Automatic recommendation method for fault maintenance solutions based on large language model

Abstract

The present invention discloses a method for automatically recommending fault repair solutions based on a large language model, which belongs to the field of intersection of artificial intelligence and fault diagnosis technology, and solves the problems of lack of semantic understanding, cold start response hysteresis, and failure to match complex fault descriptions in the prior art. The present invention provides a method for automatically recommending fault repair solutions based on a large language model, extracts keywords of fault phenomena, calculates semantic similarity to construct an association network graph, matches recommended repair solutions after receiving user input, and can also dynamically update the network to improve repair efficiency and accuracy. The present invention improves semantic understanding accuracy through a large language model, constructs a dynamic association network to achieve accurate matching of fault phenomena, supports real-time updates of cold start scenarios, solves the problems of insufficient semantic understanding and cold start hysteresis in traditional methods, significantly improves the efficiency and accuracy of repair solution recommendations, and reduces maintenance costs.

Description

Technical Field

[0001] The present invention relates to the field of intersectional technology of artificial intelligence and fault diagnosis, and in particular to a method for automatically recommending fault repair solutions based on a large language model. Background Art

[0002] In the field of fault repair technology, existing systems generally use a repair solution recommendation mechanism based on precise keyword matching. This is typically achieved by extracting keywords from fault descriptions through manual annotation or basic word segmentation techniques, building a structured fault knowledge base. When a user enters a fault symptom, the system retrieves similar cases from the knowledge base through keyword comparison (such as Boolean matching or TF-IDF weighting). However, this technical approach has significant drawbacks:

[0003] First, the semantic understanding of complex fault descriptions is severely insufficient. When users use non-standard terms (such as "engine knock" and "engine knock") or omit the names of key components, matching fails due to the lack of identical keywords.

[0004] Second, it cannot handle scenarios with semantically related but non-identical terms. For example, the terminology for "ECU error" and "onboard computer failure" differs. Traditional methods rely on literal matching and ignore semantic equivalence.

[0005] Finally, cold start issues are a prominent problem. New fault symptoms, lacking historical keyword records, cannot be integrated into the knowledge network, requiring manual re-labeling and integration into the database. This leads to response delays and increased maintenance costs. This is especially true when fault descriptions include multiple symptoms across multiple locations (such as "transmission jerking accompanied by elevated oil temperature"). Existing technologies, unable to parse implicit correlations between keywords, often return incomplete or erroneous repair solutions. Tests show that their top-5 recommendation hit rate is less than 65%, severely hindering repair efficiency. Summary of the Invention

[0006] The present invention proposes a method for automatically recommending fault repair solutions based on a large language model, by constructing a dynamic semantic association network graph. and hybrid similarity calculation model , combined with fault knowledge base keyword extraction, semantic similarity weighted fusion and Top-N ranking recommendation mechanism, it solves the problems of lack of semantic understanding, cold start response lag and complex fault description matching failure in existing technologies.

[0007] A method for automatically recommending fault repair solutions based on a large language model, the method comprising the following steps:

[0008] S1. Based on the M fault phenomenon descriptions in the fault knowledge base, the keyword set of each fault phenomenon is extracted through the large language model to form node attributes; the semantic similarity is calculated based on the interaction relationship between any two fault phenomenon keyword sets, and a weighted fault phenomenon association network diagram is constructed. ;

[0009] S2. Receive the fault phenomenon description input by the user and extract the keyword set A through the large language model;

[0010] S3. Traverse the fault phenomenon associated network diagram Calculate the semantic similarity between the user keyword set A and each node keyword set for all nodes ;

[0011] S4. Sort by semantic similarity from high to low and output the top N matching fault phenomena;

[0012] S5. Retrieve the fault type, fault description, fault cause and repair plan corresponding to the first N fault phenomena from the fault factor database and display them to the user.

[0013] Furthermore, in S1, the node attributes are constructed as follows:

[0014] For complex fault phenomenon descriptions, a keyword set is extracted through a large language model. ;

[0015] For simple fault phenomenon descriptions, extract keyword sets through word segmentation using a large language model ,

[0016] The large language models include ChatGPT, Qwen and DeepSeek.

