Knowledge enhancement large model development method for high-end equipment fault diagnosis

Abstract

The invention discloses a high-end equipment fault diagnosis-oriented knowledge enhancement large model development method, which comprises the following steps of: constructing a professional knowledge base covering field professional knowledge, historical maintenance practice records and expert experience data, carrying out parameter fine tuning on a pre-trained large model on the basis of the professional knowledge base, optimizing fine tuning samples and data distribution, and carrying out fault diagnosis on the pre-trained large model; enabling the professional large model to adapt to different fault types; a knowledge retrieval enhanced question and answer mechanism is established, accurate matching and dynamic retrieval optimization are realized based on a knowledge index system of a high-end equipment fault phenomenon and troubleshooting logic, secondary knowledge injection is performed in combination with questions of a user, and the question and answer precision and practicability are improved. The method can be widely applied to operation and maintenance of high-end equipment, and the intelligent diagnosis capability and the operation and maintenance efficiency are effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence, and in particular to a knowledge-enhanced large model development method for high-end equipment fault diagnosis. Background Art

[0002] The demand for intelligent diagnosis, predictive maintenance, and optimized decision-making in the field of high-end equipment is growing. Large language models (LLMs) have demonstrated strong capabilities in natural language understanding, knowledge generalization, and reasoning. In the field of high-end equipment fault diagnosis, how to use large language models to promote the intelligent development of equipment operation and maintenance is a major research challenge. Currently, the methods for developing large language models in vertical fields to improve their domain adaptability mainly include the following two categories:

[0003] (1) Retrieval-Augmented Generation (RAG): RAG is a technology that combines information retrieval and generative models, aiming to improve the knowledge capacity and factuality of pre-trained large models. Pre-trained large models mainly rely on pre-trained corpora for text generation, which has problems such as knowledge obsolescence and hallucination. RAG technology can effectively enhance the accuracy and interpretability of generated results by retrieving external knowledge in real time. The implementation of RAG functions mainly includes two core components: retriever and generator. The retriever uses vector search, BM25 or other indexing methods to extract the most relevant information to the user query from the knowledge base; the generator performs enhanced reasoning and text generation based on the retrieved content. This technology is widely used in multiple vertical fields such as law and medicine, and can improve the authenticity, professionalism and contextual relevance of the output content while ensuring fluency.

[0004] (2) Fine-tuning of large model parameters: Fine-tuning of large model parameters refers to adjusting the model weights based on a small amount of data based on a pre-trained language model to adapt to a specific task or domain. Compared with training from scratch, fine-tuning of large models can reduce computational costs while improving the performance of the model in specific scenarios. The fine-tuning methods mainly include: ① Full parameter fine-tuning: Adjust all model parameters to make them fully adapt to the new task, but the computational overhead is large. ② Adapter fine-tuning: Insert a small adaptation layer into the original model and only train the adapter parameters to reduce storage requirements. ③ LoRA (Low-Rank Adaptation): Use low-rank matrix decomposition to only update the weights of specific layers, significantly reducing computational and storage costs. ④Prefix-Tuning / Prompt-Tuning: Freeze the original parameters and only fine-tune the input prompt words to make the model efficiently adapt to new tasks. This method represents the weight update by adding a low-rank decomposition next to the original weight matrix. Specifically, the weight matrix to be learned is decomposed into the product of two smaller matrices, which significantly reduces the number of parameters that need to be optimized. This method maintains good performance while greatly reducing storage and computational overhead.

[0005] In summary, the main advantage of RAG is that it can dynamically supplement information from external knowledge bases without modifying model parameters, ensuring timely and controllable answers. Its disadvantages include strong retrieval dependency. If the knowledge base is of poor quality or the indexing mechanism is imperfect, the retrieved information may be inaccurate, affecting model output. Furthermore, the computational overhead is high, and the retrieval-generation process increases inference latency. Parameter fine-tuning enhances the model's endogenous knowledge by adjusting its weights, improving generalization across domain tasks and reducing external dependencies during inference. However, its disadvantages are high computational cost, requiring large amounts of labeled data, requiring retraining to update knowledge, and the potential for catastrophic forgetting, where the model forgets old knowledge after new knowledge is injected. Therefore, how to combine the advantages of both approaches to meet the practical needs of this field remains under investigation. Summary of the Invention

[0006] In view of the shortcomings of the existing technology, the present invention proposes a knowledge-enhanced large model development method for high-end equipment fault diagnosis.

