User question answer output method of hot rolling production line intelligent operation and maintenance online question answering system

Abstract

The invention discloses a user question answer output method for a hot rolling production line intelligent operation and maintenance online question-answering system. The method comprises the following steps: obtaining operation and maintenance documents in the iron and steel industry, and screening to generate expert question-answering pairs; carrying out vectorization processing on the expert question-answer pairs and then storing the expert question-answer pairs in a vector database; performing vectorization processing on professional documents of the iron and steel industry and fault maintenance and disposal reports of a production line over the years, and loading the vectorized documents and the fault maintenance and disposal reports into a vector database; when a user asks a question, vectorizing the question of the user, comparing the question of the user with an expert question-answer pair pre-stored in a vector database, and directly outputting an answer of the expert question-answer pair as an answer of the question of the user when the similarity is high; if the similarity is low, inputting the user question and the similar expert question and answer pairs and paragraph texts in the vector database into a large language model, and summarizing and outputting a congruent answer of the user question by the large language model according to the user question and the similar expert question and answer pairs and paragraph texts. According to the method, the accuracy of outputting answers for the user questions is greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of online question-answering consultation for intelligent operation and maintenance of hot rolling production lines, and in particular to a method for outputting answers to user questions in an online question-answering system for intelligent operation and maintenance of hot rolling production lines. Background Art

[0002] In the steel hot rolling industry, daily maintenance, problem feedback and fault recording are already arduous tasks. Frontline employees are often unable to trace problems in real time. In addition, industry expert resources are limited and cannot provide Q&A support at any time. In view of this, with the advancement of large language model technology and its wide application in many fields such as internal enterprise training, knowledge sharing and customer service, the steel industry urgently needs the support of such question-answering systems.

[0003] At present, the question-answering system mainly relies on large language models and vector retrieval technology, which have shown great flexibility and applicability to specific business areas of enterprises. Nevertheless, the existing text recall technology, as an important factor affecting the accuracy of the question-answering system, has certain limitations in the actual question-answering application in the field of hot rolling of steel. In particular, when performing text vector retrieval on professional issues such as the hydraulic system, press roll, auxiliary roll, and layer cold roller of the hot rolling production line based on the existing documents of front-line employees and experts, paragraphs containing redundant or erroneous information may be retrieved, and the experience of hot rolling experts cannot be effectively integrated, which limits the accuracy of the output results of the question-answering system.

[0004] In order to solve these problems, combined with the steel industry's own factors (a lot of maintenance and fault records are accumulated on a daily basis), it is particularly important to develop a new generation of user question answer output technology. This technology needs to be able to improve the accuracy of daily hot rolling field questions and be able to combine with the experience of hot rolling experts, thereby significantly improving the overall performance and accuracy of the question answering system. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for outputting answers to user questions in an online question-and-answer system for intelligent operation and maintenance of a hot rolling production line. The method is aimed at the question-and-answer system in the hot rolling field of the steel industry, overcomes the defects of the existing answer output, converts expert experience into the form of question-and-answer pairs, and uses vectorization technology to store the question-and-answer pairs in a vector database. When a user asks a question, the question-and-answer system not only retrieves document fragments related to the question, but also combines the expert experience data stored in the vector database, and performs comprehensive analysis and summary through a large language model to improve the accuracy of the output answer.

[0006] In order to solve the above technical problems, the method for outputting answers to user questions of the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines of the present invention comprises the following steps: Step 1: Obtain the operation and maintenance documents of the steel industry, and generate expert Q&A pairs for the steel industry based on the documents; Step 2: vectorize the expert question and answer pairs through the Embedding model, and convert the expert question and answer pairs into vectorized display. The vectorized expert question and answer pair data is stored in the vector database; Step 3: For professional documents of the steel industry and the fault inspection and disposal reports of production lines over the years, semantic recognition and segmentation are performed according to paragraph texts, and the recognized and segmented paragraph texts are vectorized through the Embedding model and loaded into the vector database; Step 4: When a user asks a question about the hot rolling production line, the Embedding model is used to vectorize the user question. The vectorized user question is compared with the expert question and answer pairs pre-stored in the vector database, and the text similar to the user question is output to determine the vector similarity between them. If the vector similarity is ≥ 0.95, the answer to the expert question and answer pair is directly output as the answer to the user question; if the vector similarity is < 0.95, a secondary screening is performed, and the user question and similar expert question and answer pairs and paragraph texts in the vector database are input into the large language model. The large language model summarizes and outputs a summary answer to the user question based on the user question and similar expert question and answer pairs and paragraph texts.

