Operation and maintenance question answering method and device and related equipment

Abstract

The invention provides an operation and maintenance question answering method and device and related equipment, and relates to the technical field of equipment operation and maintenance. The method comprises the steps that retrieval is conducted in a vector database based on user feature vectors of operation and maintenance questions, and target questions and similarity are obtained; when the similarity is greater than a first preset threshold value, inputting the operation and maintenance question into a large language model, and outputting an operation and maintenance answer; when the similarity is smaller than a second preset threshold value, determining that an operation and maintenance answer cannot be generated; when the similarity is greater than or equal to a second preset threshold value and less than or equal to a first preset threshold value, obtaining a determination result of the user; if a determination result is that the target question is the same as the operation and maintenance question, inputting the target question into a large language model, and outputting an operation and maintenance answer; and if the determined result is a new operation and maintenance question, repeatedly executing the steps until an operation and maintenance answer is generated. By means of the technical means, the problems that in the related technology, the equipment operation and maintenance intelligent level is low, and the operation and maintenance scheme lacks detailed technical data support are solved.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of device operation and maintenance, and in particular to an operation and maintenance problem solving method and device and related equipment. BACKGROUND

[0002] Currently, train equipment early warning and health management mainly relies on sensor technology to collect running state and fault information, combines historical data and environmental factors, and realizes state monitoring, fault prediction, health evaluation and maintenance decision support through specific algorithms. Current train equipment operation and maintenance still needs the participation of technical personnel, and the finally determined operation and maintenance scheme does not have more detailed technical material support. SUMMARY

[0003] The present disclosure provides an operation and maintenance problem solving method, device and related equipment, which improves the intelligent level of device operation and maintenance and generates an operation and maintenance scheme containing detailed technical material support.

[0004] According to one aspect of the present disclosure, an operation and maintenance problem solving method is provided, which includes: receiving a user's operation and maintenance problem about a device, and generating a user feature vector of the operation and maintenance problem; searching in a vector database based on the user feature vector to obtain a target problem and a similarity between the target problem and the operation and maintenance problem; when the similarity is greater than a first preset threshold, inputting the operation and maintenance problem into a large language model to output an operation and maintenance answer about the device; when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result of the user; if the determination result is that the target problem and the operation and maintenance problem are the same, inputting the target problem into the large language model to output the operation and maintenance answer; and if the determination result is a new operation and maintenance problem, repeating the above steps until the operation and maintenance answer is generated.

[0005] In one embodiment of the present disclosure, before receiving the user's operation and maintenance problem about the device, the method further includes: obtaining device material, wherein the device material includes device basic technical material, device operation and maintenance material, and device running and health material; preprocessing the device material to obtain a directory and a first text; generating instruction data based on the directory and the first text, wherein the instruction data includes multiple pairs of questions and answers; and fine-tuning the large language model using the instruction data.

[0006] In one embodiment of the present disclosure, generating instruction data based on the directory and the first text includes: extracting keywords of the first text; performing word segmentation processing on the directory, the keywords and the first text to obtain a second text; performing sentence-level clustering on the second text to obtain multiple text blocks; and generating instruction data based on the multiple text blocks.

[0007] In an embodiment of the present disclosure, the device profile is preprocessed to obtain the catalog and the first text, including: extracting the catalog, text content, table content and pictures of the device profile; generating first descriptive text and second descriptive text of the table content and the pictures respectively; dividing the third text according to the catalog to obtain the fourth text, wherein the third text includes the text content, the first descriptive text and the second descriptive text; performing data cleaning and sentence segmentation on the fourth text to obtain the first text.

[0008] In an embodiment of the present disclosure, before extracting the catalog, text content, table content and pictures of the device profile, the method further includes: determining that a character string with a frequency greater than a preset frequency in the device profile is a header or footer; and deleting the header or footer in the device profile.

[0009] In an embodiment of the present disclosure, after generating the instruction data based on the catalog and the first text, the method further includes: converting the instruction data into an embedding vector to construct a vector database.

[0010] In an embodiment of the present disclosure, the method further includes: receiving a user's operation and maintenance question about the device, generating a user feature vector of the operation and maintenance question, and executing the following loop: searching in the vector database based on the user feature vector to obtain a target question and a similarity; when the similarity is greater than a first preset threshold, inputting the operation and maintenance question into a large language model to output an operation and maintenance answer and exiting the loop; when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated and exiting the loop; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result; if the determination result is that the target question and the operation and maintenance question are the same, inputting the target question into the large language model to output the operation and maintenance answer and exiting the loop; and if the determination result is a new operation and maintenance question, generating a user feature vector of the new operation and maintenance question.

[0011] According to another aspect of the present disclosure, there is provided an operation and maintenance question answering apparatus, comprising: a receiving unit configured to receive a user's operation and maintenance question about a device and generate a user feature vector of the operation and maintenance question; a retrieving unit configured to retrieve in a vector database based on the user feature vector to obtain a target question and a similarity between the target question and the operation and maintenance question; a first processing unit configured to, when the similarity is greater than a first preset threshold, input the operation and maintenance question into a large language model to output an operation and maintenance answer about the device; a second processing unit configured to, when the similarity is less than a second preset threshold, determine that the operation and maintenance answer cannot be generated; a third processing unit configured to, when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtain a determination result of the user; a fourth processing unit configured to, if the determination result is that the target question and the operation and maintenance question are the same, input the target question into the large language model to output the operation and maintenance answer; and a fifth processing unit configured to, if the determination result is a new operation and maintenance question, repeatedly execute the above steps until the operation and maintenance answer is generated.