[0007] The specific technical solutions are as follows:

[0008] A method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis includes the following steps:

[0009] S1: Extract domain expertise, historical maintenance practice records, and expert experience data to obtain corresponding question-answer pairs. These pairs are stored collectively to form a professional knowledge base for high-end equipment fault diagnosis. The data in the professional knowledge base is transformed to form a vectorized database.

[0010] S2. Fine-tune parameters based on the pre-trained large model and the professional knowledge base to obtain a professional large model: Build a fine-tuning sample set for different types of faults based on the professional knowledge base to optimize the distribution of training data; the fine-tuning sample set includes common faults, rare faults, and complex fault scenarios for different types of equipment;

[0011] S3: Establish a knowledge retrieval enhanced question-answering mechanism to achieve secondary knowledge injection of answers from professional large models, and obtain a knowledge-enhanced large model for high-end equipment fault diagnosis; the knowledge retrieval enhanced question-answering mechanism achieves secondary knowledge injection by: constructing a knowledge index system based on high-end equipment fault phenomena and troubleshooting logic, and performing semantic matching and information screening on user questions based on the knowledge index system to obtain retrieval results; using a knowledge enhancement-based verification mechanism to verify the initial answers generated by the professional large model based on user questions. If the initial answer is verified to be correct, the initial answer and the retrieval results are combined as the final answer; otherwise, the initial answer is corrected through the retrieval results to obtain the final answer.

[0012] Furthermore, the domain expertise, i.e., professional manuals on high-end equipment related to fault diagnosis, includes: high-end equipment structure, operation manuals, and troubleshooting manuals;

[0013] The historical maintenance practice records, i.e., historical maintenance and operation data, include: sensor data, maintenance support records, and customer relationship management systems;

[0014] The expert experience data, namely the fault troubleshooting rule base, includes: a fault judgment rule base and a logic troubleshooting rule base.

[0015] Furthermore, the operations for extracting knowledge from the domain expertise specifically include: extracting maintenance-related content from professional manuals of high-end equipment using high-precision text recognition and conversion tools, processing the maintenance-related content using a pre-trained large model, identifying, filtering, and condensing relevant sentences containing key information, such as fault symptoms, fault diagnosis, and maintenance solutions; assembling the relevant sentences into question-answer pairs in the format of "fault symptoms, fault diagnosis + maintenance solution";

[0016] The specific operation of knowledge extraction for the historical maintenance practice records is as follows: the historical maintenance data in the structured database is cleaned according to the field name and content, the fault diagnosis-related content is retained, and the fault diagnosis-related content is pre-processed using regular expressions to obtain a maintenance manual. The pre-processing includes: denoising, format standardization, and special symbol filtering; the maintenance manual is deeply processed using a pre-trained large model to identify, filter, and condense relevant sentences containing key information, such as fault phenomenon, fault diagnosis, and maintenance plan; the relevant sentences are assembled into question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan";

[0017] The specific operations for knowledge extraction from the expert experience data are as follows: the expert experience data includes a fault judgment rule base and a logical troubleshooting rule base; the fault judgment rule base and the logical troubleshooting rule base are assembled, the hierarchical relationship therein is retained and a tree-like retrieval method is adopted; and the hierarchy and content of the fault judgment rule base and the logical troubleshooting rule base are understood through a pre-trained large model, and question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan" and containing rule base hierarchical information are assembled.

[0018] Furthermore, in S1, the data in the professional knowledge base is transformed to obtain a vectorized database. The specific operations are as follows:

[0019] For different data types, we use AI similarity search tools to create different types of vector indexes. The process of creating a vector index includes vector neighbor search and index writing and loading. We also determine the optimal parameter values for the vector index based on the requirements of the application scenario, including query speed, storage space, and accuracy. During this process, we adjust the number of cluster centers in the quantized index to balance retrieval time, training costs, and index storage space.

[0020] The basic functions of the vectorized database are implemented through asynchronous functions, including writing, modifying, and deleting. Concurrent operations are handled through asynchronous tools. Feature selection and dimensionality reduction are used to control the dimension of vectorization.

