Automatic FMEA construction method and system based on large language model
By combining a large language model with the RAG knowledge base, potential failure modes are automatically identified and evaluated, solving the problems of low efficiency and insufficient data processing capabilities of existing FMEA methods, and achieving efficient and accurate FMEA generation and risk assessment.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ANHUI POLYTECHNIC UNIV
- Filing Date
- 2025-12-23
- Publication Date
- 2026-04-28
AI Technical Summary
Existing FMEA methods rely on expert experience, are inefficient and prone to missing key information, struggle to handle unstructured data, especially user comments, and lack automation and self-learning capabilities.
By employing a Large Language Model (LLM) combined with the RAG knowledge base, and through preprocessing, retrieval, and generation modules, potential failure modes are automatically identified and FMEA entries are generated. Multi-source knowledge is then integrated for risk assessment and visualization.
It achieves automated and efficient generation of FMEA, improves the accuracy and comprehensiveness of judgment, can process unstructured data, dynamically updates the knowledge base, and improves the system's adaptability and interpretability.
Smart Images
Figure CN121935640A_ABST
Abstract
Description
Technical Field
[0002] This invention belongs to the field of language processing. Specifically, this invention relates to an automated FMEA construction method and system based on a large language model. Background Technology
[0003] Enterprises face increasingly complex challenges in product design, manufacturing processes, and user demands, urgently requiring them to leverage advanced information technology and intelligent tools to transform and upgrade their quality management models. To ensure product reliability throughout the entire design and manufacturing process, enterprises must systematically identify and assess potential failure risks and proactively deploy targeted preventative measures. Failure Mode and Effects Analysis (FMEA), as a systematic quality and reliability analysis tool, assists engineers and technicians in comprehensively identifying and assessing potential failure modes (FMs) and their impact on product performance, thereby ensuring products meet stringent quality standards. Traditionally, FMEA relies on expert teams performing manual execution, employing structured methods to identify and analyze potential failures in products or their processes. However, this method heavily depends on expert experience and manual processing, dealing with a large volume of documents and user feedback, resulting in a time-consuming and inefficient process, and is prone to information overload leading to the omission of critical information.
[0004] In recent years, researchers have proposed various improved methods based on automation and artificial intelligence to enhance the efficiency and accuracy of FMEA. Rule-based extraction methods can only cover preset patterns and struggle to handle complex semantics and contextual relationships in text; conventional models have limited understanding capabilities and are also unable to effectively process unstructured data such as user comments. Although machine learning techniques can achieve a certain degree of automatic extraction of fault information, they typically rely on large amounts of labeled data and are mostly applicable to the identification of known fault types, limiting their ability to discover potential new faults.
[0005] The rise of generative artificial intelligence and Large Language Models (LLMs) has opened up new possibilities for the automated construction of FMEAs. Large Language Models, trained on massive amounts of text data, possess powerful natural language understanding capabilities and can automatically process and analyze structured and unstructured documents. Furthermore, by building domain knowledge bases, the performance of LLMs can be further optimized, enhancing their applicability in specialized scenarios.
[0006] Therefore, this invention proposes an automated FMEA construction method and system based on a large language model. Summary of the Invention
[0007] This invention aims to overcome the shortcomings of existing technologies and proposes an automated FMEA construction method and system based on a large language model to achieve the following objectives: integrating LLM into the FMEA process to achieve automatic identification of potential failure modes and in-depth analysis of the causes and effects of failures, thereby improving the efficiency of FMEA construction while enhancing the accuracy and comprehensiveness of its judgment.
[0008] To achieve the above objectives, the technical solution adopted by this invention is: an automated FMEA construction method based on a large language model, the method comprising the following steps: Step S1: Obtain user comment data and FMEA domain knowledge and perform preprocessing; Step S2: Construct a RAG-based LLM, including: a RAG knowledge base built based on FMEA domain knowledge; Step S3: Using pre-designed prompts and pre-processed user comment data as input to the LLM, the LLM analyzes the user comment data through the RAG knowledge base built on FMEA domain knowledge and outputs FMEA items related to the user comment data. Step S4: Summarize the FMEA entries and visualize them.