[0012] According to still another aspect of the present disclosure, there is provided an electronic device, comprising: a processor; and a memory storing executable instructions of the processor; wherein the processor is configured to execute any of the above methods via executing the executable instructions.

[0013] According to yet another aspect of the present disclosure, there is provided a computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements any of the above methods.

[0014] According to yet another aspect of the present disclosure, there is provided a computer program product comprising computer instructions stored in a computer readable storage medium, which, when executed by a processor, implement the operational instructions of any of the above methods.

[0015] In the embodiments of the present disclosure, the user feature vector of the operation and maintenance question is generated, the target question and the similarity are determined based on the user feature vector, the double thresholds of the similarity are set, the operation and maintenance answer containing detailed technical material support is directly generated by calling the large language model when the similarity is high, it is determined that the operation and maintenance answer cannot be generated when the similarity is low, the user confirmation mechanism is introduced when the similarity is medium, and the problem understanding is ensured to be accurate, thereby solving the problems of low intelligent level of device operation and maintenance, and lack of detailed technical material support for operation and maintenance scheme in the related art, and further improving the intelligent level of device operation and maintenance and generating the operation and maintenance scheme containing detailed technical material support.

[0016] It should be understood that the foregoing general description and the following detailed description are only exemplary and explanatory, and are not limiting to the present disclosure. BRIEF DESCRIPTION OF DRAWINGS

[0017] The accompanying drawings, which are incorporated herein and constitute part of this specification, illustrate implementations of the present disclosure and, together with the description, further serve to explain the principles of the present disclosure. It is apparent that the drawing in the following description is only some embodiments of the present disclosure, and other drawings can be obtained according to these drawings without creative labor for those skilled in the art.

[0018] Figure 1 A schematic diagram of an operation and maintenance problem solving system in an embodiment of the present disclosure is shown.

[0019] Figure 2 A flowchart of an operation and maintenance problem solving method in an embodiment of the present disclosure is shown.

[0020] Figure 3 A flowchart of an operation and maintenance model fine-tuning method in an embodiment of the present disclosure is shown.

[0021] Figure 4 A flowchart of an instruction data generation method in an embodiment of the present disclosure is shown.

[0022] Figure 5 A flowchart of a data preprocessing method in an embodiment of the present disclosure is shown.

[0023] Figure 6 A flowchart of another operation and maintenance problem solving method in an embodiment of the present disclosure is shown.

[0024] Figure 7 A schematic diagram of an operation and maintenance problem solving device in an embodiment of the present disclosure is shown.

[0025] Figure 8 A schematic diagram of an electronic device in an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0026] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the implementations set forth herein; rather, these implementations are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example implementations to those skilled in the art. The described features, structures, or characteristics can be combined in one or more implementations.

[0027] In addition, the accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments of the present disclosure and, as such, a change in the drawings should be construed as causing a change in the specification. In the drawings:

[0028] It should be understood that each of the steps recited in the method embodiments of the present disclosure can be performed in different orders and / or in parallel. In addition, the method embodiments can include additional steps and / or omit performing the steps shown. The scope of the present disclosure is not limited in this respect.

[0029] It should be noted that the concepts of "first", "second", etc. mentioned in the present disclosure are only used to distinguish different devices, modules or units, and are not used to limit the order or interdependence of the functions performed by these devices, modules or units.

[0030] It should be noted that the adjectives "one", "multiple" mentioned in the present disclosure are illustrative and not limiting, and those skilled in the art should understand that "one or more" should be understood unless the context clearly indicates otherwise.

[0031] It should be noted that the embodiments of the present disclosure and the technical features in the embodiments can be combined with each other without conflict.

[0032] For ease of understanding, the following first explains several terms related to the present disclosure as follows: PHM (Prognostics and Health Management): A comprehensive management method that evaluates the health status of equipment or systems, diagnoses faults, and predicts the remaining useful life through real-time monitoring, data analysis and modeling techniques, widely used in aerospace, high-end manufacturing and energy industries.

[0033] ANN (Approximate Nearest Neighbor) search: A technique for efficiently finding the most similar items to the query sample in large-scale high-dimensional data sets, which significantly improves search speed and reduces computation and storage overhead by sacrificing a small amount of retrieval accuracy, widely used in vector databases, recommendation systems, image retrieval and semantic search scenarios.

[0034] Large Language Model: A natural language processing model with a large number of parameters, trained on a large amount of text data, with strong language understanding, generation and reasoning capabilities, capable of completing various tasks such as question answering, creation, translation, and code generation based on context.

[0035] SFT (Supervised Fine-Tuning): A model optimization method that uses high-quality small-scale data sets with clear input-output labels to further adjust model parameters based on pre-trained large models through supervised learning, making them perform better in specific tasks or fields.

[0036] Prompt: In large language model interaction, it refers to user input instructions, questions or context information, used to guide the model to generate expected output; good prompt design can significantly improve the accuracy, relevance and controllability of model response, which is the core element of prompt engineering.

[0037] Token: The basic unit of text processing by language models, which can be a word, a word, a punctuation mark or a sub-word unit. The model encodes and calculates the input text after dividing it into a series of tokens. Different languages and tokenization strategies will affect the number and form of tokens.

[0038] Self-Instruct Framework: A self-supervised fine-tuning method that expands training data by generating instruction-response pairs through the model itself. First, use a small amount of seed examples to guide the model to generate new samples, then filter high-quality data for subsequent training, thereby improving the model's instruction-following ability without the need for large amounts of human annotation.

[0039] OCR (Optical Character Recognition): A technology that automatically recognizes and converts text in images (such as scanned documents, photos, screenshots) into editable and searchable text format, widely used in bill processing, document digitization, license plate recognition and other scenarios.