[0021] Furthermore, in S2, breakpoint resume training and weight decay strategies are adopted during parameter fine-tuning to improve resource utilization efficiency and training efficiency; and a training optimization method of gradient accumulation and mixed precision training is adopted to reduce resource usage.

[0022] Furthermore, in S3, in the case where the user asks multiple questions, it is first determined whether the correlation between the contextual user questions and the current user questions reaches a threshold. If not, no context understanding is performed; if so, the knowledge retrieval is adjusted in combination with the contextual information of the user questions to enhance the search scope of the knowledge base.

[0023] Furthermore, in S3, semantic matching and information screening are performed on user questions based on the knowledge index system, as follows:

[0024] The characteristic words in the user's questions and the keywords in the knowledge retrieval enhanced knowledge base are uniformly encoded to obtain word vectors. The text similarity of the word vectors is calculated and sorted based on the probability of fault occurrence according to the similarity.

[0025] By filtering the fault tree, equipment type, and fault phenomenon description, the search scope is narrowed, so that the answers of the large model can be focused on feasible solutions and troubleshooting logic.

[0026] Furthermore, the screening fault tree is based on a fault tree structure, listing common causes and related components step by step; storing fault characteristics of different granularities and related troubleshooting paths at nodes at different levels, and searching through the parent-child node relationship of nodes at all levels; cooperating with the knowledge indexing system, automatically jumping to or recommending lower-level troubleshooting strategies based on the upper-level fault conclusions.

[0027] The beneficial effects of the present invention are:

[0028] The method proposed in this paper builds a professional knowledge base for high-end equipment fault repair, ensuring high data quality and logical integrity, and providing in-depth and reliable reference information during the question-and-answer process. The resulting knowledge-enhanced large model for high-end equipment fault diagnosis can better adapt to different types of fault scenarios, reducing the occurrence of irrelevant information and incorrect answers. The designed knowledge retrieval-enhanced question-and-answer mechanism uses a knowledge index system for precise retrieval and screening, significantly improving the accuracy and practicality of answers while balancing explainability and customization requirements, thereby effectively reducing the operation and maintenance costs of high-end equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of a method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis in an embodiment of the present invention.

[0030] Figure 2 1 is a flow chart of question-answer pair output in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention will be described in detail below based on the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become more apparent. The present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only for explaining the present invention and are not intended to limit the present invention.

[0032] Before further describing the embodiments of the present invention in detail, the nouns and terms involved in the embodiments of the present invention are explained. The nouns and terms involved in the embodiments of the present invention are subject to the following explanations.

[0033] (1) Checkpoint Resuming: Checkpoint resuming means saving the state in the middle of model training. After the training is interrupted (such as equipment failure or unexpected stop), the training can be resumed from the most recent storage point instead of starting from the beginning. Usually, when training a neural network, the model's weight parameters, optimizer state (such as Adam and SGD's momentum parameters) and learning rate and other information are regularly saved and stored as checkpoint files (such as .ckpt, .pth format). When training is terminated for some reason, the latest checkpoint can be reloaded to continue training without repeating the completed calculations. The advantage of checkpoint resuming is that it improves the utilization of computing resources and prevents the loss of training progress, which is extremely important in large-scale deep learning tasks.

[0034] (2) Weight Decay: Weight decay is a regularization technique used to prevent neural networks from overfitting. The core idea is to add an L2 regularization term to the model parameters in the loss function to constrain the size of the weights and avoid overfitting the training data. Mathematically, weight decay adds a penalty term proportional to the square of the weight to the loss function: Weight decay is often used in conjunction with optimizers such as SGD and AdamW, and is suitable for deep learning models such as Transformer to improve generalization capabilities. Weight decay is similar to L2 regularization, but through an independent weight update mechanism in optimizers such as AdamW, it is more suitable for large-scale training tasks.

[0035] (3) Gradient Accumulation: Gradient accumulation is a method to improve the training efficiency of pre-trained large models. It is suitable for situations where computing resources are limited or large-batch training cannot be used directly. Typically, deep learning models use batch gradient descent, and the parameters updated in one training are calculated from a batch of samples. However, when the video memory is insufficient, a large batch will cause video memory overflow. Gradient accumulation accumulates gradients over multiple small batches and updates the model parameters at one time after a certain number of update steps, thereby simulating the effect of large-batch training. The advantage of gradient accumulation is that it relieves video memory pressure and improves model training stability. At the same time, it can imitate the effect of large-batch training and improve model convergence. However, it may cause training time to increase and require the optimizer's learning rate to be adjusted to adapt to the accumulated gradient updates.