[0009] Preferably, the preprocessing includes: First, regular expressions are used to process the comment text, removing URL links, advertising tags, and meaningless phrases, as well as invalid characters and advertising information. Second, the jieba word segmentation tool in the Python library is used in conjunction with a custom automotive industry dictionary for Chinese word segmentation. Finally, stop word filtering and normalization are used to improve semantic density.
[0010] Preferably, the RAG knowledge base built based on FMEA domain knowledge includes a preprocessing module, a retrieval module, and a generation module; The preprocessing module is used to convert FMEA domain knowledge into semantic vectors and store them in a vector database through an embedding model, and to convert the input user comment data into query vectors through an embedding model. The retrieval module is used to locate and extract similar documents related to user comment data in the vector database according to the query vector; The generation module is used to combine prompts with similar documents using a planning model, and finally outputs FMEA entries related to user review data.
[0011] Preferably, the embedding model includes the multilingual embedding model provided by Baidu Smart Cloud Big Model Platform.
[0012] Preferably, the planning model includes DeepSeek-R1.
[0013] Preferably, the FMEA domain knowledge includes user-generated content from social media platforms, internal company documents, and industry standards and literature.
[0014] Preferably, the input user comment data is automatically added to the RAG knowledge base after being converted into query vectors.
[0015] Preferably, the FMEA items preset by the prompt words include the identification of relevant components, the professional description of failure modes and the basis for their identification, the description of the consequences of failure modes, and the systematic analysis of potential causes.
[0016] Preferably, the FMEA items preset by the prompt words also include the quantitative assessment results of failure mode risk factors by LLM, namely severity S, occurrence O and detectability D, and the risk priority number is calculated by the formula RPN = S×O×D; and the actionable control measures and improvement suggestions proposed by LLM based on this and combined with knowledge base or engineering reasoning.
[0017] This application also provides an automated FMEA construction system based on a large language model. Using the aforementioned automated FMEA construction method based on a large language model, the system includes: The data acquisition module is used to acquire user comment data and perform preprocessing. LLM build modules: used to build RAG-based LLMs, including: a RAG knowledge base built based on FMEA domain knowledge; The prompt word construction module is used to construct prompt words to drive the LLM to analyze user comment data through the RAG knowledge base built on FMEA domain knowledge, and output FMEA entries related to the user comment data; The visualization module is used to summarize and visualize FMEA entries.
[0018] The technical effects of this invention are as follows: This invention combines a local domain knowledge base with LLM (Local Level Management) to automatically identify potential failure modes in user reviews, infer the causes and impacts of failures, assess risks, calculate Risk Priority Numbers (RPNs), and provide improvement suggestions. Ultimately, it achieves automatic generation of FMEA (Factors for Optimized Factor Analysis) tables, improving FMEA construction efficiency while enhancing the accuracy and comprehensiveness of its judgments. The LLM includes a RAG (Rapid Area Knowledge Base) built based on FMEA domain knowledge, integrating multi-source knowledge and semantic retrieval to improve the engineering relevance and reliability of the LLM-generated results. Attached Figure Description
[0019] Figure 1 A flowchart illustrating an automated FMEA construction method based on a large language model; Figure 2 A workflow diagram for the RAG knowledge base built based on FMEA domain knowledge; Figure 3 This is a schematic diagram of an FMEA table provided for an embodiment of the present invention. Detailed Implementation
[0020] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. The purpose is to help those skilled in the art to have a more complete, accurate, and in-depth understanding of the inventive concept and technical solution of the present invention, and to facilitate its implementation. To make the technical solution of the present invention clearer, the present invention will be explained and illustrated through the following embodiments.
[0021] This embodiment provides an automated FMEA construction method based on a large language model, aiming to solve problems in existing technologies such as: single and outdated knowledge sources, inability to adapt to rapidly changing real-world application scenarios; need for manual screening of model-generated results, resulting in low automation; lack of semantic recognition and processing capabilities for unstructured text (such as user comments and maintenance records); and low self-learning capabilities, lacking dynamic expansion and updating of the knowledge base. Figure 1 As shown, the method in this embodiment includes the following steps: Step S1: Obtain user comment data and FMEA domain knowledge and perform preprocessing; Step S2: Construct a RAG-based LLM, including: a RAG knowledge base built based on FMEA domain knowledge; Step S3: Using pre-designed prompts and pre-processed user comment data as input to the LLM, the LLM analyzes the user comment data through the RAG knowledge base built on FMEA domain knowledge and outputs FMEA items related to the user comment data. Step S4: Summarize the FMEA entries and visualize them.