[0040] Zero-shot learning model: A machine learning model that can identify or classify a class without seeing any samples of that class during training, only through class semantic description (such as attributes, text definition). It is suitable for tasks in scenarios where labeled data is scarce or classes are dynamically expanding.

[0041] text2vec-large-chinese: A large-scale text vectorization model for Chinese, capable of mapping Chinese sentences or paragraphs into high-dimensional semantic vectors, preserving rich semantic information, commonly used for text similarity calculation, clustering, retrieval and other tasks.

[0042] Figure 1 A schematic diagram of an operation and maintenance problem solving system in an embodiment of the present disclosure is shown, which includes a receiving module 101, a processing module 102, and a large language model 103.

[0043] The receiving module 101 can receive a user's operation and maintenance problem about a device.

[0044] The processing module 102 can process the operation and maintenance problem received by the receiving module 101, and during the processing, the processing module 102 can call the large language model 103.

[0045] The processing module 102 can be installed with an application program to perform the following: generating a user feature vector of the operation and maintenance problem; searching in a vector database based on the user feature vector to obtain a target problem and a similarity between the target problem and the operation and maintenance problem; when the similarity is greater than a first preset threshold, inputting the operation and maintenance problem into the large language model to output an operation and maintenance answer about the device; when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result of the user; if the determination result is that the target problem and the operation and maintenance problem are the same, inputting the target problem into the large language model to output the operation and maintenance answer; and if the determination result is a new operation and maintenance problem, repeating the above steps until the operation and maintenance answer is generated.

[0046] Figure 2 A flowchart of an operation and maintenance problem solving method in an embodiment of the present disclosure is shown, which includes the following steps as shown in Figure 2 S201, receiving a user's operation and maintenance problem about a device, and generating a user feature vector of the operation and maintenance problem.

[0047] Exemplarily, the operation and maintenance problem is a natural language question raised by the user about the operation, maintenance, or fault handling of the device.

[0048] Exemplarily, the user feature vector is a numerical vector obtained after the operation and maintenance problem is vectorized, which is used to represent the semantic information of the problem in a vector space.

[0049] S202, searching in a vector database based on the user feature vector to obtain a target problem and a similarity between the target problem and the operation and maintenance problem.

[0050] Exemplarily, the vector database is a data structure that stores a plurality of questions and their corresponding vector representations.

[0051] Exemplarily, the target problem is an existing question entry in the vector database that is most similar in semantics to the current operation and maintenance problem.

[0052] ​Exemplarily, the similarity can be calculated by cosine similarity or inner product between the user feature vector and the vector corresponding to the target question.

[0053] Exemplarily, the nearest neighbor algorithm can also be used to calculate the target question closest to the user feature vector in the vector database and the semantic distance between the target question and the operation and maintenance question. The semantic distance can replace the similarity and play the role of the similarity below.

[0054] S203, when the similarity is greater than the first preset threshold, input the operation and maintenance question into the large language model to output the operation and maintenance answer about the equipment.

[0055] Exemplarily, the first preset threshold is an upper threshold for determining that the operation and maintenance question and the target question are highly semantically matched. When the similarity exceeds the value, it is confirmed that the answer can be directly generated.

[0056] Exemplarily, the operation and maintenance answer is a natural language response output by the large language model for the operation and maintenance question, and the content includes diagnostic basis, operation steps, maintenance suggestions, and supported technical materials.

[0057] S204, when the similarity is less than the second preset threshold, it is determined that the operation and maintenance answer cannot be generated.

[0058] Exemplarily, the second preset threshold is a lower threshold for determining that the operation and maintenance question and the target question are not semantically matched. When the similarity is lower than the value, it is determined that an effective answer cannot be generated.

[0059] S205, when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, the user's determination result is obtained.

[0060] Exemplarily, the determination result is the confirmation feedback made by the user on whether the target question is equal to the operation and maintenance question of the user. The determination result includes the target question and the same or new operation and maintenance question.

[0061] S206, if the determination result is that the target question and the operation and maintenance question are the same, input the target question into the large language model to output the operation and maintenance answer.

[0062] S207, if the determination result is a new operation and maintenance question, repeat the above steps until the operation and maintenance answer is generated.

[0063] In this embodiment, a user feature vector of the operation and maintenance problem is generated; based on the user feature vector, retrieval is performed in the vector database to obtain a target problem and a similarity; a double threshold of the similarity is set, and then when the similarity is high, a large language model is directly called to generate an operation and maintenance answer containing detailed technical material support, when the similarity is low, it is determined that the operation and maintenance answer cannot be generated, and when the similarity is medium, a determination result of the user is obtained; if the user feedback target problem and the operation and maintenance problem are the same, the target problem is input into the large language model, and an operation and maintenance answer is output; if the user inputs a new operation and maintenance problem, the above steps are repeatedly executed until an operation and maintenance answer is generated. The problems that the intelligent level of device operation and maintenance is low and the operation and maintenance scheme lacks detailed technical material support in the related art are solved, and the intelligent level of device operation and maintenance is improved, and an operation and maintenance scheme containing detailed technical material support is generated.

[0064] Figure 3 A flowchart of a method for fine-tuning an operation and maintenance model in an embodiment of the present disclosure is shown. As shown in Figure 3 , the method comprises the following steps: S301, obtaining device material, wherein the device material comprises device basic technical material, device operation and maintenance material, and device running and health material; S302, preprocessing the device material to obtain a directory and a first text; S303, generating instruction data based on the directory and the first text, wherein the instruction data comprises multiple pairs of questions and answers; S304, fine-tuning a large language model using the instruction data.

[0065] Exemplarily, the device material is a collection of information for supporting device operation and maintenance decision and knowledge service.

