CN-121996676-A - Intelligent question-answering auxiliary method for fault diagnosis

CN121996676ACN 121996676 ACN121996676 ACN 121996676ACN-121996676-A

Abstract

The application provides an intelligent question-answering auxiliary method for fault diagnosis, which comprises the steps of constructing a multi-mode input processing module, supporting text and voice input modes by the multi-mode input processing module, realizing flexible collection of fault information, providing comprehensive and accurate initial data for follow-up intelligent retrieval and question-answering, constructing a vector knowledge base supporting efficient retrieval by taking maintenance domain knowledge as a core, providing a solid knowledge base for intelligent question-answering, constructing an intelligent Agent decision module, carrying out user intention recognition, task scheduling and data calling on the basis of the intelligent Agent decision module, combining semantic retrieval with a generated question-answering model to obtain a retrieval enhancement generated question-answering module, realizing multi-round reasoning and knowledge deduction capability of complex problems, constructing a multi-mode feedback module, and converting an intelligent question-answering result into text or voice form feedback by the multi-mode feedback module to adapt to maintenance scene requirements.

Inventors

HUANG ZHUN
YAN ZHONGSHAN
YANG CHAODONG
LIU PAN
CHEN HAIYANG
Wan Yinwei
ZHOU JUN

Assignees

中国飞行试验研究院

Dates

Publication Date: 20260508
Application Date: 20251227

Claims (8)

1. An intelligent question-answering auxiliary method for fault diagnosis is characterized by comprising the following steps: Step 1, constructing a multi-mode input processing module, wherein the multi-mode input processing module supports two input modes of text and voice, realizes flexible collection of fault information, and provides comprehensive and accurate initial data for subsequent intelligent retrieval and question answering; Step 2, constructing a vector knowledge base supporting efficient retrieval by taking maintenance domain knowledge as a core, and providing a solid knowledge base for intelligent question and answer; Step 3, an intelligent Agent decision module is constructed, and user intention recognition, task scheduling and data calling are carried out based on the intelligent Agent decision module; step 4, combining semantic retrieval with a generated question-answering model to obtain a retrieval enhancement generated question-answering module, and realizing multi-round reasoning and knowledge deduction capability of complex problems; And 5, constructing a multi-mode feedback module, wherein the multi-mode feedback module converts the intelligent question-answering result into text or voice form for feedback, and adapts to the maintenance scene requirement.
2. The method according to claim 1, wherein the step 1 comprises: Text input processing, namely receiving fault description text directly input by a user; the voice input process, which integrates the automatic voice recognition technology to convert the fault information dictated by the user into characters; The input intention unified conversion is that the fault information input by words or voices is standardized into a structured query statement, and the structured query statement comprises core fields of 'equipment type', 'fault phenomenon', 'fault part', 'running state', and a unified and standard format is provided for subsequent vector knowledge base retrieval, so that the accuracy and high efficiency of retrieval are ensured.
3. The method according to claim 1, wherein the step 2 comprises: Collecting and cleaning knowledge, namely collecting an equipment maintenance manual, a historical fault case, manufacturer technical documents, maintenance rules and industry standards, performing deduplication processing on the collected data, performing format unification, converting documents with different formats into unified text formats, and performing structural processing on form data; Cutting and vectorizing the document, namely cutting the processed document into 500-1000 word fragments by adopting a sliding window method, and converting each fragment into 768 dimension vectors by a pre-training model; vector storage and optimization, namely storing vector data into a Chroma vector database, adopting ONNX Runtime deployment, supporting GPU batch reasoning, and greatly improving the retrieval accuracy and throughput of the model; And dynamically updating the knowledge base, namely setting a periodic updating mechanism, automatically incorporating new maintenance cases, technical documents and industry standards, and updating the vector model through incremental training.
4. The method according to claim 1, wherein the step 3 comprises: The intention recognition engine is used for judging multi-classification intention of the input query based on a fine-tuned BERT/RoBERTa model and recognizing whether the input query is a knowledge retrieval, data calling or flow control type request; The semantic judgment engine is used for deconstructing and mapping tasks on the complex semantic structure by combining the context, the entity label and the semantic vector alignment information; The data calling and knowledge matching engine is used for automatically generating a structured query statement by the Agent when judging the intention of data calling, dynamically scheduling the operation parameters of equipment, maintenance logs and fault statistics business data; And unifying abstract data interfaces, namely solving the difference between semantic mapping and table structure of multi-source data and improving the universality of data calling.
5. The method according to claim 1, wherein the step 4 comprises: The multi-stage retrieval mechanism is that keyword and semantic mixed retrieval is realized through an elastic search, the confidence coefficient is more than 0.75, semantic neighbor search top-k=2 is executed through a vector database, double retrieval results are fused for embedding reordering, and top-3 optimal paragraphs are reserved; LLM reasoning enhancement, namely inputting the paragraphs as context prompt words into a generating type large model, executing multi-round question-answer generation, outputting structured contents and covering the dimension of 'cause analysis-maintenance suggestion-notice'; and (3) unknown fault processing mechanism, namely when the vector library fails to match the high-similarity paragraph, the system is retracted to a large model internal knowledge generation module, a speculative solution is built autonomously, and a reference suggestion mark is added to prompt manual review.
6. The method according to claim 1, wherein the step 5 comprises: text feedback, namely displaying answers in the form of a structured card, highlighting key steps and important information, and supporting image-text combination; The voice feedback is that integrating text-to-voice technology, converting text answers into natural voice, supporting speech speed adjustment and key rereading, adapting to the external requirement of noisy environment; and the feedback form is self-adaptive, namely the feedback form is automatically recommended according to the input mode, and meanwhile, the manual switching of a user is supported.
7. The method of claim 2, wherein the text input processing further comprises: Redundant information is removed through word segmentation, entity identification and semantic disambiguation operation, and core content of fault description is defined.
8. The method of claim 2, wherein the voice input process further comprises: And taking the noisy environment existing in the maintenance site into consideration, optimizing the recognition accuracy by adopting a noise suppression algorithm, and simultaneously supporting the self-adaptive recognition of dialects and technical terms.