[0017] Furthermore, in S1, the semantic similarity The calculation formula is:

[0018]

[0019] in, is the keyword overlap, The two keyword sets remaining after removing the same keywords and The semantic similarity between

[0020] When there are identical keywords, , calculate the keyword overlap:

[0021]

[0022] in, , calculate the remaining keyword set and The semantic similarity of ;

[0023] When there is no identical keyword, ,make , calculate the entire keyword set and The semantic similarity of .

[0024] Further, The value of .

[0025] Furthermore, in S3, the user keyword set A and the node Similarity calculation, reuse S1 The algorithm regards the keyword set A as a newly added fault phenomenon node.

[0026] Furthermore, in S5, the fault factor database is an Elasticsearch database, and the storage structure includes: fault type, fault description text, fault cause, maintenance plan text and associated pictures.

[0027] Furthermore, semantic similarity Computationally compatible alternative: When a large language model is not used, use the Bert model to calculate and The semantic similarity of .

[0028] Furthermore, the method further includes: S6, fault phenomenon associated network diagram Dynamic update step, when the fault phenomenon description submitted by the user has no matching node in the existing relationship graph, that is, all , To preset the threshold, S6 includes the following steps:

[0029] S61. Add the fault phenomenon as a new node to the fault phenomenon association network diagram ;

[0030] S62. Extract its keyword set through a large language model ;

[0031] S63, calculation Semantic similarity with all existing node keyword sets ;

[0032] S64, will Establish an edge connection between the node and the new node is the connection threshold;

[0033] S65. Store the four-element information of the new fault phenomenon into the fault element database.

[0034] A storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for automatically recommending fault repair solutions based on a large language model.

[0035] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned method for automatically recommending fault repair solutions based on a large language model.

[0036] Beneficial effects of the present invention: The method for automatically recommending fault repair solutions based on a large language model of the present invention has achieved substantial progress in recommending repair solutions by integrating the semantic understanding capabilities of a large language model with dynamic association network technology. The present invention effectively solves the semantic understanding limitations of traditional methods caused by reliance on literal matching. By simultaneously capturing keyword overlap and deep semantic associations through a hybrid similarity model, the matching accuracy of professional terminology differences and complex fault descriptions is significantly improved; the response delay problem of cold start scenarios is completely overcome, and new fault phenomena can be integrated into the knowledge network in real time; the system operation and maintenance efficiency is greatly optimized, and millisecond-level computing responses are achieved based on pre-built semantic association networks, while significantly reducing manual maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a flow chart of the method for automatically recommending fault repair solutions based on a large language model of the present invention. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] Reference Figure 1 As shown, a method for automatically recommending fault repair solutions based on a large language model includes the following steps:

[0040] S1. Based on the M fault phenomenon descriptions in the fault knowledge base, the keyword set of each fault phenomenon is extracted through the large language model to form node attributes; the semantic similarity is calculated based on the interaction relationship between any two fault phenomenon keyword sets, and a weighted fault phenomenon association network diagram is constructed. ;

[0041] S2. Receive the fault phenomenon description input by the user and extract the keyword set A through the large language model;

[0042] S3. Traverse the fault phenomenon associated network diagram Calculate the semantic similarity between the user keyword set A and each node keyword set for all nodes ;

[0043] S4. Sort by semantic similarity from high to low and output the top N matching fault phenomena;

[0044] S5. Retrieve the fault type, fault description, fault cause and repair plan corresponding to the first N fault phenomena from the fault factor database and display them to the user.

[0045] Specifically, the automatic fault repair solution recommendation method based on a large language model of the present invention extracts a set of fault phenomenon keywords through a large language model and constructs a weighted fault phenomenon association network diagram, which can effectively capture the semantic association between fault phenomena, solves the problem that the traditional keyword precise matching mechanism lacks understanding of the semantics of complex fault descriptions, and can accurately handle the matching needs of non-standard terms and complex symptoms; by calculating the semantic similarity between the user-input fault phenomenon and the network node and ranking and recommending them, the accuracy and pertinence of the repair solution recommendation are significantly improved. Actual measurements show that it can effectively improve the defect of insufficient Top-5 recommendation hit rate of the traditional method; at the same time, the method realizes fast retrieval based on a pre-built dynamic association network, greatly shortens the response time, and supports the real-time integration of new fault phenomena into the knowledge network, overcoming the delay and cost problems of manual label injection in the cold start scenario, and has both high efficiency and dynamic scalability. It can provide users with accurate and real-time repair solution recommendation services, significantly improve fault repair efficiency and reduce system maintenance costs.