[0066] Exemplarily, the device basic technical material is document content describing the structure, principle, technical parameter and design specification of the device.

[0067] Exemplarily, the device operation and maintenance material is a procedural document guiding users on how to operate the device, perform daily maintenance and handle faults.

[0068] Exemplarily, the device running and health material is a collection of data recording state data, diagnosis results, early warning logs and maintenance closed-loop information generated during actual operation of the device.

[0069] Exemplarily, preprocessing is a process of text extraction, header and footer removal, directory identification, content cleaning and sentence segmentation on the device material, aiming to generate a structured first text.

[0070] Exemplarily, the directory is a chapter title hierarchical structure extracted from the device material, used for logical segmentation of the main text content.

[0071] Exemplarily, the first text is a clean text content obtained after preprocessing and organized by directory, serving as a basic material for subsequent instruction data generation.

[0072] Exemplarily, the instruction data is a training sample set composed of multiple pairs of questions and answers, used for supervised fine-tuning of the large language model.

[0073] In this embodiment, after obtaining the device data, the device data is preprocessed to obtain a directory and a first text, and instruction data is generated based on the directory and the first text. The fine-tuned large language model is fine-tuned using the instruction data. Through the above technical means, the large language model can refer to real, accurate and structured technical data when answering operation and maintenance questions, thereby ensuring that the operation and maintenance answers are professional and traceable.

[0074] In one exemplary embodiment, fine-tuning the large language model using the instruction data includes: organizing each pair of question and answer in the instruction data into an "input-target output" format, updating the parameters of the large language model by minimizing the cross-entropy loss of the model in the answer part; during the training process, the question part is used as a prompt, and the answer part is used as an expected generated sequence, and the model only calculates the loss for the token corresponding to the answer.

[0075] In an optional embodiment, a device knowledge graph is constructed, wherein the device knowledge graph contains semantic relationships between device entities, failure modes, maintenance actions and technical parameters; the device knowledge graph is converted into natural language description triples; the instruction data is generated based on the natural language description triples and historical operation and maintenance work orders; the large language model is fine-tuned using the instruction data, and the generation process is enhanced in the reasoning stage by combining graph retrieval. Through the above technical means, the answering ability of the large language model for complex causal relationships and multi-hop reasoning problems is improved.

[0076] Figure 4 A flowchart of a method for generating instruction data in an embodiment of the present disclosure is shown. As shown in Figure 4 , the method comprises the following steps: S401, extracting keywords from the first text; S402, performing word segmentation processing on the directory, the keywords and the first text to obtain a second text; S403, performing sentence-level clustering on the second text to obtain multiple text blocks; S404, generating instruction data based on the multiple text blocks.

[0077] Exemplarily, the keywords are words or phrases extracted from the first text that can represent the core semantics of the paragraph, which can be extracted using a zero-shot learning model.

[0078] Exemplarily, the word segmentation processing is an operation of segmenting the catalog, the keywords, and the first text by using a Chinese word segmentation tool to obtain a second text composed of a sequence of words.

[0079] Exemplarily, the second text is a structured text representation composed of words obtained after the word segmentation processing, and is used for subsequent clustering analysis.

[0080] Exemplarily, the sentence-level clustering is an operation of clustering sentences in the second text according to semantic similarity, aiming to classify sentences with consistent topics into the same category.

[0081] Exemplarily, the text block is a text unit with coherent semantics and focused topics formed after the sentence-level clustering, and each text block contains a plurality of semantically related sentences.

[0082] In this embodiment, the word segmentation processing is performed on the catalog, the keywords, and the first text to obtain the second text, the sentence-level clustering is performed on the second text to obtain a plurality of text blocks, and finally the instruction data is generated based on the plurality of text blocks. Through the above technical means, it is ensured that the generated instruction data covers complete technical topics and has semantic cohesion, thereby improving the accuracy and consistency of the fine-tuned large language model in answering operation and maintenance problems.

[0083] In one exemplary embodiment, generating the instruction data based on the plurality of text blocks can be to use the Self-Instruct framework to guide the large language model to automatically generate high-quality instruction-answer pairs. First, a small number of seed examples written by humans are selected from the plurality of text blocks, and each example contains a typical operation and maintenance problem and an accurate answer generated based on the content of the text block; then a prompt template is constructed to instruct the large language model to simulate different roles (such as operators, maintenance technicians, and review experts) to ask diversified questions based on the text block as a knowledge source, and to generate reference answers strictly according to the content of the text block; inputting the prompt and a text block into the large language model to generate new instruction-answer pairs in batches; filtering the generated results to remove samples that do not reference the original text, have factual errors, or have inconsistent formats; retaining high-quality samples as new instruction data, and adding them to the seed set for iterative expansion; and finally forming an instruction data set covering task types such as parameter query, fault diagnosis, operation guidance, and training examination, which is used for subsequent fine-tuning of the large language model.

[0084] In an optional embodiment, the instruction data is generated based on the catalog and the first text, including: using a large language model to generate an abstract of the first text to obtain a simplified technical description; assigning a task type label to each abstract in combination with the catalog hierarchy, the task type label including parameter query, fault diagnosis, operation guidance, and training examination; constructing a corresponding instruction template according to the task type label, and filling the abstract into the template to generate multiple pairs of questions and answers; and performing deduplication and fact consistency verification on the generated instruction data. Through the above technical means, the task diversity and field adaptability of the instruction data are improved.

