 4.6: Vector storage: Save the dimensionally normalized vectors into the vector database. To save the dimensionally normalized vectors, each vector can be stored as a record in the vector database in text or binary form. In text form, each record can be stored in a row with one vector, and the data of each dimension is separated by commas or spaces. In binary form, the vector can be stored directly in binary encoding. This ensures that the vector has a consistent scale during storage and query, which is convenient for subsequent use.

[0068] The form of text after vectorization and dimension normalization is one of the basic components of the equipment operation and maintenance large model plug-in knowledge base.

[0069] Step 5: Create and save vector index (semantic index)

[0070] In the external knowledge base of large models, text vectors are usually saved in the form of vector indexes. Vector index is a data structure used to store and organize vector data to support efficient similarity search and retrieval. The following are the steps to create and save vector indexes:

[0071] Step 5.1: Build vector index: Use vector index structure (such as KD-Tree, Ball Tree, inverted index, etc.) to build index for the calculated text vector. Vector index organizes vectors according to the similarity between them for fast similarity search and retrieval;

[0072] When constructing a vector index, you need to select a suitable index structure based on the task requirements. For example, KD-Tree is a binary tree structure that can divide the vector space into multiple subspaces to efficiently search for the nearest neighbor vector; Ball Tree is an index structure based on spherical regions that can better handle non-uniformly distributed vectors; and inverted index is a commonly used index structure for quickly finding vectors containing specific features.

[0073] During the index construction process, the calculated text vectors need to be inserted into the corresponding index structure. According to the similarity between vectors, similar vectors are clustered together to form index nodes or inverted lists. In this way, when performing similarity search and retrieval, similar vector sets can be quickly located by comparing the similarity between the query vector and the vector in the index.

[0074] Step 5.2: Store vector index: Store the constructed vector index structure in the form of a file in the vector database so that it can be quickly loaded and used when needed.

[0075] The present invention saves the vector index in the form of a text file or a binary file;

[0076] If you choose the text file format, you can save the index structure in an easy-to-read text format. Each line represents a node or inverted list of a vector index, and each node or list has a corresponding vector and related information. Such a text file can be easily viewed and debugged.

[0077] If you choose binary file format, you can save the index structure in binary data. This method can reduce storage space usage and is faster to read. Binary files can use a specific serialization method to convert the index structure into binary data and save it to a specified path.

[0078] Regardless of whether you choose a text file or a binary file, you need to store the constructed vector index structure in a specific format and specification. In this way, in subsequent applications, you can reconstruct the vector index by loading the file for similarity search and retrieval.

[0079] Vector index is another basic component of the equipment operation and maintenance large model plug-in knowledge base.

[0080] Step 6: Knowledge platform application

[0081] Step 6.1: Preprocess user questions: Preprocess user questions, such as word segmentation, removal of stop words, punctuation marks and special characters, etc., to ensure that the format of the question meets the processing requirements.

[0082] Step 6.2: Text vectorization: Use the text vectorization method in step 4.4 and the dimension normalization method in step 4.5 to convert the user question into a vector representation, so that the question can be represented in a form that matches the vector in the knowledge base.

[0083] Step 6.3: Vector similarity matching: Use a vector similarity calculation method (such as cosine similarity) to perform similarity matching between the vector of the user's question and the vector in the knowledge base, and find the knowledge base vector that is most similar to the user's question.

[0084] Step 6.4: Question response: This process is that the system searches for relevant information from the knowledge base based on the user's question, and uses the large language model for further parsing, generation and processing, so as to provide more accurate and useful answers or responses to meet the user's needs. The detailed steps are as follows:

[0085] Step 6.4.1: Extract relevant information: After retrieving the knowledge base vector that is most similar to the user's question from the knowledge base, use the vector index to extract information related to the user's question from the knowledge base, which can be the answer to the question, solution, related text paragraphs, etc.

[0086] Step 6.4.2: Parse, construct and generate a new prompt: Based on the results matched from the knowledge base in step 6.3, the system parses and constructs to generate a new prompt. The function of the prompt is mainly to prompt the AI ​​model with the context of the input information and the parameter information of the input model. It can be a question, a prompt or a text for further interaction with the language model. This fusion method can be adjusted and optimized according to specific needs and tasks to improve the accuracy and relevance of the answer.

[0087] Step 6.4.3: prompt calls the device operation and maintenance model: The device operation and maintenance model generates a response with relevant content based on the new prompt received. The device operation and maintenance model here is a model with strong language understanding and generation capabilities after pre-training and fine-tuning of professional knowledge in the field of device operation and maintenance. The device operation and maintenance model will parse and understand the information in the prompt that effectively integrates the user question and the knowledge base, and generate a response related to the knowledge base.

[0088] The above are the detailed steps of knowledge base application. In this process, the key is to effectively integrate user questions and knowledge base information so that the big model can better understand the context and semantics. The integration method can be adjusted and optimized according to specific needs and tasks to improve the accuracy and relevance of the answers.

[0089] The above implementation modes are not limitations of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the scope of the technical solution of the present invention also belong to the protection scope of the present invention.