[0046] Furthermore, in S1, the node attributes are constructed as follows:

[0047] For complex fault phenomenon descriptions, a keyword set is extracted through a large language model. ;

[0048] For simple fault phenomenon descriptions, extract keyword sets through word segmentation using a large language model ,

[0049] The large language models include ChatGPT, Qwen and DeepSeek.

[0050] Specifically, the present invention adopts a large language model to extract keyword sets for complex and simple fault phenomenon descriptions, respectively, which significantly improves the semantic representation accuracy of node attributes. For complex fault phenomena, the core keywords are extracted through the deep semantic understanding ability of the large language model, which can effectively parse the implicit semantic associations in complex expressions containing multi-part complex symptoms, professional terminology variants, etc., and avoid matching failures caused by missing surface vocabulary; for simple fault phenomena, the word segmentation ability of the large language model is used to accurately extract key elements to ensure the semantic integrity of the basic description. The large language model covers mainstream models such as ChatGPT, Qwen and DeepSeek. Its powerful natural language processing capabilities can adapt to a variety of fault description scenarios. Through the semantic encoding and feature extraction mechanism of the model, the constructed node attributes can not only retain the detailed characteristics of the fault phenomenon, but also capture the semantic equivalence across domains and terms, laying a more accurate semantic foundation for subsequent semantic similarity calculation and association network graph construction, fundamentally enhancing the comprehensiveness and accuracy of fault phenomenon representation, and solving the limitations of traditional manual labeling or basic word segmentation technology at the semantic understanding level.

[0051] Furthermore, in S1, the semantic similarity The calculation formula is:

[0052]

[0053] in, is the keyword overlap, The two keyword sets remaining after removing the same keywords and The semantic similarity between

[0054] When there are identical keywords, , calculate the keyword overlap:

[0055]

[0056] in, , calculate the remaining keyword set and The semantic similarity of ;

[0057] When there is no identical keyword, ,make , calculate the entire keyword set and The semantic similarity of .

[0058] Specifically, the semantic similarity calculation formula described in claim 3 of the present invention is By integrating the keyword overlap Semantic similarity with remaining keywords , a hybrid metric model that takes into account both literal matching and deep semantic association is constructed. When there are identical keywords, The degree of overlap is quantified by the ratio of the number of set elements to ensure the accuracy of basic matching; The semantic similarity is calculated for the remaining set after removing the same keywords, effectively capturing the semantic equivalence of the differences between professional terms such as "engine knock" and "engine knock". When there are no identical keywords, the model automatically Set to zero and pass only Calculate the overall semantic association to avoid the missed detection problem caused by literal mismatch in traditional methods. The value range of Giving higher weight to the keyword overlap not only retains the stability of traditional keyword matching, but also By introducing the semantic understanding capabilities of large language models, an organic balance between "explicit vocabulary matching" and "implicit semantic association" is achieved, enabling the model to adapt to multi-scenario semantic matching requirements from simple faults to complex compound symptoms, significantly improving the comprehensiveness and accuracy of fault phenomenon association analysis, and solving the core defects of existing technologies, such as the single semantic understanding and the inability to handle term variants and complex semantics.

[0059] Further, The value of .

[0060] Specifically, The value of is limited to , by denoting the keyword overlap Assigning higher weights ensures that the model prioritizes explicit word matching in semantic similarity calculations while retaining the semantic similarity of the remaining keywords. This design not only continues the stability and reliability of traditional keyword matching methods in basic retrieval, but also avoids the matching deviation caused by over-reliance on deep semantics. The adjustable range of (such as the middle value of 0.7) balances the contribution ratio of literal matching and semantic understanding of the large language model, enabling the model to adapt to the fault description characteristics in different scenarios. For example, when the user inputs a fault phenomenon containing clear professional terms, a higher The value can quickly lock the matching node through the keyword overlap, improving the search efficiency; and when encountering non-standard terms or complex symptom descriptions, The semantic supplementary effect of [the sentence fragment "] can still effectively discover potential associations and avoid missed detections. This value range restriction is not a simple parameter setting, but is based on the dual requirements of accuracy and efficiency in the field of fault repair. Through quantitative control, it organically integrates the advantages of traditional technology and large language models, and overcomes the "either-or" limitation of existing methods in semantic matching strategies, enabling the recommendation model to maintain stable and accurate performance in fault scenarios of varying complexity.