[0085] Figure 5 A flowchart of a data preprocessing method in an embodiment of the present disclosure is shown, which includes the following steps as shown in Figure 5 S501, extracting a catalog, text content, table content, and pictures of the equipment data; S502, generating first descriptive text and second descriptive text of the table content and the pictures, respectively; S503, dividing the third text according to the catalog to obtain the fourth text, wherein the third text includes the text content, the first descriptive text, and the second descriptive text; S504, performing data cleaning and sentence segmentation on the fourth text to obtain the first text.

[0086] Exemplarily, the text content is original literal information extracted from the equipment data.

[0087] Exemplarily, the table content is a structured data table identified and extracted from the equipment data.

[0088] Exemplarily, the pictures are image elements extracted from the equipment data, including schematic diagrams, waveform diagrams, assembly diagrams, or interface screenshots, etc.

[0089] Exemplarily, the first descriptive text is a natural language description generated after semantic conversion of the table content, used to express key information in the table.

[0090] Exemplarily, the second descriptive text is a textual description generated by manual annotation or in combination with OCR and semantic understanding of the pictures, used to describe the technical meaning expressed by the pictures.

[0091] Exemplarily, the third text is a complete text set formed by merging the text content, the first descriptive text, and the second descriptive text.

[0092] Exemplarily, the fourth text is a paragraph set corresponding to each chapter title obtained by logically dividing the third text according to the catalog structure.

[0093] ​Exemplarily, the data cleaning is an operation of identifying and removing special characters, extra spaces, line breaks, non-semantic noise and the like contained in the fourth text, and a regular matching algorithm machine can be used for data cleaning.

[0094] Exemplarily, the sentence segmentation is an operation of cutting the fourth text after data cleaning into independent sentences according to Chinese semantic boundaries, ensuring that each sentence expresses a complete semantic unit, and providing a basic granularity for subsequent word segmentation processing and clustering analysis.

[0095] In this embodiment, the directory, the text content, the table content and the picture of the equipment material are extracted, the first descriptive text of the table content and the second descriptive text of the picture are generated, the text content, the first descriptive text and the second descriptive text are divided according to the directory to obtain the fourth text, and finally the data cleaning and the sentence segmentation are performed on the fourth text to obtain the first text. Through the above technical means, the unified structured processing of the multi-modal equipment material is realized, the key information in the table and the picture can be included in the subsequent instruction generation and model fine-tuning process, and thus the coverage capability of the operation and maintenance answer to the non-pure text technical material is enhanced.

[0096] In an optional embodiment, the equipment material is preprocessed to obtain the directory and the first text, including: automatically identifying the directory items based on the document hierarchical labels, and constructing a directory tree; mapping all extracted contents to structured segments according to the directory tree nodes; performing special character filtering, line break normalization and term standardization based on a domain dictionary on each segment; and finally using a rule-guided sentence boundary detector to complete the sentence segmentation and output the first text. Through the above technical means, the robustness and semantic fidelity of the preprocessing of complex format documents (such as scanned manuals) are improved.

[0097] In an embodiment of the present disclosure, before extracting the directory, the text content, the table content and the picture of the equipment material, the method further includes: determining a string with a frequency of occurrence greater than a preset frequency in the equipment material as a header or a footer; and deleting the header or the footer in the equipment material.

[0098] Exemplarily, the header is a repetitive text area at the top of each page in the equipment material, which usually contains document titles, chapter names or version information, etc.

[0099] Exemplarily, the footer is a repetitive text area at the bottom of each page in the equipment material, which usually contains page numbers, company logos or confidentiality statements, etc.

[0100] In this embodiment, the frequency of occurrence of the string in the equipment material is counted, and the string with the frequency of occurrence greater than the preset frequency is determined as the header or the footer. Through the above technical means, the text noise caused by the repeated appearance of the header and the footer is eliminated, and the accuracy of the subsequent vectorization representation is avoided.

[0101] In an optional embodiment, the candidate header and footer regions are identified based on the page position and font features; the cross-page co-occurrence rate of the text in each candidate region in the full text is counted; the text with a co-occurrence rate higher than a preset frequency and a stable position is marked as a header or footer; the page number mode is filtered in combination with regular rules; the marked header and footer content is removed from the original text stream, and the remaining body text is reserved for subsequent processing. Through the above technical means, the recognition accuracy of the header and footer in a document with complex format or non-standard layout is improved.

[0102] In an embodiment of the present disclosure, after generating the instruction data based on the directory and the first text, the method further comprises: converting the instruction data into embedding vectors to construct a vector database.

[0103] Illustratively, the embedding vector is a numerical representation obtained by mapping the question text in the instruction data to a high-dimensional semantic space through a vector encoding model.

[0104] In this embodiment, the questions in the instruction data are converted into embedding vectors, and the questions and embedding vectors are saved to the vector database. Through the above technical means, when the user operation and maintenance question is received, the system can quickly retrieve the most relevant known question based on semantic similarity, thereby providing accurate context support for subsequent answer generation.

[0105] In an optional embodiment, the text2vec-large-chinese is used to convert the questions in the instruction data into embedding vectors; each embedding vector is indexed and associated with the corresponding question; all embedding vectors are loaded into the vector database; and the vector database is periodically updated to include newly generated instruction data. Through the above technical means, it is ensured that the vector retrieval stage efficiently returns the high-relevance historical question and its authoritative answer as the basis for generation.

[0106] Figure 6 A flowchart of another operation and maintenance question answering method in an embodiment of the present disclosure is shown. As shown in Figure 6 the flowchart, the method comprises the following steps: S601, receiving a user operation and maintenance question about a device, generating a user feature vector of the operation and maintenance question, and executing the following loop: S602, performing retrieval in the vector database based on the user feature vector to obtain a target question and a similarity; S603, when the similarity is greater than a first preset threshold, inputting the operation and maintenance question into a large language model to output an operation and maintenance answer, and exiting the loop; S604, when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated, and exiting the loop; S605, when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result. S606, if the determination result is that the target problem is the same as the operation and maintenance problem, input the target problem into the large language model, output the operation and maintenance answer, and exit the loop; S607, if the determination result is a new operation and maintenance problem, generate a user feature vector of the new operation and maintenance problem.