[0036] (4) Mixed Precision Training: Mixed precision training is a technique for accelerating deep learning model training by using different numerical precisions simultaneously to reduce computational costs and video memory usage. Typically, pre-trained large models use 32-bit floating point numbers (FP32) for parameter calculations to achieve more accurate results, but FP32 has a large computational load and occupies a lot of video memory. Mixed precision training mainly uses 16-bit floating point numbers (FP16, BF16) to perform most calculations, and only converts back to FP32 during key numerical operations (such as gradient calculations and weight updates) to ensure computational stability. Its core components include Automatic Mixed Precision (AMP), which can automatically determine when to use FP16 to reduce computational load and perform dynamic numerical scaling when necessary to prevent numerical underflow. The advantage of mixed precision is that it reduces video memory usage, improves computational throughput, and enables the same hardware to train larger models.

[0037] like Figure 1 As shown in FIG, a method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis includes the following steps:

[0038] S1. Construct a professional knowledge base for high-end equipment fault diagnosis: perform knowledge extraction on different data types, extract the corresponding question and answer pairs respectively, and store all the question and answer pairs to obtain the professional knowledge base. This ensures the high quality and logical structure of the data, and improves the interpretability and accuracy of fault diagnosis questions and answers. The data is divided into types including: domain expertise, historical maintenance practice records, and expert experience data. Domain expertise refers to professional manuals of high-end equipment related to fault diagnosis, including: high-end equipment structure, operation manuals, and troubleshooting manuals; historical maintenance practice records refer to historical maintenance and operation data, including: sensor data, maintenance support records, and customer relationship management system (Customer Relationship Management, hereinafter referred to as CRM); expert experience data refers to the fault troubleshooting rule base, including: fault judgment rule base and logical troubleshooting rule base.

[0039] like Figure 2 As shown in the figure, knowledge extraction is performed for different data types, and the corresponding question-answer pairs are extracted respectively, as follows:

[0040] (1) The extraction of question-answer pairs based on domain expertise is specifically achieved in the following ways: using high-precision text recognition and conversion tools to extract maintenance-related content from professional manuals of high-end equipment, using a pre-trained large model with strong semantic understanding capabilities to process the obtained maintenance-related content (text), and combining the embedded representation of the pre-trained large model to identify, screen, and condense relevant sentences containing key information such as fault phenomenon, fault diagnosis, and maintenance plan. Finally, the relevant sentences are assembled into question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan".

[0041] In this embodiment, high-precision optical character recognition (OCR) technology is used to identify maintenance-related content in professional manuals of high-end equipment, and is combined with layout analysis technology to improve the ability to parse complex format text. Secondly, a natural language processing model is called to preprocess the recognition results, and the recognized text is subjected to operations such as denoising, format standardization, and special symbol filtering to optimize data quality. Then, a text block automatic segmentation algorithm is used to divide the preprocessed recognition result content into structured units such as titles, chapters, tables, and fault descriptions to obtain maintenance-related content, providing support for subsequent semantic analysis.

[0042] (2) The extraction of question-answer pairs based on historical maintenance practice records is specifically achieved in the following way: the historical maintenance data in the structured database is cleaned according to the field name and content, the fault diagnosis-related content is retained, and the fault diagnosis-related content is pre-processed using regular expressions to obtain the maintenance manual. The pre-processing includes operations such as denoising, format standardization, and special symbol filtering. The maintenance manual is deeply processed using a pre-trained large model with strong semantic understanding capabilities. Combined with the embedded representation of the pre-trained large model, relevant sentences containing key information such as fault phenomenon, fault diagnosis, and maintenance plan are identified, screened, and condensed. Finally, the relevant sentences are assembled into question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan".

[0043] (3) The extraction of question-answer pairs based on expert experience data is specifically achieved in the following ways: 1. The fault judgment rule base and the logic troubleshooting rule base are assembled, the hierarchical relationship therein is retained (the hierarchical relationship is represented by encoding the parent-child relationship, making the tree-structured data storage and query more efficient), and a tree-structured retrieval method is adopted; 2. The hierarchy and content of the fault judgment rule base and the logic troubleshooting rule base are understood through the pre-trained large model, and finally assembled into question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan" and containing the rule base hierarchy information.