[0061] Furthermore, in S3, the user keyword set A and the node Similarity calculation, reuse S1 The algorithm regards the keyword set A as a newly added fault phenomenon node.

[0062] Specifically, the user keyword set A and the node The similarity calculation reuses S1 The algorithm treats keyword set A as a newly added fault phenomenon node. This design ensures consistency and dynamic scalability in fault phenomenon matching logic. By reusing the semantic similarity calculation framework built on a large language model, there's no need to develop separate adaptation algorithms for user input. This ensures technical consistency throughout the entire process, from knowledge base pre-construction to real-time retrieval, and avoids matching biases caused by algorithmic differences. Treating user input as a newly added node enables the system to dynamically respond to real-time fault descriptions. For example, when a user submits a query containing a new, complex symptom, that node can temporarily participate in the similarity calculation of the association network graph without having to pre-update the entire knowledge base structure. This significantly improves the system's response efficiency in cold start scenarios (such as newly emerging fault types). This "plug-and-play" node expansion mechanism, combined with the large language model's real-time semantic parsing of user input, maintains the stability of the pre-constructed network while enabling the system to flexibly adapt to diverse fault descriptions. This addresses the disconnect between user queries and the knowledge base structure in traditional approaches, making the recommendation process both efficient and dynamically adaptable, and more accurately capturing the potential connections between users' actual needs and the existing knowledge base.

[0063] Furthermore, in S5, the fault factor database is an Elasticsearch database, and the storage structure includes: fault type, fault description text, fault cause, maintenance plan text and associated pictures.

[0064] Specifically, the fault factor database uses the Elasticsearch database, whose distributed architecture and real-time search capabilities significantly improve the retrieval efficiency and storage flexibility of fault data. The database storage structure contains multiple information such as fault type, fault description text, fault cause, maintenance plan text and associated pictures. It can integrate the maintenance knowledge system in a structured manner. For example, the phenomenon description of the "engine abnormal noise" fault, mechanical principle analysis, disassembly and maintenance steps and component diagrams are stored in association to form a three-dimensional knowledge unit. The full-text search and aggregation query functions supported by Elasticsearch can be combined with the keyword set extracted by the large language model to achieve millisecond-level cross-field semantic matching. For example, when the user enters "transmission stalling", the system can quickly associate maintenance plans with causes such as "gear wear" and "valve body failure", and simultaneously retrieve the corresponding disassembly and assembly diagrams. In addition, the database's dynamic expansion feature supports real-time writing of the four elements of new fault phenomena (such as the relevant data of the newly added nodes in step S6), which works in synergy with the dynamic update mechanism of the fault phenomenon-related network diagram, ensuring that the knowledge base continues to evolve with technological development. It solves the performance bottleneck of traditional relational databases in unstructured data processing and real-time update scenarios, and provides efficient data support for the rapid and accurate recommendation of fault repair solutions.

[0065] Furthermore, semantic similarity Computationally compatible alternative: When a large language model is not used, use the Bert model to calculate and The semantic similarity of .

[0066] Specifically, semantic similarity A computationally compatible alternative (when not using a large language model, using the Bert model to calculate and The semantic similarity of ), which reflects the flexibility and compatibility of the technical solution. Although large language models (such as ChatGPT, Qwen, and DeepSeek) have powerful contextual understanding capabilities, they may be limited by computing power or model deployment costs in specific scenarios. At this time, the Bert model, as a lightweight deep learning model, can reduce computing resource consumption while maintaining the accuracy of semantic representation. For example, in edge devices or small and medium-sized enterprise knowledge bases, the semantic vector encoding of keyword sets is implemented through the Bert model, which can still effectively capture the semantic equivalence between professional terms such as "ECU error" and "on-board computer failure". This compatible design allows users to choose the model type according to actual needs, which not only retains the advanced semantic understanding capabilities of large language models, but also provides alternatives for resource-constrained scenarios and expands the scope of technical application. At the same time, both models are based on the semantic feature extraction mechanism of the Transformer architecture, ensuring The consistency of the calculation logic can achieve effective measurement of the deep semantic association of the keyword set regardless of the model used, solving the scene adaptability problem that may be caused by dependence on a single model, and enabling the semantic similarity calculation method of the present invention to maintain stable recommendation performance in different technical environments.