[0022] Specifically, referring to step S1, not only does the user comment data input by the user for FMEA analysis need to be preprocessed before being input into the LLM (Large Language Model), but the FMEA domain knowledge used for knowledge base construction also needs to be preprocessed to ensure data reliability and thus improve the accuracy of the LLM model. The preprocessing includes data cleaning, word segmentation, and stop word removal. The specific steps in this embodiment are as follows: First, regular expressions are used to process the comment text, removing URLs, advertising tags, and meaningless phrases, as well as invalid characters and advertising information. Second, the jieba word segmentation tool from the Python library, combined with a custom automotive industry dictionary, is used for Chinese word segmentation, as there are no spaces between words in Chinese text. Finally, stop word filtering and normalization are applied to improve semantic density. Stop words are frequently occurring but meaningless words; removing them reduces feature dimensionality, decreases workload, and improves feature extraction quality.
[0023] Referring to step S2, this embodiment constructs an LLM based on RAG (Retrieval Enhanced Generation), the key of which lies in the RAG knowledge base built based on FMEA domain knowledge. FMEA domain knowledge includes user-generated content (UGC) from social media platforms, internal company documents (manuals, repair records, after-sales data), and industry standards and literature, etc., integrating multi-source data to provide rich external knowledge sources for LLM and improve the relevance and accuracy of LLM output.
[0024] In this embodiment, the RAG knowledge base built based on FMEA domain knowledge includes a preprocessing module, a retrieval module, and a generation module, wherein: The preprocessing module is used to transform FMEA domain knowledge into semantic vectors through an embedding model to preserve semantic features and store them in a vector database, and to transform input user comment data into query vectors (which are also a type of semantic vector) through an embedding model. The retrieval module is used to locate and extract similar documents related to user comment data in the vector database according to the query vector, so as to provide contextual support for LLM generation; The generation module is used to combine prompt words with similar documents using a planning model, and finally output FMEA entries related to user comment data; FMEA entries can be constructed based on prompt words, usually in the form of a structured document in JSON format.
[0025] The RAG knowledge base, built upon FMEA domain knowledge, integrates multi-source knowledge and semantic retrieval, enhancing the engineering relevance and reliability of LLM-generated results. This allows LLM to generate answers based on specific background knowledge without retraining, facilitating integration into enterprise knowledge bases and reducing the risk of misinformation due to a lack of domain knowledge. Furthermore, its modular design enables synergy among various stages, efficiently extracting and analyzing information from unstructured user comments to automatically generate structured entries conforming to FMEA annotations. Its specific workflow is as follows: Figure 2 As shown.
[0026] In this embodiment, the embedding model uses the Multilingual Embedding model provided by Baidu AI Cloud's Big Model Platform. This model supports cross-language semantic matching, efficiently converts text into numerical vectors, and has advantages in Chinese semantic representation; at the same time, its maintenance-free and low-cost characteristics reduce deployment difficulty. With this tool, the retrieval module can process mixed Chinese and English or cross-language content, providing semantically rich contextual information for downstream generation.
[0027] This embodiment uses DeepSeek-R1 as the planning model, responsible for task planning and invocation. DeepSeek-R1 is a type of LLM (Local Language Management) model, which has strong advantages in natural language processing and is suitable for mining and analyzing user reviews in the new energy vehicle field. DeepSeek-R1 can accurately identify technical terms, implicit failure modes, and related failure components in reviews, and understand multi-level issues through contextual modeling, thus providing comprehensive structured input for FMEA (Factors for External Detection and Analysis).
[0028] Furthermore, the RAG knowledge base in this embodiment has adaptive and continuous learning capabilities, which can absorb new comment information and update the knowledge base in real time. That is, after the input user comment data is converted into query vectors, it is automatically added to the RAG knowledge base to ensure the dynamism and timeliness of the analysis results.