[0044] Part of the data in the professional knowledge base (mainly question-and-answer pairs corresponding to historical maintenance practice records) was transformed to obtain the Weaviate vector search engine database (hereinafter referred to as the vectorized database). Specifically, for different data types, FAISS (Facebook AI Similarity Search) was used to create vector indexes, and different index types were tried to be selected to improve retrieval accuracy and efficiency; the process of creating a vector index includes: vector neighbor search and index writing and loading. The vector index was then optimized: different parameter values were experimented with, combined with specific application scenarios (such as query speed, storage space, and accuracy requirements) to determine the optimal parameter settings; in addition, the number of cluster centers of the quantitative index (nlist, used to determine how many clusters the data is divided into) was adjusted during the optimization process to balance retrieval time, training cost, and index storage space. Finally, asynchronous and concurrent search functions based on FAISS were developed to achieve the goal of quickly processing a large number of search requests. When creating a vectorized database, the focus is on retrieval business logic design, including: first, implementing the basic functions of the vectorized database such as writing, modifying, and deleting through asynchronous functions; second, processing concurrent operations through asynchronous tools (async tools) to improve the performance of vectorized database operations; finally, reasonably controlling the dimension of vectorization through feature selection, dimensionality reduction and other technologies, optimizing the database storage structure and adding a cache mechanism to improve the performance of the vectorized database. The present invention has been tested on the performance of vector similarity calculation, data retrieval, writing, and deletion, proving that the vectorized database supports efficient similarity search, can quickly locate the knowledge content most relevant to the query from massive data, improve information retrieval efficiency, and verify the effectiveness of the method.

[0045] Through the above operations, sensitive information is stored in a local vectorized database, and access is controlled through permission management, reducing the risk of data leakage. The vectorized database can support complex queries based on vector operations, providing richer functionality and flexibility for question-answering systems; built-in vector operations support real-time knowledge updates and queries.

[0046] S2: Fine-tune parameters based on the pre-trained large model and the professional knowledge base to create a professional large model, achieving deep infusion of knowledge in the field of high-end equipment fault diagnosis. Specifically, based on the professional knowledge base, construct a fine-tuning sample set for different types of faults, adjust some parameters of the pre-trained large model, and optimize the distribution of training data so that the adjusted professional large model can more accurately match the actual application scenarios. The fine-tuning sample set must cover common faults, rare faults, and complex fault scenarios for different types of equipment.

[0047] When fine-tuning parameters, it is necessary to try a variety of fine-tuning methods, including: LoRa, Q-LoRa, Adapter-Tuning, Prefix-tuning, and other efficient parameter fine-tuning methods. The model's performance in answering various faults can be improved by optimizing model training methods and resource usage. Specifically, during fine-tuning, breakpoint resume training and weight decay strategies can be used to improve training stability, support efficient processing of large quantities of training data, increase model convergence speed, and reduce the risk of overfitting, thereby improving resource utilization and training efficiency. To address the computing resource limitations during the fine-tuning process, training optimization methods such as gradient accumulation and mixed precision training can be used to reduce resource usage.

[0048] S3: Establish a knowledge retrieval-enhanced question-answering mechanism to achieve secondary knowledge injection into the answers of the professional large model, and obtain a knowledge-enhanced large model for high-end equipment fault diagnosis. Secondary knowledge injection optimizes the reasoning process of the professional large model by retrieving supplementary information, making the answers it generates more comprehensive and reliable.

[0049] The knowledge retrieval-enhanced question-answering mechanism implements secondary knowledge injection, which includes the following three steps: constructing a knowledge index system based on high-end equipment fault phenomena and troubleshooting logic; performing semantic matching and information screening on user questions based on the knowledge index system to obtain retrieval results; and employing a knowledge-enhanced verification mechanism to verify the initial response generated by the professional large model based on the user's question. If the initial response is verified to be correct, the initial response and the retrieval results are combined as the final response; otherwise, the initial response is corrected using the retrieval results to obtain the final response. In both cases, new and old knowledge are integrated. Semantic matching and information screening based on user questions are used to narrow the scope of fault retrieval. During this step, if the user asks multiple questions, the correlation between the contextual user questions and the current user question is determined to have reached a threshold. If not, context understanding is not performed. If so, the contextual information of the user questions is combined to adjust the scope of the knowledge retrieval-enhanced knowledge base search to improve the accuracy of the question-answering (primarily focusing on the current user's question, supplemented by the contextual user questions).