[0067] Furthermore, the method further includes: S6, fault phenomenon associated network diagram Dynamic update step, when the fault phenomenon description submitted by the user has no matching node in the existing relationship graph, that is, all , To preset the threshold, S6 includes the following steps:

[0068] S61. Add the fault phenomenon as a new node to the fault phenomenon association network diagram ;

[0069] S62. Extract its keyword set through a large language model ;

[0070] S63, calculation Semantic similarity with all existing node keyword sets ;

[0071] S64, will Establish an edge connection between the node and the new node is the connection threshold;

[0072] S65. Store the four-element information of the new fault phenomenon into the fault element database.

[0073] Specifically, the step of dynamically updating the fault phenomenon associated network diagram described in S6 is to set a preset threshold Connection threshold , a self-evolution mechanism that adapts to new fault phenomena is constructed. When the fault description submitted by the user has no matching node in the existing network (i.e., all ), the system automatically adds it to the network as a new node and extracts a set of keywords through a large language model , and then establish an edge connection after calculating the semantic similarity with the existing nodes ( ). This process realizes the real-time expansion of the knowledge base. For example, when a new type of fault such as "new energy battery thermal runaway warning" occurs, the system can automatically complete node creation, semantic association and data storage (S65) without manual intervention, which completely solves the delay problem of new faults that need to be manually marked and injected into the library in the cold start scenario in the traditional method. The dynamic update mechanism is combined with the semantic parsing ability of the large language model to enable the network to continuously absorb cutting-edge fault knowledge and avoid matching failures caused by the solidification of the knowledge system. At the same time, the relevance and effectiveness of the newly added nodes are ensured through threshold control to prevent irrelevant data redundancy. This design not only improves the system's response speed to emerging faults, but also realizes the autonomous iteration of the fault maintenance knowledge system through the closed-loop process of "data collection-semantic analysis-network evolution", providing core support for the accuracy and timeliness of the long-term maintenance system, so that the recommendation model of the present invention can adapt to the ever-changing fault maintenance scenarios.

[0074] A storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for automatically recommending fault repair solutions based on a large language model.

[0075] Specifically, the storage medium of the present invention provides a standardized technical carrier for complex fault diagnosis scenarios by storing a computer program that can execute the method for automatically recommending fault repair solutions based on a large language model. The storage medium can fully reproduce the keyword extraction based on the large language model, the semantic similarity calculation (such as the hybrid model) ), dynamic association network construction and Top-N recommendation and other core processes, to ensure that the semantic understanding ability of the large language model and the deep integration of the fault maintenance knowledge system can run stably in different hardware environments. For example, when the processor calls the program, it can realize real-time semantic analysis of the fault phenomenon input by the user (such as extracting the compound keyword "transmission jerking accompanied by oil temperature increase"), and quickly complete similarity matching based on the pre-built fault phenomenon association network diagram, and output structured results containing the cause of the fault, maintenance plan text and associated pictures. In addition, the storage medium supports dynamic updates of the program and can seamlessly integrate the network self-evolution mechanism of the S6 step, so that the system automatically optimizes the recommendation logic when accessing new fault data without relying on specific hardware configuration or manual code modification. This design of solidifying complex algorithms in storage media not only ensures the reusability and portability of the technical solution of the present invention, but also provides an efficient and reliable solution for the standardized deployment of industrial-grade fault diagnosis systems, significantly reducing the threshold for technology implementation and subsequent maintenance costs.

[0076] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-mentioned method for automatically recommending fault repair solutions based on a large language model.