[0107] In this embodiment, the operation and maintenance problems raised by the user are processed through a loop mechanism: first, the problem is converted into a user feature vector, and the most similar target problem and its similarity are retrieved in the vector database; if the similarity is higher than a first preset threshold, the original problem is directly used to drive the large language model to generate an operation and maintenance answer; if the similarity is lower than a second preset threshold, it is determined that there is no effective answer and the process is terminated; if the similarity is between the two thresholds, the user is requested to confirm whether the target problem is the same as his intention, and if the confirmation is consistent, the target problem is used to generate an answer, otherwise the user's modified new problem is re-vectorized and enters the retrieval loop again until an answer is output or a clear refusal is given. Through the above technical means, interactive clarification and accurate matching of user's fuzzy or biased problems are realized, and the generation of wrong answers due to semantic ambiguity is avoided, while the safety refusal ability of the system in uncertain scenarios is guaranteed.

[0108] In one application scenario, taking the train equipment early warning and health management (PHM) scenario as an example: In the preparation stage, obtain train equipment information, including equipment basic technical information (such as traction motor technical specification), equipment operation and maintenance information (such as temperature early warning disposal process in HXD3C maintenance manual), and equipment operation and health information (such as historical temperature early warning log and maintenance record); then preprocess these information: identify and delete page headers (such as "HXD3C locomotive maintenance manual") and page footers (such as "page X") with a frequency higher than a preset frequency, extract the directory, text content, table content and pictures; convert the table content into the first descriptive text (such as the allowable temperature rise limit of the traction motor is 120K), and generate the second descriptive text from the pictures (such as the HMI interface displays the "traction motor 1 temperature high" warning); merge the text content, the first descriptive text and the second descriptive text into the third text, and cut the third text into the fourth text according to the directory structure; perform data cleaning (remove special symbols and redundant spaces) and sentence segmentation (cut according to Chinese sentence boundaries) on the fourth text to obtain the first text. Then, extract the keywords of the first text, perform word segmentation on the directory, the keywords and the first text to obtain the second text, and perform sentence-level clustering on the second text to form a plurality of semantically coherent text blocks; based on these text blocks, use the Self-Instruct framework combined with a large language model to generate instruction data composed of multiple pairs of questions and answers, for example, the question "how to handle the traction motor temperature anomaly" corresponds to the answer "1. Check if the cooling fan is running; 2. Check the HMI warning code; 3. If the temperature continues to be too high, request to run at reduced power … According to: Section 4.3 of HXD3C Maintenance Manual"; use the instruction data to supervise and fine-tune the large language model; and convert the question part in the instruction data into an embedding vector to construct a vector database.

[0109] Receive the operation and maintenance question "how to handle the traction motor temperature anomaly", generate its user feature vector, and retrieve the target question "traction motor high temperature early warning disposal steps" and the similarity 0.82 in the vector database; since the similarity is greater than or equal to the second preset threshold (0.3) and less than the first preset threshold (0.85), the target question is displayed to the user and the user is requested to confirm; the user confirms that they are the same, inputs the target question into the fine-tuned large language model, and the model outputs an operation and maintenance answer containing specific operation steps, reference sources and safety prompts based on the context of the knowledge base, completing the entire processing flow.

[0110] Based on the same inventive concept, the operation and maintenance question answering device is also provided in the embodiments of the present disclosure, as follows. Since the operation and maintenance question answering device solves problems in a similar way to the above-mentioned method embodiments, the implementation of the operation and maintenance question answering device can be referred to the implementation of the above-mentioned method embodiments, and the repeated parts will not be described again.

[0111] Figure 7 An operation and maintenance question answering device is shown in the embodiments of the present disclosure, as follows.Figure 7 As shown, the converged service management network element can include: The receiving unit 701 is configured to receive a user's operation and maintenance question about a device, and generate a user feature vector of the operation and maintenance question. The searching unit 702 is configured to search in a vector database based on the user feature vector, to obtain a target question and a similarity between the target question and the operation and maintenance question. The first processing unit 703 is configured to, when the similarity is greater than a first preset threshold, input the operation and maintenance question into a large language model, and output an operation and maintenance answer about the device. The second processing unit 704 is configured to, when the similarity is less than a second preset threshold, determine that the operation and maintenance answer cannot be generated. The third processing unit 705 is configured to, when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtain a determination result of the user. The fourth processing unit 706 is configured to, if the determination result is that the target question and the operation and maintenance question are the same, input the target question into the large language model, and output the operation and maintenance answer. The fifth processing unit 707 is configured to, if the determination result is a new operation and maintenance question, repeatedly execute the above steps until the operation and maintenance answer is generated.

[0112] In some embodiments, the first processing unit 703 is further configured to obtain device data, wherein the device data includes device basic technical data, device operation and maintenance data, and device running and health data; preprocess the device data to obtain a catalog and a first text; generate instruction data based on the catalog and the first text, wherein the instruction data includes multiple pairs of questions and answers; and fine-tune the large language model using the instruction data.

[0113] In some embodiments, the first processing unit 703 is further configured to extract keywords of the first text; perform word segmentation processing on the catalog, the keywords, and the first text to obtain a second text; perform sentence-level clustering on the second text to obtain multiple text blocks; and generate the instruction data based on the multiple text blocks.