Description

Intelligent question-answering auxiliary method for fault diagnosis Technical Field The application belongs to the technical field of intelligent diagnosis and artificial intelligence intersection, and particularly relates to an intelligent question-answering auxiliary method for fault diagnosis. Background At present, equipment fault diagnosis mainly depends on experience accumulation of maintenance personnel and manual inquiry data, and has the following typical problems: (1) The information input is single, and the complex field environment (such as noisy, shielding and the like) is difficult to adapt; (2) Maintenance knowledge is distributed in heterogeneous carriers such as manuals, databases, historical cases and the like, cross-source retrieval efficiency is low, and complex pattern matching is difficult; (3) Most of output results are in a text form, lack of intuitiveness and interactivity, and seriously affect maintenance efficiency and accuracy; (4) The lack of reasoning about unrecorded new faults does not provide effective assistance. The prior art is lack of a set of intelligent question-answering method integrating multi-mode input, semantic understanding, cross-source knowledge fusion and generation type reasoning, and is difficult to meet the requirements of quick, accurate and efficient maintenance and guarantee of modern complex equipment. Disclosure of Invention The invention aims to provide an intelligent question-answering auxiliary method for fault diagnosis, which realizes high-efficiency input, accurate matching and multi-form feedback of fault information by combining multi-mode input processing, vector knowledge base construction and generation type large model reasoning and assists maintenance personnel to quickly complete fault diagnosis and maintenance operation. The application provides an intelligent question-answering auxiliary method for fault diagnosis, which comprises the following steps: Step 1, constructing a multi-mode input processing module, wherein the multi-mode input processing module supports two input modes of text and voice, realizes flexible collection of fault information, and provides comprehensive and accurate initial data for subsequent intelligent retrieval and question answering; Step 2, constructing a vector knowledge base supporting efficient retrieval by taking maintenance domain knowledge as a core, and providing a solid knowledge base for intelligent question and answer; Step 3, an intelligent Agent decision module is constructed, and user intention recognition, task scheduling and data calling are carried out based on the intelligent Agent decision module; step 4, combining semantic retrieval with a generated question-answering model to obtain a retrieval enhancement generated question-answering module, and realizing multi-round reasoning and knowledge deduction capability of complex problems; And 5, constructing a multi-mode feedback module, wherein the multi-mode feedback module converts the intelligent question-answering result into text or voice form for feedback, and adapts to the maintenance scene requirement. Preferably, the step 1 includes: Text input processing, namely receiving fault description text directly input by a user; the voice input process, which integrates the automatic voice recognition technology to convert the fault information dictated by the user into characters; The input intention unified conversion is that the fault information input by words or voices is standardized into a structured query statement, and the structured query statement comprises core fields of 'equipment type', 'fault phenomenon', 'fault part', 'running state', and a unified and standard format is provided for subsequent vector knowledge base retrieval, so that the accuracy and high efficiency of retrieval are ensured. Preferably, the step 2 includes: Collecting and cleaning knowledge, namely collecting an equipment maintenance manual, a historical fault case, manufacturer technical documents, maintenance rules and industry standards, performing deduplication processing on the collected data, performing format unification, converting documents with different formats into unified text formats, and performing structural processing on form data; Cutting and vectorizing the document, namely cutting the processed document into 500-1000 word fragments by adopting a sliding window method, and converting each fragment into 768 dimension vectors by a pre-training model; vector storage and optimization, namely storing vector data into a Chroma vector database, adopting ONNX Runtime deployment, supporting GPU batch reasoning, and greatly improving the retrieval accuracy and throughput of the model; And dynamically updating the knowledge base, namely setting a periodic updating mechanism, automatically incorporating new maintenance cases, technical documents and industry standards, and updating the vector model through incremental training. Preferably, the step 3 i