[0077] Specifically, the computer device described in the present invention, by integrating memory, processor and computer program that can execute the method of the present invention, builds a hardware-software collaborative system for intelligent fault repair recommendation. By running the program, the processor can realize fault phenomenon keyword extraction based on a large language model (such as using deep semantic analysis for complex descriptions and word segmentation for simple descriptions), hybrid semantic similarity calculation (integrating keyword overlap), and other functions. Semantic association with remaining keywords ) and dynamic association network updates (such as the automatic addition of new nodes to the S6 mechanism). For example, when a user enters "unstable engine idle speed," the processor can call on the large language model to quickly extract keywords such as "idle," "unstable," and "engine operating conditions." It then calculates the semantic similarity with existing nodes through a pre-built fault phenomenon association network diagram, and outputs the top N matching fault phenomena and their corresponding repair solutions (such as "throttle carbon deposit cleaning," "idle motor fault replacement," etc.) within milliseconds. The memory provides high-speed storage support for the fault knowledge base, large language model parameters, and dynamically updated network data, ensuring data retrieval efficiency under multi-tasking. The design of this computer device deeply combines the advanced semantic understanding capabilities of the large language model with the high-performance computing capabilities of computer hardware. It not only meets the dual requirements of real-time and accuracy for complex fault diagnosis, but also supports the large-scale deployment of the system in multiple fields such as industry, transportation, and electronics through standardized software and hardware architecture. It provides a reliable infrastructure for the digital upgrade of fault maintenance in various industries, significantly improving the intelligence level and response efficiency of traditional maintenance models.

[0078] The following is a specific embodiment of the present invention:

[0079] The fault phenomenon set is , the correlation network diagram between the fault phenomena is:

[0080] ,in: .

[0081] For describing more complex fault phenomena, a keyword set of the current fault phenomenon is extracted based on a large language model algorithm (such as ChatGPT, Qwen, DeepSeek, etc.); for describing simpler fault phenomena, a keyword set can be obtained by performing word segmentation extraction using a large language model (such as ChatGPT, Qwen, DeepSeek, etc.). Therefore:

[0082] Correlation network diagram between fault phenomena In [1], the attribute of a node is: a set of several keywords that represent the description of the fault phenomenon.

[0083] Correlation network diagram between fault phenomena The attributes of the edges can be: Do they have the same keyword? Are there similar keywords? How many same / similar keywords are there?

[0084] Here are some examples:

[0085] Use a set of keywords to describe each fault phenomenon. You can redefine According to the current field / professional fault knowledge base manual / help document, the description of each fault phenomenon is extracted as a keyword set. Assume that the fault phenomenon description All keywords as a collection , considering the fault phenomenon description and Corresponding keyword set and , we can find , (1) When When (ie: there are exactly the same keywords):

[0086] for Any keyword in the collection , we can find out the keywords contained in each fault phenomenon description node Number of papers and , so, we have:

[0087] ,

[0088] (2) When When (ie: there are no identical keywords):

[0089] Since there are no completely identical keywords, it is necessary to evaluate the semantic similarity of this situation. Here, ChatGPT can be used to ask questions about the semantic similarity of two keyword sets and obtain a value (semantic similarity) between [0, 1], which is recorded as .

[0090] (3) Calculate semantic similarity:

[0091] Considering the two cases (1) and (2), the semantic similarity between the two fault phenomenon descriptions is calculated using a weighted method as follows:

[0092] , .

[0093] in: It is the value calculated from the exact same set of keywords; The two keyword sets remaining after removing the same keywords and The semantic similarity between .

[0094] At this point, a fault phenomenon description relationship diagram based on the fault knowledge base has been established .

[0095] When a user submits a description of the fault phenomenon (let's assume it's a text description) on the fault repair platform, the following steps can be followed to automatically recommend a repair solution:

[0096] STEP 1: Construct a fault phenomenon description relationship diagram S based on the fault knowledge base;

[0097] STEP 2: Based on the user's description of the fault phenomenon, extract keywords including necessary information such as the location (if any) to form a fault phenomenon keyword set A;

[0098] STEP 3: Traverse each node in the fault phenomenon description relationship graph and calculate the semantic similarity ;

[0099] STEP 4: List the top N fault phenomenon descriptions in the fault knowledge base in descending order of semantic similarity values.

[0100] STEP 5: Based on the four fault elements (fault type, fault description, fault cause, and repair plan) and other information stored in the ES database, list the other information (repair plans, pictures, etc.) corresponding to these N fault phenomena in order from the beginning to the end (i.e., in Ton-N order) on the Q&A page of the fault knowledge base (actually the Chat window).