[0114] In some embodiments, the first processing unit 703 is further configured to extract a catalog, text content, table content, and pictures of the device data; generate first descriptive text and second descriptive text of the table content and the pictures, respectively; divide a third text according to the catalog to obtain a fourth text, wherein the third text includes the text content, the first descriptive text, and the second descriptive text; and perform data cleaning and sentence segmentation on the fourth text to obtain the first text.

[0115] In some embodiments, the first processing unit 703 is further configured to determine a string with a frequency of occurrence in the device profile greater than a preset frequency as a header or a footer; and delete the header or the footer in the device profile.

[0116] In some embodiments, the first processing unit 703 is further configured to convert the instruction data into embedding vectors to construct a vector database.

[0117] In some embodiments, the fifth processing unit 707 is further configured to receive a user's operation and maintenance question about the device, generate a user feature vector of the operation and maintenance question, and perform a loop: search in the vector database based on the user feature vector to obtain a target question and a similarity; when the similarity is greater than a first preset threshold, input the operation and maintenance question into the large language model to output an operation and maintenance answer and exit the loop; when the similarity is less than a second preset threshold, determine that the operation and maintenance answer cannot be generated and exit the loop; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtain a determination result; if the determination result is that the target question and the operation and maintenance question are the same, input the target question into the large language model to output the operation and maintenance answer and exit the loop; and if the determination result is a new operation and maintenance question, generate a user feature vector of the new operation and maintenance question.

[0118] Those skilled in the art can understand that various aspects of the present disclosure can be implemented as a system, a method or a program product. Therefore, various aspects of the present disclosure can be embodied as a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or an implementation combining hardware and software aspects, which can be collectively referred to as "circuitry", "module" or "system" herein.

[0119] The electronic device 800 according to this implementation of the present disclosure will be described below with reference to Figure 8 Figure 8 The displayed electronic device 800 is merely an example and should not impose any limitation on the functions and use range of the embodiments of the present disclosure.

[0120] As shown in Figure 8 The electronic device 800 is in the form of a general computing device. The components of the electronic device 800 can include, but are not limited to, at least one processor 810, at least one memory 820, and a bus 830 connecting different system components (including the memory 820 and the processor 810).

[0121] ​The memory stores program codes which can be executed by the processor 810 to enable the processor 810 to perform the steps described in the above “Exemplary Method” section according to various exemplary embodiments of the present disclosure. For example, the processor 810 can perform the following steps of the above method embodiments: generating a user feature vector of the operation and maintenance question; performing retrieval in the vector database based on the user feature vector to obtain a target question and a similarity between the target question and the operation and maintenance question; when the similarity is greater than a first preset threshold, inputting the operation and maintenance question into the large language model to output an operation and maintenance answer about the device; when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result of the user; if the determination result is that the target question and the operation and maintenance question are the same, inputting the target question into the large language model to output the operation and maintenance answer; and if the determination result is a new operation and maintenance question, repeating the above steps until the operation and maintenance answer is generated.

[0122] The memory 820 can include a readable medium in the form of volatile memory, such as a random access memory (RAM) 8201 and / or a cache memory 8202, and can further include a read-only memory (ROM) 8203.

[0123] The memory 820 can further include program / utility 8204 having a set of programs / modules 8205, including but not limited to, an operating system, one or more applications, other program modules, and program data, and each of these examples, or some combination thereof, can include implementation of a network environment.

[0124] The bus 830 can represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration bus, a processor or local bus using any of a variety of bus architectures.

[0125] The electronic device 800 can also communicate with one or more external devices 840 such as a keyboard or pointing device, a Bluetooth device, or a database via I / O interface 850. The communication can be facilitated via a network adapter 860. As depicted, the network adapter 860 is in communication with the other components of the electronic device 800 through bus 830. It should be appreciated that the network adapter 860 can be implemented as part of the electronic device 800 or can be a separate and unique component. It is also contemplated that the network adapter 860 can be implemented as multiple network adapters, such as a first network adapter for communicating with any external devices 840 and a second network adapter for communicating with one or more networks. In addition, it is contemplated that the network adapter 860 can be implemented as any wireless or wired communication interface, including, but not limited to, a modem, a LAN port, a wireless port, a communication port, a serial port, or a USB port.

[0126] In the disclosed exemplary embodiments, a computer readable storage medium is also provided, which can be a readable signal medium or a readable storage medium.

[0127] In some possible implementation manners, various aspects of the present disclosure can also be implemented as a program product, which includes a program code for causing an end device to perform the steps described in the above “specific embodiments” section according to various exemplary embodiments of the present disclosure when the program product is run on the end device.

[0128] More specific examples of the computer readable storage medium in the present disclosure can include, but are not limited to, an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any appropriate combination of the foregoing.

[0129] In the present disclosure, the computer readable storage medium can include a data signal carried in a baseband or as part of a carrier wave propagating through a transmission medium, in which the readable program code is carried. Such a propagated data signal can take a variety of forms, including, but not limited to, an electromagnetic signal, an optical signal, or any appropriate combination thereof. The readable signal medium can also be any readable medium that can be used to carry, propagate, or transmit the program for use by or in connection with an instruction execution system, apparatus, or device.

[0130] Optionally, program code embodied on a computer readable storage medium can be transmitted by way of any appropriate medium, including, but not limited to wireless, wired, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0131] In particular embodiments, the program code implementing the methods of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, or the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the remote computing device, as a stand-alone software package, partly on the user's computing device and partly on the remote computing device, or entirely on the remote or server. In the latter scenario, the remote computing device can be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device such as through the Internet using an Internet Service Provider. In some embodiments, electronic circuitry including, for example, programmable logic circuitry, application specific circuitry, or field programmable gate array (FPGA) circuitry can execute the program code. In some embodiments, multiple computing devices can be used to execute the program code.