[0050] The knowledge indexing system performs semantic matching and information screening for user questions. First, semantic matching is performed. Question feature words and keywords in the knowledge retrieval-enhanced knowledge base are uniformly encoded using the BERT-Tokenizer to generate word embeddings. Text similarity is then calculated for these embeddings, and the similarities are used to rank the results based on fault likelihood. Compared to traditional keyword retrieval, semantic matching can better understand user intent when user descriptions are incomplete or ambiguous. Next, information screening is performed. By filtering specialized information such as fault trees, device types, and fault symptom descriptions, the search scope is narrowed, allowing the professional model's answers to focus on feasible solutions and troubleshooting logic, reducing the injection of duplicate or irrelevant information and accelerating the professional model's response. The fault tree screening process, based on the fault tree structure, lists common causes and related components at each level. Nodes at different levels store fault characteristics of varying granularity and related troubleshooting paths, and search is performed based on the parent-child relationships at each level. Based on the upper-level fault conclusion, the system automatically jumps to or recommends troubleshooting strategies for the lower levels, generating relevant answers to the fault tree search process and enhancing the interpretability of the professional model.

[0051] The multimodality of the knowledge retrieval system is realized through a knowledge extraction model. It supports multimodal content parsing and text understanding in formats such as Word, PDF, Excel, PPT, and CSV, and links multimedia such as pictures to the corresponding nodes of the corresponding fault tree. When a specific fault node is found in the fault tree based on user questions, the picture link of this node will be automatically returned to realize the comprehensive management and call of pictures and videos.

[0052] The present invention improves the diagnostic capability and question-answering accuracy of the pre-trained large model through high-quality knowledge injection. First, a professional knowledge base covering domain expertise, historical maintenance practice records and expert experience data is constructed, and knowledge extraction is used to ensure data accuracy and logical consistency. Secondly, parameter fine-tuning is performed based on the pre-trained large model and the professional knowledge base, fine-tuning samples are made for different fault types, and data distribution and training resources are optimized so that the obtained professional large model can accurately adapt to different fault types and actual scenarios. Finally, a knowledge retrieval enhanced question-answering mechanism is established, and accurate matching and dynamic retrieval optimization are achieved based on the knowledge indexing system of high-end equipment failure phenomena and troubleshooting logic. Secondary knowledge injection is performed in combination with user questions to improve question-answering accuracy and practicality. The present invention can be widely used in the operation and maintenance of high-end equipment, effectively improving intelligent diagnostic capabilities and operation and maintenance efficiency.

[0053] Those skilled in the art will understand that the foregoing descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art will still be able to modify the technical solutions described in the foregoing examples or substitute equivalents for some of the technical features therein. Any modifications, equivalent substitutions, etc. made within the spirit and principles of the invention shall be included within the scope of protection of the invention.

Claims

1. A knowledge-enhanced large model development method for high-end equipment fault diagnosis, characterized by: The following steps are involved: S1: Extract domain expertise, historical maintenance practice records, and expert experience data to obtain corresponding question-answer pairs. These pairs are then stored together to form a professional knowledge base for high-end equipment fault diagnosis. Transform the data in the professional knowledge base to obtain a vectorized database; S2. Fine-tune parameters based on the pre-trained large model and the professional knowledge base to obtain a professional large model: Build a fine-tuning sample set for different types of faults based on the professional knowledge base to optimize the distribution of training data; the fine-tuning sample set includes common faults, rare faults, and complex fault scenarios for different types of equipment; S3: Establish a knowledge retrieval-enhanced question-answering mechanism to achieve secondary knowledge injection into professional large-scale model answers, and obtain a knowledge-enhanced large-scale model for high-end equipment fault diagnosis; The knowledge retrieval-enhanced question-answering mechanism achieves secondary knowledge injection by: building a knowledge index system based on high-end equipment failure phenomena and troubleshooting logic, performing semantic matching and information screening on user questions based on the knowledge index system, and obtaining retrieval results; using a knowledge-enhanced verification mechanism to verify the initial answers generated by the professional large model based on user questions. If the initial answer is verified to be correct, the initial answer and the retrieval results are combined as the final answer; Otherwise, the initial answer is corrected through the search results to obtain the final answer.

2. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: The domain expertise mentioned above refers to professional manuals on high-end equipment related to fault diagnosis, including: high-end equipment structure, operation manuals, and troubleshooting manuals; The historical maintenance practice records, i.e., historical maintenance and operation data, include: sensor data, maintenance support records, and customer relationship management systems; The expert experience data, namely the fault troubleshooting rule base, includes: a fault judgment rule base and a logic troubleshooting rule base.

3. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: The specific operations for extracting knowledge from the domain expertise are as follows: using high-precision text recognition and conversion tools to extract maintenance-related content from professional manuals for high-end equipment; using a pre-trained large model to process the maintenance-related content, identifying, filtering, and condensing relevant sentences containing key information, such as fault symptoms, fault diagnosis, and maintenance solutions; and assembling these relevant sentences into question-and-answer pairs in the format of "fault symptoms, fault diagnosis + maintenance solution"; The specific operation of knowledge extraction from the historical maintenance practice records is as follows: the historical maintenance data in the structured database is cleaned according to the field name and content, the fault diagnosis-related content is retained, and the fault diagnosis-related content is pre-processed using regular expressions to obtain a maintenance manual. The pre-processing includes: denoising, format standardization, and special symbol filtering; the maintenance manual is deeply processed using a pre-trained large model to identify, filter, and condense relevant sentences containing key information, such as fault phenomenon, fault diagnosis, and maintenance plan; the relevant sentences are assembled into question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan"; The specific operations for knowledge extraction from the expert experience data are as follows: the expert experience data includes a fault judgment rule base and a logical troubleshooting rule base; the fault judgment rule base and the logical troubleshooting rule base are assembled, the hierarchical relationship therein is retained, and a tree-like retrieval method is adopted; and the hierarchy and content of the fault judgment rule base and the logical troubleshooting rule base are understood through a pre-trained large model, and question-answer pairs in the format of "fault phenomenon, fault diagnosis + maintenance plan" and containing rule base hierarchical information are assembled.

4. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: In S1, the data in the professional knowledge base is transformed to obtain a vectorized database. The specific operations are as follows: For different data types, we use AI similarity search tools to create different types of vector indexes. The process of creating a vector index includes vector neighbor search and index writing and loading. We also determine the optimal parameter values for the vector index based on the requirements of the application scenario, including query speed, storage space, and accuracy. During this process, we adjust the number of cluster centers in the quantized index to balance retrieval time, training costs, and index storage space. The basic functions of the vectorized database are implemented through asynchronous functions, including writing, modifying, and deleting. Concurrent operations are handled through asynchronous tools. Feature selection and dimensionality reduction are used to control the dimension of vectorization.

5. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: In S2, breakpoint resume training and weight decay strategies are adopted during parameter fine-tuning to improve resource utilization efficiency and training efficiency; a training optimization method of gradient accumulation and mixed precision training is adopted to reduce resource usage.

6. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: In S3, when a user asks multiple questions, it is first determined whether the correlation between the contextual user questions and the current user questions reaches a threshold. If not, no context understanding is performed; if so, the knowledge retrieval is adjusted based on the contextual information of the user questions to enhance the search scope of the knowledge base.

7. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 1 is characterized in that: In S3, semantic matching and information screening are performed on user questions based on the knowledge indexing system, as follows: The characteristic words in the user's questions and the keywords in the knowledge retrieval enhanced knowledge base are uniformly encoded to obtain word vectors. The text similarity of the word vectors is calculated and sorted based on the probability of fault occurrence according to the similarity. By filtering the fault tree, equipment type, and fault phenomenon description, the search scope is narrowed, so that the answers of the large model can be focused on feasible solutions and troubleshooting logic.

8. The method for developing a knowledge-enhanced large model for high-end equipment fault diagnosis according to claim 7 is characterized in that: The screening fault tree is based on a fault tree structure, listing common causes and related components level by level; storing fault characteristics of different granularities and related troubleshooting paths at nodes at different levels, and searching through the parent-child node relationship of nodes at all levels; cooperating with the knowledge indexing system, automatically jumping to or recommending lower-level troubleshooting strategies based on the upper-level fault conclusions.

Images