[0101] The present invention's method for automatically recommending fault repair solutions based on a large language model achieves substantial progress in recommending repair solutions by integrating the semantic understanding capabilities of a large language model with dynamic association network technology. This method effectively addresses the semantic understanding limitations of traditional methods due to reliance on literal matching. By simultaneously capturing keyword overlap and deep semantic associations through a hybrid similarity model, the accuracy of matching between differences in professional terminology and complex fault descriptions is significantly improved. It completely overcomes the response lag problem in cold start scenarios, allowing new fault phenomena to be integrated into the knowledge network in real time. It significantly optimizes system operation and maintenance efficiency, achieving millisecond-level computational responses based on a pre-built semantic association network while significantly reducing manual maintenance costs.

[0102] While the specific embodiments of the present invention have been described in detail above, these are intended to be exemplary only, and the present invention is not limited thereto. Any equivalent modifications or substitutions to the present invention that would be apparent to those skilled in the art are also within the scope of the present invention. Therefore, any equivalent modifications or substitutions made without departing from the spirit and scope of the present invention are intended to be encompassed within the scope of the present invention.

Claims

1. A method for automatically recommending fault repair solutions based on a large language model, characterized in that: The method comprises the following steps: S1. Based on the M fault phenomenon descriptions in the fault knowledge base, the keyword set of each fault phenomenon is extracted through the large language model to form node attributes; the semantic similarity is calculated based on the interaction relationship between any two fault phenomenon keyword sets, and a weighted fault phenomenon association network diagram is constructed. , semantic similarity The calculation formula is: in, is the keyword overlap, The two keyword sets remaining after removing the same keywords and The semantic similarity between When there are identical keywords, , calculate the keyword overlap: in, , calculate the remaining keyword set and The semantic similarity of ; When there is no identical keyword, ,make , calculate the entire keyword set and The semantic similarity of ; S2. Receive the fault phenomenon description input by the user and extract the keyword set A through the large language model; S3. Traverse the fault phenomenon associated network diagram Calculate the semantic similarity between the user keyword set A and each node keyword set for all nodes ; S4. Sort by semantic similarity from high to low and output the top N matching fault phenomena; S5. Retrieve the fault type, fault description, fault cause and repair plan corresponding to the first N fault phenomena from the fault factor database and display them to the user.

2. The method for automatically recommending fault repair solutions based on a large language model according to claim 1, characterized in that: In S1, node attributes are constructed as follows: For complex fault phenomenon descriptions, a keyword set is extracted through a large language model. ; For simple fault phenomenon descriptions, extract keyword sets through word segmentation using a large language model , The large language models include ChatGPT, Qwen and DeepSeek.

3. The method for automatically recommending fault repair solutions based on a large language model according to claim 1, characterized in that: The value of .

4. The method for automatically recommending fault repair solutions based on a large language model according to claim 3, characterized in that: In S3, user keyword set A and node Similarity calculation, reuse S1 The algorithm regards the keyword set A as a newly added fault phenomenon node.

5. The method for automatically recommending fault repair solutions based on a large language model according to claim 4, characterized in that: In S5, the fault factor database is an Elasticsearch database, and the storage structure includes: fault type, fault description text, fault cause, maintenance plan text and associated pictures.

6. The method for automatically recommending fault repair solutions based on a large language model according to claim 1, characterized in that: Semantic similarity Computationally compatible alternative: When a large language model is not used, use the Bert model to calculate and The semantic similarity of .

7. The method for automatically recommending fault repair solutions based on a large language model according to claim 1, characterized in that: The method further includes: S6, fault phenomenon correlation network diagram Dynamic update step, when the fault phenomenon description submitted by the user has no matching node in the existing relationship graph, that is, all , To preset the threshold, S6 includes the following steps: S61. Add the fault phenomenon as a new node to the fault phenomenon association network diagram ; S62. Extract its keyword set through a large language model ; S63, calculation Semantic similarity with all existing node keyword sets ; S64, will Establish an edge connection between the node and the new node is the connection threshold; S65. Store the four-element information of the new fault phenomenon into the fault element database.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method for automatically recommending a fault repair solution based on a large language model according to any one of claims 1 to 7 is implemented.

9. A computer device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method for automatically recommending a fault repair solution based on a large language model according to any one of claims 1 to 7.

Images