[0132] The computer program product or computer program of the present disclosure includes computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the computer device executes the operation and maintenance problem solving method provided in any of the various optional manners in the embodiments of the present disclosure.

[0133] It should be noted that although several modules or units of devices for action execution are mentioned in the foregoing detailed description, such division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into a plurality of modules or units.

[0134] In addition, although the various steps of the methods of the present disclosure are described in a particular order in the accompanying drawings, this does not require or imply that the steps must be performed in this particular order, or that all of the steps shown must be performed to achieve the desired result. Additionally or alternatively, certain steps can be omitted, a plurality of steps can be combined into one step, one step can be divided into a plurality of steps, and the like.

[0135] Through the description of the above embodiments, those skilled in the art can easily understand that the example embodiments described herein can be implemented by software, or by software in combination with necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash disk, a mobile hard disk, or the like) or a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a mobile terminal, or a network device, etc.) execute the methods according to the embodiments of the present disclosure.

[0136] Other embodiments of the present disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the features disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure following the general principles thereof and including such departures from the present disclosure that come within known use or custom in the art to which the present disclosure pertains. The specification and examples are to be regarded as illustrative only, and the true scope of the present disclosure is indicated by the appended claims.

Claims

1. An operation and maintenance problem solving method, characterized by, The method comprises the following steps: receiving a user's operation and maintenance problem about a device, and generating a user feature vector of the operation and maintenance problem; based on the user feature vector, searching in a vector database to obtain a target problem and a similarity between the target problem and the operation and maintenance problem; when the similarity is greater than a first preset threshold, inputting the operation and maintenance problem into a large language model to output an operation and maintenance answer about the device; when the similarity is less than a second preset threshold, determining that the operation and maintenance answer cannot be generated; when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtaining a determination result of the user; if the determination result is that the target problem and the operation and maintenance problem are the same, inputting the target problem into the large language model to output the operation and maintenance answer; if the determination result is a new operation and maintenance problem, repeating the above steps until the operation and maintenance answer is generated.

2. The method of claim 1, wherein, Before the step of receiving a user's operation and maintenance problem about a device, the method further comprises: obtaining device data, wherein the device data comprises device basic technical data, device operation and maintenance data, and device operation and health data; preprocessing the device data to obtain a directory and a first text; generating instruction data based on the directory and the first text, wherein the instruction data comprises multiple pairs of questions and answers; fine-tuning the large language model using the instruction data.

3. The method of claim 2, wherein, The step of generating instruction data based on the directory and the first text comprises: extracting keywords from the first text; performing word segmentation processing on the directory, the keywords, and the first text to obtain a second text; performing sentence-level clustering on the second text to obtain multiple text blocks; generating the instruction data based on the multiple text blocks.

4. The method of claim 2, wherein, The step of preprocessing the device data to obtain a directory and a first text comprises: extracting a directory, text content, table content, and pictures from the device data; generating first descriptive text and second descriptive text for the table content and the pictures, respectively; dividing a third text into fourth text according to the directory, wherein the third text comprises the text content, the first descriptive text, and the second descriptive text; performing data cleaning and sentence segmentation on the fourth text to obtain the first text.

5. The method of claim 4, wherein, Before the step of extracting a directory, text content, table content, and pictures from the device data, the method further comprises: determining that a string with a frequency greater than a preset frequency in the device data is a header or a footer; deleting the header or the footer in the device data.

6. The method of claim 2, wherein, After the step of generating instruction data based on the directory and the first text, the method further comprises: converting the instruction data into an embedding vector to construct a vector database.

7. The method of claim 1, wherein, The method further comprises: receiving a user's operation and maintenance problem about a device, and generating a user feature vector of the operation and maintenance problem, and performing the following loop: based on the user feature vector, searching in the vector database to obtain the target problem and the similarity; when the similarity is greater than the first preset threshold, inputting the operation and maintenance problem into a large language model to output the operation and maintenance answer, and exiting the loop; When the similarity is less than the second preset threshold, it is determined that the operation and maintenance answer cannot be generated, and the loop is exited. When the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, the determination result is obtained. If the determination result is that the target question and the operation and maintenance question are the same, the target question is input into the large language model, the operation and maintenance answer is output, and the loop is exited. If the determination result is a new operation and maintenance question, a user feature vector of the new operation and maintenance question is generated. 8.An operation and maintenance problem solving apparatus characterized by comprising: Comprise: a receiving unit configured to receive a user's operation and maintenance question about a device and generate a user feature vector of the operation and maintenance question; a retrieving unit configured to retrieve in a vector database based on the user feature vector to obtain a target question and a similarity between the target question and the operation and maintenance question; a first processing unit configured to, when the similarity is greater than a first preset threshold, input the operation and maintenance question into a large language model and output an operation and maintenance answer about the device; a second processing unit configured to, when the similarity is less than a second preset threshold, determine that the operation and maintenance answer cannot be generated; a third processing unit configured to, when the similarity is greater than or equal to the second preset threshold and less than or equal to the first preset threshold, obtain a user's determination result; a fourth processing unit configured to, if the determination result is that the target question and the operation and maintenance question are the same, input the target question into the large language model and output the operation and maintenance answer; a fifth processing unit configured to, if the determination result is a new operation and maintenance question, repeatedly execute the above steps until the operation and maintenance answer is generated.

9. An electronic device, comprising: Comprise: a processor; and a memory for storing executable instructions of the processor; wherein the processor is configured to execute the method of any one of claims 1-7 by executing the executable instructions.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-7. The computer program is executed by the processor to implement the method of any one of claims 1-7.

Images