[0007] Furthermore, the key prompt words of the expert question and answer pairs in step one are refined and input into the large language model. The generation ability of the large language model is used to enrich and expand the content of the original expert question and answer pairs, increase the diversity and coverage of the expert question and answer pairs, and enhance the depth and breadth of the large language model in dealing with complex user questions.

[0008] Furthermore, when performing secondary screening in step 4, the Reranker intelligent text sorting tool is used to evaluate the relevance of similar expert question-answer pairs and paragraph texts to user questions, and re-sort the similar expert question-answer pairs and paragraph texts to output several document paragraphs that are most relevant to the user questions. The large language model summarizes several document paragraphs based on the user questions and outputs a summary answer to the user questions.

[0009] Furthermore, the vector similarity judgment in step 4 adopts cosine similarity calculation, and the similarity of the two vectors is evaluated by calculating the cosine value of the angle between the two vectors.

[0010] Since the method for outputting answers to user questions in the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines of the present invention adopts the above-mentioned technical scheme, that is, the method obtains operation and maintenance documents of the steel industry, screens and generates expert question-and-answer pairs; vectorizes the expert question-and-answer pairs through the Embedding model and stores them in the vector database; the professional documents of the steel industry and the fault inspection and disposal reports of the production lines over the years are vectorized and loaded into the vector database through the Embedding model; when the user asks questions related to the hot rolling production line, the user questions are vectorized and compared with the expert question-and-answer pairs pre-stored in the vector database. If the vector similarity is ≥0.95, the answer to the expert question-and-answer pair is directly output as the answer to the user question; if the vector similarity is <0.95, a secondary screening is performed, and the user questions and similar expert question-and-answer pairs and paragraph texts in the vector database are input into the large language model, and the large language model summarizes and outputs a summary answer to the user question based on the user questions and similar expert question-and-answer pairs and paragraph texts. This method is aimed at the question-answering system in the hot rolling field of the steel industry. When a user asks a question, the question-answering system not only retrieves document fragments related to the question, but also combines the expert experience data stored in the vector database and performs comprehensive analysis and summary through a large language model to improve the accuracy of the output answer. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments: Figure 1 The present invention is a flowchart of a method for outputting answers to user questions in an online question-and-answer system for intelligent operation and maintenance of a hot rolling production line. DETAILED DESCRIPTION

[0012] Implementation example Figure 1 As shown, the method for outputting answers to user questions in the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines of the present invention comprises the following steps: Step 1: Obtain the operation and maintenance documents of the steel industry, and generate expert Q&A pairs for the steel industry based on the documents; Step 2: vectorize the expert question and answer pairs through the Embedding model, and convert the expert question and answer pairs into vectorized display. The vectorized expert question and answer pair data is stored in the vector database; Step 3: For professional documents of the steel industry and the fault inspection and disposal reports of production lines over the years, semantic recognition and segmentation are performed according to paragraph texts, and the recognized and segmented paragraph texts are vectorized through the Embedding model and loaded into the vector database; Step 4: When a user asks a question about the hot rolling production line, the Embedding model is used to vectorize the user question. The vectorized user question is compared with the expert question and answer pairs pre-stored in the vector database, and the text similar to the user question is output to determine the vector similarity between them. If the vector similarity is ≥ 0.95, the answer to the expert question and answer pair is directly output as the answer to the user question; if the vector similarity is < 0.95, a secondary screening is performed, and the user question and similar expert question and answer pairs and paragraph texts in the vector database are input into the large language model. The large language model summarizes and outputs a summary answer to the user question based on the user question and similar expert question and answer pairs and paragraph texts.

[0013] Preferably, the key prompt words of the expert question and answer pairs in step one are extracted and input into the large language model, and the generation ability of the large language model is used to enrich and expand the content of the original expert question and answer pairs, increase the diversity and coverage of the expert question and answer pairs, and enhance the depth and breadth of the large language model in processing complex user questions.

[0014] Preferably, when performing secondary screening in step 4, the Reranker intelligent text sorting tool is used to evaluate the relevance of similar expert question and answer pairs and paragraph texts to user questions, and re-sort the similar expert question and answer pairs and paragraph texts, output several document paragraphs that are most relevant to the user question, and the large language model summarizes several document paragraphs based on the user question, and outputs a summary answer to the user question.

[0015] Preferably, the vector similarity judgment in step 4 adopts cosine similarity calculation, and the similarity of the two vectors is evaluated by calculating the cosine value of the angle between the two vectors.

[0016] This method collects information reported by production line equipment, hot rolling failure cases and other documents, parses the information recorded in the documents, identifies and splits them according to semantics, and generates initial expert question and answer pairs such as original consultations and failure questions based on relevant questions. Subsequently, independent experts in the field conduct a detailed review to select highly valuable and representative expert insights. In addition, experts can also organize targeted experiences for difficult points frequently encountered in daily life. These experiences summarized in production are organized into the form of question and answer pairs to ensure their practicality and effectiveness in the system. Through this process, it can be ensured that the expert experience adopted by the question and answer system is selected and verified, so as to provide users with more accurate and professional services.