Claims

1. A method for constructing a device operation and maintenance knowledge platform based on a large model, characterized by: The following steps are involved: Step 1: Prepare a text database to collect knowledge documents in the field of equipment operation and maintenance; Step 2: Read text data from the text database and preprocess the read text data; Step 3: Text segmentation: Split the preprocessed text data into text paragraphs; Step 4: Text vectorization: embed the segmented text paragraphs and vectorize them so that the semantic information of the text is encoded into a numerical vector; then the text vector is dimensionally normalized; Step 5: Create and save vector indexes, including: Step 5.1: Build vector index: Use the vector index structure to build the index of the calculated text vector; Step 5.2: Store vector index: Store the constructed vector index structure in the form of a file in the vector database, and save the vector index in the form of a text file or a binary file; Step 6: The form of the text data after vectorization and dimension normalization, and the constructed vector index together form the equipment operation and maintenance large model plug-in knowledge base; Apply knowledge middle platform based on the equipment operation and maintenance big model.

2. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 1 is characterized in that: In step 1, the text database is composed of professional text materials in the field of equipment operation and maintenance that are uploaded to the specified path, including but not limited to equipment technical data, equipment failure analysis reports, and equipment maintenance records. The formats of the text materials include but are not limited to pdf, txt, md, and docx.

3. The method for constructing a large-model-based equipment operation and maintenance knowledge platform according to claim 2 is characterized in that: In step 1, the preprocessing method includes text cleaning and screening, and only valid text information is retained in the end.

4. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 1 is characterized in that: In step 3, the text segmentation methods include: segmentation by rules, segmentation by length, and segmentation by semantics.

5. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 1 is characterized in that: In step 4, the pre-trained text2vec model is used to perform text vectorization. The specific process is as follows: Step 4.1: Text preprocessing: Preprocess each text to convert it into a word sequence; Step 4.2: Build a vocabulary: Based on the text data preprocessed in 4.1, build a vocabulary to record all the words that have appeared. Each word will be associated with a unique index value. Step 4.3: Train the embedding model: Use the embedding algorithm provided by text2vec to train the vocabulary and obtain the embedding vector of the word. During the training process, the model will learn the semantic information of the word based on the context. Step 4.4: Text segment embedding: Select a text vectorization method based on the specific task to perform text segment embedding processing; text vectorization methods include word frequency-based representation and word embedding-based representation; Step 4.5: Dimension normalization: Dimension normalization is performed on the text segment vector to make the vector consistent in numerical range; Step 4.6: Vector storage: Store the normalized vectors in a vector database in text or binary form, taking each vector as a record.

6. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 5 is characterized by: In step 4.5, the dimension normalization method includes L2 norm normalization, which is performed by calculating the L2 norm, i.e., the Euclidean norm, of the vector and then dividing each dimension of the vector by the L2 norm, thereby scaling the length of the vector to 1.

7. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 1 is characterized by: In step 5.1, when constructing the vector index, you need to select a suitable index structure according to the task requirements. The index structures include KD-Tree, BallTree, and inverted index. In the process of building the index, the calculated text vectors are inserted into the corresponding index structure, and similar vectors are clustered together according to the similarity between the vectors to form the nodes or inverted lists of the index.

8. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 7 is characterized in that: In step 5.2, when selecting text file storage: save the index structure in an easy-to-read text format, with each line representing a node or inverted list of a vector index, and the corresponding vector and related information for each node or list; When choosing to store in binary file format: use the serialization method to convert the index structure into binary data and save it to the specified path.

9. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 5 is characterized by: In step 6, the knowledge platform application process includes the following: Step 6.1: Pre-process the user's questions to make the format of the questions meet the processing requirements; Step 6.2: Text vectorization: Use the text vectorization method in step 4.4 and the dimension normalization method in step 4.5 to convert the user question into a vector representation; Step 6.3: Vector similarity matching: Use the vector similarity calculation method to perform similarity matching between the vector of the user's question and the vector in the knowledge base, and find the knowledge base vector that is most similar to the user's question; Step 6.4: Question response: Based on the user's question, relevant information is searched from the knowledge base, and a large language model is used for further parsing, generation and processing to provide an answer or response.

10. The method for constructing a large model-based equipment operation and maintenance knowledge platform according to claim 9 is characterized in that: In step 6.4, the question response includes the following processing steps: Step 6.4.1: Extract relevant information: After retrieving the knowledge base vector that is most similar to the user's question from the knowledge base, extract information related to the user's question from the knowledge base with the help of the vector index. The relevant information includes the answer to the question, the solution, and the relevant text paragraphs; Step 6.4.2: Parse, construct and generate a new prompt: Based on the results matched from the knowledge base in step 6.3, the system parses and constructs to generate a new prompt; the prompt is used to indicate the context of the input information and the parameter information of the input model to further interact with the large language model; Step 6.4.3: prompt calls the equipment operation and maintenance model: the equipment operation and maintenance model generates a reply with relevant content based on the new prompt received.

Images