[0029] Referring to step S3, after the RAG-based LLM is built, it can be used for actual FMEA construction. During use, the LLM requires pre-designed prompts to guide the generation of targeted natural language input or instructions. Guided by these prompts, the LLM can utilize the FMEA method, combined with the linguistic features of new energy vehicle user reviews, to design and optimize prompts for failure mode identification and contextual semantic understanding. Through continuous iteration and verification of the prompts, the model can accurately capture potential failure information in complex contexts and improve the interpretability and professionalism of the generated results. For example, the FMEA entries in this embodiment can be constructed based on the prompts.
[0030] The prompt word design strategy in this embodiment is as follows: The prompt word sets the large language model as an experienced FMEA expert. First, using "new energy vehicle user review data" as input, the text is preprocessed and analyzed line by line to extract technical clauses. Then, through semantic understanding and similarity clustering, equivalent or near-synonymous expressions are grouped into a unified set of questions, while retaining all original clauses to ensure the traceability of the results. Next, LLM implements retrieval enhancement generation technology for each set of questions, retrieving relevant supporting evidence from the knowledge base. If direct evidence is lacking, "inference" is explicitly marked in the results, and background information is supplemented with engineering experience. In this context of evidence enhancement, the model generates structured entries for each question according to the FMEA paradigm, including: identification of relevant components, professional description of failure modes and their identification basis, description of the consequences of failure modes, and systematic analysis of potential causes. Simultaneously, the model quantifies the risk factors of failure modes, namely severity (S), occurrence (O), and detection (D), and calculates the risk priority number using the formula RPN = S × O × D. Based on this, the model further incorporates knowledge base or engineering reasoning to propose actionable control measures and improvement suggestions. Finally, all potential failure modes are summarized in descending order of RPN values and output as a structured JSON array.
[0031] Finally, referring to step S4, this embodiment summarizes all FMEA items, visualizes them, and generates an FMEA report. For example, an FMEA chart constructed using comment text data and FMEA items generated based on LLM is shown below. Figure 3 As shown.
[0032] This embodiment applies the LLM-based automated FMEA construction method to fault analysis in new energy vehicles, illustrating the practical role of LLM in improving the FMEA process. First, LLM performs semantic understanding analysis on initial user comments such as "This car's power performance is average" and "The car's acceleration is a bit delayed at start-up, not very powerful," identifying the potential failure mode as delayed power output and slow acceleration response. Second, LLM, combined with the RAG knowledge base, further generates potential failure consequences and potential failure causes, providing sources of analysis evidence for management review. Finally, it calculates the Risk Priority Number (RPN) and proposes corresponding control recommendations for this failure mode. Issues such as delayed power output and slow acceleration response; unsuitable seat firmness and lack of ventilation and heating functions; and overly stiff suspension with poor shock absorption have high risk priority. Given limited company resources, management should prioritize improving high-risk failure modes. FMEA data, derived from unsatisfactory complaints, provides valuable reference for improving product quality.
[0033] This embodiment also provides an automated FMEA construction system based on a large language model. Using the above-described automated FMEA construction method based on a large language model, the system includes: The data acquisition module is used to acquire user comment data and perform preprocessing. LLM build modules: used to build RAG-based LLMs, including: a RAG knowledge base built based on FMEA domain knowledge; The prompt word construction module is used to construct prompt words to drive the LLM to analyze user comment data through the RAG knowledge base built on FMEA domain knowledge, and output FMEA entries related to the user comment data; The visualization module is used to summarize and visualize FMEA entries.
[0034] This invention constructs a dataset using real user reviews and company data for new energy vehicles, enabling automated generation of the entire FME (Fulfilled Quality Management) process. It extends LLM (Limited Management Model) applications from theoretical verification to a practically deployable quality management system, enhancing the engineering applicability and validation value of the research.
[0035] The beneficial effects of this invention include: 1. Richer Knowledge Base Data Sources: It integrates multi-source data, including user-generated content (UGC), maintenance records, industry standards, and academic literature, and dynamically updates knowledge through the RAG mechanism. This improves the system's adaptability to new problems and scenarios, reducing the risk of knowledge gaps and obsolescence.
[0036] 2. Higher degree of automation: Achieve end-to-end automation from text recognition and knowledge retrieval to FMEA report generation, reducing manual intervention and improving efficiency.