[0017] The unique experience of experts is converted into cleaned and verified question-answer pairs, that is, a concise and clear question-answer format. In view of the scarcity and value of expert experience, this method further adopts the strategy of expanding the question-answer pairs. This expansion process is achieved by inputting key prompt words into a large language model, and using the model's generative ability to enrich and expand the content of the original question-answer pairs. It can not only retain the rich connotations of expert experience, but also increase the diversity and coverage of question-answer pairs, thereby improving the depth and breadth of the question-answering system when processing complex queries.

[0018] This method uses the Embedding model (open source bge-large-zh-v1.5, which processes the data into 1024-dimensional data) to vectorize the expanded expert question and answer pairs. The function of the Embedding model is to convert a natural language text that is less than the set maximum length into a vector of fixed dimension, and store these vectors in a vector database. The vector database is a special database designed to store a large number of vector elements and optimized for vector calculations. A rich expert experience knowledge base is built in the form of a vector database.

[0019] After the professional texts and expert experience questions and answers in the steel industry are vectorized and stored in the vector database, when users use the large language model question-answering system, they ask questions, and the system vectorizes the sentences of the user's questions. Through the vector recall algorithm in the vector database (the vector database uses cosine similarity, and evaluates the similarity of two vectors by calculating the cosine value of the angle between two vectors), several document paragraphs with the highest similarity are recalled.

[0020] If the similarity score is higher than the set threshold, the system will directly provide an answer based on expert experience. If the similarity score fails to reach the set threshold, the system will use the Reranker intelligent text sorting tool, which evaluates the relevance of documents to user questions and re-ranks them to output several document paragraphs that are most relevant to the user's question.

[0021] Ultimately, users will receive answers from expert experience question-answer pairs, or summary answers generated from several document paragraphs after comprehensive sorting by a large language model. This dual-track output answer strategy ensures that users can obtain accurate and comprehensive answers to their questions, while fully leveraging the advantages of expert experience and system intelligent sorting.

Claims

1. A method for outputting answers to user questions in an online question-and-answer system for intelligent operation and maintenance of a hot rolling production line, characterized in that The steps include: Step 1: Obtain the operation and maintenance documents of the steel industry, and generate expert Q&A pairs for the steel industry based on the documents; Step 2: vectorize the expert question and answer pairs through the Embedding model, and convert the expert question and answer pairs into vectorized display. The vectorized expert question and answer pair data is stored in the vector database; Step 3: For professional documents of the steel industry and the fault inspection and disposal reports of production lines over the years, semantic recognition and segmentation are performed according to paragraph texts, and the recognized and segmented paragraph texts are vectorized through the Embedding model and loaded into the vector database; Step 4: When a user asks a question about the hot rolling production line, the Embedding model is used to vectorize the user's question. The vectorized user question is compared with the expert question-answer pair pre-stored in the vector database, and the text similar to the user's question is output to determine the vector similarity between them. If the vector similarity is ≥ 0.95, the answer to the expert question-answer pair is directly output as the answer to the user's question. If the vector similarity is less than 0.95, a secondary screening is performed, and the user question and similar expert question-answer pairs and paragraph texts in the vector database are input into the large language model. The large language model summarizes the user question and similar expert question-answer pairs and paragraph texts and outputs a summary answer to the user question.

2. The method for outputting answers to user questions in the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines according to claim 1, characterized in that: Extract the key prompt words of the expert question and answer pairs in step 1 and input them into the large language model. Use the generation ability of the large language model to enrich and expand the content of the original expert question and answer pairs, increase the diversity and coverage of the expert question and answer pairs, and enhance the depth and breadth of the large language model in dealing with complex user questions.

3. The method for outputting answers to user questions in the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines according to claim 1 or 2, characterized in that: When performing secondary screening in step 4, the Reranker intelligent text sorting tool is used to evaluate the relevance of similar expert question-answer pairs and paragraph texts to user questions, and re-sort the similar expert question-answer pairs and paragraph texts to output several document paragraphs that are most relevant to the user questions. The large language model summarizes several document paragraphs based on the user questions and outputs a summary answer to the user questions.

4. The method for outputting answers to user questions in the online question-and-answer system for intelligent operation and maintenance of hot rolling production lines according to claim 3 is characterized by: In the step 4, the vector similarity is determined by using cosine similarity calculation, and the similarity of the two vectors is evaluated by calculating the cosine value of the angle between the two vectors.

Images