[0037] 3. More accurate semantic understanding: The DeepSeek-R1 model is optimized for the domain, and prompt word templates for car reviews are designed to improve the semantic parsing ability of unstructured text.
[0038] 4. Knowledge has the ability to continuously evolve: The knowledge base is dynamically updated through an adaptive learning mechanism, supporting continuous system optimization.
[0039] 5. More interpretable results: Outputs structured JSON format and annotates the reasoning basis, ensuring that the results are traceable and verifiable.
[0040] 6. More comprehensive risk identification: It can automatically discover potential new failure modes from massive user feedback, making up for the shortcomings of traditional FMEA in that it is easy to miss hidden dangers.
[0041] The present invention has been described above by way of example with reference to the accompanying drawings. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution; or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.
Claims
1. An automated FMEA construction method based on a large language model, characterized in that: The method includes the following steps: Step S1: Obtain user comment data and FMEA domain knowledge and perform preprocessing; Step S2: Construct a RAG-based LLM, including: a RAG knowledge base built based on FMEA domain knowledge; Step S3: Using pre-designed prompts and pre-processed user comment data as input to the LLM, the LLM analyzes the user comment data through the RAG knowledge base built on FMEA domain knowledge and outputs FMEA items related to the user comment data. Step S4: Summarize the FMEA entries and visualize them.
2. The automated FMEA construction method based on a large language model according to claim 1, characterized in that: The preprocessing includes: First, regular expressions are used to process the comment text, removing URL links, advertising tags, and meaningless phrases, as well as invalid characters and advertising information. Second, the jieba word segmentation tool in the Python library is used in conjunction with a custom automotive industry dictionary for Chinese word segmentation. Finally, stop word filtering and normalization are used to improve semantic density.
3. The automated FMEA construction method based on a large language model according to claim 1, characterized in that: The RAG knowledge base, built based on FMEA domain knowledge, includes a preprocessing module, a retrieval module, and a generation module. The preprocessing module is used to convert FMEA domain knowledge into semantic vectors and store them in a vector database through an embedding model, and to convert the input user comment data into query vectors through an embedding model. The retrieval module is used to locate and extract similar documents related to user comment data in the vector database according to the query vector; The generation module is used to combine prompts with similar documents using a planning model, and finally outputs FMEA entries related to user review data.
4. The automated FMEA construction method based on a large language model according to claim 3, characterized in that: The embedding model includes the multilingual embedding model provided by Baidu Smart Cloud Big Model Platform.
5. The automated FMEA construction method based on a large language model according to claim 3, characterized in that: The planning model includes DeepSeek-R1.
6. The automated FMEA construction method based on a large language model according to claim 3, characterized in that: The FMEA domain knowledge includes user-generated content on social media platforms, internal company documents, and industry standards and literature.
7. The automated FMEA construction method based on a large language model according to claim 3, characterized in that: The input user comment data is automatically added to the RAG knowledge base after being converted into query vectors.
8. The automated FMEA construction method based on a large language model according to claim 1, characterized in that: The FMEA items preset by the prompts include the identification of relevant components, the professional description of failure modes and the basis for their identification, the description of the consequences of failure modes, and the systematic analysis of potential causes.
9. The automated FMEA construction method based on a large language model according to claim 1 or 8, characterized in that: The FMEA items pre-set by the prompts also include the LLM's quantitative assessment results of failure mode risk factors, namely severity (S), occurrence (O), and detectability (D), and the risk priority number is calculated using the formula RPN = S × O × D; as well as the actionable control measures and improvement suggestions proposed by the LLM based on this and combined with knowledge base or engineering reasoning.
10. An automated FMEA construction system based on a large language model, using the automated FMEA construction method based on a large language model according to claims 1-9, characterized in that: The system includes: The data acquisition module is used to acquire user comment data and perform preprocessing. LLM build modules: used to build RAG-based LLMs, including: a RAG knowledge base built based on FMEA domain knowledge; The prompt word construction module is used to construct prompt words to drive the LLM to analyze user comment data through the RAG knowledge base built on FMEA domain knowledge, and output FMEA entries related to the user comment data; The visualization module is used to summarize and visualize FMEA entries.